1/*
2 * Copyright © 2016 Bas Nieuwenhuizen
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24#include "ac_nir_to_llvm.h"
25#include "ac_gpu_info.h"
26#include "ac_binary.h"
27#include "ac_llvm_build.h"
28#include "ac_llvm_util.h"
29#include "ac_shader_abi.h"
30#include "ac_shader_util.h"
31#include "nir/nir.h"
32#include "nir/nir_deref.h"
33#include "sid.h"
34#include "util/bitscan.h"
35#include "util/u_math.h"
36#include <llvm/Config/llvm-config.h>
37
38struct ac_nir_context {
39   struct ac_llvm_context ac;
40   struct ac_shader_abi *abi;
41   const struct ac_shader_args *args;
42
43   gl_shader_stage stage;
44   shader_info *info;
45
46   LLVMValueRef *ssa_defs;
47
48   LLVMValueRef scratch;
49   LLVMValueRef constant_data;
50
51   struct hash_table *defs;
52   struct hash_table *phis;
53   struct hash_table *vars;
54   struct hash_table *verified_interp;
55
56   LLVMValueRef main_function;
57   LLVMBasicBlockRef continue_block;
58   LLVMBasicBlockRef break_block;
59
60   LLVMValueRef vertex_id_replaced;
61   LLVMValueRef instance_id_replaced;
62   LLVMValueRef tes_u_replaced;
63   LLVMValueRef tes_v_replaced;
64   LLVMValueRef tes_rel_patch_id_replaced;
65   LLVMValueRef tes_patch_id_replaced;
66};
67
68static LLVMValueRef get_sampler_desc_index(struct ac_nir_context *ctx, nir_deref_instr *deref_instr,
69                                           const nir_instr *instr, bool image);
70
71static LLVMValueRef get_sampler_desc(struct ac_nir_context *ctx, nir_deref_instr *deref_instr,
72                                     enum ac_descriptor_type desc_type, const nir_instr *instr,
73                                     LLVMValueRef index, bool image, bool write);
74
75static LLVMTypeRef get_def_type(struct ac_nir_context *ctx, const nir_ssa_def *def)
76{
77   LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, def->bit_size);
78   if (def->num_components > 1) {
79      type = LLVMVectorType(type, def->num_components);
80   }
81   return type;
82}
83
84static LLVMValueRef get_src(struct ac_nir_context *nir, nir_src src)
85{
86   assert(src.is_ssa);
87   return nir->ssa_defs[src.ssa->index];
88}
89
90static LLVMValueRef get_memory_ptr_t(struct ac_nir_context *ctx, nir_src src, LLVMTypeRef elem_type, unsigned c_off)
91{
92   LLVMValueRef ptr = get_src(ctx, src);
93   LLVMValueRef lds_i8 = ctx->ac.lds;
94   if (ctx->stage != MESA_SHADER_COMPUTE)
95      lds_i8 = LLVMBuildBitCast(ctx->ac.builder, ctx->ac.lds, LLVMPointerType(ctx->ac.i8, AC_ADDR_SPACE_LDS), "");
96
97   ptr = LLVMBuildAdd(ctx->ac.builder, ptr, LLVMConstInt(ctx->ac.i32, c_off, 0), "");
98   ptr = LLVMBuildGEP2(ctx->ac.builder, ctx->ac.i8, lds_i8, &ptr, 1, "");
99   int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
100
101   return LLVMBuildBitCast(ctx->ac.builder, ptr, LLVMPointerType(elem_type, addr_space), "");
102}
103
104static LLVMValueRef get_memory_ptr(struct ac_nir_context *ctx, nir_src src, unsigned bit_size, unsigned c_off)
105{
106   return get_memory_ptr_t(ctx, src, LLVMIntTypeInContext(ctx->ac.context, bit_size), c_off);
107}
108
109static LLVMBasicBlockRef get_block(struct ac_nir_context *nir, const struct nir_block *b)
110{
111   struct hash_entry *entry = _mesa_hash_table_search(nir->defs, b);
112   return (LLVMBasicBlockRef)entry->data;
113}
114
115static LLVMValueRef get_alu_src(struct ac_nir_context *ctx, nir_alu_src src,
116                                unsigned num_components)
117{
118   LLVMValueRef value = get_src(ctx, src.src);
119   bool need_swizzle = false;
120
121   assert(value);
122   unsigned src_components = ac_get_llvm_num_components(value);
123   for (unsigned i = 0; i < num_components; ++i) {
124      assert(src.swizzle[i] < src_components);
125      if (src.swizzle[i] != i)
126         need_swizzle = true;
127   }
128
129   if (need_swizzle || num_components != src_components) {
130      LLVMValueRef masks[] = {LLVMConstInt(ctx->ac.i32, src.swizzle[0], false),
131                              LLVMConstInt(ctx->ac.i32, src.swizzle[1], false),
132                              LLVMConstInt(ctx->ac.i32, src.swizzle[2], false),
133                              LLVMConstInt(ctx->ac.i32, src.swizzle[3], false)};
134
135      if (src_components > 1 && num_components == 1) {
136         value = LLVMBuildExtractElement(ctx->ac.builder, value, masks[0], "");
137      } else if (src_components == 1 && num_components > 1) {
138         LLVMValueRef values[] = {value, value, value, value};
139         value = ac_build_gather_values(&ctx->ac, values, num_components);
140      } else {
141         LLVMValueRef swizzle = LLVMConstVector(masks, num_components);
142         value = LLVMBuildShuffleVector(ctx->ac.builder, value, value, swizzle, "");
143      }
144   }
145   assert(!src.negate);
146   assert(!src.abs);
147   return value;
148}
149
150static LLVMValueRef emit_int_cmp(struct ac_llvm_context *ctx, LLVMIntPredicate pred,
151                                 LLVMValueRef src0, LLVMValueRef src1)
152{
153   src0 = ac_to_integer(ctx, src0);
154   src1 = ac_to_integer(ctx, src1);
155   return LLVMBuildICmp(ctx->builder, pred, src0, src1, "");
156}
157
158static LLVMValueRef emit_float_cmp(struct ac_llvm_context *ctx, LLVMRealPredicate pred,
159                                   LLVMValueRef src0, LLVMValueRef src1)
160{
161   src0 = ac_to_float(ctx, src0);
162   src1 = ac_to_float(ctx, src1);
163   return LLVMBuildFCmp(ctx->builder, pred, src0, src1, "");
164}
165
166static LLVMValueRef emit_intrin_1f_param(struct ac_llvm_context *ctx, const char *intrin,
167                                         LLVMTypeRef result_type, LLVMValueRef src0)
168{
169   char name[64], type[64];
170   LLVMValueRef params[] = {
171      ac_to_float(ctx, src0),
172   };
173
174   ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
175   ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
176   assert(length < sizeof(name));
177   return ac_build_intrinsic(ctx, name, result_type, params, 1, AC_FUNC_ATTR_READNONE);
178}
179
180static LLVMValueRef emit_intrin_1f_param_scalar(struct ac_llvm_context *ctx, const char *intrin,
181                                                LLVMTypeRef result_type, LLVMValueRef src0)
182{
183   if (LLVMGetTypeKind(result_type) != LLVMVectorTypeKind)
184      return emit_intrin_1f_param(ctx, intrin, result_type, src0);
185
186   LLVMTypeRef elem_type = LLVMGetElementType(result_type);
187   LLVMValueRef ret = LLVMGetUndef(result_type);
188
189   /* Scalarize the intrinsic, because vectors are not supported. */
190   for (unsigned i = 0; i < LLVMGetVectorSize(result_type); i++) {
191      char name[64], type[64];
192      LLVMValueRef params[] = {
193         ac_to_float(ctx, ac_llvm_extract_elem(ctx, src0, i)),
194      };
195
196      ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
197      ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
198      assert(length < sizeof(name));
199      ret = LLVMBuildInsertElement(
200         ctx->builder, ret,
201         ac_build_intrinsic(ctx, name, elem_type, params, 1, AC_FUNC_ATTR_READNONE),
202         LLVMConstInt(ctx->i32, i, 0), "");
203   }
204   return ret;
205}
206
207static LLVMValueRef emit_intrin_2f_param(struct ac_llvm_context *ctx, const char *intrin,
208                                         LLVMTypeRef result_type, LLVMValueRef src0,
209                                         LLVMValueRef src1)
210{
211   char name[64], type[64];
212   LLVMValueRef params[] = {
213      ac_to_float(ctx, src0),
214      ac_to_float(ctx, src1),
215   };
216
217   ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
218   ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
219   assert(length < sizeof(name));
220   return ac_build_intrinsic(ctx, name, result_type, params, 2, AC_FUNC_ATTR_READNONE);
221}
222
223static LLVMValueRef emit_intrin_3f_param(struct ac_llvm_context *ctx, const char *intrin,
224                                         LLVMTypeRef result_type, LLVMValueRef src0,
225                                         LLVMValueRef src1, LLVMValueRef src2)
226{
227   char name[64], type[64];
228   LLVMValueRef params[] = {
229      ac_to_float(ctx, src0),
230      ac_to_float(ctx, src1),
231      ac_to_float(ctx, src2),
232   };
233
234   ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
235   ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
236   assert(length < sizeof(name));
237   return ac_build_intrinsic(ctx, name, result_type, params, 3, AC_FUNC_ATTR_READNONE);
238}
239
240static LLVMValueRef emit_bcsel(struct ac_llvm_context *ctx, LLVMValueRef src0, LLVMValueRef src1,
241                               LLVMValueRef src2)
242{
243   LLVMTypeRef src1_type = LLVMTypeOf(src1);
244   LLVMTypeRef src2_type = LLVMTypeOf(src2);
245
246   if (LLVMGetTypeKind(src1_type) == LLVMPointerTypeKind &&
247       LLVMGetTypeKind(src2_type) != LLVMPointerTypeKind) {
248      src2 = LLVMBuildIntToPtr(ctx->builder, src2, src1_type, "");
249   } else if (LLVMGetTypeKind(src2_type) == LLVMPointerTypeKind &&
250              LLVMGetTypeKind(src1_type) != LLVMPointerTypeKind) {
251      src1 = LLVMBuildIntToPtr(ctx->builder, src1, src2_type, "");
252   }
253
254   return LLVMBuildSelect(ctx->builder, src0, ac_to_integer_or_pointer(ctx, src1),
255                          ac_to_integer_or_pointer(ctx, src2), "");
256}
257
258static LLVMValueRef emit_iabs(struct ac_llvm_context *ctx, LLVMValueRef src0)
259{
260   return ac_build_imax(ctx, src0, LLVMBuildNeg(ctx->builder, src0, ""));
261}
262
263static LLVMValueRef emit_uint_carry(struct ac_llvm_context *ctx, const char *intrin,
264                                    LLVMValueRef src0, LLVMValueRef src1)
265{
266   LLVMTypeRef ret_type;
267   LLVMTypeRef types[] = {ctx->i32, ctx->i1};
268   LLVMValueRef res;
269   LLVMValueRef params[] = {src0, src1};
270   ret_type = LLVMStructTypeInContext(ctx->context, types, 2, true);
271
272   res = ac_build_intrinsic(ctx, intrin, ret_type, params, 2, AC_FUNC_ATTR_READNONE);
273
274   res = LLVMBuildExtractValue(ctx->builder, res, 1, "");
275   res = LLVMBuildZExt(ctx->builder, res, ctx->i32, "");
276   return res;
277}
278
279static LLVMValueRef emit_b2f(struct ac_llvm_context *ctx, LLVMValueRef src0, unsigned bitsize)
280{
281   assert(ac_get_elem_bits(ctx, LLVMTypeOf(src0)) == 1);
282
283   switch (bitsize) {
284   case 16:
285      if (LLVMGetTypeKind(LLVMTypeOf(src0)) == LLVMVectorTypeKind) {
286         assert(LLVMGetVectorSize(LLVMTypeOf(src0)) == 2);
287         LLVMValueRef f[] = {
288            LLVMBuildSelect(ctx->builder, ac_llvm_extract_elem(ctx, src0, 0),
289                            ctx->f16_1, ctx->f16_0, ""),
290            LLVMBuildSelect(ctx->builder, ac_llvm_extract_elem(ctx, src0, 1),
291                            ctx->f16_1, ctx->f16_0, ""),
292         };
293         return ac_build_gather_values(ctx, f, 2);
294      }
295      return LLVMBuildSelect(ctx->builder, src0, ctx->f16_1, ctx->f16_0, "");
296   case 32:
297      return LLVMBuildSelect(ctx->builder, src0, ctx->f32_1, ctx->f32_0, "");
298   case 64:
299      return LLVMBuildSelect(ctx->builder, src0, ctx->f64_1, ctx->f64_0, "");
300   default:
301      unreachable("Unsupported bit size.");
302   }
303}
304
305static LLVMValueRef emit_f2b(struct ac_llvm_context *ctx, LLVMValueRef src0)
306{
307   src0 = ac_to_float(ctx, src0);
308   LLVMValueRef zero = LLVMConstNull(LLVMTypeOf(src0));
309   return LLVMBuildFCmp(ctx->builder, LLVMRealUNE, src0, zero, "");
310}
311
312static LLVMValueRef emit_b2i(struct ac_llvm_context *ctx, LLVMValueRef src0, unsigned bitsize)
313{
314   switch (bitsize) {
315   case 8:
316      return LLVMBuildSelect(ctx->builder, src0, ctx->i8_1, ctx->i8_0, "");
317   case 16:
318      return LLVMBuildSelect(ctx->builder, src0, ctx->i16_1, ctx->i16_0, "");
319   case 32:
320      return LLVMBuildSelect(ctx->builder, src0, ctx->i32_1, ctx->i32_0, "");
321   case 64:
322      return LLVMBuildSelect(ctx->builder, src0, ctx->i64_1, ctx->i64_0, "");
323   default:
324      unreachable("Unsupported bit size.");
325   }
326}
327
328static LLVMValueRef emit_i2b(struct ac_llvm_context *ctx, LLVMValueRef src0)
329{
330   LLVMValueRef zero = LLVMConstNull(LLVMTypeOf(src0));
331   return LLVMBuildICmp(ctx->builder, LLVMIntNE, src0, zero, "");
332}
333
334static LLVMValueRef emit_f2f16(struct ac_llvm_context *ctx, LLVMValueRef src0)
335{
336   LLVMValueRef result;
337   LLVMValueRef cond = NULL;
338
339   src0 = ac_to_float(ctx, src0);
340   result = LLVMBuildFPTrunc(ctx->builder, src0, ctx->f16, "");
341
342   if (ctx->gfx_level >= GFX8) {
343      LLVMValueRef args[2];
344      /* Check if the result is a denormal - and flush to 0 if so. */
345      args[0] = result;
346      args[1] = LLVMConstInt(ctx->i32, N_SUBNORMAL | P_SUBNORMAL, false);
347      cond =
348         ac_build_intrinsic(ctx, "llvm.amdgcn.class.f16", ctx->i1, args, 2, AC_FUNC_ATTR_READNONE);
349   }
350
351   /* need to convert back up to f32 */
352   result = LLVMBuildFPExt(ctx->builder, result, ctx->f32, "");
353
354   if (ctx->gfx_level >= GFX8)
355      result = LLVMBuildSelect(ctx->builder, cond, ctx->f32_0, result, "");
356   else {
357      /* for GFX6-GFX7 */
358      /* 0x38800000 is smallest half float value (2^-14) in 32-bit float,
359       * so compare the result and flush to 0 if it's smaller.
360       */
361      LLVMValueRef temp, cond2;
362      temp = emit_intrin_1f_param(ctx, "llvm.fabs", ctx->f32, result);
363      cond = LLVMBuildFCmp(
364         ctx->builder, LLVMRealOGT,
365         LLVMBuildBitCast(ctx->builder, LLVMConstInt(ctx->i32, 0x38800000, false), ctx->f32, ""),
366         temp, "");
367      cond2 = LLVMBuildFCmp(ctx->builder, LLVMRealONE, temp, ctx->f32_0, "");
368      cond = LLVMBuildAnd(ctx->builder, cond, cond2, "");
369      result = LLVMBuildSelect(ctx->builder, cond, ctx->f32_0, result, "");
370   }
371   return result;
372}
373
374static LLVMValueRef emit_umul_high(struct ac_llvm_context *ctx, LLVMValueRef src0,
375                                   LLVMValueRef src1)
376{
377   LLVMValueRef dst64, result;
378   src0 = LLVMBuildZExt(ctx->builder, src0, ctx->i64, "");
379   src1 = LLVMBuildZExt(ctx->builder, src1, ctx->i64, "");
380
381   dst64 = LLVMBuildMul(ctx->builder, src0, src1, "");
382   dst64 = LLVMBuildLShr(ctx->builder, dst64, LLVMConstInt(ctx->i64, 32, false), "");
383   result = LLVMBuildTrunc(ctx->builder, dst64, ctx->i32, "");
384   return result;
385}
386
387static LLVMValueRef emit_imul_high(struct ac_llvm_context *ctx, LLVMValueRef src0,
388                                   LLVMValueRef src1)
389{
390   LLVMValueRef dst64, result;
391   src0 = LLVMBuildSExt(ctx->builder, src0, ctx->i64, "");
392   src1 = LLVMBuildSExt(ctx->builder, src1, ctx->i64, "");
393
394   dst64 = LLVMBuildMul(ctx->builder, src0, src1, "");
395   dst64 = LLVMBuildAShr(ctx->builder, dst64, LLVMConstInt(ctx->i64, 32, false), "");
396   result = LLVMBuildTrunc(ctx->builder, dst64, ctx->i32, "");
397   return result;
398}
399
400static LLVMValueRef emit_bfm(struct ac_llvm_context *ctx, LLVMValueRef bits, LLVMValueRef offset)
401{
402   /* mask = ((1 << bits) - 1) << offset */
403   return LLVMBuildShl(
404      ctx->builder,
405      LLVMBuildSub(ctx->builder, LLVMBuildShl(ctx->builder, ctx->i32_1, bits, ""), ctx->i32_1, ""),
406      offset, "");
407}
408
409static LLVMValueRef emit_bitfield_select(struct ac_llvm_context *ctx, LLVMValueRef mask,
410                                         LLVMValueRef insert, LLVMValueRef base)
411{
412   /* Calculate:
413    *   (mask & insert) | (~mask & base) = base ^ (mask & (insert ^ base))
414    * Use the right-hand side, which the LLVM backend can convert to V_BFI.
415    */
416   return LLVMBuildXor(
417      ctx->builder, base,
418      LLVMBuildAnd(ctx->builder, mask, LLVMBuildXor(ctx->builder, insert, base, ""), ""), "");
419}
420
421static LLVMValueRef emit_pack_2x16(struct ac_llvm_context *ctx, LLVMValueRef src0,
422                                   LLVMValueRef (*pack)(struct ac_llvm_context *ctx,
423                                                        LLVMValueRef args[2]))
424{
425   LLVMValueRef comp[2];
426
427   src0 = ac_to_float(ctx, src0);
428   comp[0] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32_0, "");
429   comp[1] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32_1, "");
430
431   return LLVMBuildBitCast(ctx->builder, pack(ctx, comp), ctx->i32, "");
432}
433
434static LLVMValueRef emit_unpack_half_2x16(struct ac_llvm_context *ctx, LLVMValueRef src0)
435{
436   LLVMValueRef const16 = LLVMConstInt(ctx->i32, 16, false);
437   LLVMValueRef temps[2], val;
438   int i;
439
440   for (i = 0; i < 2; i++) {
441      val = i == 1 ? LLVMBuildLShr(ctx->builder, src0, const16, "") : src0;
442      val = LLVMBuildTrunc(ctx->builder, val, ctx->i16, "");
443      val = LLVMBuildBitCast(ctx->builder, val, ctx->f16, "");
444      temps[i] = LLVMBuildFPExt(ctx->builder, val, ctx->f32, "");
445   }
446   return ac_build_gather_values(ctx, temps, 2);
447}
448
449static LLVMValueRef emit_ddxy(struct ac_nir_context *ctx, nir_op op, LLVMValueRef src0)
450{
451   unsigned mask;
452   int idx;
453   LLVMValueRef result;
454
455   if (op == nir_op_fddx_fine)
456      mask = AC_TID_MASK_LEFT;
457   else if (op == nir_op_fddy_fine)
458      mask = AC_TID_MASK_TOP;
459   else
460      mask = AC_TID_MASK_TOP_LEFT;
461
462   /* for DDX we want to next X pixel, DDY next Y pixel. */
463   if (op == nir_op_fddx_fine || op == nir_op_fddx_coarse || op == nir_op_fddx)
464      idx = 1;
465   else
466      idx = 2;
467
468   result = ac_build_ddxy(&ctx->ac, mask, idx, src0);
469   return result;
470}
471
472struct waterfall_context {
473   LLVMBasicBlockRef phi_bb[2];
474   bool use_waterfall;
475};
476
477/* To deal with divergent descriptors we can create a loop that handles all
478 * lanes with the same descriptor on a given iteration (henceforth a
479 * waterfall loop).
480 *
481 * These helper create the begin and end of the loop leaving the caller
482 * to implement the body.
483 *
484 * params:
485 *  - ctx is the usal nir context
486 *  - wctx is a temporary struct containing some loop info. Can be left uninitialized.
487 *  - value is the possibly divergent value for which we built the loop
488 *  - divergent is whether value is actually divergent. If false we just pass
489 *     things through.
490 */
491static LLVMValueRef enter_waterfall(struct ac_nir_context *ctx, struct waterfall_context *wctx,
492                                    LLVMValueRef value, bool divergent)
493{
494   /* If the app claims the value is divergent but it is constant we can
495    * end up with a dynamic index of NULL. */
496   if (!value)
497      divergent = false;
498
499   wctx->use_waterfall = divergent;
500   if (!divergent)
501      return value;
502
503   ac_build_bgnloop(&ctx->ac, 6000);
504
505   LLVMValueRef active = LLVMConstInt(ctx->ac.i1, 1, false);
506   LLVMValueRef scalar_value[NIR_MAX_VEC_COMPONENTS];
507
508   for (unsigned i = 0; i < ac_get_llvm_num_components(value); i++) {
509      LLVMValueRef comp = ac_llvm_extract_elem(&ctx->ac, value, i);
510      scalar_value[i] = ac_build_readlane(&ctx->ac, comp, NULL);
511      active = LLVMBuildAnd(ctx->ac.builder, active,
512                            LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, comp, scalar_value[i], ""), "");
513   }
514
515   wctx->phi_bb[0] = LLVMGetInsertBlock(ctx->ac.builder);
516   ac_build_ifcc(&ctx->ac, active, 6001);
517
518   return ac_build_gather_values(&ctx->ac, scalar_value, ac_get_llvm_num_components(value));
519}
520
521static LLVMValueRef exit_waterfall(struct ac_nir_context *ctx, struct waterfall_context *wctx,
522                                   LLVMValueRef value)
523{
524   LLVMValueRef ret = NULL;
525   LLVMValueRef phi_src[2];
526   LLVMValueRef cc_phi_src[2] = {
527      LLVMConstInt(ctx->ac.i32, 0, false),
528      LLVMConstInt(ctx->ac.i32, 0xffffffff, false),
529   };
530
531   if (!wctx->use_waterfall)
532      return value;
533
534   wctx->phi_bb[1] = LLVMGetInsertBlock(ctx->ac.builder);
535
536   ac_build_endif(&ctx->ac, 6001);
537
538   if (value) {
539      phi_src[0] = LLVMGetUndef(LLVMTypeOf(value));
540      phi_src[1] = value;
541
542      ret = ac_build_phi(&ctx->ac, LLVMTypeOf(value), 2, phi_src, wctx->phi_bb);
543   }
544
545   /*
546    * By using the optimization barrier on the exit decision, we decouple
547    * the operations from the break, and hence avoid LLVM hoisting the
548    * opteration into the break block.
549    */
550   LLVMValueRef cc = ac_build_phi(&ctx->ac, ctx->ac.i32, 2, cc_phi_src, wctx->phi_bb);
551   ac_build_optimization_barrier(&ctx->ac, &cc, false);
552
553   LLVMValueRef active =
554      LLVMBuildICmp(ctx->ac.builder, LLVMIntNE, cc, ctx->ac.i32_0, "uniform_active2");
555   ac_build_ifcc(&ctx->ac, active, 6002);
556   ac_build_break(&ctx->ac);
557   ac_build_endif(&ctx->ac, 6002);
558
559   ac_build_endloop(&ctx->ac, 6000);
560   return ret;
561}
562
563static void visit_alu(struct ac_nir_context *ctx, const nir_alu_instr *instr)
564{
565   LLVMValueRef src[16], result = NULL;
566   unsigned num_components = instr->dest.dest.ssa.num_components;
567   unsigned src_components;
568   LLVMTypeRef def_type = get_def_type(ctx, &instr->dest.dest.ssa);
569
570   assert(nir_op_infos[instr->op].num_inputs <= ARRAY_SIZE(src));
571   switch (instr->op) {
572   case nir_op_vec2:
573   case nir_op_vec3:
574   case nir_op_vec4:
575   case nir_op_vec5:
576   case nir_op_vec8:
577   case nir_op_vec16:
578   case nir_op_unpack_32_2x16:
579   case nir_op_unpack_64_2x32:
580   case nir_op_unpack_64_4x16:
581      src_components = 1;
582      break;
583   case nir_op_pack_half_2x16:
584   case nir_op_pack_snorm_2x16:
585   case nir_op_pack_unorm_2x16:
586   case nir_op_pack_uint_2x16:
587   case nir_op_pack_sint_2x16:
588   case nir_op_pack_32_2x16:
589   case nir_op_pack_64_2x32:
590      src_components = 2;
591      break;
592   case nir_op_unpack_half_2x16:
593      src_components = 1;
594      break;
595   case nir_op_cube_face_coord_amd:
596   case nir_op_cube_face_index_amd:
597      src_components = 3;
598      break;
599   case nir_op_pack_32_4x8:
600   case nir_op_pack_64_4x16:
601      src_components = 4;
602      break;
603   default:
604      src_components = num_components;
605      break;
606   }
607   for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
608      src[i] = get_alu_src(ctx, instr->src[i], src_components);
609
610   switch (instr->op) {
611   case nir_op_mov:
612      result = src[0];
613      break;
614   case nir_op_fneg:
615      src[0] = ac_to_float(&ctx->ac, src[0]);
616      result = LLVMBuildFNeg(ctx->ac.builder, src[0], "");
617      if (ctx->ac.float_mode == AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO) {
618         /* fneg will be optimized by backend compiler with sign
619          * bit removed via XOR. This is probably a LLVM bug.
620          */
621         result = ac_build_canonicalize(&ctx->ac, result, instr->dest.dest.ssa.bit_size);
622      }
623      break;
624   case nir_op_ineg:
625      if (instr->no_unsigned_wrap)
626         result = LLVMBuildNUWNeg(ctx->ac.builder, src[0], "");
627      else if (instr->no_signed_wrap)
628         result = LLVMBuildNSWNeg(ctx->ac.builder, src[0], "");
629      else
630         result = LLVMBuildNeg(ctx->ac.builder, src[0], "");
631      break;
632   case nir_op_inot:
633      result = LLVMBuildNot(ctx->ac.builder, src[0], "");
634      break;
635   case nir_op_iadd:
636      if (instr->no_unsigned_wrap)
637         result = LLVMBuildNUWAdd(ctx->ac.builder, src[0], src[1], "");
638      else if (instr->no_signed_wrap)
639         result = LLVMBuildNSWAdd(ctx->ac.builder, src[0], src[1], "");
640      else
641         result = LLVMBuildAdd(ctx->ac.builder, src[0], src[1], "");
642      break;
643   case nir_op_uadd_sat:
644   case nir_op_iadd_sat: {
645      char name[64], type[64];
646      ac_build_type_name_for_intr(def_type, type, sizeof(type));
647      snprintf(name, sizeof(name), "llvm.%cadd.sat.%s",
648               instr->op == nir_op_uadd_sat ? 'u' : 's', type);
649      result = ac_build_intrinsic(&ctx->ac, name, def_type, src, 2, AC_FUNC_ATTR_READNONE);
650      break;
651   }
652   case nir_op_usub_sat:
653   case nir_op_isub_sat: {
654      char name[64], type[64];
655      ac_build_type_name_for_intr(def_type, type, sizeof(type));
656      snprintf(name, sizeof(name), "llvm.%csub.sat.%s",
657               instr->op == nir_op_usub_sat ? 'u' : 's', type);
658      result = ac_build_intrinsic(&ctx->ac, name, def_type, src, 2, AC_FUNC_ATTR_READNONE);
659      break;
660   }
661   case nir_op_fadd:
662      src[0] = ac_to_float(&ctx->ac, src[0]);
663      src[1] = ac_to_float(&ctx->ac, src[1]);
664      result = LLVMBuildFAdd(ctx->ac.builder, src[0], src[1], "");
665      break;
666   case nir_op_fsub:
667      src[0] = ac_to_float(&ctx->ac, src[0]);
668      src[1] = ac_to_float(&ctx->ac, src[1]);
669      result = LLVMBuildFSub(ctx->ac.builder, src[0], src[1], "");
670      break;
671   case nir_op_isub:
672      if (instr->no_unsigned_wrap)
673         result = LLVMBuildNUWSub(ctx->ac.builder, src[0], src[1], "");
674      else if (instr->no_signed_wrap)
675         result = LLVMBuildNSWSub(ctx->ac.builder, src[0], src[1], "");
676      else
677         result = LLVMBuildSub(ctx->ac.builder, src[0], src[1], "");
678      break;
679   case nir_op_imul:
680      if (instr->no_unsigned_wrap)
681         result = LLVMBuildNUWMul(ctx->ac.builder, src[0], src[1], "");
682      else if (instr->no_signed_wrap)
683         result = LLVMBuildNSWMul(ctx->ac.builder, src[0], src[1], "");
684      else
685         result = LLVMBuildMul(ctx->ac.builder, src[0], src[1], "");
686      break;
687   case nir_op_imod:
688      result = LLVMBuildSRem(ctx->ac.builder, src[0], src[1], "");
689      break;
690   case nir_op_umod:
691      result = LLVMBuildURem(ctx->ac.builder, src[0], src[1], "");
692      break;
693   case nir_op_irem:
694      result = LLVMBuildSRem(ctx->ac.builder, src[0], src[1], "");
695      break;
696   case nir_op_idiv:
697      result = LLVMBuildSDiv(ctx->ac.builder, src[0], src[1], "");
698      break;
699   case nir_op_udiv:
700      result = LLVMBuildUDiv(ctx->ac.builder, src[0], src[1], "");
701      break;
702   case nir_op_fmul:
703      src[0] = ac_to_float(&ctx->ac, src[0]);
704      src[1] = ac_to_float(&ctx->ac, src[1]);
705      result = LLVMBuildFMul(ctx->ac.builder, src[0], src[1], "");
706      break;
707   case nir_op_fmulz:
708      assert(LLVM_VERSION_MAJOR >= 12);
709      src[0] = ac_to_float(&ctx->ac, src[0]);
710      src[1] = ac_to_float(&ctx->ac, src[1]);
711      result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.fmul.legacy", ctx->ac.f32,
712                                  src, 2, AC_FUNC_ATTR_READNONE);
713      break;
714   case nir_op_frcp:
715      /* For doubles, we need precise division to pass GLCTS. */
716      if (ctx->ac.float_mode == AC_FLOAT_MODE_DEFAULT_OPENGL && ac_get_type_size(def_type) == 8) {
717         result = LLVMBuildFDiv(ctx->ac.builder, ctx->ac.f64_1, ac_to_float(&ctx->ac, src[0]), "");
718      } else {
719         result = emit_intrin_1f_param_scalar(&ctx->ac, "llvm.amdgcn.rcp",
720                                              ac_to_float_type(&ctx->ac, def_type), src[0]);
721      }
722      if (ctx->abi->clamp_div_by_zero)
723         result = ac_build_fmin(&ctx->ac, result,
724                                LLVMConstReal(ac_to_float_type(&ctx->ac, def_type), FLT_MAX));
725      break;
726   case nir_op_iand:
727      result = LLVMBuildAnd(ctx->ac.builder, src[0], src[1], "");
728      break;
729   case nir_op_ior:
730      result = LLVMBuildOr(ctx->ac.builder, src[0], src[1], "");
731      break;
732   case nir_op_ixor:
733      result = LLVMBuildXor(ctx->ac.builder, src[0], src[1], "");
734      break;
735   case nir_op_ishl:
736      if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) <
737          ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
738         src[1] = LLVMBuildZExt(ctx->ac.builder, src[1], LLVMTypeOf(src[0]), "");
739      else if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) >
740               ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
741         src[1] = LLVMBuildTrunc(ctx->ac.builder, src[1], LLVMTypeOf(src[0]), "");
742      result = LLVMBuildShl(ctx->ac.builder, src[0], src[1], "");
743      break;
744   case nir_op_ishr:
745      if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) <
746          ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
747         src[1] = LLVMBuildZExt(ctx->ac.builder, src[1], LLVMTypeOf(src[0]), "");
748      else if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) >
749               ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
750         src[1] = LLVMBuildTrunc(ctx->ac.builder, src[1], LLVMTypeOf(src[0]), "");
751      result = LLVMBuildAShr(ctx->ac.builder, src[0], src[1], "");
752      break;
753   case nir_op_ushr:
754      if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) <
755          ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
756         src[1] = LLVMBuildZExt(ctx->ac.builder, src[1], LLVMTypeOf(src[0]), "");
757      else if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) >
758               ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
759         src[1] = LLVMBuildTrunc(ctx->ac.builder, src[1], LLVMTypeOf(src[0]), "");
760      result = LLVMBuildLShr(ctx->ac.builder, src[0], src[1], "");
761      break;
762   case nir_op_ilt:
763      result = emit_int_cmp(&ctx->ac, LLVMIntSLT, src[0], src[1]);
764      break;
765   case nir_op_ine:
766      result = emit_int_cmp(&ctx->ac, LLVMIntNE, src[0], src[1]);
767      break;
768   case nir_op_ieq:
769      result = emit_int_cmp(&ctx->ac, LLVMIntEQ, src[0], src[1]);
770      break;
771   case nir_op_ige:
772      result = emit_int_cmp(&ctx->ac, LLVMIntSGE, src[0], src[1]);
773      break;
774   case nir_op_ult:
775      result = emit_int_cmp(&ctx->ac, LLVMIntULT, src[0], src[1]);
776      break;
777   case nir_op_uge:
778      result = emit_int_cmp(&ctx->ac, LLVMIntUGE, src[0], src[1]);
779      break;
780   case nir_op_feq:
781      result = emit_float_cmp(&ctx->ac, LLVMRealOEQ, src[0], src[1]);
782      break;
783   case nir_op_fneu:
784      result = emit_float_cmp(&ctx->ac, LLVMRealUNE, src[0], src[1]);
785      break;
786   case nir_op_flt:
787      result = emit_float_cmp(&ctx->ac, LLVMRealOLT, src[0], src[1]);
788      break;
789   case nir_op_fge:
790      result = emit_float_cmp(&ctx->ac, LLVMRealOGE, src[0], src[1]);
791      break;
792   case nir_op_fabs:
793      result =
794         emit_intrin_1f_param(&ctx->ac, "llvm.fabs", ac_to_float_type(&ctx->ac, def_type), src[0]);
795      if (ctx->ac.float_mode == AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO) {
796         /* fabs will be optimized by backend compiler with sign
797          * bit removed via AND.
798          */
799         result = ac_build_canonicalize(&ctx->ac, result, instr->dest.dest.ssa.bit_size);
800      }
801      break;
802   case nir_op_fsat:
803      src[0] = ac_to_float(&ctx->ac, src[0]);
804      result = ac_build_fsat(&ctx->ac, src[0],
805                             ac_to_float_type(&ctx->ac, def_type));
806      break;
807   case nir_op_iabs:
808      result = emit_iabs(&ctx->ac, src[0]);
809      break;
810   case nir_op_imax:
811      result = ac_build_imax(&ctx->ac, src[0], src[1]);
812      break;
813   case nir_op_imin:
814      result = ac_build_imin(&ctx->ac, src[0], src[1]);
815      break;
816   case nir_op_umax:
817      result = ac_build_umax(&ctx->ac, src[0], src[1]);
818      break;
819   case nir_op_umin:
820      result = ac_build_umin(&ctx->ac, src[0], src[1]);
821      break;
822   case nir_op_isign:
823      result = ac_build_isign(&ctx->ac, src[0]);
824      break;
825   case nir_op_fsign:
826      src[0] = ac_to_float(&ctx->ac, src[0]);
827      result = ac_build_fsign(&ctx->ac, src[0]);
828      break;
829   case nir_op_ffloor:
830      result =
831         emit_intrin_1f_param(&ctx->ac, "llvm.floor", ac_to_float_type(&ctx->ac, def_type), src[0]);
832      break;
833   case nir_op_ftrunc:
834      result =
835         emit_intrin_1f_param(&ctx->ac, "llvm.trunc", ac_to_float_type(&ctx->ac, def_type), src[0]);
836      break;
837   case nir_op_fceil:
838      result =
839         emit_intrin_1f_param(&ctx->ac, "llvm.ceil", ac_to_float_type(&ctx->ac, def_type), src[0]);
840      break;
841   case nir_op_fround_even:
842      result =
843         emit_intrin_1f_param(&ctx->ac, "llvm.rint", ac_to_float_type(&ctx->ac, def_type), src[0]);
844      break;
845   case nir_op_ffract:
846      result = emit_intrin_1f_param_scalar(&ctx->ac, "llvm.amdgcn.fract",
847                                           ac_to_float_type(&ctx->ac, def_type), src[0]);
848      break;
849   case nir_op_fsin:
850      result =
851         emit_intrin_1f_param(&ctx->ac, "llvm.sin", ac_to_float_type(&ctx->ac, def_type), src[0]);
852      break;
853   case nir_op_fcos:
854      result =
855         emit_intrin_1f_param(&ctx->ac, "llvm.cos", ac_to_float_type(&ctx->ac, def_type), src[0]);
856      break;
857   case nir_op_fsin_amd:
858   case nir_op_fcos_amd:
859      /* before GFX9, v_sin_f32 and v_cos_f32 had a valid input domain of [-256, +256] */
860      if (ctx->ac.gfx_level < GFX9)
861         src[0] = emit_intrin_1f_param_scalar(&ctx->ac, "llvm.amdgcn.fract",
862                                              ac_to_float_type(&ctx->ac, def_type), src[0]);
863      result =
864         emit_intrin_1f_param(&ctx->ac, instr->op == nir_op_fsin_amd ? "llvm.amdgcn.sin" : "llvm.amdgcn.cos",
865                              ac_to_float_type(&ctx->ac, def_type), src[0]);
866      break;
867   case nir_op_fsqrt:
868      result =
869         emit_intrin_1f_param(&ctx->ac, "llvm.sqrt", ac_to_float_type(&ctx->ac, def_type), src[0]);
870      break;
871   case nir_op_fexp2:
872      result =
873         emit_intrin_1f_param(&ctx->ac, "llvm.exp2", ac_to_float_type(&ctx->ac, def_type), src[0]);
874      break;
875   case nir_op_flog2:
876      result =
877         emit_intrin_1f_param(&ctx->ac, "llvm.log2", ac_to_float_type(&ctx->ac, def_type), src[0]);
878      break;
879   case nir_op_frsq:
880      result = emit_intrin_1f_param_scalar(&ctx->ac, "llvm.amdgcn.rsq",
881                                           ac_to_float_type(&ctx->ac, def_type), src[0]);
882      if (ctx->abi->clamp_div_by_zero)
883         result = ac_build_fmin(&ctx->ac, result,
884                                LLVMConstReal(ac_to_float_type(&ctx->ac, def_type), FLT_MAX));
885      break;
886   case nir_op_frexp_exp:
887      src[0] = ac_to_float(&ctx->ac, src[0]);
888      result = ac_build_frexp_exp(&ctx->ac, src[0], ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])));
889      if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) == 16)
890         result = LLVMBuildSExt(ctx->ac.builder, result, ctx->ac.i32, "");
891      break;
892   case nir_op_frexp_sig:
893      src[0] = ac_to_float(&ctx->ac, src[0]);
894      result = ac_build_frexp_mant(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
895      break;
896   case nir_op_fpow:
897      if (instr->dest.dest.ssa.bit_size != 32) {
898         /* 16 and 64 bits */
899         result = emit_intrin_1f_param(&ctx->ac, "llvm.log2",
900                                       ac_to_float_type(&ctx->ac, def_type), src[0]);
901         result = LLVMBuildFMul(ctx->ac.builder, result, ac_to_float(&ctx->ac, src[1]), "");
902         result = emit_intrin_1f_param(&ctx->ac, "llvm.exp2",
903                                       ac_to_float_type(&ctx->ac, def_type), result);
904         break;
905      }
906      if (LLVM_VERSION_MAJOR >= 12) {
907         result = emit_intrin_1f_param(&ctx->ac, "llvm.log2",
908                                       ac_to_float_type(&ctx->ac, def_type), src[0]);
909         result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.fmul.legacy", ctx->ac.f32,
910                                     (LLVMValueRef[]){result, ac_to_float(&ctx->ac, src[1])},
911                                     2, AC_FUNC_ATTR_READNONE);
912         result = emit_intrin_1f_param(&ctx->ac, "llvm.exp2",
913                                       ac_to_float_type(&ctx->ac, def_type), result);
914         break;
915      }
916      /* Older LLVM doesn't have fmul.legacy. */
917      result = emit_intrin_2f_param(&ctx->ac, "llvm.pow", ac_to_float_type(&ctx->ac, def_type),
918                                    src[0], src[1]);
919      break;
920   case nir_op_fmax:
921      result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum", ac_to_float_type(&ctx->ac, def_type),
922                                    src[0], src[1]);
923      if (ctx->ac.gfx_level < GFX9 && instr->dest.dest.ssa.bit_size == 32) {
924         /* Only pre-GFX9 chips do not flush denorms. */
925         result = ac_build_canonicalize(&ctx->ac, result, instr->dest.dest.ssa.bit_size);
926      }
927      break;
928   case nir_op_fmin:
929      result = emit_intrin_2f_param(&ctx->ac, "llvm.minnum", ac_to_float_type(&ctx->ac, def_type),
930                                    src[0], src[1]);
931      if (ctx->ac.gfx_level < GFX9 && instr->dest.dest.ssa.bit_size == 32) {
932         /* Only pre-GFX9 chips do not flush denorms. */
933         result = ac_build_canonicalize(&ctx->ac, result, instr->dest.dest.ssa.bit_size);
934      }
935      break;
936   case nir_op_ffma:
937      /* FMA is slow on gfx6-8, so it shouldn't be used. */
938      assert(instr->dest.dest.ssa.bit_size != 32 || ctx->ac.gfx_level >= GFX9);
939      result = emit_intrin_3f_param(&ctx->ac, "llvm.fma", ac_to_float_type(&ctx->ac, def_type),
940                                    src[0], src[1], src[2]);
941      break;
942   case nir_op_ffmaz:
943      assert(LLVM_VERSION_MAJOR >= 12 && ctx->ac.gfx_level >= GFX10_3);
944      src[0] = ac_to_float(&ctx->ac, src[0]);
945      src[1] = ac_to_float(&ctx->ac, src[1]);
946      src[2] = ac_to_float(&ctx->ac, src[2]);
947      result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.fma.legacy", ctx->ac.f32,
948                                  src, 3, AC_FUNC_ATTR_READNONE);
949      break;
950   case nir_op_ldexp:
951      src[0] = ac_to_float(&ctx->ac, src[0]);
952      if (ac_get_elem_bits(&ctx->ac, def_type) == 32)
953         result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ldexp.f32", ctx->ac.f32, src, 2,
954                                     AC_FUNC_ATTR_READNONE);
955      else if (ac_get_elem_bits(&ctx->ac, def_type) == 16)
956         result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ldexp.f16", ctx->ac.f16, src, 2,
957                                     AC_FUNC_ATTR_READNONE);
958      else
959         result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ldexp.f64", ctx->ac.f64, src, 2,
960                                     AC_FUNC_ATTR_READNONE);
961      break;
962   case nir_op_bfm:
963      result = emit_bfm(&ctx->ac, src[0], src[1]);
964      break;
965   case nir_op_bitfield_select:
966      result = emit_bitfield_select(&ctx->ac, src[0], src[1], src[2]);
967      break;
968   case nir_op_ubfe:
969      result = ac_build_bfe(&ctx->ac, src[0], src[1], src[2], false);
970      break;
971   case nir_op_ibfe:
972      result = ac_build_bfe(&ctx->ac, src[0], src[1], src[2], true);
973      break;
974   case nir_op_bitfield_reverse:
975      result = ac_build_bitfield_reverse(&ctx->ac, src[0]);
976      break;
977   case nir_op_bit_count:
978      result = ac_build_bit_count(&ctx->ac, src[0]);
979      break;
980   case nir_op_vec2:
981   case nir_op_vec3:
982   case nir_op_vec4:
983   case nir_op_vec5:
984   case nir_op_vec8:
985   case nir_op_vec16:
986      for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
987         src[i] = ac_to_integer(&ctx->ac, src[i]);
988      result = ac_build_gather_values(&ctx->ac, src, num_components);
989      break;
990   case nir_op_f2i8:
991   case nir_op_f2i16:
992   case nir_op_f2imp:
993   case nir_op_f2i32:
994   case nir_op_f2i64:
995      src[0] = ac_to_float(&ctx->ac, src[0]);
996      result = LLVMBuildFPToSI(ctx->ac.builder, src[0], def_type, "");
997      break;
998   case nir_op_f2u8:
999   case nir_op_f2u16:
1000   case nir_op_f2ump:
1001   case nir_op_f2u32:
1002   case nir_op_f2u64:
1003      src[0] = ac_to_float(&ctx->ac, src[0]);
1004      result = LLVMBuildFPToUI(ctx->ac.builder, src[0], def_type, "");
1005      break;
1006   case nir_op_i2f16:
1007   case nir_op_i2fmp:
1008   case nir_op_i2f32:
1009   case nir_op_i2f64:
1010      result = LLVMBuildSIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
1011      break;
1012   case nir_op_u2f16:
1013   case nir_op_u2fmp:
1014   case nir_op_u2f32:
1015   case nir_op_u2f64:
1016      result = LLVMBuildUIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
1017      break;
1018   case nir_op_f2f16_rtz:
1019   case nir_op_f2f16:
1020   case nir_op_f2fmp:
1021      src[0] = ac_to_float(&ctx->ac, src[0]);
1022
1023      /* For OpenGL, we want fast packing with v_cvt_pkrtz_f16, but if we use it,
1024       * all f32->f16 conversions have to round towards zero, because both scalar
1025       * and vec2 down-conversions have to round equally.
1026       */
1027      if (ctx->ac.float_mode == AC_FLOAT_MODE_DEFAULT_OPENGL || instr->op == nir_op_f2f16_rtz) {
1028         src[0] = ac_to_float(&ctx->ac, src[0]);
1029
1030         if (LLVMTypeOf(src[0]) == ctx->ac.f64)
1031            src[0] = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ctx->ac.f32, "");
1032
1033         /* Fast path conversion. This only works if NIR is vectorized
1034          * to vec2 16.
1035          */
1036         if (LLVMTypeOf(src[0]) == ctx->ac.v2f32) {
1037            LLVMValueRef args[] = {
1038               ac_llvm_extract_elem(&ctx->ac, src[0], 0),
1039               ac_llvm_extract_elem(&ctx->ac, src[0], 1),
1040            };
1041            result = ac_build_cvt_pkrtz_f16(&ctx->ac, args);
1042            break;
1043         }
1044
1045         assert(ac_get_llvm_num_components(src[0]) == 1);
1046         LLVMValueRef param[2] = {src[0], LLVMGetUndef(ctx->ac.f32)};
1047         result = ac_build_cvt_pkrtz_f16(&ctx->ac, param);
1048         result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
1049      } else {
1050         if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
1051            result =
1052               LLVMBuildFPExt(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
1053         else
1054            result =
1055               LLVMBuildFPTrunc(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
1056      }
1057      break;
1058   case nir_op_f2f16_rtne:
1059   case nir_op_f2f32:
1060   case nir_op_f2f64:
1061      src[0] = ac_to_float(&ctx->ac, src[0]);
1062      if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
1063         result = LLVMBuildFPExt(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
1064      else
1065         result =
1066            LLVMBuildFPTrunc(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
1067      break;
1068   case nir_op_u2u8:
1069   case nir_op_u2u16:
1070   case nir_op_u2u32:
1071   case nir_op_u2u64:
1072      if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
1073         result = LLVMBuildZExt(ctx->ac.builder, src[0], def_type, "");
1074      else
1075         result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
1076      break;
1077   case nir_op_i2i8:
1078   case nir_op_i2i16:
1079   case nir_op_i2imp:
1080   case nir_op_i2i32:
1081   case nir_op_i2i64:
1082      if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
1083         result = LLVMBuildSExt(ctx->ac.builder, src[0], def_type, "");
1084      else
1085         result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
1086      break;
1087   case nir_op_bcsel:
1088      result = emit_bcsel(&ctx->ac, src[0], src[1], src[2]);
1089      break;
1090   case nir_op_find_lsb:
1091      result = ac_find_lsb(&ctx->ac, ctx->ac.i32, src[0]);
1092      break;
1093   case nir_op_ufind_msb:
1094      result = ac_build_umsb(&ctx->ac, src[0], ctx->ac.i32);
1095      break;
1096   case nir_op_ifind_msb:
1097      result = ac_build_imsb(&ctx->ac, src[0], ctx->ac.i32);
1098      break;
1099  case nir_op_uclz: {
1100      LLVMValueRef params[2] = {
1101         src[0],
1102         ctx->ac.i1false,
1103      };
1104      result = ac_build_intrinsic(&ctx->ac, "llvm.ctlz.i32", ctx->ac.i32, params, 2, AC_FUNC_ATTR_READNONE);
1105      break;
1106  }
1107   case nir_op_uadd_carry:
1108      result = emit_uint_carry(&ctx->ac, "llvm.uadd.with.overflow.i32", src[0], src[1]);
1109      break;
1110   case nir_op_usub_borrow:
1111      result = emit_uint_carry(&ctx->ac, "llvm.usub.with.overflow.i32", src[0], src[1]);
1112      break;
1113   case nir_op_b2f16:
1114   case nir_op_b2f32:
1115   case nir_op_b2f64:
1116      result = emit_b2f(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
1117      break;
1118   case nir_op_f2b1:
1119      result = emit_f2b(&ctx->ac, src[0]);
1120      break;
1121   case nir_op_b2i8:
1122   case nir_op_b2i16:
1123   case nir_op_b2i32:
1124   case nir_op_b2i64:
1125      result = emit_b2i(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
1126      break;
1127   case nir_op_i2b1:
1128   case nir_op_b2b1: /* after loads */
1129      result = emit_i2b(&ctx->ac, src[0]);
1130      break;
1131   case nir_op_b2b16: /* before stores */
1132      result = LLVMBuildZExt(ctx->ac.builder, src[0], ctx->ac.i16, "");
1133      break;
1134   case nir_op_b2b32: /* before stores */
1135      result = LLVMBuildZExt(ctx->ac.builder, src[0], ctx->ac.i32, "");
1136      break;
1137   case nir_op_fquantize2f16:
1138      result = emit_f2f16(&ctx->ac, src[0]);
1139      break;
1140   case nir_op_umul_high:
1141      result = emit_umul_high(&ctx->ac, src[0], src[1]);
1142      break;
1143   case nir_op_imul_high:
1144      result = emit_imul_high(&ctx->ac, src[0], src[1]);
1145      break;
1146   case nir_op_pack_half_2x16:
1147      result = emit_pack_2x16(&ctx->ac, src[0], ac_build_cvt_pkrtz_f16);
1148      break;
1149   case nir_op_pack_half_2x16_split:
1150      src[0] = ac_to_float(&ctx->ac, src[0]);
1151      src[1] = ac_to_float(&ctx->ac, src[1]);
1152      result = LLVMBuildBitCast(ctx->ac.builder,
1153                                ac_build_cvt_pkrtz_f16(&ctx->ac, src),
1154                                ctx->ac.i32, "");
1155      break;
1156   case nir_op_pack_snorm_2x16:
1157      result = emit_pack_2x16(&ctx->ac, src[0], ac_build_cvt_pknorm_i16);
1158      break;
1159   case nir_op_pack_unorm_2x16:
1160      result = emit_pack_2x16(&ctx->ac, src[0], ac_build_cvt_pknorm_u16);
1161      break;
1162   case nir_op_pack_uint_2x16: {
1163      LLVMValueRef comp[2];
1164
1165      comp[0] = LLVMBuildExtractElement(ctx->ac.builder, src[0], ctx->ac.i32_0, "");
1166      comp[1] = LLVMBuildExtractElement(ctx->ac.builder, src[0], ctx->ac.i32_1, "");
1167
1168      result = ac_build_cvt_pk_u16(&ctx->ac, comp, 16, false);
1169      break;
1170   }
1171   case nir_op_pack_sint_2x16: {
1172      LLVMValueRef comp[2];
1173
1174      comp[0] = LLVMBuildExtractElement(ctx->ac.builder, src[0], ctx->ac.i32_0, "");
1175      comp[1] = LLVMBuildExtractElement(ctx->ac.builder, src[0], ctx->ac.i32_1, "");
1176
1177      result = ac_build_cvt_pk_i16(&ctx->ac, comp, 16, false);
1178      break;
1179   }
1180   case nir_op_unpack_half_2x16:
1181      result = emit_unpack_half_2x16(&ctx->ac, src[0]);
1182      break;
1183   case nir_op_unpack_half_2x16_split_x: {
1184      assert(ac_get_llvm_num_components(src[0]) == 1);
1185      LLVMValueRef tmp = emit_unpack_half_2x16(&ctx->ac, src[0]);
1186      result = LLVMBuildExtractElement(ctx->ac.builder, tmp, ctx->ac.i32_0, "");
1187      break;
1188   }
1189   case nir_op_unpack_half_2x16_split_y: {
1190      assert(ac_get_llvm_num_components(src[0]) == 1);
1191      LLVMValueRef tmp = emit_unpack_half_2x16(&ctx->ac, src[0]);
1192      result = LLVMBuildExtractElement(ctx->ac.builder, tmp, ctx->ac.i32_1, "");
1193      break;
1194   }
1195   case nir_op_fddx:
1196   case nir_op_fddy:
1197   case nir_op_fddx_fine:
1198   case nir_op_fddy_fine:
1199   case nir_op_fddx_coarse:
1200   case nir_op_fddy_coarse:
1201      result = emit_ddxy(ctx, instr->op, src[0]);
1202      break;
1203
1204   case nir_op_unpack_64_4x16: {
1205      result = LLVMBuildBitCast(ctx->ac.builder, src[0], ctx->ac.v4i16, "");
1206      break;
1207   }
1208   case nir_op_pack_64_4x16: {
1209      result = LLVMBuildBitCast(ctx->ac.builder, src[0], ctx->ac.i64, "");
1210      break;
1211   }
1212
1213   case nir_op_unpack_64_2x32: {
1214      result = LLVMBuildBitCast(ctx->ac.builder, src[0],
1215            ctx->ac.v2i32, "");
1216      break;
1217   }
1218   case nir_op_unpack_64_2x32_split_x: {
1219      assert(ac_get_llvm_num_components(src[0]) == 1);
1220      LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0], ctx->ac.v2i32, "");
1221      result = LLVMBuildExtractElement(ctx->ac.builder, tmp, ctx->ac.i32_0, "");
1222      break;
1223   }
1224   case nir_op_unpack_64_2x32_split_y: {
1225      assert(ac_get_llvm_num_components(src[0]) == 1);
1226      LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0], ctx->ac.v2i32, "");
1227      result = LLVMBuildExtractElement(ctx->ac.builder, tmp, ctx->ac.i32_1, "");
1228      break;
1229   }
1230
1231   case nir_op_pack_64_2x32: {
1232      result = LLVMBuildBitCast(ctx->ac.builder, src[0],
1233            ctx->ac.i64, "");
1234      break;
1235   }
1236   case nir_op_pack_64_2x32_split: {
1237      LLVMValueRef tmp = ac_build_gather_values(&ctx->ac, src, 2);
1238      result = LLVMBuildBitCast(ctx->ac.builder, tmp, ctx->ac.i64, "");
1239      break;
1240   }
1241
1242   case nir_op_pack_32_4x8:
1243   case nir_op_pack_32_2x16: {
1244      result = LLVMBuildBitCast(ctx->ac.builder, src[0],
1245            ctx->ac.i32, "");
1246      break;
1247   }
1248   case nir_op_pack_32_2x16_split: {
1249      LLVMValueRef tmp = ac_build_gather_values(&ctx->ac, src, 2);
1250      result = LLVMBuildBitCast(ctx->ac.builder, tmp, ctx->ac.i32, "");
1251      break;
1252   }
1253
1254   case nir_op_unpack_32_2x16: {
1255      result = LLVMBuildBitCast(ctx->ac.builder, src[0],
1256            ctx->ac.v2i16, "");
1257      break;
1258   }
1259   case nir_op_unpack_32_2x16_split_x: {
1260      LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0], ctx->ac.v2i16, "");
1261      result = LLVMBuildExtractElement(ctx->ac.builder, tmp, ctx->ac.i32_0, "");
1262      break;
1263   }
1264   case nir_op_unpack_32_2x16_split_y: {
1265      LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0], ctx->ac.v2i16, "");
1266      result = LLVMBuildExtractElement(ctx->ac.builder, tmp, ctx->ac.i32_1, "");
1267      break;
1268   }
1269
1270   case nir_op_cube_face_coord_amd: {
1271      src[0] = ac_to_float(&ctx->ac, src[0]);
1272      LLVMValueRef results[2];
1273      LLVMValueRef in[3];
1274      for (unsigned chan = 0; chan < 3; chan++)
1275         in[chan] = ac_llvm_extract_elem(&ctx->ac, src[0], chan);
1276      results[0] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubesc", ctx->ac.f32, in, 3,
1277                                      AC_FUNC_ATTR_READNONE);
1278      results[1] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubetc", ctx->ac.f32, in, 3,
1279                                      AC_FUNC_ATTR_READNONE);
1280      LLVMValueRef ma = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubema", ctx->ac.f32, in, 3,
1281                                           AC_FUNC_ATTR_READNONE);
1282      results[0] = ac_build_fdiv(&ctx->ac, results[0], ma);
1283      results[1] = ac_build_fdiv(&ctx->ac, results[1], ma);
1284      LLVMValueRef offset = LLVMConstReal(ctx->ac.f32, 0.5);
1285      results[0] = LLVMBuildFAdd(ctx->ac.builder, results[0], offset, "");
1286      results[1] = LLVMBuildFAdd(ctx->ac.builder, results[1], offset, "");
1287      result = ac_build_gather_values(&ctx->ac, results, 2);
1288      break;
1289   }
1290
1291   case nir_op_cube_face_index_amd: {
1292      src[0] = ac_to_float(&ctx->ac, src[0]);
1293      LLVMValueRef in[3];
1294      for (unsigned chan = 0; chan < 3; chan++)
1295         in[chan] = ac_llvm_extract_elem(&ctx->ac, src[0], chan);
1296      result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubeid", ctx->ac.f32, in, 3,
1297                                  AC_FUNC_ATTR_READNONE);
1298      break;
1299   }
1300
1301   case nir_op_extract_u8:
1302   case nir_op_extract_i8:
1303   case nir_op_extract_u16:
1304   case nir_op_extract_i16: {
1305      bool is_signed = instr->op == nir_op_extract_i16 || instr->op == nir_op_extract_i8;
1306      unsigned size = instr->op == nir_op_extract_u8 || instr->op == nir_op_extract_i8 ? 8 : 16;
1307      LLVMValueRef offset = LLVMConstInt(LLVMTypeOf(src[0]), nir_src_as_uint(instr->src[1].src) * size, false);
1308      result = LLVMBuildLShr(ctx->ac.builder, src[0], offset, "");
1309      result = LLVMBuildTrunc(ctx->ac.builder, result, LLVMIntTypeInContext(ctx->ac.context, size), "");
1310      if (is_signed)
1311         result = LLVMBuildSExt(ctx->ac.builder, result, LLVMTypeOf(src[0]), "");
1312      else
1313         result = LLVMBuildZExt(ctx->ac.builder, result, LLVMTypeOf(src[0]), "");
1314      break;
1315   }
1316
1317   case nir_op_insert_u8:
1318   case nir_op_insert_u16: {
1319      unsigned size = instr->op == nir_op_insert_u8 ? 8 : 16;
1320      LLVMValueRef offset = LLVMConstInt(LLVMTypeOf(src[0]), nir_src_as_uint(instr->src[1].src) * size, false);
1321      LLVMValueRef mask = LLVMConstInt(LLVMTypeOf(src[0]), u_bit_consecutive(0, size), false);
1322      result = LLVMBuildShl(ctx->ac.builder, LLVMBuildAnd(ctx->ac.builder, src[0], mask, ""), offset, "");
1323      break;
1324   }
1325
1326   case nir_op_sdot_4x8_iadd:
1327   case nir_op_udot_4x8_uadd:
1328   case nir_op_sdot_4x8_iadd_sat:
1329   case nir_op_udot_4x8_uadd_sat: {
1330      const char *name = instr->op == nir_op_sdot_4x8_iadd ||
1331                         instr->op == nir_op_sdot_4x8_iadd_sat
1332                         ? "llvm.amdgcn.sdot4" : "llvm.amdgcn.udot4";
1333      src[3] = LLVMConstInt(ctx->ac.i1, instr->op == nir_op_sdot_4x8_iadd_sat ||
1334                                        instr->op == nir_op_udot_4x8_uadd_sat, false);
1335      result = ac_build_intrinsic(&ctx->ac, name, def_type, src, 4, AC_FUNC_ATTR_READNONE);
1336      break;
1337   }
1338
1339   case nir_op_sdot_2x16_iadd:
1340   case nir_op_udot_2x16_uadd:
1341   case nir_op_sdot_2x16_iadd_sat:
1342   case nir_op_udot_2x16_uadd_sat: {
1343      const char *name = instr->op == nir_op_sdot_2x16_iadd ||
1344                         instr->op == nir_op_sdot_2x16_iadd_sat
1345                         ? "llvm.amdgcn.sdot2" : "llvm.amdgcn.udot2";
1346      src[0] = LLVMBuildBitCast(ctx->ac.builder, src[0], ctx->ac.v2i16, "");
1347      src[1] = LLVMBuildBitCast(ctx->ac.builder, src[1], ctx->ac.v2i16, "");
1348      src[3] = LLVMConstInt(ctx->ac.i1, instr->op == nir_op_sdot_2x16_iadd_sat ||
1349                                        instr->op == nir_op_udot_2x16_uadd_sat, false);
1350      result = ac_build_intrinsic(&ctx->ac, name, def_type, src, 4, AC_FUNC_ATTR_READNONE);
1351      break;
1352   }
1353
1354   case nir_op_sad_u8x4:
1355      result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.sad.u8", ctx->ac.i32,
1356                                  (LLVMValueRef[]){src[0], src[1], src[2]}, 3,
1357                                  AC_FUNC_ATTR_READNONE);
1358      break;
1359
1360   default:
1361      fprintf(stderr, "Unknown NIR alu instr: ");
1362      nir_print_instr(&instr->instr, stderr);
1363      fprintf(stderr, "\n");
1364      abort();
1365   }
1366
1367   if (result) {
1368      assert(instr->dest.dest.is_ssa);
1369      result = ac_to_integer_or_pointer(&ctx->ac, result);
1370      ctx->ssa_defs[instr->dest.dest.ssa.index] = result;
1371   }
1372}
1373
1374static void visit_load_const(struct ac_nir_context *ctx, const nir_load_const_instr *instr)
1375{
1376   LLVMValueRef values[4], value = NULL;
1377   LLVMTypeRef element_type = LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
1378
1379   for (unsigned i = 0; i < instr->def.num_components; ++i) {
1380      switch (instr->def.bit_size) {
1381      case 1:
1382         values[i] = LLVMConstInt(element_type, instr->value[i].b, false);
1383         break;
1384      case 8:
1385         values[i] = LLVMConstInt(element_type, instr->value[i].u8, false);
1386         break;
1387      case 16:
1388         values[i] = LLVMConstInt(element_type, instr->value[i].u16, false);
1389         break;
1390      case 32:
1391         values[i] = LLVMConstInt(element_type, instr->value[i].u32, false);
1392         break;
1393      case 64:
1394         values[i] = LLVMConstInt(element_type, instr->value[i].u64, false);
1395         break;
1396      default:
1397         fprintf(stderr, "unsupported nir load_const bit_size: %d\n", instr->def.bit_size);
1398         abort();
1399      }
1400   }
1401   if (instr->def.num_components > 1) {
1402      value = LLVMConstVector(values, instr->def.num_components);
1403   } else
1404      value = values[0];
1405
1406   ctx->ssa_defs[instr->def.index] = value;
1407}
1408
1409static LLVMValueRef get_buffer_size(struct ac_nir_context *ctx, LLVMValueRef descriptor,
1410                                    bool in_elements)
1411{
1412   LLVMValueRef size =
1413      LLVMBuildExtractElement(ctx->ac.builder, descriptor, LLVMConstInt(ctx->ac.i32, 2, false), "");
1414
1415   /* GFX8 only */
1416   if (ctx->ac.gfx_level == GFX8 && in_elements) {
1417      /* On GFX8, the descriptor contains the size in bytes,
1418       * but TXQ must return the size in elements.
1419       * The stride is always non-zero for resources using TXQ.
1420       */
1421      LLVMValueRef stride = LLVMBuildExtractElement(ctx->ac.builder, descriptor, ctx->ac.i32_1, "");
1422      stride = LLVMBuildLShr(ctx->ac.builder, stride, LLVMConstInt(ctx->ac.i32, 16, false), "");
1423      stride = LLVMBuildAnd(ctx->ac.builder, stride, LLVMConstInt(ctx->ac.i32, 0x3fff, false), "");
1424
1425      size = LLVMBuildUDiv(ctx->ac.builder, size, stride, "");
1426   }
1427   return size;
1428}
1429
1430/* Gather4 should follow the same rules as bilinear filtering, but the hardware
1431 * incorrectly forces nearest filtering if the texture format is integer.
1432 * The only effect it has on Gather4, which always returns 4 texels for
1433 * bilinear filtering, is that the final coordinates are off by 0.5 of
1434 * the texel size.
1435 *
1436 * The workaround is to subtract 0.5 from the unnormalized coordinates,
1437 * or (0.5 / size) from the normalized coordinates.
1438 *
1439 * However, cube textures with 8_8_8_8 data formats require a different
1440 * workaround of overriding the num format to USCALED/SSCALED. This would lose
1441 * precision in 32-bit data formats, so it needs to be applied dynamically at
1442 * runtime. In this case, return an i1 value that indicates whether the
1443 * descriptor was overridden (and hence a fixup of the sampler result is needed).
1444 */
1445static LLVMValueRef lower_gather4_integer(struct ac_llvm_context *ctx, struct ac_image_args *args,
1446                                          const nir_tex_instr *instr)
1447{
1448   nir_alu_type stype = nir_alu_type_get_base_type(instr->dest_type);
1449   LLVMValueRef wa_8888 = NULL;
1450   LLVMValueRef half_texel[2];
1451   LLVMValueRef result;
1452
1453   assert(stype == nir_type_int || stype == nir_type_uint);
1454
1455   if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
1456      LLVMValueRef formats;
1457      LLVMValueRef data_format;
1458      LLVMValueRef wa_formats;
1459
1460      formats = LLVMBuildExtractElement(ctx->builder, args->resource, ctx->i32_1, "");
1461
1462      data_format = LLVMBuildLShr(ctx->builder, formats, LLVMConstInt(ctx->i32, 20, false), "");
1463      data_format =
1464         LLVMBuildAnd(ctx->builder, data_format, LLVMConstInt(ctx->i32, (1u << 6) - 1, false), "");
1465      wa_8888 = LLVMBuildICmp(ctx->builder, LLVMIntEQ, data_format,
1466                              LLVMConstInt(ctx->i32, V_008F14_IMG_DATA_FORMAT_8_8_8_8, false), "");
1467
1468      uint32_t wa_num_format = stype == nir_type_uint
1469                                  ? S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_USCALED)
1470                                  : S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_SSCALED);
1471      wa_formats = LLVMBuildAnd(ctx->builder, formats,
1472                                LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT, false), "");
1473      wa_formats =
1474         LLVMBuildOr(ctx->builder, wa_formats, LLVMConstInt(ctx->i32, wa_num_format, false), "");
1475
1476      formats = LLVMBuildSelect(ctx->builder, wa_8888, wa_formats, formats, "");
1477      args->resource =
1478         LLVMBuildInsertElement(ctx->builder, args->resource, formats, ctx->i32_1, "");
1479   }
1480
1481   if (instr->sampler_dim == GLSL_SAMPLER_DIM_RECT) {
1482      assert(!wa_8888);
1483      half_texel[0] = half_texel[1] = LLVMConstReal(ctx->f32, -0.5);
1484   } else {
1485      struct ac_image_args resinfo = {0};
1486      LLVMBasicBlockRef bbs[2];
1487
1488      LLVMValueRef unnorm = NULL;
1489      LLVMValueRef default_offset = ctx->f32_0;
1490      if (instr->sampler_dim == GLSL_SAMPLER_DIM_2D && !instr->is_array) {
1491         /* In vulkan, whether the sampler uses unnormalized
1492          * coordinates or not is a dynamic property of the
1493          * sampler. Hence, to figure out whether or not we
1494          * need to divide by the texture size, we need to test
1495          * the sampler at runtime. This tests the bit set by
1496          * radv_init_sampler().
1497          */
1498         LLVMValueRef sampler0 =
1499            LLVMBuildExtractElement(ctx->builder, args->sampler, ctx->i32_0, "");
1500         sampler0 = LLVMBuildLShr(ctx->builder, sampler0, LLVMConstInt(ctx->i32, 15, false), "");
1501         sampler0 = LLVMBuildAnd(ctx->builder, sampler0, ctx->i32_1, "");
1502         unnorm = LLVMBuildICmp(ctx->builder, LLVMIntEQ, sampler0, ctx->i32_1, "");
1503         default_offset = LLVMConstReal(ctx->f32, -0.5);
1504      }
1505
1506      bbs[0] = LLVMGetInsertBlock(ctx->builder);
1507      if (wa_8888 || unnorm) {
1508         assert(!(wa_8888 && unnorm));
1509         LLVMValueRef not_needed = wa_8888 ? wa_8888 : unnorm;
1510         /* Skip the texture size query entirely if we don't need it. */
1511         ac_build_ifcc(ctx, LLVMBuildNot(ctx->builder, not_needed, ""), 2000);
1512         bbs[1] = LLVMGetInsertBlock(ctx->builder);
1513      }
1514
1515      /* Query the texture size. */
1516      resinfo.dim = ac_get_sampler_dim(ctx->gfx_level, instr->sampler_dim, instr->is_array);
1517      resinfo.opcode = ac_image_get_resinfo;
1518      resinfo.dmask = 0xf;
1519      resinfo.lod = ctx->i32_0;
1520      resinfo.resource = args->resource;
1521      resinfo.attributes = AC_FUNC_ATTR_READNONE;
1522      LLVMValueRef size = ac_build_image_opcode(ctx, &resinfo);
1523
1524      /* Compute -0.5 / size. */
1525      for (unsigned c = 0; c < 2; c++) {
1526         half_texel[c] =
1527            LLVMBuildExtractElement(ctx->builder, size, LLVMConstInt(ctx->i32, c, 0), "");
1528         half_texel[c] = LLVMBuildUIToFP(ctx->builder, half_texel[c], ctx->f32, "");
1529         half_texel[c] = ac_build_fdiv(ctx, ctx->f32_1, half_texel[c]);
1530         half_texel[c] =
1531            LLVMBuildFMul(ctx->builder, half_texel[c], LLVMConstReal(ctx->f32, -0.5), "");
1532      }
1533
1534      if (wa_8888 || unnorm) {
1535         ac_build_endif(ctx, 2000);
1536
1537         for (unsigned c = 0; c < 2; c++) {
1538            LLVMValueRef values[2] = {default_offset, half_texel[c]};
1539            half_texel[c] = ac_build_phi(ctx, ctx->f32, 2, values, bbs);
1540         }
1541      }
1542   }
1543
1544   for (unsigned c = 0; c < 2; c++) {
1545      LLVMValueRef tmp;
1546      tmp = LLVMBuildBitCast(ctx->builder, args->coords[c], ctx->f32, "");
1547      args->coords[c] = LLVMBuildFAdd(ctx->builder, tmp, half_texel[c], "");
1548   }
1549
1550   args->attributes = AC_FUNC_ATTR_READNONE;
1551   result = ac_build_image_opcode(ctx, args);
1552
1553   if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
1554      LLVMValueRef tmp, tmp2;
1555
1556      /* if the cube workaround is in place, f2i the result. */
1557      for (unsigned c = 0; c < 4; c++) {
1558         tmp = LLVMBuildExtractElement(ctx->builder, result, LLVMConstInt(ctx->i32, c, false), "");
1559         if (stype == nir_type_uint)
1560            tmp2 = LLVMBuildFPToUI(ctx->builder, tmp, ctx->i32, "");
1561         else
1562            tmp2 = LLVMBuildFPToSI(ctx->builder, tmp, ctx->i32, "");
1563         tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->i32, "");
1564         tmp2 = LLVMBuildBitCast(ctx->builder, tmp2, ctx->i32, "");
1565         tmp = LLVMBuildSelect(ctx->builder, wa_8888, tmp2, tmp, "");
1566         tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->f32, "");
1567         result =
1568            LLVMBuildInsertElement(ctx->builder, result, tmp, LLVMConstInt(ctx->i32, c, false), "");
1569      }
1570   }
1571   return result;
1572}
1573
1574static LLVMValueRef build_tex_intrinsic(struct ac_nir_context *ctx, const nir_tex_instr *instr,
1575                                        struct ac_image_args *args)
1576{
1577   assert((!args->tfe || !args->d16) && "unsupported");
1578
1579   if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
1580      unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa);
1581
1582      assert(instr->dest.is_ssa);
1583
1584      /* Buffers don't support A16. */
1585      if (args->a16)
1586         args->coords[0] = LLVMBuildZExt(ctx->ac.builder, args->coords[0], ctx->ac.i32, "");
1587
1588      return ac_build_buffer_load_format(&ctx->ac, args->resource, args->coords[0], ctx->ac.i32_0,
1589                                         util_last_bit(mask), 0, true,
1590                                         instr->dest.ssa.bit_size == 16,
1591                                         args->tfe);
1592   }
1593
1594   args->opcode = ac_image_sample;
1595
1596   switch (instr->op) {
1597   case nir_texop_txf:
1598   case nir_texop_txf_ms:
1599   case nir_texop_samples_identical:
1600      args->opcode = args->level_zero || instr->sampler_dim == GLSL_SAMPLER_DIM_MS
1601                        ? ac_image_load
1602                        : ac_image_load_mip;
1603      args->level_zero = false;
1604      break;
1605   case nir_texop_txs:
1606   case nir_texop_query_levels:
1607      args->opcode = ac_image_get_resinfo;
1608      if (!args->lod)
1609         args->lod = ctx->ac.i32_0;
1610      args->level_zero = false;
1611      break;
1612   case nir_texop_tex:
1613      if (ctx->stage != MESA_SHADER_FRAGMENT &&
1614          (ctx->stage != MESA_SHADER_COMPUTE ||
1615           ctx->info->cs.derivative_group == DERIVATIVE_GROUP_NONE)) {
1616         assert(!args->lod);
1617         args->level_zero = true;
1618      }
1619      break;
1620   case nir_texop_tg4:
1621      args->opcode = ac_image_gather4;
1622      if (!args->lod && !args->bias)
1623         args->level_zero = true;
1624      break;
1625   case nir_texop_lod:
1626      args->opcode = ac_image_get_lod;
1627      break;
1628   case nir_texop_fragment_fetch_amd:
1629   case nir_texop_fragment_mask_fetch_amd:
1630      args->opcode = ac_image_load;
1631      args->level_zero = false;
1632      break;
1633   default:
1634      break;
1635   }
1636
1637   /* Aldebaran doesn't have image_sample_lz, but image_sample behaves like lz. */
1638   if (!ctx->ac.has_3d_cube_border_color_mipmap)
1639      args->level_zero = false;
1640
1641   if (instr->op == nir_texop_tg4 && ctx->ac.gfx_level <= GFX8 &&
1642       (instr->dest_type & (nir_type_int | nir_type_uint))) {
1643      return lower_gather4_integer(&ctx->ac, args, instr);
1644   }
1645
1646   /* Fixup for GFX9 which allocates 1D textures as 2D. */
1647   if (instr->op == nir_texop_lod && ctx->ac.gfx_level == GFX9) {
1648      if ((args->dim == ac_image_2darray || args->dim == ac_image_2d) && !args->coords[1]) {
1649         args->coords[1] = ctx->ac.i32_0;
1650      }
1651   }
1652
1653   args->attributes = AC_FUNC_ATTR_READNONE;
1654   bool cs_derivs =
1655      ctx->stage == MESA_SHADER_COMPUTE && ctx->info->cs.derivative_group != DERIVATIVE_GROUP_NONE;
1656   if (ctx->stage == MESA_SHADER_FRAGMENT || cs_derivs) {
1657      /* Prevent texture instructions with implicit derivatives from being
1658       * sinked into branches. */
1659      switch (instr->op) {
1660      case nir_texop_tex:
1661      case nir_texop_txb:
1662      case nir_texop_lod:
1663         args->attributes |= AC_FUNC_ATTR_CONVERGENT;
1664         break;
1665      default:
1666         break;
1667      }
1668   }
1669
1670   return ac_build_image_opcode(&ctx->ac, args);
1671}
1672
1673static LLVMValueRef visit_load_push_constant(struct ac_nir_context *ctx, nir_intrinsic_instr *instr)
1674{
1675   LLVMValueRef ptr, addr;
1676   LLVMValueRef src0 = get_src(ctx, instr->src[0]);
1677   unsigned index = nir_intrinsic_base(instr);
1678
1679   addr = LLVMConstInt(ctx->ac.i32, index, 0);
1680   addr = LLVMBuildAdd(ctx->ac.builder, addr, src0, "");
1681
1682   /* Load constant values from user SGPRS when possible, otherwise
1683    * fallback to the default path that loads directly from memory.
1684    */
1685   if (LLVMIsConstant(src0) && instr->dest.ssa.bit_size >= 32) {
1686      unsigned count = instr->dest.ssa.num_components;
1687      unsigned offset = index;
1688
1689      if (instr->dest.ssa.bit_size == 64)
1690         count *= 2;
1691
1692      offset += LLVMConstIntGetZExtValue(src0);
1693      offset /= 4;
1694
1695      uint64_t mask = BITFIELD64_MASK(count) << offset;
1696      if ((ctx->args->inline_push_const_mask | mask) == ctx->args->inline_push_const_mask &&
1697          offset + count <= (sizeof(ctx->args->inline_push_const_mask) * 8u)) {
1698         LLVMValueRef *const push_constants = alloca(count * sizeof(LLVMValueRef));
1699         unsigned arg_index =
1700            util_bitcount64(ctx->args->inline_push_const_mask & BITFIELD64_MASK(offset));
1701         for (unsigned i = 0; i < count; i++)
1702            push_constants[i] = ac_get_arg(&ctx->ac, ctx->args->inline_push_consts[arg_index++]);
1703         LLVMValueRef res = ac_build_gather_values(&ctx->ac, push_constants, count);
1704         return instr->dest.ssa.bit_size == 64
1705                   ? LLVMBuildBitCast(ctx->ac.builder, res, get_def_type(ctx, &instr->dest.ssa), "")
1706                   : res;
1707      }
1708   }
1709
1710   ptr = LLVMBuildGEP(ctx->ac.builder, ac_get_arg(&ctx->ac, ctx->args->push_constants), &addr, 1, "");
1711
1712   if (instr->dest.ssa.bit_size == 8) {
1713      unsigned load_dwords = instr->dest.ssa.num_components > 1 ? 2 : 1;
1714      LLVMTypeRef vec_type = LLVMVectorType(ctx->ac.i8, 4 * load_dwords);
1715      ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
1716      LLVMValueRef res = LLVMBuildLoad2(ctx->ac.builder, vec_type, ptr, "");
1717
1718      LLVMValueRef params[3];
1719      if (load_dwords > 1) {
1720         LLVMValueRef res_vec = LLVMBuildBitCast(ctx->ac.builder, res, ctx->ac.v2i32, "");
1721         params[0] = LLVMBuildExtractElement(ctx->ac.builder, res_vec,
1722                                             LLVMConstInt(ctx->ac.i32, 1, false), "");
1723         params[1] = LLVMBuildExtractElement(ctx->ac.builder, res_vec,
1724                                             LLVMConstInt(ctx->ac.i32, 0, false), "");
1725      } else {
1726         res = LLVMBuildBitCast(ctx->ac.builder, res, ctx->ac.i32, "");
1727         params[0] = ctx->ac.i32_0;
1728         params[1] = res;
1729      }
1730      params[2] = addr;
1731      res = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.alignbyte", ctx->ac.i32, params, 3, 0);
1732
1733      res = LLVMBuildTrunc(
1734         ctx->ac.builder, res,
1735         LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.num_components * 8), "");
1736      if (instr->dest.ssa.num_components > 1)
1737         res = LLVMBuildBitCast(ctx->ac.builder, res,
1738                                LLVMVectorType(ctx->ac.i8, instr->dest.ssa.num_components), "");
1739      return res;
1740   } else if (instr->dest.ssa.bit_size == 16) {
1741      unsigned load_dwords = instr->dest.ssa.num_components / 2 + 1;
1742      LLVMTypeRef vec_type = LLVMVectorType(ctx->ac.i16, 2 * load_dwords);
1743      ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
1744      LLVMValueRef res = LLVMBuildLoad2(ctx->ac.builder, vec_type, ptr, "");
1745      res = LLVMBuildBitCast(ctx->ac.builder, res, vec_type, "");
1746      LLVMValueRef cond = LLVMBuildLShr(ctx->ac.builder, addr, ctx->ac.i32_1, "");
1747      cond = LLVMBuildTrunc(ctx->ac.builder, cond, ctx->ac.i1, "");
1748      LLVMValueRef mask[] = {
1749         LLVMConstInt(ctx->ac.i32, 0, false), LLVMConstInt(ctx->ac.i32, 1, false),
1750         LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
1751         LLVMConstInt(ctx->ac.i32, 4, false)};
1752      LLVMValueRef swizzle_aligned = LLVMConstVector(&mask[0], instr->dest.ssa.num_components);
1753      LLVMValueRef swizzle_unaligned = LLVMConstVector(&mask[1], instr->dest.ssa.num_components);
1754      LLVMValueRef shuffle_aligned =
1755         LLVMBuildShuffleVector(ctx->ac.builder, res, res, swizzle_aligned, "");
1756      LLVMValueRef shuffle_unaligned =
1757         LLVMBuildShuffleVector(ctx->ac.builder, res, res, swizzle_unaligned, "");
1758      res = LLVMBuildSelect(ctx->ac.builder, cond, shuffle_unaligned, shuffle_aligned, "");
1759      return LLVMBuildBitCast(ctx->ac.builder, res, get_def_type(ctx, &instr->dest.ssa), "");
1760   }
1761
1762   LLVMTypeRef ptr_type = get_def_type(ctx, &instr->dest.ssa);
1763   ptr = ac_cast_ptr(&ctx->ac, ptr, ptr_type);
1764
1765   return LLVMBuildLoad2(ctx->ac.builder, ptr_type, ptr, "");
1766}
1767
1768static LLVMValueRef visit_get_ssbo_size(struct ac_nir_context *ctx,
1769                                        const nir_intrinsic_instr *instr)
1770{
1771   bool non_uniform = nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM;
1772   LLVMValueRef rsrc = ctx->abi->load_ssbo(ctx->abi, get_src(ctx, instr->src[0]), false, non_uniform);
1773   return get_buffer_size(ctx, rsrc, false);
1774}
1775
1776static LLVMValueRef extract_vector_range(struct ac_llvm_context *ctx, LLVMValueRef src,
1777                                         unsigned start, unsigned count)
1778{
1779   LLVMValueRef mask[] = {ctx->i32_0, ctx->i32_1, LLVMConstInt(ctx->i32, 2, false),
1780                          LLVMConstInt(ctx->i32, 3, false)};
1781
1782   unsigned src_elements = ac_get_llvm_num_components(src);
1783
1784   if (count == src_elements) {
1785      assert(start == 0);
1786      return src;
1787   } else if (count == 1) {
1788      assert(start < src_elements);
1789      return LLVMBuildExtractElement(ctx->builder, src, mask[start], "");
1790   } else {
1791      assert(start + count <= src_elements);
1792      assert(count <= 4);
1793      LLVMValueRef swizzle = LLVMConstVector(&mask[start], count);
1794      return LLVMBuildShuffleVector(ctx->builder, src, src, swizzle, "");
1795   }
1796}
1797
1798static unsigned get_cache_policy(struct ac_nir_context *ctx, enum gl_access_qualifier access,
1799                                 bool may_store_unaligned, bool writeonly_memory)
1800{
1801   unsigned cache_policy = 0;
1802
1803   /* GFX6 has a TC L1 bug causing corruption of 8bit/16bit stores.  All
1804    * store opcodes not aligned to a dword are affected. The only way to
1805    * get unaligned stores is through shader images.
1806    */
1807   if (((may_store_unaligned && ctx->ac.gfx_level == GFX6) ||
1808        /* If this is write-only, don't keep data in L1 to prevent
1809         * evicting L1 cache lines that may be needed by other
1810         * instructions.
1811         */
1812        writeonly_memory || access & (ACCESS_COHERENT | ACCESS_VOLATILE))) {
1813      cache_policy |= ac_glc;
1814   }
1815
1816   if (access & ACCESS_STREAM_CACHE_POLICY)
1817      cache_policy |= ac_slc | ac_glc;
1818
1819   return cache_policy;
1820}
1821
1822static LLVMValueRef enter_waterfall_ssbo(struct ac_nir_context *ctx, struct waterfall_context *wctx,
1823                                         const nir_intrinsic_instr *instr, nir_src src)
1824{
1825   return enter_waterfall(ctx, wctx, get_src(ctx, src),
1826                          nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
1827}
1828
1829static void visit_store_ssbo(struct ac_nir_context *ctx, nir_intrinsic_instr *instr)
1830{
1831   if (ctx->ac.postponed_kill) {
1832      LLVMValueRef cond = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.i1, ctx->ac.postponed_kill, "");
1833      ac_build_ifcc(&ctx->ac, cond, 7000);
1834   }
1835
1836   LLVMValueRef src_data = get_src(ctx, instr->src[0]);
1837   int elem_size_bytes = ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src_data)) / 8;
1838   unsigned writemask = nir_intrinsic_write_mask(instr);
1839   enum gl_access_qualifier access = nir_intrinsic_access(instr);
1840   bool writeonly_memory = access & ACCESS_NON_READABLE;
1841   unsigned cache_policy = get_cache_policy(ctx, access, false, writeonly_memory);
1842
1843   struct waterfall_context wctx;
1844   LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[1]);
1845
1846   LLVMValueRef rsrc = ctx->abi->load_ssbo(ctx->abi, rsrc_base, true, false);
1847   LLVMValueRef base_data = src_data;
1848   base_data = ac_trim_vector(&ctx->ac, base_data, instr->num_components);
1849   LLVMValueRef base_offset = get_src(ctx, instr->src[2]);
1850
1851   while (writemask) {
1852      int start, count;
1853      LLVMValueRef data, offset;
1854      LLVMTypeRef data_type;
1855
1856      u_bit_scan_consecutive_range(&writemask, &start, &count);
1857
1858      if (count == 3 && elem_size_bytes != 4) {
1859         writemask |= 1 << (start + 2);
1860         count = 2;
1861      }
1862      int num_bytes = count * elem_size_bytes; /* count in bytes */
1863
1864      /* we can only store 4 DWords at the same time.
1865       * can only happen for 64 Bit vectors. */
1866      if (num_bytes > 16) {
1867         writemask |= ((1u << (count - 2)) - 1u) << (start + 2);
1868         count = 2;
1869         num_bytes = 16;
1870      }
1871
1872      /* check alignment of 16 Bit stores */
1873      if (elem_size_bytes == 2 && num_bytes > 2 && (start % 2) == 1) {
1874         writemask |= ((1u << (count - 1)) - 1u) << (start + 1);
1875         count = 1;
1876         num_bytes = 2;
1877      }
1878
1879      /* Due to alignment issues, split stores of 8-bit/16-bit
1880       * vectors.
1881       */
1882      if (ctx->ac.gfx_level == GFX6 && count > 1 && elem_size_bytes < 4) {
1883         writemask |= ((1u << (count - 1)) - 1u) << (start + 1);
1884         count = 1;
1885         num_bytes = elem_size_bytes;
1886      }
1887
1888      data = extract_vector_range(&ctx->ac, base_data, start, count);
1889
1890      offset = LLVMBuildAdd(ctx->ac.builder, base_offset,
1891                            LLVMConstInt(ctx->ac.i32, start * elem_size_bytes, false), "");
1892
1893      if (num_bytes == 1) {
1894         ac_build_buffer_store_byte(&ctx->ac, rsrc, data, offset, ctx->ac.i32_0, cache_policy);
1895      } else if (num_bytes == 2) {
1896         ac_build_buffer_store_short(&ctx->ac, rsrc, data, offset, ctx->ac.i32_0, cache_policy);
1897      } else {
1898         switch (num_bytes) {
1899         case 16: /* v4f32 */
1900            data_type = ctx->ac.v4f32;
1901            break;
1902         case 12: /* v3f32 */
1903            data_type = ctx->ac.v3f32;
1904            break;
1905         case 8: /* v2f32 */
1906            data_type = ctx->ac.v2f32;
1907            break;
1908         case 4: /* f32 */
1909            data_type = ctx->ac.f32;
1910            break;
1911         default:
1912            unreachable("Malformed vector store.");
1913         }
1914         data = LLVMBuildBitCast(ctx->ac.builder, data, data_type, "");
1915
1916         ac_build_buffer_store_dword(&ctx->ac, rsrc, data, NULL, offset,
1917                                     ctx->ac.i32_0, cache_policy);
1918      }
1919   }
1920
1921   exit_waterfall(ctx, &wctx, NULL);
1922
1923   if (ctx->ac.postponed_kill)
1924      ac_build_endif(&ctx->ac, 7000);
1925}
1926
1927static LLVMValueRef emit_ssbo_comp_swap_64(struct ac_nir_context *ctx, LLVMValueRef descriptor,
1928                                           LLVMValueRef offset, LLVMValueRef compare,
1929                                           LLVMValueRef exchange, bool image)
1930{
1931   LLVMBasicBlockRef start_block = NULL, then_block = NULL;
1932   if (ctx->abi->robust_buffer_access || image) {
1933      LLVMValueRef size = ac_llvm_extract_elem(&ctx->ac, descriptor, 2);
1934
1935      LLVMValueRef cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, offset, size, "");
1936      start_block = LLVMGetInsertBlock(ctx->ac.builder);
1937
1938      ac_build_ifcc(&ctx->ac, cond, -1);
1939
1940      then_block = LLVMGetInsertBlock(ctx->ac.builder);
1941   }
1942
1943   if (image)
1944      offset = LLVMBuildMul(ctx->ac.builder, offset, LLVMConstInt(ctx->ac.i32, 8, false), "");
1945
1946   LLVMValueRef ptr_parts[2] = {
1947      ac_llvm_extract_elem(&ctx->ac, descriptor, 0),
1948      LLVMBuildAnd(ctx->ac.builder, ac_llvm_extract_elem(&ctx->ac, descriptor, 1),
1949                   LLVMConstInt(ctx->ac.i32, 65535, 0), "")};
1950
1951   ptr_parts[1] = LLVMBuildTrunc(ctx->ac.builder, ptr_parts[1], ctx->ac.i16, "");
1952   ptr_parts[1] = LLVMBuildSExt(ctx->ac.builder, ptr_parts[1], ctx->ac.i32, "");
1953
1954   offset = LLVMBuildZExt(ctx->ac.builder, offset, ctx->ac.i64, "");
1955
1956   LLVMValueRef ptr = ac_build_gather_values(&ctx->ac, ptr_parts, 2);
1957   ptr = LLVMBuildBitCast(ctx->ac.builder, ptr, ctx->ac.i64, "");
1958   ptr = LLVMBuildAdd(ctx->ac.builder, ptr, offset, "");
1959   ptr = LLVMBuildIntToPtr(ctx->ac.builder, ptr, LLVMPointerType(ctx->ac.i64, AC_ADDR_SPACE_GLOBAL),
1960                           "");
1961
1962   LLVMValueRef result =
1963      ac_build_atomic_cmp_xchg(&ctx->ac, ptr, compare, exchange, "singlethread-one-as");
1964   result = LLVMBuildExtractValue(ctx->ac.builder, result, 0, "");
1965
1966   if (ctx->abi->robust_buffer_access || image) {
1967      ac_build_endif(&ctx->ac, -1);
1968
1969      LLVMBasicBlockRef incoming_blocks[2] = {
1970         start_block,
1971         then_block,
1972      };
1973
1974      LLVMValueRef incoming_values[2] = {
1975         LLVMConstInt(ctx->ac.i64, 0, 0),
1976         result,
1977      };
1978      LLVMValueRef ret = LLVMBuildPhi(ctx->ac.builder, ctx->ac.i64, "");
1979      LLVMAddIncoming(ret, incoming_values, incoming_blocks, 2);
1980      return ret;
1981   } else {
1982      return result;
1983   }
1984}
1985
1986static LLVMValueRef visit_atomic_ssbo(struct ac_nir_context *ctx, nir_intrinsic_instr *instr)
1987{
1988   if (ctx->ac.postponed_kill) {
1989      LLVMValueRef cond = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.i1, ctx->ac.postponed_kill, "");
1990      ac_build_ifcc(&ctx->ac, cond, 7001);
1991   }
1992
1993   LLVMTypeRef return_type = LLVMTypeOf(get_src(ctx, instr->src[2]));
1994   const char *op;
1995   char name[64], type[8];
1996   LLVMValueRef params[6], descriptor;
1997   LLVMValueRef result;
1998   int arg_count = 0;
1999
2000   struct waterfall_context wctx;
2001   LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[0]);
2002
2003   switch (instr->intrinsic) {
2004   case nir_intrinsic_ssbo_atomic_add:
2005      op = "add";
2006      break;
2007   case nir_intrinsic_ssbo_atomic_imin:
2008      op = "smin";
2009      break;
2010   case nir_intrinsic_ssbo_atomic_umin:
2011      op = "umin";
2012      break;
2013   case nir_intrinsic_ssbo_atomic_imax:
2014      op = "smax";
2015      break;
2016   case nir_intrinsic_ssbo_atomic_umax:
2017      op = "umax";
2018      break;
2019   case nir_intrinsic_ssbo_atomic_and:
2020      op = "and";
2021      break;
2022   case nir_intrinsic_ssbo_atomic_or:
2023      op = "or";
2024      break;
2025   case nir_intrinsic_ssbo_atomic_xor:
2026      op = "xor";
2027      break;
2028   case nir_intrinsic_ssbo_atomic_exchange:
2029      op = "swap";
2030      break;
2031   case nir_intrinsic_ssbo_atomic_comp_swap:
2032      op = "cmpswap";
2033      break;
2034   case nir_intrinsic_ssbo_atomic_fmin:
2035      op = "fmin";
2036      break;
2037   case nir_intrinsic_ssbo_atomic_fmax:
2038      op = "fmax";
2039      break;
2040   default:
2041      abort();
2042   }
2043
2044   descriptor = ctx->abi->load_ssbo(ctx->abi, rsrc_base, true, false);
2045
2046   if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap && return_type == ctx->ac.i64) {
2047      result = emit_ssbo_comp_swap_64(ctx, descriptor, get_src(ctx, instr->src[1]),
2048                                      get_src(ctx, instr->src[2]), get_src(ctx, instr->src[3]), false);
2049   } else {
2050      LLVMValueRef data = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
2051
2052      if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap) {
2053         params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[3]), 0);
2054      }
2055      if (instr->intrinsic == nir_intrinsic_ssbo_atomic_fmin ||
2056          instr->intrinsic == nir_intrinsic_ssbo_atomic_fmax) {
2057         data = ac_to_float(&ctx->ac, data);
2058         return_type = LLVMTypeOf(data);
2059      }
2060      params[arg_count++] = data;
2061      params[arg_count++] = descriptor;
2062      params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
2063      params[arg_count++] = ctx->ac.i32_0;               /* soffset */
2064      params[arg_count++] = ctx->ac.i32_0;               /* slc */
2065
2066      ac_build_type_name_for_intr(return_type, type, sizeof(type));
2067      snprintf(name, sizeof(name), "llvm.amdgcn.raw.buffer.atomic.%s.%s", op, type);
2068
2069      result = ac_build_intrinsic(&ctx->ac, name, return_type, params, arg_count, 0);
2070
2071      if (instr->intrinsic == nir_intrinsic_ssbo_atomic_fmin ||
2072          instr->intrinsic == nir_intrinsic_ssbo_atomic_fmax) {
2073         result = ac_to_integer(&ctx->ac, result);
2074      }
2075   }
2076
2077   result = exit_waterfall(ctx, &wctx, result);
2078   if (ctx->ac.postponed_kill)
2079      ac_build_endif(&ctx->ac, 7001);
2080   return result;
2081}
2082
2083static LLVMValueRef visit_load_buffer(struct ac_nir_context *ctx, nir_intrinsic_instr *instr)
2084{
2085   struct waterfall_context wctx;
2086   LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[0]);
2087
2088   int elem_size_bytes = instr->dest.ssa.bit_size / 8;
2089   int num_components = instr->num_components;
2090   enum gl_access_qualifier access = nir_intrinsic_access(instr);
2091   unsigned cache_policy = get_cache_policy(ctx, access, false, false);
2092
2093   LLVMValueRef offset = get_src(ctx, instr->src[1]);
2094   LLVMValueRef rsrc = ctx->abi->load_ssbo(ctx->abi, rsrc_base, false, false);
2095   LLVMValueRef vindex = ctx->ac.i32_0;
2096
2097   LLVMTypeRef def_type = get_def_type(ctx, &instr->dest.ssa);
2098   LLVMTypeRef def_elem_type = num_components > 1 ? LLVMGetElementType(def_type) : def_type;
2099
2100   LLVMValueRef results[4];
2101   for (int i = 0; i < num_components;) {
2102      int num_elems = num_components - i;
2103      if (elem_size_bytes < 4 && nir_intrinsic_align(instr) % 4 != 0)
2104         num_elems = 1;
2105      if (num_elems * elem_size_bytes > 16)
2106         num_elems = 16 / elem_size_bytes;
2107      int load_bytes = num_elems * elem_size_bytes;
2108
2109      LLVMValueRef immoffset = LLVMConstInt(ctx->ac.i32, i * elem_size_bytes, false);
2110      LLVMValueRef voffset = LLVMBuildAdd(ctx->ac.builder, offset, immoffset, "");
2111
2112      LLVMValueRef ret;
2113
2114      if (load_bytes == 1) {
2115         ret = ac_build_buffer_load_byte(&ctx->ac, rsrc, voffset, ctx->ac.i32_0,
2116                                          cache_policy);
2117      } else if (load_bytes == 2) {
2118         ret = ac_build_buffer_load_short(&ctx->ac, rsrc, voffset, ctx->ac.i32_0,
2119                                           cache_policy);
2120      } else {
2121         int num_channels = util_next_power_of_two(load_bytes) / 4;
2122         bool can_speculate = access & ACCESS_CAN_REORDER;
2123
2124         ret = ac_build_buffer_load(&ctx->ac, rsrc, num_channels, vindex, voffset, ctx->ac.i32_0,
2125                                    ctx->ac.f32, cache_policy, can_speculate, false);
2126      }
2127
2128      LLVMTypeRef byte_vec = LLVMVectorType(ctx->ac.i8, ac_get_type_size(LLVMTypeOf(ret)));
2129      ret = LLVMBuildBitCast(ctx->ac.builder, ret, byte_vec, "");
2130      ret = ac_trim_vector(&ctx->ac, ret, load_bytes);
2131
2132      LLVMTypeRef ret_type = LLVMVectorType(def_elem_type, num_elems);
2133      ret = LLVMBuildBitCast(ctx->ac.builder, ret, ret_type, "");
2134
2135      for (unsigned j = 0; j < num_elems; j++) {
2136         results[i + j] =
2137            LLVMBuildExtractElement(ctx->ac.builder, ret, LLVMConstInt(ctx->ac.i32, j, false), "");
2138      }
2139      i += num_elems;
2140   }
2141
2142   LLVMValueRef ret = ac_build_gather_values(&ctx->ac, results, num_components);
2143   return exit_waterfall(ctx, &wctx, ret);
2144}
2145
2146static LLVMValueRef enter_waterfall_ubo(struct ac_nir_context *ctx, struct waterfall_context *wctx,
2147                                        const nir_intrinsic_instr *instr)
2148{
2149   return enter_waterfall(ctx, wctx, get_src(ctx, instr->src[0]),
2150                          nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
2151}
2152
2153static LLVMValueRef get_global_address(struct ac_nir_context *ctx,
2154                                       nir_intrinsic_instr *instr,
2155                                       LLVMTypeRef type)
2156{
2157   bool is_store = instr->intrinsic == nir_intrinsic_store_global ||
2158                   instr->intrinsic == nir_intrinsic_store_global_amd;
2159   LLVMValueRef addr = get_src(ctx, instr->src[is_store ? 1 : 0]);
2160
2161   LLVMTypeRef ptr_type = LLVMPointerType(type, AC_ADDR_SPACE_GLOBAL);
2162
2163   if (nir_intrinsic_has_base(instr)) {
2164      /* _amd variants */
2165      uint32_t base = nir_intrinsic_base(instr);
2166      unsigned num_src = nir_intrinsic_infos[instr->intrinsic].num_srcs;
2167      LLVMValueRef offset = get_src(ctx, instr->src[num_src - 1]);
2168      offset = LLVMBuildAdd(ctx->ac.builder, offset, LLVMConstInt(ctx->ac.i32, base, false), "");
2169
2170      LLVMTypeRef i8_ptr_type = LLVMPointerType(ctx->ac.i8, AC_ADDR_SPACE_GLOBAL);
2171      addr = LLVMBuildIntToPtr(ctx->ac.builder, addr, i8_ptr_type, "");
2172      addr = LLVMBuildGEP(ctx->ac.builder, addr, &offset, 1, "");
2173      return type == ctx->ac.i8 ? addr : LLVMBuildBitCast(ctx->ac.builder, addr, ptr_type, "");
2174   } else {
2175      return LLVMBuildIntToPtr(ctx->ac.builder, addr, ptr_type, "");
2176   }
2177}
2178
2179static LLVMValueRef visit_load_global(struct ac_nir_context *ctx,
2180                                      nir_intrinsic_instr *instr)
2181{
2182   LLVMTypeRef result_type = get_def_type(ctx, &instr->dest.ssa);
2183   LLVMValueRef val;
2184   LLVMValueRef addr = get_global_address(ctx, instr, result_type);
2185
2186   val = LLVMBuildLoad2(ctx->ac.builder, result_type, addr, "");
2187
2188   if (nir_intrinsic_access(instr) & (ACCESS_COHERENT | ACCESS_VOLATILE)) {
2189      LLVMSetOrdering(val, LLVMAtomicOrderingMonotonic);
2190      LLVMSetAlignment(val, ac_get_type_size(result_type));
2191   }
2192
2193   return val;
2194}
2195
2196static void visit_store_global(struct ac_nir_context *ctx,
2197				     nir_intrinsic_instr *instr)
2198{
2199   if (ctx->ac.postponed_kill) {
2200      LLVMValueRef cond = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.i1, ctx->ac.postponed_kill, "");
2201      ac_build_ifcc(&ctx->ac, cond, 7002);
2202   }
2203
2204   LLVMValueRef data = get_src(ctx, instr->src[0]);
2205   LLVMTypeRef type = LLVMTypeOf(data);
2206   LLVMValueRef addr = get_global_address(ctx, instr, type);
2207   LLVMValueRef val;
2208
2209   val = LLVMBuildStore(ctx->ac.builder, data, addr);
2210
2211   if (nir_intrinsic_access(instr) & (ACCESS_COHERENT | ACCESS_VOLATILE)) {
2212      LLVMSetOrdering(val, LLVMAtomicOrderingMonotonic);
2213      LLVMSetAlignment(val, ac_get_type_size(type));
2214   }
2215
2216   if (ctx->ac.postponed_kill)
2217      ac_build_endif(&ctx->ac, 7002);
2218}
2219
2220static LLVMValueRef visit_global_atomic(struct ac_nir_context *ctx,
2221					nir_intrinsic_instr *instr)
2222{
2223   if (ctx->ac.postponed_kill) {
2224      LLVMValueRef cond = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.i1, ctx->ac.postponed_kill, "");
2225      ac_build_ifcc(&ctx->ac, cond, 7002);
2226   }
2227
2228   LLVMValueRef data = get_src(ctx, instr->src[1]);
2229   LLVMAtomicRMWBinOp op;
2230   LLVMValueRef result;
2231
2232   /* use "singlethread" sync scope to implement relaxed ordering */
2233   const char *sync_scope = "singlethread-one-as";
2234
2235   if (instr->intrinsic == nir_intrinsic_global_atomic_fmin ||
2236       instr->intrinsic == nir_intrinsic_global_atomic_fmax ||
2237       instr->intrinsic == nir_intrinsic_global_atomic_fmin_amd ||
2238       instr->intrinsic == nir_intrinsic_global_atomic_fmax_amd) {
2239      data = ac_to_float(&ctx->ac, data);
2240   }
2241
2242   LLVMTypeRef data_type = LLVMTypeOf(data);
2243
2244   LLVMValueRef addr = get_global_address(ctx, instr, data_type);
2245
2246   if (instr->intrinsic == nir_intrinsic_global_atomic_comp_swap ||
2247       instr->intrinsic == nir_intrinsic_global_atomic_comp_swap_amd) {
2248      LLVMValueRef data1 = get_src(ctx, instr->src[2]);
2249      result = ac_build_atomic_cmp_xchg(&ctx->ac, addr, data, data1, sync_scope);
2250      result = LLVMBuildExtractValue(ctx->ac.builder, result, 0, "");
2251   } else if (instr->intrinsic == nir_intrinsic_global_atomic_fmin ||
2252              instr->intrinsic == nir_intrinsic_global_atomic_fmax ||
2253              instr->intrinsic == nir_intrinsic_global_atomic_fmin_amd ||
2254              instr->intrinsic == nir_intrinsic_global_atomic_fmax_amd) {
2255      const char *op = instr->intrinsic == nir_intrinsic_global_atomic_fmin ? "fmin" : "fmax";
2256      char name[64], type[8];
2257      LLVMValueRef params[2];
2258      int arg_count = 0;
2259
2260      params[arg_count++] = addr;
2261      params[arg_count++] = data;
2262
2263      ac_build_type_name_for_intr(data_type, type, sizeof(type));
2264      snprintf(name, sizeof(name), "llvm.amdgcn.global.atomic.%s.%s.p1%s.%s", op, type, type, type);
2265
2266      result = ac_build_intrinsic(&ctx->ac, name, data_type, params, arg_count, 0);
2267      result = ac_to_integer(&ctx->ac, result);
2268   } else {
2269      switch (instr->intrinsic) {
2270      case nir_intrinsic_global_atomic_add:
2271      case nir_intrinsic_global_atomic_add_amd:
2272         op = LLVMAtomicRMWBinOpAdd;
2273         break;
2274      case nir_intrinsic_global_atomic_umin:
2275      case nir_intrinsic_global_atomic_umin_amd:
2276         op = LLVMAtomicRMWBinOpUMin;
2277         break;
2278      case nir_intrinsic_global_atomic_umax:
2279      case nir_intrinsic_global_atomic_umax_amd:
2280         op = LLVMAtomicRMWBinOpUMax;
2281         break;
2282      case nir_intrinsic_global_atomic_imin:
2283      case nir_intrinsic_global_atomic_imin_amd:
2284         op = LLVMAtomicRMWBinOpMin;
2285         break;
2286      case nir_intrinsic_global_atomic_imax:
2287      case nir_intrinsic_global_atomic_imax_amd:
2288         op = LLVMAtomicRMWBinOpMax;
2289         break;
2290      case nir_intrinsic_global_atomic_and:
2291      case nir_intrinsic_global_atomic_and_amd:
2292         op = LLVMAtomicRMWBinOpAnd;
2293         break;
2294      case nir_intrinsic_global_atomic_or:
2295      case nir_intrinsic_global_atomic_or_amd:
2296         op = LLVMAtomicRMWBinOpOr;
2297         break;
2298      case nir_intrinsic_global_atomic_xor:
2299      case nir_intrinsic_global_atomic_xor_amd:
2300         op = LLVMAtomicRMWBinOpXor;
2301         break;
2302      case nir_intrinsic_global_atomic_exchange:
2303      case nir_intrinsic_global_atomic_exchange_amd:
2304         op = LLVMAtomicRMWBinOpXchg;
2305         break;
2306      default:
2307         unreachable("Invalid global atomic operation");
2308      }
2309
2310      result = ac_build_atomic_rmw(&ctx->ac, op, addr, ac_to_integer(&ctx->ac, data), sync_scope);
2311   }
2312
2313   if (ctx->ac.postponed_kill)
2314      ac_build_endif(&ctx->ac, 7002);
2315
2316   return result;
2317}
2318
2319static LLVMValueRef visit_load_ubo_buffer(struct ac_nir_context *ctx, nir_intrinsic_instr *instr)
2320{
2321   struct waterfall_context wctx;
2322   LLVMValueRef rsrc_base = enter_waterfall_ubo(ctx, &wctx, instr);
2323
2324   LLVMValueRef ret;
2325   LLVMValueRef rsrc = rsrc_base;
2326   LLVMValueRef offset = get_src(ctx, instr->src[1]);
2327   int num_components = instr->num_components;
2328
2329   if (ctx->abi->load_ubo)
2330      rsrc = ctx->abi->load_ubo(ctx->abi, rsrc);
2331
2332   /* Convert to a scalar 32-bit load. */
2333   if (instr->dest.ssa.bit_size == 64)
2334      num_components *= 2;
2335   else if (instr->dest.ssa.bit_size == 16)
2336      num_components = DIV_ROUND_UP(num_components, 2);
2337   else if (instr->dest.ssa.bit_size == 8)
2338      num_components = DIV_ROUND_UP(num_components, 4);
2339
2340   ret =
2341      ac_build_buffer_load(&ctx->ac, rsrc, num_components, NULL, offset, NULL,
2342                           ctx->ac.f32, 0, true, true);
2343
2344   /* Convert to the original type. */
2345   if (instr->dest.ssa.bit_size == 64) {
2346      ret = LLVMBuildBitCast(ctx->ac.builder, ret,
2347                             LLVMVectorType(ctx->ac.i64, num_components / 2), "");
2348   } else if (instr->dest.ssa.bit_size == 16) {
2349      ret = LLVMBuildBitCast(ctx->ac.builder, ret,
2350                             LLVMVectorType(ctx->ac.i16, num_components * 2), "");
2351   } else if (instr->dest.ssa.bit_size == 8) {
2352      ret = LLVMBuildBitCast(ctx->ac.builder, ret,
2353                             LLVMVectorType(ctx->ac.i8, num_components * 4), "");
2354   }
2355
2356   ret = ac_trim_vector(&ctx->ac, ret, instr->num_components);
2357   ret = LLVMBuildBitCast(ctx->ac.builder, ret, get_def_type(ctx, &instr->dest.ssa), "");
2358
2359   return exit_waterfall(ctx, &wctx, ret);
2360}
2361
2362static unsigned type_scalar_size_bytes(const struct glsl_type *type)
2363{
2364   assert(glsl_type_is_vector_or_scalar(type) || glsl_type_is_matrix(type));
2365   return glsl_type_is_boolean(type) ? 4 : glsl_get_bit_size(type) / 8;
2366}
2367
2368static void visit_store_output(struct ac_nir_context *ctx, nir_intrinsic_instr *instr)
2369{
2370   if (ctx->ac.postponed_kill) {
2371      LLVMValueRef cond = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.i1, ctx->ac.postponed_kill, "");
2372      ac_build_ifcc(&ctx->ac, cond, 7002);
2373   }
2374
2375   unsigned base = nir_intrinsic_base(instr);
2376   unsigned writemask = nir_intrinsic_write_mask(instr);
2377   unsigned component = nir_intrinsic_component(instr);
2378   LLVMValueRef src = ac_to_float(&ctx->ac, get_src(ctx, instr->src[0]));
2379   nir_src offset = *nir_get_io_offset_src(instr);
2380
2381   /* No indirect indexing is allowed here. */
2382   assert(nir_src_is_const(offset) && nir_src_as_uint(offset) == 0);
2383
2384   switch (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src))) {
2385   case 16:
2386   case 32:
2387      break;
2388   case 64:
2389      unreachable("64-bit IO should have been lowered to 32 bits");
2390      return;
2391   default:
2392      unreachable("unhandled store_output bit size");
2393      return;
2394   }
2395
2396   writemask <<= component;
2397
2398   for (unsigned chan = 0; chan < 8; chan++) {
2399      if (!(writemask & (1 << chan)))
2400         continue;
2401
2402      LLVMValueRef value = ac_llvm_extract_elem(&ctx->ac, src, chan - component);
2403      LLVMValueRef output_addr = ctx->abi->outputs[base * 4 + chan];
2404
2405      if (!ctx->abi->is_16bit[base * 4 + chan] &&
2406          LLVMTypeOf(value) == ctx->ac.f16) {
2407         LLVMValueRef output, index;
2408
2409         /* Insert the 16-bit value into the low or high bits of the 32-bit output
2410          * using read-modify-write.
2411          */
2412         index = LLVMConstInt(ctx->ac.i32, nir_intrinsic_io_semantics(instr).high_16bits, 0);
2413         output = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.v2f16, output_addr, "");
2414         output = LLVMBuildInsertElement(ctx->ac.builder, output, value, index, "");
2415         value = LLVMBuildBitCast(ctx->ac.builder, output, ctx->ac.f32, "");
2416      }
2417      LLVMBuildStore(ctx->ac.builder, value, output_addr);
2418   }
2419
2420   if (ctx->ac.postponed_kill)
2421      ac_build_endif(&ctx->ac, 7002);
2422}
2423
2424static int image_type_to_components_count(enum glsl_sampler_dim dim, bool array)
2425{
2426   switch (dim) {
2427   case GLSL_SAMPLER_DIM_BUF:
2428      return 1;
2429   case GLSL_SAMPLER_DIM_1D:
2430      return array ? 2 : 1;
2431   case GLSL_SAMPLER_DIM_2D:
2432      return array ? 3 : 2;
2433   case GLSL_SAMPLER_DIM_MS:
2434      return array ? 4 : 3;
2435   case GLSL_SAMPLER_DIM_3D:
2436   case GLSL_SAMPLER_DIM_CUBE:
2437      return 3;
2438   case GLSL_SAMPLER_DIM_RECT:
2439   case GLSL_SAMPLER_DIM_SUBPASS:
2440      return 2;
2441   case GLSL_SAMPLER_DIM_SUBPASS_MS:
2442      return 3;
2443   default:
2444      break;
2445   }
2446   return 0;
2447}
2448
2449static LLVMValueRef adjust_sample_index_using_fmask(struct ac_llvm_context *ctx,
2450                                                    LLVMValueRef coord_x, LLVMValueRef coord_y,
2451                                                    LLVMValueRef coord_z, LLVMValueRef sample_index,
2452                                                    LLVMValueRef fmask_desc_ptr)
2453{
2454   if (!fmask_desc_ptr)
2455      return sample_index;
2456
2457   unsigned sample_chan = coord_z ? 3 : 2;
2458   LLVMValueRef addr[4] = {coord_x, coord_y, coord_z};
2459   addr[sample_chan] = sample_index;
2460
2461   ac_apply_fmask_to_sample(ctx, fmask_desc_ptr, addr, coord_z != NULL);
2462   return addr[sample_chan];
2463}
2464
2465static nir_deref_instr *get_image_deref(const nir_intrinsic_instr *instr)
2466{
2467   assert(instr->src[0].is_ssa);
2468   return nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2469}
2470
2471static LLVMValueRef get_image_descriptor(struct ac_nir_context *ctx,
2472                                         const nir_intrinsic_instr *instr,
2473                                         LLVMValueRef dynamic_index,
2474                                         enum ac_descriptor_type desc_type, bool write)
2475{
2476   nir_deref_instr *deref_instr = instr->src[0].ssa->parent_instr->type == nir_instr_type_deref
2477                                     ? nir_instr_as_deref(instr->src[0].ssa->parent_instr)
2478                                     : NULL;
2479
2480   return get_sampler_desc(ctx, deref_instr, desc_type, &instr->instr, dynamic_index, true, write);
2481}
2482
2483static void get_image_coords(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr,
2484                             LLVMValueRef dynamic_desc_index, struct ac_image_args *args,
2485                             enum glsl_sampler_dim dim, bool is_array)
2486{
2487   LLVMValueRef src0 = get_src(ctx, instr->src[1]);
2488   LLVMValueRef masks[] = {
2489      LLVMConstInt(ctx->ac.i32, 0, false),
2490      LLVMConstInt(ctx->ac.i32, 1, false),
2491      LLVMConstInt(ctx->ac.i32, 2, false),
2492      LLVMConstInt(ctx->ac.i32, 3, false),
2493   };
2494   LLVMValueRef sample_index = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
2495
2496   int count;
2497   ASSERTED bool add_frag_pos =
2498      (dim == GLSL_SAMPLER_DIM_SUBPASS || dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
2499   bool is_ms = (dim == GLSL_SAMPLER_DIM_MS || dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
2500   bool gfx9_1d = ctx->ac.gfx_level == GFX9 && dim == GLSL_SAMPLER_DIM_1D;
2501   assert(!add_frag_pos && "Input attachments should be lowered by this point.");
2502   count = image_type_to_components_count(dim, is_array);
2503
2504   if (ctx->ac.gfx_level < GFX11 &&
2505       is_ms && (instr->intrinsic == nir_intrinsic_image_deref_load ||
2506                 instr->intrinsic == nir_intrinsic_bindless_image_load ||
2507                 instr->intrinsic == nir_intrinsic_image_deref_sparse_load ||
2508                 instr->intrinsic == nir_intrinsic_bindless_image_sparse_load)) {
2509      LLVMValueRef fmask_load_address[3];
2510
2511      fmask_load_address[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
2512      fmask_load_address[1] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[1], "");
2513      if (is_array)
2514         fmask_load_address[2] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[2], "");
2515      else
2516         fmask_load_address[2] = NULL;
2517
2518      sample_index = adjust_sample_index_using_fmask(
2519         &ctx->ac, fmask_load_address[0], fmask_load_address[1], fmask_load_address[2],
2520         sample_index, get_image_descriptor(ctx, instr, dynamic_desc_index, AC_DESC_FMASK, false));
2521   }
2522   if (count == 1 && !gfx9_1d) {
2523      if (instr->src[1].ssa->num_components)
2524         args->coords[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
2525      else
2526         args->coords[0] = src0;
2527   } else {
2528      int chan;
2529      if (is_ms)
2530         count--;
2531      for (chan = 0; chan < count; ++chan) {
2532         args->coords[chan] = ac_llvm_extract_elem(&ctx->ac, src0, chan);
2533      }
2534
2535      if (gfx9_1d) {
2536         if (is_array) {
2537            args->coords[2] = args->coords[1];
2538            args->coords[1] = ctx->ac.i32_0;
2539         } else
2540            args->coords[1] = ctx->ac.i32_0;
2541         count++;
2542      }
2543      if (ctx->ac.gfx_level == GFX9 && dim == GLSL_SAMPLER_DIM_2D && !is_array) {
2544         /* The hw can't bind a slice of a 3D image as a 2D
2545          * image, because it ignores BASE_ARRAY if the target
2546          * is 3D. The workaround is to read BASE_ARRAY and set
2547          * it as the 3rd address operand for all 2D images.
2548          */
2549         LLVMValueRef first_layer, const5, mask;
2550
2551         const5 = LLVMConstInt(ctx->ac.i32, 5, 0);
2552         mask = LLVMConstInt(ctx->ac.i32, S_008F24_BASE_ARRAY(~0), 0);
2553         first_layer = LLVMBuildExtractElement(ctx->ac.builder, args->resource, const5, "");
2554         first_layer = LLVMBuildAnd(ctx->ac.builder, first_layer, mask, "");
2555
2556         args->coords[count] = first_layer;
2557         count++;
2558      }
2559
2560      if (is_ms) {
2561         args->coords[count] = sample_index;
2562         count++;
2563      }
2564   }
2565}
2566
2567static LLVMValueRef enter_waterfall_image(struct ac_nir_context *ctx,
2568                                          struct waterfall_context *wctx,
2569                                          const nir_intrinsic_instr *instr)
2570{
2571   nir_deref_instr *deref_instr = NULL;
2572
2573   if (instr->src[0].ssa->parent_instr->type == nir_instr_type_deref)
2574      deref_instr = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2575
2576   LLVMValueRef value = get_sampler_desc_index(ctx, deref_instr, &instr->instr, true);
2577   return enter_waterfall(ctx, wctx, value, nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
2578}
2579
2580static LLVMValueRef visit_image_load(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr,
2581                                     bool bindless)
2582{
2583   LLVMValueRef res;
2584
2585   enum glsl_sampler_dim dim;
2586   enum gl_access_qualifier access = nir_intrinsic_access(instr);
2587   bool is_array;
2588   if (bindless) {
2589      dim = nir_intrinsic_image_dim(instr);
2590      is_array = nir_intrinsic_image_array(instr);
2591   } else {
2592      const nir_deref_instr *image_deref = get_image_deref(instr);
2593      const struct glsl_type *type = image_deref->type;
2594      const nir_variable *var = nir_deref_instr_get_variable(image_deref);
2595      dim = glsl_get_sampler_dim(type);
2596      access |= var->data.access;
2597      is_array = glsl_sampler_type_is_array(type);
2598   }
2599
2600   struct waterfall_context wctx;
2601   LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
2602
2603   struct ac_image_args args = {0};
2604
2605   args.cache_policy = get_cache_policy(ctx, access, false, false);
2606   args.tfe = instr->intrinsic == nir_intrinsic_image_deref_sparse_load ||
2607              instr->intrinsic == nir_intrinsic_bindless_image_sparse_load;
2608
2609   if (dim == GLSL_SAMPLER_DIM_BUF) {
2610      unsigned num_channels = util_last_bit(nir_ssa_def_components_read(&instr->dest.ssa));
2611      if (instr->dest.ssa.bit_size == 64)
2612         num_channels = num_channels < 4 ? 2 : 4;
2613      LLVMValueRef rsrc, vindex;
2614
2615      rsrc = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_BUFFER, false);
2616      vindex =
2617         LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]), ctx->ac.i32_0, "");
2618
2619      assert(instr->dest.is_ssa);
2620      bool can_speculate = access & ACCESS_CAN_REORDER;
2621      res = ac_build_buffer_load_format(&ctx->ac, rsrc, vindex, ctx->ac.i32_0, num_channels,
2622                                        args.cache_policy, can_speculate,
2623                                        instr->dest.ssa.bit_size == 16,
2624                                        args.tfe);
2625      res = ac_build_expand(&ctx->ac, res, num_channels, args.tfe ? 5 : 4);
2626
2627      res = ac_trim_vector(&ctx->ac, res, instr->dest.ssa.num_components);
2628      res = ac_to_integer(&ctx->ac, res);
2629   } else {
2630      bool level_zero = nir_src_is_const(instr->src[3]) && nir_src_as_uint(instr->src[3]) == 0;
2631
2632      args.opcode = level_zero ? ac_image_load : ac_image_load_mip;
2633      args.resource = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_IMAGE, false);
2634      get_image_coords(ctx, instr, dynamic_index, &args, dim, is_array);
2635      args.dim = ac_get_image_dim(ctx->ac.gfx_level, dim, is_array);
2636      if (!level_zero)
2637         args.lod = get_src(ctx, instr->src[3]);
2638      args.dmask = 15;
2639      args.attributes = AC_FUNC_ATTR_READONLY;
2640
2641      assert(instr->dest.is_ssa);
2642      args.d16 = instr->dest.ssa.bit_size == 16;
2643
2644      res = ac_build_image_opcode(&ctx->ac, &args);
2645   }
2646
2647   if (instr->dest.ssa.bit_size == 64) {
2648      LLVMValueRef code = NULL;
2649      if (args.tfe) {
2650         code = ac_llvm_extract_elem(&ctx->ac, res, 4);
2651         res = ac_trim_vector(&ctx->ac, res, 4);
2652      }
2653
2654      res = LLVMBuildBitCast(ctx->ac.builder, res, LLVMVectorType(ctx->ac.i64, 2), "");
2655      LLVMValueRef x = LLVMBuildExtractElement(ctx->ac.builder, res, ctx->ac.i32_0, "");
2656      LLVMValueRef w = LLVMBuildExtractElement(ctx->ac.builder, res, ctx->ac.i32_1, "");
2657
2658      if (code)
2659         code = LLVMBuildZExt(ctx->ac.builder, code, ctx->ac.i64, "");
2660      LLVMValueRef values[5] = {x, ctx->ac.i64_0, ctx->ac.i64_0, w, code};
2661      res = ac_build_gather_values(&ctx->ac, values, 4 + args.tfe);
2662   }
2663
2664   return exit_waterfall(ctx, &wctx, res);
2665}
2666
2667static void visit_image_store(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr,
2668                              bool bindless)
2669{
2670   if (ctx->ac.postponed_kill) {
2671      LLVMValueRef cond = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.i1, ctx->ac.postponed_kill, "");
2672      ac_build_ifcc(&ctx->ac, cond, 7003);
2673   }
2674
2675   enum glsl_sampler_dim dim;
2676   enum gl_access_qualifier access = nir_intrinsic_access(instr);
2677   bool is_array;
2678
2679   if (bindless) {
2680      dim = nir_intrinsic_image_dim(instr);
2681      is_array = nir_intrinsic_image_array(instr);
2682   } else {
2683      const nir_deref_instr *image_deref = get_image_deref(instr);
2684      const struct glsl_type *type = image_deref->type;
2685      const nir_variable *var = nir_deref_instr_get_variable(image_deref);
2686      dim = glsl_get_sampler_dim(type);
2687      access |= var->data.access;
2688      is_array = glsl_sampler_type_is_array(type);
2689   }
2690
2691   struct waterfall_context wctx;
2692   LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
2693
2694   bool writeonly_memory = access & ACCESS_NON_READABLE;
2695   struct ac_image_args args = {0};
2696
2697   args.cache_policy = get_cache_policy(ctx, access, true, writeonly_memory);
2698
2699   LLVMValueRef src = get_src(ctx, instr->src[3]);
2700   if (instr->src[3].ssa->bit_size == 64) {
2701      /* only R64_UINT and R64_SINT supported */
2702      src = ac_llvm_extract_elem(&ctx->ac, src, 0);
2703      src = LLVMBuildBitCast(ctx->ac.builder, src, ctx->ac.v2f32, "");
2704   } else {
2705      src = ac_to_float(&ctx->ac, src);
2706   }
2707
2708   if (dim == GLSL_SAMPLER_DIM_BUF) {
2709      LLVMValueRef rsrc = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_BUFFER, true);
2710      unsigned src_channels = ac_get_llvm_num_components(src);
2711      LLVMValueRef vindex;
2712
2713      if (src_channels == 3)
2714         src = ac_build_expand_to_vec4(&ctx->ac, src, 3);
2715
2716      vindex =
2717         LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]), ctx->ac.i32_0, "");
2718
2719      ac_build_buffer_store_format(&ctx->ac, rsrc, src, vindex, ctx->ac.i32_0, args.cache_policy);
2720   } else {
2721      bool level_zero = nir_src_is_const(instr->src[4]) && nir_src_as_uint(instr->src[4]) == 0;
2722
2723      args.opcode = level_zero ? ac_image_store : ac_image_store_mip;
2724      args.data[0] = src;
2725      args.resource = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_IMAGE, true);
2726      get_image_coords(ctx, instr, dynamic_index, &args, dim, is_array);
2727      args.dim = ac_get_image_dim(ctx->ac.gfx_level, dim, is_array);
2728      if (!level_zero)
2729         args.lod = get_src(ctx, instr->src[4]);
2730      args.dmask = 15;
2731      args.d16 = ac_get_elem_bits(&ctx->ac, LLVMTypeOf(args.data[0])) == 16;
2732
2733      ac_build_image_opcode(&ctx->ac, &args);
2734   }
2735
2736   exit_waterfall(ctx, &wctx, NULL);
2737   if (ctx->ac.postponed_kill)
2738      ac_build_endif(&ctx->ac, 7003);
2739}
2740
2741static LLVMValueRef visit_image_atomic(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr,
2742                                       bool bindless)
2743{
2744   if (ctx->ac.postponed_kill) {
2745      LLVMValueRef cond = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.i1, ctx->ac.postponed_kill, "");
2746      ac_build_ifcc(&ctx->ac, cond, 7004);
2747   }
2748
2749   LLVMValueRef params[7];
2750   int param_count = 0;
2751
2752   bool cmpswap = instr->intrinsic == nir_intrinsic_image_deref_atomic_comp_swap ||
2753                  instr->intrinsic == nir_intrinsic_bindless_image_atomic_comp_swap;
2754   const char *atomic_name;
2755   char intrinsic_name[64];
2756   enum ac_atomic_op atomic_subop;
2757   ASSERTED int length;
2758
2759   enum glsl_sampler_dim dim;
2760   bool is_array;
2761   if (bindless) {
2762      dim = nir_intrinsic_image_dim(instr);
2763      is_array = nir_intrinsic_image_array(instr);
2764   } else {
2765      const struct glsl_type *type = get_image_deref(instr)->type;
2766      dim = glsl_get_sampler_dim(type);
2767      is_array = glsl_sampler_type_is_array(type);
2768   }
2769
2770   struct waterfall_context wctx;
2771   LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
2772
2773   switch (instr->intrinsic) {
2774   case nir_intrinsic_bindless_image_atomic_add:
2775   case nir_intrinsic_image_deref_atomic_add:
2776      atomic_name = "add";
2777      atomic_subop = ac_atomic_add;
2778      break;
2779   case nir_intrinsic_bindless_image_atomic_imin:
2780   case nir_intrinsic_image_deref_atomic_imin:
2781      atomic_name = "smin";
2782      atomic_subop = ac_atomic_smin;
2783      break;
2784   case nir_intrinsic_bindless_image_atomic_umin:
2785   case nir_intrinsic_image_deref_atomic_umin:
2786      atomic_name = "umin";
2787      atomic_subop = ac_atomic_umin;
2788      break;
2789   case nir_intrinsic_bindless_image_atomic_imax:
2790   case nir_intrinsic_image_deref_atomic_imax:
2791      atomic_name = "smax";
2792      atomic_subop = ac_atomic_smax;
2793      break;
2794   case nir_intrinsic_bindless_image_atomic_umax:
2795   case nir_intrinsic_image_deref_atomic_umax:
2796      atomic_name = "umax";
2797      atomic_subop = ac_atomic_umax;
2798      break;
2799   case nir_intrinsic_bindless_image_atomic_and:
2800   case nir_intrinsic_image_deref_atomic_and:
2801      atomic_name = "and";
2802      atomic_subop = ac_atomic_and;
2803      break;
2804   case nir_intrinsic_bindless_image_atomic_or:
2805   case nir_intrinsic_image_deref_atomic_or:
2806      atomic_name = "or";
2807      atomic_subop = ac_atomic_or;
2808      break;
2809   case nir_intrinsic_bindless_image_atomic_xor:
2810   case nir_intrinsic_image_deref_atomic_xor:
2811      atomic_name = "xor";
2812      atomic_subop = ac_atomic_xor;
2813      break;
2814   case nir_intrinsic_bindless_image_atomic_exchange:
2815   case nir_intrinsic_image_deref_atomic_exchange:
2816      atomic_name = "swap";
2817      atomic_subop = ac_atomic_swap;
2818      break;
2819   case nir_intrinsic_bindless_image_atomic_comp_swap:
2820   case nir_intrinsic_image_deref_atomic_comp_swap:
2821      atomic_name = "cmpswap";
2822      atomic_subop = 0; /* not used */
2823      break;
2824   case nir_intrinsic_bindless_image_atomic_inc_wrap:
2825   case nir_intrinsic_image_deref_atomic_inc_wrap: {
2826      atomic_name = "inc";
2827      atomic_subop = ac_atomic_inc_wrap;
2828      break;
2829   }
2830   case nir_intrinsic_bindless_image_atomic_dec_wrap:
2831   case nir_intrinsic_image_deref_atomic_dec_wrap:
2832      atomic_name = "dec";
2833      atomic_subop = ac_atomic_dec_wrap;
2834      break;
2835   case nir_intrinsic_image_deref_atomic_fmin:
2836      atomic_name = "fmin";
2837      atomic_subop = ac_atomic_fmin;
2838      break;
2839   case nir_intrinsic_image_deref_atomic_fmax:
2840      atomic_name = "fmax";
2841      atomic_subop = ac_atomic_fmax;
2842      break;
2843   default:
2844      abort();
2845   }
2846
2847   if (cmpswap)
2848      params[param_count++] = get_src(ctx, instr->src[4]);
2849   params[param_count++] = get_src(ctx, instr->src[3]);
2850
2851   if (atomic_subop == ac_atomic_fmin || atomic_subop == ac_atomic_fmax)
2852      params[0] = ac_to_float(&ctx->ac, params[0]);
2853
2854   LLVMValueRef result;
2855   if (dim == GLSL_SAMPLER_DIM_BUF) {
2856      params[param_count++] = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_BUFFER, true);
2857      params[param_count++] = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]),
2858                                                      ctx->ac.i32_0, ""); /* vindex */
2859      params[param_count++] = ctx->ac.i32_0;                              /* voffset */
2860      if (cmpswap && instr->dest.ssa.bit_size == 64) {
2861         result = emit_ssbo_comp_swap_64(ctx, params[2], params[3], params[1], params[0], true);
2862      } else {
2863         LLVMTypeRef data_type = LLVMTypeOf(params[0]);
2864         char type[8];
2865
2866         params[param_count++] = ctx->ac.i32_0; /* soffset */
2867         params[param_count++] = ctx->ac.i32_0; /* slc */
2868
2869         ac_build_type_name_for_intr(data_type, type, sizeof(type));
2870         length = snprintf(intrinsic_name, sizeof(intrinsic_name),
2871                           "llvm.amdgcn.struct.buffer.atomic.%s.%s",
2872                           atomic_name, type);
2873
2874         assert(length < sizeof(intrinsic_name));
2875         result = ac_build_intrinsic(&ctx->ac, intrinsic_name, LLVMTypeOf(params[0]), params, param_count, 0);
2876      }
2877   } else {
2878      struct ac_image_args args = {0};
2879      args.opcode = cmpswap ? ac_image_atomic_cmpswap : ac_image_atomic;
2880      args.atomic = atomic_subop;
2881      args.data[0] = params[0];
2882      if (cmpswap)
2883         args.data[1] = params[1];
2884      args.resource = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_IMAGE, true);
2885      get_image_coords(ctx, instr, dynamic_index, &args, dim, is_array);
2886      args.dim = ac_get_image_dim(ctx->ac.gfx_level, dim, is_array);
2887
2888      result = ac_build_image_opcode(&ctx->ac, &args);
2889   }
2890
2891   result = exit_waterfall(ctx, &wctx, result);
2892   if (ctx->ac.postponed_kill)
2893      ac_build_endif(&ctx->ac, 7004);
2894   return result;
2895}
2896
2897static LLVMValueRef visit_image_samples(struct ac_nir_context *ctx, nir_intrinsic_instr *instr)
2898{
2899   struct waterfall_context wctx;
2900   LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
2901   LLVMValueRef rsrc = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_IMAGE, false);
2902
2903   LLVMValueRef ret = ac_build_image_get_sample_count(&ctx->ac, rsrc);
2904   if (ctx->abi->robust_buffer_access) {
2905      LLVMValueRef dword1, is_null_descriptor;
2906
2907      /* Extract the second dword of the descriptor, if it's
2908       * all zero, then it's a null descriptor.
2909       */
2910      dword1 =
2911         LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 1, false), "");
2912      is_null_descriptor = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, dword1,
2913                                         LLVMConstInt(ctx->ac.i32, 0, false), "");
2914      ret = LLVMBuildSelect(ctx->ac.builder, is_null_descriptor, ctx->ac.i32_0, ret, "");
2915   }
2916
2917   return exit_waterfall(ctx, &wctx, ret);
2918}
2919
2920static LLVMValueRef visit_image_size(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr,
2921                                     bool bindless)
2922{
2923   LLVMValueRef res;
2924
2925   enum glsl_sampler_dim dim;
2926   bool is_array;
2927   if (bindless) {
2928      dim = nir_intrinsic_image_dim(instr);
2929      is_array = nir_intrinsic_image_array(instr);
2930   } else {
2931      const struct glsl_type *type = get_image_deref(instr)->type;
2932      dim = glsl_get_sampler_dim(type);
2933      is_array = glsl_sampler_type_is_array(type);
2934   }
2935
2936   struct waterfall_context wctx;
2937   LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
2938
2939   if (dim == GLSL_SAMPLER_DIM_BUF) {
2940      res = get_buffer_size(
2941         ctx, get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_BUFFER, false), true);
2942   } else {
2943
2944      struct ac_image_args args = {0};
2945
2946      args.dim = ac_get_image_dim(ctx->ac.gfx_level, dim, is_array);
2947      args.dmask = 0xf;
2948      args.resource = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_IMAGE, false);
2949      args.opcode = ac_image_get_resinfo;
2950      assert(nir_src_as_uint(instr->src[1]) == 0);
2951      args.lod = ctx->ac.i32_0;
2952      args.attributes = AC_FUNC_ATTR_READNONE;
2953
2954      res = ac_build_image_opcode(&ctx->ac, &args);
2955
2956      if (ctx->ac.gfx_level == GFX9 && dim == GLSL_SAMPLER_DIM_1D && is_array) {
2957         LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
2958         LLVMValueRef layers = LLVMBuildExtractElement(ctx->ac.builder, res, two, "");
2959         res = LLVMBuildInsertElement(ctx->ac.builder, res, layers, ctx->ac.i32_1, "");
2960      }
2961   }
2962   return exit_waterfall(ctx, &wctx, res);
2963}
2964
2965static void emit_discard(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr)
2966{
2967   LLVMValueRef cond;
2968
2969   if (instr->intrinsic == nir_intrinsic_discard_if ||
2970       instr->intrinsic == nir_intrinsic_terminate_if) {
2971      cond = LLVMBuildNot(ctx->ac.builder, get_src(ctx, instr->src[0]), "");
2972   } else {
2973      assert(instr->intrinsic == nir_intrinsic_discard ||
2974             instr->intrinsic == nir_intrinsic_terminate);
2975      cond = ctx->ac.i1false;
2976   }
2977
2978   ac_build_kill_if_false(&ctx->ac, cond);
2979}
2980
2981static void emit_demote(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr)
2982{
2983   LLVMValueRef cond;
2984
2985   if (instr->intrinsic == nir_intrinsic_demote_if) {
2986      cond = LLVMBuildNot(ctx->ac.builder, get_src(ctx, instr->src[0]), "");
2987   } else {
2988      assert(instr->intrinsic == nir_intrinsic_demote);
2989      cond = ctx->ac.i1false;
2990   }
2991
2992   if (LLVM_VERSION_MAJOR >= 13) {
2993      /* This demotes the pixel if the condition is false. */
2994      ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.wqm.demote", ctx->ac.voidt, &cond, 1, 0);
2995      return;
2996   }
2997
2998   LLVMValueRef mask = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.i1, ctx->ac.postponed_kill, "");
2999   mask = LLVMBuildAnd(ctx->ac.builder, mask, cond, "");
3000   LLVMBuildStore(ctx->ac.builder, mask, ctx->ac.postponed_kill);
3001
3002   if (!ctx->info->fs.needs_all_helper_invocations) {
3003      /* This is an optional optimization that only kills whole inactive quads.
3004       * It's not used when subgroup operations can possibly use all helper
3005       * invocations.
3006       */
3007      if (ctx->ac.flow->depth == 0) {
3008         ac_build_kill_if_false(&ctx->ac, ac_build_wqm_vote(&ctx->ac, cond));
3009      } else {
3010         /* amdgcn.wqm.vote doesn't work inside conditional blocks. Here's why.
3011          *
3012          * The problem is that kill(wqm.vote(0)) kills all active threads within
3013          * the block, which breaks the whole quad mode outside the block if
3014          * the conditional block has partially active quads (2x2 pixel blocks).
3015          * E.g. threads 0-3 are active outside the block, but only thread 0 is
3016          * active inside the block. Thread 0 shouldn't be killed by demote,
3017          * because threads 1-3 are still active outside the block.
3018          *
3019          * The fix for amdgcn.wqm.vote would be to return S_WQM((live & ~exec) | cond)
3020          * instead of S_WQM(cond).
3021          *
3022          * The less efficient workaround we do here is to save the kill condition
3023          * to a temporary (postponed_kill) and do kill(wqm.vote(cond)) after we
3024          * exit the conditional block.
3025          */
3026         ctx->ac.conditional_demote_seen = true;
3027      }
3028   }
3029}
3030
3031static LLVMValueRef visit_load_local_invocation_index(struct ac_nir_context *ctx)
3032{
3033   if (ctx->args->tcs_wave_id.used) {
3034      return ac_build_imad(&ctx->ac,
3035                           ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->tcs_wave_id), 0, 3),
3036                           LLVMConstInt(ctx->ac.i32, ctx->ac.wave_size, 0),
3037                           ac_get_thread_id(&ctx->ac));
3038   } else if (ctx->args->vs_rel_patch_id.used) {
3039      return ac_get_arg(&ctx->ac, ctx->args->vs_rel_patch_id);
3040   } else if (ctx->args->merged_wave_info.used) {
3041      /* Thread ID in threadgroup in merged ESGS. */
3042      LLVMValueRef wave_id = ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->merged_wave_info), 24, 4);
3043      LLVMValueRef wave_size = LLVMConstInt(ctx->ac.i32, ctx->ac.wave_size, false);
3044      LLVMValueRef threads_before = LLVMBuildMul(ctx->ac.builder, wave_id, wave_size, "");
3045      return LLVMBuildAdd(ctx->ac.builder, threads_before, ac_get_thread_id(&ctx->ac), "");
3046   }
3047
3048   LLVMValueRef result;
3049   LLVMValueRef thread_id = ac_get_thread_id(&ctx->ac);
3050   result = LLVMBuildAnd(ctx->ac.builder, ac_get_arg(&ctx->ac, ctx->args->tg_size),
3051                         LLVMConstInt(ctx->ac.i32, 0xfc0, false), "");
3052
3053   if (ctx->ac.wave_size == 32)
3054      result = LLVMBuildLShr(ctx->ac.builder, result, LLVMConstInt(ctx->ac.i32, 1, false), "");
3055
3056   return LLVMBuildAdd(ctx->ac.builder, result, thread_id, "");
3057}
3058
3059static LLVMValueRef visit_load_subgroup_id(struct ac_nir_context *ctx)
3060{
3061   if (ctx->stage == MESA_SHADER_COMPUTE) {
3062      LLVMValueRef result;
3063      result = LLVMBuildAnd(ctx->ac.builder, ac_get_arg(&ctx->ac, ctx->args->tg_size),
3064                            LLVMConstInt(ctx->ac.i32, 0xfc0, false), "");
3065      return LLVMBuildLShr(ctx->ac.builder, result, LLVMConstInt(ctx->ac.i32, 6, false), "");
3066   } else if (ctx->args->merged_wave_info.used) {
3067      return ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->merged_wave_info), 24, 4);
3068   } else {
3069      return LLVMConstInt(ctx->ac.i32, 0, false);
3070   }
3071}
3072
3073static LLVMValueRef visit_load_num_subgroups(struct ac_nir_context *ctx)
3074{
3075   if (ctx->stage == MESA_SHADER_COMPUTE) {
3076      return LLVMBuildAnd(ctx->ac.builder, ac_get_arg(&ctx->ac, ctx->args->tg_size),
3077                          LLVMConstInt(ctx->ac.i32, 0x3f, false), "");
3078   } else if (ctx->args->merged_wave_info.used) {
3079      return ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->merged_wave_info), 28, 4);
3080   } else {
3081      return LLVMConstInt(ctx->ac.i32, 1, false);
3082   }
3083}
3084
3085static LLVMValueRef visit_first_invocation(struct ac_nir_context *ctx)
3086{
3087   LLVMValueRef active_set = ac_build_ballot(&ctx->ac, ctx->ac.i32_1);
3088   const char *intr = ctx->ac.wave_size == 32 ? "llvm.cttz.i32" : "llvm.cttz.i64";
3089
3090   /* The second argument is whether cttz(0) should be defined, but we do not care. */
3091   LLVMValueRef args[] = {active_set, ctx->ac.i1false};
3092   LLVMValueRef result = ac_build_intrinsic(&ctx->ac, intr, ctx->ac.iN_wavemask, args, 2,
3093                                            AC_FUNC_ATTR_NOUNWIND | AC_FUNC_ATTR_READNONE);
3094
3095   return LLVMBuildTrunc(ctx->ac.builder, result, ctx->ac.i32, "");
3096}
3097
3098static LLVMValueRef visit_load_shared(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr)
3099{
3100   LLVMValueRef values[4], derived_ptr, index, ret;
3101   unsigned const_off = nir_intrinsic_base(instr);
3102
3103   LLVMTypeRef elem_type = LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
3104   LLVMValueRef ptr = get_memory_ptr_t(ctx, instr->src[0], elem_type, const_off);
3105
3106   for (int chan = 0; chan < instr->num_components; chan++) {
3107      index = LLVMConstInt(ctx->ac.i32, chan, 0);
3108      derived_ptr = LLVMBuildGEP2(ctx->ac.builder, elem_type, ptr, &index, 1, "");
3109      values[chan] = LLVMBuildLoad2(ctx->ac.builder, elem_type, derived_ptr, "");
3110   }
3111
3112   ret = ac_build_gather_values(&ctx->ac, values, instr->num_components);
3113
3114   return LLVMBuildBitCast(ctx->ac.builder, ret, get_def_type(ctx, &instr->dest.ssa), "");
3115}
3116
3117static void visit_store_shared(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr)
3118{
3119   LLVMValueRef derived_ptr, data, index;
3120   LLVMBuilderRef builder = ctx->ac.builder;
3121
3122   unsigned const_off = nir_intrinsic_base(instr);
3123   LLVMTypeRef elem_type = LLVMIntTypeInContext(ctx->ac.context, instr->src[0].ssa->bit_size);
3124   LLVMValueRef ptr = get_memory_ptr_t(ctx, instr->src[1], elem_type, const_off);
3125   LLVMValueRef src = get_src(ctx, instr->src[0]);
3126
3127   int writemask = nir_intrinsic_write_mask(instr);
3128   for (int chan = 0; chan < 4; chan++) {
3129      if (!(writemask & (1 << chan))) {
3130         continue;
3131      }
3132      data = ac_llvm_extract_elem(&ctx->ac, src, chan);
3133      index = LLVMConstInt(ctx->ac.i32, chan, 0);
3134      derived_ptr = LLVMBuildGEP2(builder, elem_type, ptr, &index, 1, "");
3135      LLVMBuildStore(builder, data, derived_ptr);
3136   }
3137}
3138
3139static LLVMValueRef visit_load_shared2_amd(struct ac_nir_context *ctx,
3140                                           const nir_intrinsic_instr *instr)
3141{
3142   LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[0], instr->dest.ssa.bit_size, 0);
3143
3144   LLVMValueRef values[2];
3145   uint8_t offsets[] = {nir_intrinsic_offset0(instr), nir_intrinsic_offset1(instr)};
3146   unsigned stride = nir_intrinsic_st64(instr) ? 64 : 1;
3147   for (unsigned i = 0; i < 2; i++) {
3148      LLVMValueRef index = LLVMConstInt(ctx->ac.i32, offsets[i] * stride, 0);
3149      LLVMValueRef derived_ptr = LLVMBuildGEP(ctx->ac.builder, ptr, &index, 1, "");
3150      values[i] = LLVMBuildLoad(ctx->ac.builder, derived_ptr, "");
3151   }
3152
3153   LLVMValueRef ret = ac_build_gather_values(&ctx->ac, values, 2);
3154   return LLVMBuildBitCast(ctx->ac.builder, ret, get_def_type(ctx, &instr->dest.ssa), "");
3155}
3156
3157static void visit_store_shared2_amd(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr)
3158{
3159   LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[1], instr->src[0].ssa->bit_size, 0);
3160   LLVMValueRef src = get_src(ctx, instr->src[0]);
3161
3162   uint8_t offsets[] = {nir_intrinsic_offset0(instr), nir_intrinsic_offset1(instr)};
3163   unsigned stride = nir_intrinsic_st64(instr) ? 64 : 1;
3164   for (unsigned i = 0; i < 2; i++) {
3165      LLVMValueRef index = LLVMConstInt(ctx->ac.i32, offsets[i] * stride, 0);
3166      LLVMValueRef derived_ptr = LLVMBuildGEP(ctx->ac.builder, ptr, &index, 1, "");
3167      LLVMBuildStore(ctx->ac.builder, ac_llvm_extract_elem(&ctx->ac, src, i), derived_ptr);
3168   }
3169}
3170
3171static LLVMValueRef visit_var_atomic(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr,
3172                                     LLVMValueRef ptr, int src_idx)
3173{
3174   if (ctx->ac.postponed_kill) {
3175      LLVMValueRef cond = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.i1, ctx->ac.postponed_kill, "");
3176      ac_build_ifcc(&ctx->ac, cond, 7005);
3177   }
3178
3179   LLVMValueRef result;
3180   LLVMValueRef src = get_src(ctx, instr->src[src_idx]);
3181
3182   const char *sync_scope = "workgroup-one-as";
3183
3184   if (instr->intrinsic == nir_intrinsic_shared_atomic_comp_swap) {
3185      LLVMValueRef src1 = get_src(ctx, instr->src[src_idx + 1]);
3186      result = ac_build_atomic_cmp_xchg(&ctx->ac, ptr, src, src1, sync_scope);
3187      result = LLVMBuildExtractValue(ctx->ac.builder, result, 0, "");
3188   } else if (instr->intrinsic == nir_intrinsic_shared_atomic_fmin ||
3189              instr->intrinsic == nir_intrinsic_shared_atomic_fmax) {
3190      const char *op = instr->intrinsic == nir_intrinsic_shared_atomic_fmin ? "fmin" : "fmax";
3191      char name[64], type[8];
3192      LLVMValueRef params[5];
3193      LLVMTypeRef src_type;
3194      int arg_count = 0;
3195
3196      src = ac_to_float(&ctx->ac, src);
3197      src_type = LLVMTypeOf(src);
3198
3199      LLVMTypeRef ptr_type =
3200         LLVMPointerType(src_type, LLVMGetPointerAddressSpace(LLVMTypeOf(ptr)));
3201      ptr = LLVMBuildBitCast(ctx->ac.builder, ptr, ptr_type, "");
3202
3203      params[arg_count++] = ptr;
3204      params[arg_count++] = src;
3205      params[arg_count++] = ctx->ac.i32_0;
3206      params[arg_count++] = ctx->ac.i32_0;
3207      params[arg_count++] = ctx->ac.i1false;
3208
3209      ac_build_type_name_for_intr(src_type, type, sizeof(type));
3210      snprintf(name, sizeof(name), "llvm.amdgcn.ds.%s.%s", op, type);
3211
3212      result = ac_build_intrinsic(&ctx->ac, name, src_type, params, arg_count, 0);
3213      result = ac_to_integer(&ctx->ac, result);
3214   } else {
3215      LLVMAtomicRMWBinOp op;
3216      switch (instr->intrinsic) {
3217      case nir_intrinsic_shared_atomic_add:
3218         op = LLVMAtomicRMWBinOpAdd;
3219         break;
3220      case nir_intrinsic_shared_atomic_umin:
3221         op = LLVMAtomicRMWBinOpUMin;
3222         break;
3223      case nir_intrinsic_shared_atomic_umax:
3224         op = LLVMAtomicRMWBinOpUMax;
3225         break;
3226      case nir_intrinsic_shared_atomic_imin:
3227         op = LLVMAtomicRMWBinOpMin;
3228         break;
3229      case nir_intrinsic_shared_atomic_imax:
3230         op = LLVMAtomicRMWBinOpMax;
3231         break;
3232      case nir_intrinsic_shared_atomic_and:
3233         op = LLVMAtomicRMWBinOpAnd;
3234         break;
3235      case nir_intrinsic_shared_atomic_or:
3236         op = LLVMAtomicRMWBinOpOr;
3237         break;
3238      case nir_intrinsic_shared_atomic_xor:
3239         op = LLVMAtomicRMWBinOpXor;
3240         break;
3241      case nir_intrinsic_shared_atomic_exchange:
3242         op = LLVMAtomicRMWBinOpXchg;
3243         break;
3244      case nir_intrinsic_shared_atomic_fadd:
3245         op = LLVMAtomicRMWBinOpFAdd;
3246         break;
3247      default:
3248         return NULL;
3249      }
3250
3251      LLVMValueRef val;
3252
3253      if (instr->intrinsic == nir_intrinsic_shared_atomic_fadd) {
3254         val = ac_to_float(&ctx->ac, src);
3255
3256         LLVMTypeRef ptr_type =
3257            LLVMPointerType(LLVMTypeOf(val), LLVMGetPointerAddressSpace(LLVMTypeOf(ptr)));
3258         ptr = LLVMBuildBitCast(ctx->ac.builder, ptr, ptr_type, "");
3259      } else {
3260         val = ac_to_integer(&ctx->ac, src);
3261      }
3262
3263      result = ac_build_atomic_rmw(&ctx->ac, op, ptr, val, sync_scope);
3264
3265      if (instr->intrinsic == nir_intrinsic_shared_atomic_fadd ||
3266          instr->intrinsic == nir_intrinsic_deref_atomic_fadd) {
3267         result = ac_to_integer(&ctx->ac, result);
3268      }
3269   }
3270
3271   if (ctx->ac.postponed_kill)
3272      ac_build_endif(&ctx->ac, 7005);
3273   return result;
3274}
3275
3276static LLVMValueRef load_sample_pos(struct ac_nir_context *ctx)
3277{
3278   LLVMValueRef values[2];
3279   LLVMValueRef pos[2];
3280
3281   pos[0] = ac_to_float(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->frag_pos[0]));
3282   pos[1] = ac_to_float(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->frag_pos[1]));
3283
3284   values[0] = ac_build_fract(&ctx->ac, pos[0], 32);
3285   values[1] = ac_build_fract(&ctx->ac, pos[1], 32);
3286   return ac_build_gather_values(&ctx->ac, values, 2);
3287}
3288
3289static LLVMValueRef lookup_interp_param(struct ac_nir_context *ctx, enum glsl_interp_mode interp,
3290                                        unsigned location)
3291{
3292   switch (interp) {
3293   case INTERP_MODE_FLAT:
3294   default:
3295      return NULL;
3296   case INTERP_MODE_SMOOTH:
3297   case INTERP_MODE_NONE:
3298      if (location == INTERP_CENTER)
3299         return ac_get_arg(&ctx->ac, ctx->args->persp_center);
3300      else if (location == INTERP_CENTROID)
3301         return ctx->abi->persp_centroid;
3302      else if (location == INTERP_SAMPLE)
3303         return ac_get_arg(&ctx->ac, ctx->args->persp_sample);
3304      break;
3305   case INTERP_MODE_NOPERSPECTIVE:
3306      if (location == INTERP_CENTER)
3307         return ac_get_arg(&ctx->ac, ctx->args->linear_center);
3308      else if (location == INTERP_CENTROID)
3309         return ctx->abi->linear_centroid;
3310      else if (location == INTERP_SAMPLE)
3311         return ac_get_arg(&ctx->ac, ctx->args->linear_sample);
3312      break;
3313   }
3314   return NULL;
3315}
3316
3317static LLVMValueRef barycentric_center(struct ac_nir_context *ctx, unsigned mode)
3318{
3319   LLVMValueRef interp_param = lookup_interp_param(ctx, mode, INTERP_CENTER);
3320   return LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2i32, "");
3321}
3322
3323static LLVMValueRef barycentric_offset(struct ac_nir_context *ctx, unsigned mode,
3324                                       LLVMValueRef offset)
3325{
3326   LLVMValueRef interp_param = lookup_interp_param(ctx, mode, INTERP_CENTER);
3327   LLVMValueRef src_c0 =
3328      ac_to_float(&ctx->ac, LLVMBuildExtractElement(ctx->ac.builder, offset, ctx->ac.i32_0, ""));
3329   LLVMValueRef src_c1 =
3330      ac_to_float(&ctx->ac, LLVMBuildExtractElement(ctx->ac.builder, offset, ctx->ac.i32_1, ""));
3331
3332   LLVMValueRef ij_out[2];
3333   LLVMValueRef ddxy_out = ac_build_ddxy_interp(&ctx->ac, interp_param);
3334
3335   /*
3336    * take the I then J parameters, and the DDX/Y for it, and
3337    * calculate the IJ inputs for the interpolator.
3338    * temp1 = ddx * offset/sample.x + I;
3339    * interp_param.I = ddy * offset/sample.y + temp1;
3340    * temp1 = ddx * offset/sample.x + J;
3341    * interp_param.J = ddy * offset/sample.y + temp1;
3342    */
3343   for (unsigned i = 0; i < 2; i++) {
3344      LLVMValueRef ix_ll = LLVMConstInt(ctx->ac.i32, i, false);
3345      LLVMValueRef iy_ll = LLVMConstInt(ctx->ac.i32, i + 2, false);
3346      LLVMValueRef ddx_el = LLVMBuildExtractElement(ctx->ac.builder, ddxy_out, ix_ll, "");
3347      LLVMValueRef ddy_el = LLVMBuildExtractElement(ctx->ac.builder, ddxy_out, iy_ll, "");
3348      LLVMValueRef interp_el = LLVMBuildExtractElement(ctx->ac.builder, interp_param, ix_ll, "");
3349      LLVMValueRef temp1, temp2;
3350
3351      interp_el = LLVMBuildBitCast(ctx->ac.builder, interp_el, ctx->ac.f32, "");
3352
3353      temp1 = ac_build_fmad(&ctx->ac, ddx_el, src_c0, interp_el);
3354      temp2 = ac_build_fmad(&ctx->ac, ddy_el, src_c1, temp1);
3355
3356      ij_out[i] = LLVMBuildBitCast(ctx->ac.builder, temp2, ctx->ac.i32, "");
3357   }
3358   interp_param = ac_build_gather_values(&ctx->ac, ij_out, 2);
3359   return LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2i32, "");
3360}
3361
3362static LLVMValueRef barycentric_centroid(struct ac_nir_context *ctx, unsigned mode)
3363{
3364   LLVMValueRef interp_param = lookup_interp_param(ctx, mode, INTERP_CENTROID);
3365   return LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2i32, "");
3366}
3367
3368static LLVMValueRef barycentric_at_sample(struct ac_nir_context *ctx, unsigned mode,
3369                                          LLVMValueRef sample_id)
3370{
3371   if (ctx->abi->interp_at_sample_force_center)
3372      return barycentric_center(ctx, mode);
3373
3374   LLVMValueRef halfval = LLVMConstReal(ctx->ac.f32, 0.5f);
3375
3376   /* fetch sample ID */
3377   LLVMValueRef sample_pos = ctx->abi->load_sample_position(ctx->abi, sample_id);
3378
3379   LLVMValueRef src_c0 = LLVMBuildExtractElement(ctx->ac.builder, sample_pos, ctx->ac.i32_0, "");
3380   src_c0 = LLVMBuildFSub(ctx->ac.builder, src_c0, halfval, "");
3381   LLVMValueRef src_c1 = LLVMBuildExtractElement(ctx->ac.builder, sample_pos, ctx->ac.i32_1, "");
3382   src_c1 = LLVMBuildFSub(ctx->ac.builder, src_c1, halfval, "");
3383   LLVMValueRef coords[] = {src_c0, src_c1};
3384   LLVMValueRef offset = ac_build_gather_values(&ctx->ac, coords, 2);
3385
3386   return barycentric_offset(ctx, mode, offset);
3387}
3388
3389static LLVMValueRef barycentric_sample(struct ac_nir_context *ctx, unsigned mode)
3390{
3391   LLVMValueRef interp_param = lookup_interp_param(ctx, mode, INTERP_SAMPLE);
3392   return LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2i32, "");
3393}
3394
3395static LLVMValueRef barycentric_model(struct ac_nir_context *ctx)
3396{
3397   return LLVMBuildBitCast(ctx->ac.builder, ac_get_arg(&ctx->ac, ctx->args->pull_model),
3398                           ctx->ac.v3i32, "");
3399}
3400
3401static LLVMValueRef load_interpolated_input(struct ac_nir_context *ctx, LLVMValueRef interp_param,
3402                                            unsigned index, unsigned comp_start,
3403                                            unsigned num_components, unsigned bitsize,
3404                                            bool high_16bits)
3405{
3406   LLVMValueRef attr_number = LLVMConstInt(ctx->ac.i32, index, false);
3407   LLVMValueRef interp_param_f;
3408
3409   interp_param_f = LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2f32, "");
3410   LLVMValueRef i = LLVMBuildExtractElement(ctx->ac.builder, interp_param_f, ctx->ac.i32_0, "");
3411   LLVMValueRef j = LLVMBuildExtractElement(ctx->ac.builder, interp_param_f, ctx->ac.i32_1, "");
3412
3413   /* Workaround for issue 2647: kill threads with infinite interpolation coeffs */
3414   if (ctx->verified_interp && !_mesa_hash_table_search(ctx->verified_interp, interp_param)) {
3415      LLVMValueRef cond = ac_build_is_inf_or_nan(&ctx->ac, i);
3416      ac_build_kill_if_false(&ctx->ac, LLVMBuildNot(ctx->ac.builder, cond, ""));
3417      _mesa_hash_table_insert(ctx->verified_interp, interp_param, interp_param);
3418   }
3419
3420   LLVMValueRef values[4];
3421   assert(bitsize == 16 || bitsize == 32);
3422   for (unsigned comp = 0; comp < num_components; comp++) {
3423      LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, comp_start + comp, false);
3424      if (bitsize == 16) {
3425         values[comp] = ac_build_fs_interp_f16(&ctx->ac, llvm_chan, attr_number,
3426                                               ac_get_arg(&ctx->ac, ctx->args->prim_mask), i, j,
3427                                               high_16bits);
3428      } else {
3429         values[comp] = ac_build_fs_interp(&ctx->ac, llvm_chan, attr_number,
3430                                           ac_get_arg(&ctx->ac, ctx->args->prim_mask), i, j);
3431      }
3432   }
3433
3434   return ac_to_integer(&ctx->ac, ac_build_gather_values(&ctx->ac, values, num_components));
3435}
3436
3437static LLVMValueRef visit_load(struct ac_nir_context *ctx, nir_intrinsic_instr *instr,
3438                               bool is_output)
3439{
3440   LLVMValueRef values[8];
3441   LLVMTypeRef dest_type = get_def_type(ctx, &instr->dest.ssa);
3442   LLVMTypeRef component_type;
3443   unsigned base = nir_intrinsic_base(instr);
3444   unsigned component = nir_intrinsic_component(instr);
3445   unsigned count = instr->dest.ssa.num_components;
3446   nir_src *vertex_index_src = nir_get_io_arrayed_index_src(instr);
3447   LLVMValueRef vertex_index = vertex_index_src ? get_src(ctx, *vertex_index_src) : NULL;
3448   nir_src offset = *nir_get_io_offset_src(instr);
3449   LLVMValueRef indir_index = NULL;
3450
3451   switch (instr->dest.ssa.bit_size) {
3452   case 16:
3453   case 32:
3454      break;
3455   case 64:
3456      unreachable("64-bit IO should have been lowered");
3457      return NULL;
3458   default:
3459      unreachable("unhandled load type");
3460      return NULL;
3461   }
3462
3463   if (LLVMGetTypeKind(dest_type) == LLVMVectorTypeKind)
3464      component_type = LLVMGetElementType(dest_type);
3465   else
3466      component_type = dest_type;
3467
3468   if (nir_src_is_const(offset))
3469      assert(nir_src_as_uint(offset) == 0);
3470   else
3471      indir_index = get_src(ctx, offset);
3472
3473   if (ctx->stage == MESA_SHADER_TESS_CTRL) {
3474      LLVMValueRef result = ctx->abi->load_tess_varyings(ctx->abi, component_type,
3475                                                         vertex_index, indir_index,
3476                                                         base, component,
3477                                                         count, !is_output);
3478      if (instr->dest.ssa.bit_size == 16) {
3479         result = ac_to_integer(&ctx->ac, result);
3480         result = LLVMBuildTrunc(ctx->ac.builder, result, dest_type, "");
3481      }
3482      return LLVMBuildBitCast(ctx->ac.builder, result, dest_type, "");
3483   }
3484
3485   /* No indirect indexing is allowed after this point. */
3486   assert(!indir_index);
3487
3488   if (ctx->stage == MESA_SHADER_FRAGMENT && is_output &&
3489       nir_intrinsic_io_semantics(instr).fb_fetch_output)
3490      return ctx->abi->emit_fbfetch(ctx->abi);
3491
3492   if (ctx->stage == MESA_SHADER_VERTEX && !is_output)
3493      return ctx->abi->load_inputs(ctx->abi, base, component, count, 0, component_type);
3494
3495   /* Other non-fragment cases have outputs in temporaries. */
3496   if (is_output && (ctx->stage == MESA_SHADER_VERTEX || ctx->stage == MESA_SHADER_TESS_EVAL)) {
3497      assert(is_output);
3498
3499      for (unsigned chan = component; chan < count + component; chan++)
3500         values[chan] = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.f32,
3501                                       ctx->abi->outputs[base * 4 + chan], "");
3502
3503      LLVMValueRef result = ac_build_varying_gather_values(&ctx->ac, values, count, component);
3504      return LLVMBuildBitCast(ctx->ac.builder, result, dest_type, "");
3505   }
3506
3507   /* Fragment shader inputs. */
3508   assert(ctx->stage == MESA_SHADER_FRAGMENT);
3509   unsigned vertex_id = 2; /* P0 */
3510
3511   if (instr->intrinsic == nir_intrinsic_load_input_vertex) {
3512      nir_const_value *src0 = nir_src_as_const_value(instr->src[0]);
3513
3514      switch (src0[0].i32) {
3515      case 0:
3516         vertex_id = 2;
3517         break;
3518      case 1:
3519         vertex_id = 0;
3520         break;
3521      case 2:
3522         vertex_id = 1;
3523         break;
3524      default:
3525         unreachable("Invalid vertex index");
3526      }
3527   }
3528
3529   LLVMValueRef attr_number = LLVMConstInt(ctx->ac.i32, base, false);
3530
3531   for (unsigned chan = 0; chan < count; chan++) {
3532      LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, (component + chan) % 4, false);
3533      values[chan] =
3534         ac_build_fs_interp_mov(&ctx->ac, LLVMConstInt(ctx->ac.i32, vertex_id, false), llvm_chan,
3535                                attr_number, ac_get_arg(&ctx->ac, ctx->args->prim_mask));
3536      values[chan] = LLVMBuildBitCast(ctx->ac.builder, values[chan], ctx->ac.i32, "");
3537      if (instr->dest.ssa.bit_size == 16 &&
3538          nir_intrinsic_io_semantics(instr).high_16bits)
3539         values[chan] = LLVMBuildLShr(ctx->ac.builder, values[chan], LLVMConstInt(ctx->ac.i32, 16, 0), "");
3540      values[chan] =
3541         LLVMBuildTruncOrBitCast(ctx->ac.builder, values[chan],
3542                                 instr->dest.ssa.bit_size == 16 ? ctx->ac.i16 : ctx->ac.i32, "");
3543   }
3544
3545   LLVMValueRef result = ac_build_gather_values(&ctx->ac, values, count);
3546   return LLVMBuildBitCast(ctx->ac.builder, result, dest_type, "");
3547}
3548
3549static LLVMValueRef
3550emit_load_frag_shading_rate(struct ac_nir_context *ctx)
3551{
3552   LLVMValueRef x_rate, y_rate, cond;
3553
3554   /* VRS Rate X = Ancillary[2:3]
3555    * VRS Rate Y = Ancillary[4:5]
3556    */
3557   x_rate = ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->ancillary), 2, 2);
3558   y_rate = ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->ancillary), 4, 2);
3559
3560   /* xRate = xRate == 0x1 ? Horizontal2Pixels : None. */
3561   cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, x_rate, ctx->ac.i32_1, "");
3562   x_rate = LLVMBuildSelect(ctx->ac.builder, cond,
3563                            LLVMConstInt(ctx->ac.i32, 4, false), ctx->ac.i32_0, "");
3564
3565   /* yRate = yRate == 0x1 ? Vertical2Pixels : None. */
3566   cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, y_rate, ctx->ac.i32_1, "");
3567   y_rate = LLVMBuildSelect(ctx->ac.builder, cond,
3568                            LLVMConstInt(ctx->ac.i32, 1, false), ctx->ac.i32_0, "");
3569
3570   return LLVMBuildOr(ctx->ac.builder, x_rate, y_rate, "");
3571}
3572
3573static LLVMValueRef
3574emit_load_frag_coord(struct ac_nir_context *ctx)
3575{
3576   LLVMValueRef values[4] = {
3577      ac_get_arg(&ctx->ac, ctx->args->frag_pos[0]), ac_get_arg(&ctx->ac, ctx->args->frag_pos[1]),
3578      ac_get_arg(&ctx->ac, ctx->args->frag_pos[2]),
3579      ac_build_fdiv(&ctx->ac, ctx->ac.f32_1, ac_get_arg(&ctx->ac, ctx->args->frag_pos[3]))};
3580
3581   return ac_to_integer(&ctx->ac, ac_build_gather_values(&ctx->ac, values, 4));
3582}
3583
3584static void visit_intrinsic(struct ac_nir_context *ctx, nir_intrinsic_instr *instr)
3585{
3586   LLVMValueRef result = NULL;
3587
3588   switch (instr->intrinsic) {
3589   case nir_intrinsic_ballot:
3590      result = ac_build_ballot(&ctx->ac, get_src(ctx, instr->src[0]));
3591      if (ctx->ac.ballot_mask_bits > ctx->ac.wave_size)
3592         result = LLVMBuildZExt(ctx->ac.builder, result, ctx->ac.iN_ballotmask, "");
3593      break;
3594   case nir_intrinsic_read_invocation:
3595      result =
3596         ac_build_readlane(&ctx->ac, get_src(ctx, instr->src[0]), get_src(ctx, instr->src[1]));
3597      break;
3598   case nir_intrinsic_read_first_invocation:
3599      result = ac_build_readlane(&ctx->ac, get_src(ctx, instr->src[0]), NULL);
3600      break;
3601   case nir_intrinsic_load_subgroup_invocation:
3602      result = ac_get_thread_id(&ctx->ac);
3603      break;
3604   case nir_intrinsic_load_workgroup_id: {
3605      LLVMValueRef values[3];
3606
3607      for (int i = 0; i < 3; i++) {
3608         values[i] = ctx->args->workgroup_ids[i].used
3609                        ? ac_get_arg(&ctx->ac, ctx->args->workgroup_ids[i])
3610                        : ctx->ac.i32_0;
3611      }
3612
3613      result = ac_build_gather_values(&ctx->ac, values, 3);
3614      break;
3615   }
3616   case nir_intrinsic_load_base_vertex:
3617   case nir_intrinsic_load_first_vertex:
3618   case nir_intrinsic_load_workgroup_size:
3619   case nir_intrinsic_load_tess_level_outer_default:
3620   case nir_intrinsic_load_tess_level_inner_default:
3621   case nir_intrinsic_load_tess_rel_patch_id_amd:
3622   case nir_intrinsic_load_patch_vertices_in:
3623   case nir_intrinsic_load_sample_mask_in:
3624   case nir_intrinsic_load_ring_tess_factors_amd:
3625   case nir_intrinsic_load_ring_tess_offchip_amd:
3626   case nir_intrinsic_load_ring_tess_offchip_offset_amd:
3627   case nir_intrinsic_load_ring_esgs_amd:
3628   case nir_intrinsic_load_ring_es2gs_offset_amd:
3629   case nir_intrinsic_load_lshs_vertex_stride_amd:
3630   case nir_intrinsic_load_tcs_num_patches_amd:
3631   case nir_intrinsic_load_hs_out_patch_data_offset_amd:
3632      result = ctx->abi->intrinsic_load(ctx->abi, instr->intrinsic);
3633      break;
3634   case nir_intrinsic_load_vertex_id_zero_base: {
3635      result = ctx->vertex_id_replaced ? ctx->vertex_id_replaced : ctx->abi->vertex_id;
3636      break;
3637   }
3638   case nir_intrinsic_load_local_invocation_id: {
3639      LLVMValueRef ids = ac_get_arg(&ctx->ac, ctx->args->local_invocation_ids);
3640
3641      if (LLVMGetTypeKind(LLVMTypeOf(ids)) == LLVMIntegerTypeKind) {
3642         /* Thread IDs are packed in VGPR0, 10 bits per component. */
3643         LLVMValueRef id[3];
3644
3645         for (unsigned i = 0; i < 3; i++)
3646            id[i] = ac_unpack_param(&ctx->ac, ids, i * 10, 10);
3647
3648         result = ac_build_gather_values(&ctx->ac, id, 3);
3649      } else {
3650         result = ids;
3651      }
3652      break;
3653   }
3654   case nir_intrinsic_load_base_instance:
3655      result = ac_get_arg(&ctx->ac, ctx->args->start_instance);
3656      break;
3657   case nir_intrinsic_load_draw_id:
3658      result = ac_get_arg(&ctx->ac, ctx->args->draw_id);
3659      break;
3660   case nir_intrinsic_load_view_index:
3661      result = ac_get_arg(&ctx->ac, ctx->args->view_index);
3662      break;
3663   case nir_intrinsic_load_invocation_id:
3664      if (ctx->stage == MESA_SHADER_TESS_CTRL) {
3665         result = ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->tcs_rel_ids), 8, 5);
3666      } else {
3667         if (ctx->ac.gfx_level >= GFX10) {
3668            result =
3669               LLVMBuildAnd(ctx->ac.builder, ac_get_arg(&ctx->ac, ctx->args->gs_invocation_id),
3670                            LLVMConstInt(ctx->ac.i32, 127, 0), "");
3671         } else {
3672            result = ac_get_arg(&ctx->ac, ctx->args->gs_invocation_id);
3673         }
3674      }
3675      break;
3676   case nir_intrinsic_load_primitive_id:
3677      if (ctx->stage == MESA_SHADER_GEOMETRY) {
3678         result = ac_get_arg(&ctx->ac, ctx->args->gs_prim_id);
3679      } else if (ctx->stage == MESA_SHADER_TESS_CTRL) {
3680         result = ac_get_arg(&ctx->ac, ctx->args->tcs_patch_id);
3681      } else if (ctx->stage == MESA_SHADER_TESS_EVAL) {
3682         result = ctx->tes_patch_id_replaced ? ctx->tes_patch_id_replaced
3683                                             : ac_get_arg(&ctx->ac, ctx->args->tes_patch_id);
3684      } else if (ctx->stage == MESA_SHADER_VERTEX) {
3685         if (ctx->args->vs_prim_id.used)
3686            result = ac_get_arg(&ctx->ac, ctx->args->vs_prim_id); /* legacy */
3687         else
3688            result = ac_get_arg(&ctx->ac, ctx->args->gs_prim_id); /* NGG */
3689      } else
3690         fprintf(stderr, "Unknown primitive id intrinsic: %d", ctx->stage);
3691      break;
3692   case nir_intrinsic_load_sample_id:
3693      result = ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->ancillary), 8, 4);
3694      break;
3695   case nir_intrinsic_load_sample_pos:
3696      result = load_sample_pos(ctx);
3697      break;
3698   case nir_intrinsic_load_frag_coord:
3699      result = emit_load_frag_coord(ctx);
3700      break;
3701   case nir_intrinsic_load_frag_shading_rate:
3702      result = emit_load_frag_shading_rate(ctx);
3703      break;
3704   case nir_intrinsic_load_front_face:
3705      result = emit_i2b(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->front_face));
3706      break;
3707   case nir_intrinsic_load_helper_invocation:
3708      result = ac_build_load_helper_invocation(&ctx->ac);
3709      break;
3710   case nir_intrinsic_is_helper_invocation:
3711      result = ac_build_is_helper_invocation(&ctx->ac);
3712      break;
3713   case nir_intrinsic_load_color0:
3714      result = ctx->abi->color0;
3715      break;
3716   case nir_intrinsic_load_color1:
3717      result = ctx->abi->color1;
3718      break;
3719   case nir_intrinsic_load_user_data_amd:
3720      assert(LLVMTypeOf(ctx->abi->user_data) == ctx->ac.v4i32);
3721      result = ctx->abi->user_data;
3722      break;
3723   case nir_intrinsic_load_instance_id:
3724      result = ctx->instance_id_replaced ? ctx->instance_id_replaced : ctx->abi->instance_id;
3725      break;
3726   case nir_intrinsic_load_num_workgroups:
3727      if (ctx->abi->load_grid_size_from_user_sgpr) {
3728         result = ac_get_arg(&ctx->ac, ctx->args->num_work_groups);
3729      } else {
3730         LLVMTypeRef ptr_type = ac_array_in_const_addr_space(ctx->ac.v3i32);
3731         LLVMValueRef ptr = ac_get_arg(&ctx->ac, ctx->args->num_work_groups);
3732         ptr = LLVMBuildBitCast(ctx->ac.builder, ptr, ptr_type, "");
3733         result = ac_build_load_invariant(&ctx->ac, ptr, ctx->ac.i32_0);
3734      }
3735      break;
3736   case nir_intrinsic_load_local_invocation_index:
3737      result = visit_load_local_invocation_index(ctx);
3738      break;
3739   case nir_intrinsic_load_subgroup_id:
3740      result = visit_load_subgroup_id(ctx);
3741      break;
3742   case nir_intrinsic_load_num_subgroups:
3743      result = visit_load_num_subgroups(ctx);
3744      break;
3745   case nir_intrinsic_first_invocation:
3746      result = visit_first_invocation(ctx);
3747      break;
3748   case nir_intrinsic_load_push_constant:
3749      result = visit_load_push_constant(ctx, instr);
3750      break;
3751   case nir_intrinsic_store_ssbo:
3752      visit_store_ssbo(ctx, instr);
3753      break;
3754   case nir_intrinsic_load_ssbo:
3755      result = visit_load_buffer(ctx, instr);
3756      break;
3757   case nir_intrinsic_load_global_constant:
3758   case nir_intrinsic_load_global:
3759   case nir_intrinsic_load_global_amd:
3760      result = visit_load_global(ctx, instr);
3761      break;
3762   case nir_intrinsic_store_global:
3763   case nir_intrinsic_store_global_amd:
3764      visit_store_global(ctx, instr);
3765      break;
3766   case nir_intrinsic_global_atomic_add:
3767   case nir_intrinsic_global_atomic_imin:
3768   case nir_intrinsic_global_atomic_umin:
3769   case nir_intrinsic_global_atomic_imax:
3770   case nir_intrinsic_global_atomic_umax:
3771   case nir_intrinsic_global_atomic_and:
3772   case nir_intrinsic_global_atomic_or:
3773   case nir_intrinsic_global_atomic_xor:
3774   case nir_intrinsic_global_atomic_exchange:
3775   case nir_intrinsic_global_atomic_comp_swap:
3776   case nir_intrinsic_global_atomic_fmin:
3777   case nir_intrinsic_global_atomic_fmax:
3778   case nir_intrinsic_global_atomic_add_amd:
3779   case nir_intrinsic_global_atomic_imin_amd:
3780   case nir_intrinsic_global_atomic_umin_amd:
3781   case nir_intrinsic_global_atomic_imax_amd:
3782   case nir_intrinsic_global_atomic_umax_amd:
3783   case nir_intrinsic_global_atomic_and_amd:
3784   case nir_intrinsic_global_atomic_or_amd:
3785   case nir_intrinsic_global_atomic_xor_amd:
3786   case nir_intrinsic_global_atomic_exchange_amd:
3787   case nir_intrinsic_global_atomic_comp_swap_amd:
3788   case nir_intrinsic_global_atomic_fmin_amd:
3789   case nir_intrinsic_global_atomic_fmax_amd:
3790      result = visit_global_atomic(ctx, instr);
3791      break;
3792   case nir_intrinsic_ssbo_atomic_add:
3793   case nir_intrinsic_ssbo_atomic_imin:
3794   case nir_intrinsic_ssbo_atomic_umin:
3795   case nir_intrinsic_ssbo_atomic_imax:
3796   case nir_intrinsic_ssbo_atomic_umax:
3797   case nir_intrinsic_ssbo_atomic_and:
3798   case nir_intrinsic_ssbo_atomic_or:
3799   case nir_intrinsic_ssbo_atomic_xor:
3800   case nir_intrinsic_ssbo_atomic_exchange:
3801   case nir_intrinsic_ssbo_atomic_comp_swap:
3802   case nir_intrinsic_ssbo_atomic_fmin:
3803   case nir_intrinsic_ssbo_atomic_fmax:
3804      result = visit_atomic_ssbo(ctx, instr);
3805      break;
3806   case nir_intrinsic_load_ubo:
3807      result = visit_load_ubo_buffer(ctx, instr);
3808      break;
3809   case nir_intrinsic_get_ssbo_size:
3810      result = visit_get_ssbo_size(ctx, instr);
3811      break;
3812   case nir_intrinsic_load_input:
3813   case nir_intrinsic_load_input_vertex:
3814   case nir_intrinsic_load_per_vertex_input:
3815      result = visit_load(ctx, instr, false);
3816      break;
3817   case nir_intrinsic_load_output:
3818   case nir_intrinsic_load_per_vertex_output:
3819      result = visit_load(ctx, instr, true);
3820      break;
3821   case nir_intrinsic_store_output:
3822   case nir_intrinsic_store_per_vertex_output:
3823      visit_store_output(ctx, instr);
3824      break;
3825   case nir_intrinsic_load_shared:
3826      result = visit_load_shared(ctx, instr);
3827      break;
3828   case nir_intrinsic_store_shared:
3829      visit_store_shared(ctx, instr);
3830      break;
3831   case nir_intrinsic_load_shared2_amd:
3832      result = visit_load_shared2_amd(ctx, instr);
3833      break;
3834   case nir_intrinsic_store_shared2_amd:
3835      visit_store_shared2_amd(ctx, instr);
3836      break;
3837   case nir_intrinsic_bindless_image_samples:
3838   case nir_intrinsic_image_deref_samples:
3839      result = visit_image_samples(ctx, instr);
3840      break;
3841   case nir_intrinsic_bindless_image_load:
3842   case nir_intrinsic_bindless_image_sparse_load:
3843      result = visit_image_load(ctx, instr, true);
3844      break;
3845   case nir_intrinsic_image_deref_load:
3846   case nir_intrinsic_image_deref_sparse_load:
3847      result = visit_image_load(ctx, instr, false);
3848      break;
3849   case nir_intrinsic_bindless_image_store:
3850      visit_image_store(ctx, instr, true);
3851      break;
3852   case nir_intrinsic_image_deref_store:
3853      visit_image_store(ctx, instr, false);
3854      break;
3855   case nir_intrinsic_bindless_image_atomic_add:
3856   case nir_intrinsic_bindless_image_atomic_imin:
3857   case nir_intrinsic_bindless_image_atomic_umin:
3858   case nir_intrinsic_bindless_image_atomic_imax:
3859   case nir_intrinsic_bindless_image_atomic_umax:
3860   case nir_intrinsic_bindless_image_atomic_and:
3861   case nir_intrinsic_bindless_image_atomic_or:
3862   case nir_intrinsic_bindless_image_atomic_xor:
3863   case nir_intrinsic_bindless_image_atomic_exchange:
3864   case nir_intrinsic_bindless_image_atomic_comp_swap:
3865   case nir_intrinsic_bindless_image_atomic_inc_wrap:
3866   case nir_intrinsic_bindless_image_atomic_dec_wrap:
3867      result = visit_image_atomic(ctx, instr, true);
3868      break;
3869   case nir_intrinsic_image_deref_atomic_add:
3870   case nir_intrinsic_image_deref_atomic_imin:
3871   case nir_intrinsic_image_deref_atomic_umin:
3872   case nir_intrinsic_image_deref_atomic_imax:
3873   case nir_intrinsic_image_deref_atomic_umax:
3874   case nir_intrinsic_image_deref_atomic_and:
3875   case nir_intrinsic_image_deref_atomic_or:
3876   case nir_intrinsic_image_deref_atomic_xor:
3877   case nir_intrinsic_image_deref_atomic_exchange:
3878   case nir_intrinsic_image_deref_atomic_comp_swap:
3879   case nir_intrinsic_image_deref_atomic_inc_wrap:
3880   case nir_intrinsic_image_deref_atomic_dec_wrap:
3881   case nir_intrinsic_image_deref_atomic_fmin:
3882   case nir_intrinsic_image_deref_atomic_fmax:
3883      result = visit_image_atomic(ctx, instr, false);
3884      break;
3885   case nir_intrinsic_bindless_image_size:
3886      result = visit_image_size(ctx, instr, true);
3887      break;
3888   case nir_intrinsic_image_deref_size:
3889      result = visit_image_size(ctx, instr, false);
3890      break;
3891   case nir_intrinsic_shader_clock:
3892      result = ac_build_shader_clock(&ctx->ac, nir_intrinsic_memory_scope(instr));
3893      break;
3894   case nir_intrinsic_discard:
3895   case nir_intrinsic_discard_if:
3896   case nir_intrinsic_terminate:
3897   case nir_intrinsic_terminate_if:
3898      emit_discard(ctx, instr);
3899      break;
3900   case nir_intrinsic_demote:
3901   case nir_intrinsic_demote_if:
3902      emit_demote(ctx, instr);
3903      break;
3904   case nir_intrinsic_memory_barrier:
3905   case nir_intrinsic_group_memory_barrier:
3906      ac_build_waitcnt(&ctx->ac, AC_WAIT_LGKM | AC_WAIT_VLOAD | AC_WAIT_VSTORE);
3907      break;
3908   case nir_intrinsic_memory_barrier_buffer:
3909   case nir_intrinsic_memory_barrier_image:
3910      ac_build_waitcnt(&ctx->ac, AC_WAIT_VLOAD | AC_WAIT_VSTORE);
3911      break;
3912   case nir_intrinsic_memory_barrier_shared:
3913   case nir_intrinsic_memory_barrier_tcs_patch:
3914      ac_build_waitcnt(&ctx->ac, AC_WAIT_LGKM);
3915      break;
3916   case nir_intrinsic_scoped_barrier: {
3917      assert(!(nir_intrinsic_memory_semantics(instr) &
3918               (NIR_MEMORY_MAKE_AVAILABLE | NIR_MEMORY_MAKE_VISIBLE)));
3919
3920      nir_variable_mode modes = nir_intrinsic_memory_modes(instr);
3921
3922      unsigned wait_flags = 0;
3923      if (modes & (nir_var_mem_global | nir_var_mem_ssbo | nir_var_image))
3924         wait_flags |= AC_WAIT_VLOAD | AC_WAIT_VSTORE;
3925      if (modes & nir_var_mem_shared)
3926         wait_flags |= AC_WAIT_LGKM;
3927
3928      if (wait_flags)
3929         ac_build_waitcnt(&ctx->ac, wait_flags);
3930
3931      if (nir_intrinsic_execution_scope(instr) == NIR_SCOPE_WORKGROUP)
3932         ac_build_s_barrier(&ctx->ac, ctx->stage);
3933      break;
3934   }
3935   case nir_intrinsic_control_barrier:
3936      /* If output patches are wholly in one wave, we don't need a barrier. */
3937      if (ctx->stage == MESA_SHADER_TESS_CTRL &&
3938          ctx->ac.wave_size % ctx->info->tess.tcs_vertices_out == 0)
3939         break;
3940
3941      ac_build_s_barrier(&ctx->ac, ctx->stage);
3942      break;
3943   case nir_intrinsic_shared_atomic_add:
3944   case nir_intrinsic_shared_atomic_imin:
3945   case nir_intrinsic_shared_atomic_umin:
3946   case nir_intrinsic_shared_atomic_imax:
3947   case nir_intrinsic_shared_atomic_umax:
3948   case nir_intrinsic_shared_atomic_and:
3949   case nir_intrinsic_shared_atomic_or:
3950   case nir_intrinsic_shared_atomic_xor:
3951   case nir_intrinsic_shared_atomic_exchange:
3952   case nir_intrinsic_shared_atomic_comp_swap:
3953   case nir_intrinsic_shared_atomic_fadd:
3954   case nir_intrinsic_shared_atomic_fmin:
3955   case nir_intrinsic_shared_atomic_fmax: {
3956      LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[0], instr->src[1].ssa->bit_size, 0);
3957      result = visit_var_atomic(ctx, instr, ptr, 1);
3958      break;
3959   }
3960   case nir_intrinsic_deref_atomic_add:
3961   case nir_intrinsic_deref_atomic_imin:
3962   case nir_intrinsic_deref_atomic_umin:
3963   case nir_intrinsic_deref_atomic_imax:
3964   case nir_intrinsic_deref_atomic_umax:
3965   case nir_intrinsic_deref_atomic_and:
3966   case nir_intrinsic_deref_atomic_or:
3967   case nir_intrinsic_deref_atomic_xor:
3968   case nir_intrinsic_deref_atomic_exchange:
3969   case nir_intrinsic_deref_atomic_comp_swap:
3970   case nir_intrinsic_deref_atomic_fadd: {
3971      LLVMValueRef ptr = get_src(ctx, instr->src[0]);
3972      result = visit_var_atomic(ctx, instr, ptr, 1);
3973      break;
3974   }
3975   case nir_intrinsic_load_barycentric_pixel:
3976      result = barycentric_center(ctx, nir_intrinsic_interp_mode(instr));
3977      break;
3978   case nir_intrinsic_load_barycentric_centroid:
3979      result = barycentric_centroid(ctx, nir_intrinsic_interp_mode(instr));
3980      break;
3981   case nir_intrinsic_load_barycentric_sample:
3982      result = barycentric_sample(ctx, nir_intrinsic_interp_mode(instr));
3983      break;
3984   case nir_intrinsic_load_barycentric_model:
3985      result = barycentric_model(ctx);
3986      break;
3987   case nir_intrinsic_load_barycentric_at_offset: {
3988      LLVMValueRef offset = ac_to_float(&ctx->ac, get_src(ctx, instr->src[0]));
3989      result = barycentric_offset(ctx, nir_intrinsic_interp_mode(instr), offset);
3990      break;
3991   }
3992   case nir_intrinsic_load_barycentric_at_sample: {
3993      LLVMValueRef sample_id = get_src(ctx, instr->src[0]);
3994      result = barycentric_at_sample(ctx, nir_intrinsic_interp_mode(instr), sample_id);
3995      break;
3996   }
3997   case nir_intrinsic_load_interpolated_input: {
3998      /* We assume any indirect loads have been lowered away */
3999      ASSERTED nir_const_value *offset = nir_src_as_const_value(instr->src[1]);
4000      assert(offset);
4001      assert(offset[0].i32 == 0);
4002
4003      LLVMValueRef interp_param = get_src(ctx, instr->src[0]);
4004      unsigned index = nir_intrinsic_base(instr);
4005      unsigned component = nir_intrinsic_component(instr);
4006      result = load_interpolated_input(ctx, interp_param, index, component,
4007                                       instr->dest.ssa.num_components, instr->dest.ssa.bit_size,
4008                                       nir_intrinsic_io_semantics(instr).high_16bits);
4009      break;
4010   }
4011   case nir_intrinsic_load_point_coord_maybe_flipped: {
4012      LLVMValueRef interp_param = lookup_interp_param(ctx, INTERP_MODE_NONE, INTERP_CENTER);
4013      /* Load point coordinates (x, y) which are written by the hw after the interpolated inputs */
4014      result = load_interpolated_input(ctx, interp_param, ctx->abi->num_interp, 2,
4015                                       instr->dest.ssa.num_components, instr->dest.ssa.bit_size,
4016                                       false);
4017      break;
4018   }
4019   case nir_intrinsic_emit_vertex:
4020      ctx->abi->emit_vertex(ctx->abi, nir_intrinsic_stream_id(instr), ctx->abi->outputs);
4021      break;
4022   case nir_intrinsic_emit_vertex_with_counter: {
4023      unsigned stream = nir_intrinsic_stream_id(instr);
4024      LLVMValueRef next_vertex = get_src(ctx, instr->src[0]);
4025      ctx->abi->emit_vertex_with_counter(ctx->abi, stream, next_vertex, ctx->abi->outputs);
4026      break;
4027   }
4028   case nir_intrinsic_end_primitive:
4029   case nir_intrinsic_end_primitive_with_counter:
4030      ctx->abi->emit_primitive(ctx->abi, nir_intrinsic_stream_id(instr));
4031      break;
4032   case nir_intrinsic_load_tess_coord: {
4033      LLVMValueRef coord[] = {
4034         ctx->tes_u_replaced ? ctx->tes_u_replaced : ac_get_arg(&ctx->ac, ctx->args->tes_u),
4035         ctx->tes_v_replaced ? ctx->tes_v_replaced : ac_get_arg(&ctx->ac, ctx->args->tes_v),
4036         ctx->ac.f32_0,
4037      };
4038
4039      /* For triangles, the vector should be (u, v, 1-u-v). */
4040      if (ctx->info->tess._primitive_mode == TESS_PRIMITIVE_TRIANGLES) {
4041         coord[2] = LLVMBuildFSub(ctx->ac.builder, ctx->ac.f32_1,
4042                                  LLVMBuildFAdd(ctx->ac.builder, coord[0], coord[1], ""), "");
4043      }
4044      result = ac_build_gather_values(&ctx->ac, coord, 3);
4045      break;
4046   }
4047   case nir_intrinsic_load_gs_vertex_offset_amd:
4048      result = ac_get_arg(&ctx->ac, ctx->args->gs_vtx_offset[nir_intrinsic_base(instr)]);
4049      break;
4050   case nir_intrinsic_vote_all: {
4051      result = ac_build_vote_all(&ctx->ac, get_src(ctx, instr->src[0]));
4052      break;
4053   }
4054   case nir_intrinsic_vote_any: {
4055      result = ac_build_vote_any(&ctx->ac, get_src(ctx, instr->src[0]));
4056      break;
4057   }
4058   case nir_intrinsic_shuffle:
4059      if (ctx->ac.gfx_level == GFX8 || ctx->ac.gfx_level == GFX9 ||
4060          (ctx->ac.gfx_level >= GFX10 && ctx->ac.wave_size == 32)) {
4061         result =
4062            ac_build_shuffle(&ctx->ac, get_src(ctx, instr->src[0]), get_src(ctx, instr->src[1]));
4063      } else {
4064         LLVMValueRef src = get_src(ctx, instr->src[0]);
4065         LLVMValueRef index = get_src(ctx, instr->src[1]);
4066         LLVMTypeRef type = LLVMTypeOf(src);
4067         struct waterfall_context wctx;
4068         LLVMValueRef index_val;
4069
4070         index_val = enter_waterfall(ctx, &wctx, index, true);
4071
4072         src = LLVMBuildZExt(ctx->ac.builder, src, ctx->ac.i32, "");
4073
4074         result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.readlane", ctx->ac.i32,
4075                                     (LLVMValueRef[]){src, index_val}, 2,
4076                                     AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
4077
4078         result = LLVMBuildTrunc(ctx->ac.builder, result, type, "");
4079
4080         result = exit_waterfall(ctx, &wctx, result);
4081      }
4082      break;
4083   case nir_intrinsic_reduce:
4084      result = ac_build_reduce(&ctx->ac, get_src(ctx, instr->src[0]), instr->const_index[0],
4085                               instr->const_index[1]);
4086      break;
4087   case nir_intrinsic_inclusive_scan:
4088      result =
4089         ac_build_inclusive_scan(&ctx->ac, get_src(ctx, instr->src[0]), instr->const_index[0]);
4090      break;
4091   case nir_intrinsic_exclusive_scan:
4092      result =
4093         ac_build_exclusive_scan(&ctx->ac, get_src(ctx, instr->src[0]), instr->const_index[0]);
4094      break;
4095   case nir_intrinsic_quad_broadcast: {
4096      unsigned lane = nir_src_as_uint(instr->src[1]);
4097      result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), lane, lane, lane, lane);
4098      break;
4099   }
4100   case nir_intrinsic_quad_swap_horizontal:
4101      result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 1, 0, 3, 2);
4102      break;
4103   case nir_intrinsic_quad_swap_vertical:
4104      result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 2, 3, 0, 1);
4105      break;
4106   case nir_intrinsic_quad_swap_diagonal:
4107      result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 3, 2, 1, 0);
4108      break;
4109   case nir_intrinsic_quad_swizzle_amd: {
4110      uint32_t mask = nir_intrinsic_swizzle_mask(instr);
4111      result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), mask & 0x3,
4112                                     (mask >> 2) & 0x3, (mask >> 4) & 0x3, (mask >> 6) & 0x3);
4113      break;
4114   }
4115   case nir_intrinsic_masked_swizzle_amd: {
4116      uint32_t mask = nir_intrinsic_swizzle_mask(instr);
4117      result = ac_build_ds_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), mask);
4118      break;
4119   }
4120   case nir_intrinsic_write_invocation_amd:
4121      result = ac_build_writelane(&ctx->ac, get_src(ctx, instr->src[0]),
4122                                  get_src(ctx, instr->src[1]), get_src(ctx, instr->src[2]));
4123      break;
4124   case nir_intrinsic_mbcnt_amd:
4125      result = ac_build_mbcnt_add(&ctx->ac, get_src(ctx, instr->src[0]), get_src(ctx, instr->src[1]));
4126      break;
4127   case nir_intrinsic_load_scratch: {
4128      LLVMValueRef offset = get_src(ctx, instr->src[0]);
4129      LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->scratch, offset);
4130      LLVMTypeRef comp_type = LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
4131      LLVMTypeRef vec_type = instr->dest.ssa.num_components == 1
4132                                ? comp_type
4133                                : LLVMVectorType(comp_type, instr->dest.ssa.num_components);
4134      unsigned addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
4135      ptr = LLVMBuildBitCast(ctx->ac.builder, ptr, LLVMPointerType(vec_type, addr_space), "");
4136      result = LLVMBuildLoad2(ctx->ac.builder, vec_type, ptr, "");
4137      break;
4138   }
4139   case nir_intrinsic_store_scratch: {
4140      LLVMValueRef offset = get_src(ctx, instr->src[1]);
4141      LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->scratch, offset);
4142      LLVMTypeRef comp_type = LLVMIntTypeInContext(ctx->ac.context, instr->src[0].ssa->bit_size);
4143      unsigned addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
4144      ptr = LLVMBuildBitCast(ctx->ac.builder, ptr, LLVMPointerType(comp_type, addr_space), "");
4145      LLVMValueRef src = get_src(ctx, instr->src[0]);
4146      unsigned wrmask = nir_intrinsic_write_mask(instr);
4147      while (wrmask) {
4148         int start, count;
4149         u_bit_scan_consecutive_range(&wrmask, &start, &count);
4150
4151         LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, start, false);
4152         LLVMValueRef offset_ptr = LLVMBuildGEP2(ctx->ac.builder, comp_type, ptr, &offset, 1, "");
4153         LLVMTypeRef vec_type = count == 1 ? comp_type : LLVMVectorType(comp_type, count);
4154         offset_ptr = LLVMBuildBitCast(ctx->ac.builder, offset_ptr,
4155                                       LLVMPointerType(vec_type, addr_space), "");
4156         LLVMValueRef offset_src = ac_extract_components(&ctx->ac, src, start, count);
4157         LLVMBuildStore(ctx->ac.builder, offset_src, offset_ptr);
4158      }
4159      break;
4160   }
4161   case nir_intrinsic_load_constant: {
4162      unsigned base = nir_intrinsic_base(instr);
4163      unsigned range = nir_intrinsic_range(instr);
4164
4165      LLVMValueRef offset = get_src(ctx, instr->src[0]);
4166      offset = LLVMBuildAdd(ctx->ac.builder, offset, LLVMConstInt(ctx->ac.i32, base, false), "");
4167
4168      /* Clamp the offset to avoid out-of-bound access because global
4169       * instructions can't handle them.
4170       */
4171      LLVMValueRef size = LLVMConstInt(ctx->ac.i32, base + range, false);
4172      LLVMValueRef cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, offset, size, "");
4173      offset = LLVMBuildSelect(ctx->ac.builder, cond, offset, size, "");
4174
4175      LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->constant_data, offset);
4176      LLVMTypeRef comp_type = LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
4177      LLVMTypeRef vec_type = instr->dest.ssa.num_components == 1
4178                                ? comp_type
4179                                : LLVMVectorType(comp_type, instr->dest.ssa.num_components);
4180      unsigned addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
4181      ptr = LLVMBuildBitCast(ctx->ac.builder, ptr, LLVMPointerType(vec_type, addr_space), "");
4182      result = LLVMBuildLoad2(ctx->ac.builder, vec_type, ptr, "");
4183      break;
4184   }
4185   case nir_intrinsic_set_vertex_and_primitive_count:
4186      /* Currently ignored. */
4187      break;
4188   case nir_intrinsic_load_buffer_amd: {
4189      LLVMValueRef descriptor = get_src(ctx, instr->src[0]);
4190      LLVMValueRef addr_voffset = get_src(ctx, instr->src[1]);
4191      LLVMValueRef addr_soffset = get_src(ctx, instr->src[2]);
4192      unsigned num_components = instr->dest.ssa.num_components;
4193      unsigned const_offset = nir_intrinsic_base(instr);
4194      bool swizzled = nir_intrinsic_is_swizzled(instr);
4195      bool reorder = nir_intrinsic_can_reorder(instr);
4196      bool slc = nir_intrinsic_slc_amd(instr);
4197
4198      enum ac_image_cache_policy cache_policy = ac_glc;
4199      if (swizzled)
4200         cache_policy |= ac_swizzled;
4201      if (slc)
4202         cache_policy |= ac_slc;
4203
4204      LLVMTypeRef channel_type;
4205      if (instr->dest.ssa.bit_size == 8)
4206         channel_type = ctx->ac.i8;
4207      else if (instr->dest.ssa.bit_size == 16)
4208         channel_type = ctx->ac.i16;
4209      else if (instr->dest.ssa.bit_size == 32)
4210         channel_type = ctx->ac.i32;
4211      else if (instr->dest.ssa.bit_size == 64)
4212         channel_type = ctx->ac.i64;
4213      else if (instr->dest.ssa.bit_size == 128)
4214         channel_type = ctx->ac.i128;
4215      else
4216         unreachable("Unsupported channel type for load_buffer_amd");
4217
4218      LLVMValueRef voffset = LLVMBuildAdd(ctx->ac.builder, addr_voffset,
4219                                          LLVMConstInt(ctx->ac.i32, const_offset, 0), "");
4220      result = ac_build_buffer_load(&ctx->ac, descriptor, num_components, NULL, voffset,
4221                                    addr_soffset, channel_type, cache_policy, reorder, false);
4222      result = ac_to_integer(&ctx->ac, ac_trim_vector(&ctx->ac, result, num_components));
4223      break;
4224   }
4225   case nir_intrinsic_store_buffer_amd: {
4226      LLVMValueRef store_data = get_src(ctx, instr->src[0]);
4227      LLVMValueRef descriptor = get_src(ctx, instr->src[1]);
4228      LLVMValueRef addr_voffset = get_src(ctx, instr->src[2]);
4229      LLVMValueRef addr_soffset = get_src(ctx, instr->src[3]);
4230      unsigned const_offset = nir_intrinsic_base(instr);
4231      bool swizzled = nir_intrinsic_is_swizzled(instr);
4232      bool slc = nir_intrinsic_slc_amd(instr);
4233
4234      enum ac_image_cache_policy cache_policy = ac_glc;
4235      if (swizzled)
4236         cache_policy |= ac_swizzled;
4237      if (slc)
4238         cache_policy |= ac_slc;
4239
4240      unsigned writemask = nir_intrinsic_write_mask(instr);
4241      while (writemask) {
4242         int start, count;
4243         u_bit_scan_consecutive_range(&writemask, &start, &count);
4244
4245         LLVMValueRef voffset = LLVMBuildAdd(
4246            ctx->ac.builder, addr_voffset,
4247            LLVMConstInt(ctx->ac.i32, const_offset + start * 4, 0), "");
4248
4249         LLVMValueRef data = extract_vector_range(&ctx->ac, store_data, start, count);
4250         ac_build_buffer_store_dword(&ctx->ac, descriptor, data, NULL, voffset, addr_soffset,
4251                                     cache_policy);
4252      }
4253      break;
4254   }
4255   case nir_intrinsic_has_input_vertex_amd: {
4256      LLVMValueRef num =
4257         ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->merged_wave_info), 0, 8);
4258      result = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, ac_get_thread_id(&ctx->ac), num, "");
4259      break;
4260   }
4261   case nir_intrinsic_has_input_primitive_amd: {
4262      LLVMValueRef num =
4263         ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->merged_wave_info), 8, 8);
4264      result = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, ac_get_thread_id(&ctx->ac), num, "");
4265      break;
4266   }
4267   case nir_intrinsic_alloc_vertices_and_primitives_amd:
4268      /* The caller should only call this conditionally for wave 0, so pass NULL to disable
4269       * the wave 0 check inside this function.
4270       */
4271      ac_build_sendmsg_gs_alloc_req(&ctx->ac, NULL,
4272                                    get_src(ctx, instr->src[0]),
4273                                    get_src(ctx, instr->src[1]));
4274      break;
4275   case nir_intrinsic_overwrite_vs_arguments_amd:
4276      ctx->vertex_id_replaced = get_src(ctx, instr->src[0]);
4277      ctx->instance_id_replaced = get_src(ctx, instr->src[1]);
4278      break;
4279   case nir_intrinsic_overwrite_tes_arguments_amd:
4280      ctx->tes_u_replaced = get_src(ctx, instr->src[0]);
4281      ctx->tes_v_replaced = get_src(ctx, instr->src[1]);
4282      ctx->tes_rel_patch_id_replaced = get_src(ctx, instr->src[2]);
4283      ctx->tes_patch_id_replaced = get_src(ctx, instr->src[3]);
4284      break;
4285   case nir_intrinsic_export_primitive_amd: {
4286      struct ac_ngg_prim prim = {0};
4287      prim.passthrough = get_src(ctx, instr->src[0]);
4288      ac_build_export_prim(&ctx->ac, &prim);
4289      break;
4290   }
4291   case nir_intrinsic_gds_atomic_add_amd: {
4292      LLVMValueRef store_val = get_src(ctx, instr->src[0]);
4293      LLVMValueRef addr = get_src(ctx, instr->src[1]);
4294      LLVMTypeRef gds_ptr_type = LLVMPointerType(ctx->ac.i32, AC_ADDR_SPACE_GDS);
4295      LLVMValueRef gds_base = LLVMBuildIntToPtr(ctx->ac.builder, addr, gds_ptr_type, "");
4296      ac_build_atomic_rmw(&ctx->ac, LLVMAtomicRMWBinOpAdd, gds_base, store_val, "workgroup-one-as");
4297      break;
4298   }
4299   case nir_intrinsic_export_vertex_amd:
4300      ctx->abi->export_vertex(ctx->abi);
4301      break;
4302   case nir_intrinsic_elect:
4303      result = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, visit_first_invocation(ctx),
4304                             ac_get_thread_id(&ctx->ac), "");
4305      break;
4306   case nir_intrinsic_byte_permute_amd:
4307      if (LLVM_VERSION_MAJOR < 13) {
4308         assert("unimplemented byte_permute, LLVM 12 doesn't have amdgcn.perm");
4309         break;
4310      }
4311      result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.perm", ctx->ac.i32,
4312                                  (LLVMValueRef[]){get_src(ctx, instr->src[0]),
4313                                                   get_src(ctx, instr->src[1]),
4314                                                   get_src(ctx, instr->src[2])},
4315                                  3, AC_FUNC_ATTR_READNONE);
4316      break;
4317   case nir_intrinsic_lane_permute_16_amd:
4318      result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.permlane16", ctx->ac.i32,
4319                                  (LLVMValueRef[]){get_src(ctx, instr->src[0]),
4320                                                   get_src(ctx, instr->src[0]),
4321                                                   get_src(ctx, instr->src[1]),
4322                                                   get_src(ctx, instr->src[2]),
4323                                                   ctx->ac.i1false,
4324                                                   ctx->ac.i1false},
4325                                  6, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
4326      break;
4327   case nir_intrinsic_load_force_vrs_rates_amd:
4328      result = ac_get_arg(&ctx->ac, ctx->args->force_vrs_rates);
4329      break;
4330   case nir_intrinsic_load_scalar_arg_amd:
4331   case nir_intrinsic_load_vector_arg_amd: {
4332      assert(nir_intrinsic_base(instr) < AC_MAX_ARGS);
4333      result = ac_to_integer(&ctx->ac, LLVMGetParam(ctx->main_function, nir_intrinsic_base(instr)));
4334      break;
4335   }
4336   case nir_intrinsic_load_smem_amd: {
4337      LLVMValueRef base = get_src(ctx, instr->src[0]);
4338      LLVMValueRef offset = get_src(ctx, instr->src[1]);
4339
4340      LLVMTypeRef result_type = get_def_type(ctx, &instr->dest.ssa);
4341      LLVMTypeRef ptr_type = LLVMPointerType(result_type, AC_ADDR_SPACE_CONST);
4342      LLVMTypeRef byte_ptr_type = LLVMPointerType(ctx->ac.i8, AC_ADDR_SPACE_CONST);
4343
4344      LLVMValueRef addr = LLVMBuildIntToPtr(ctx->ac.builder, base, byte_ptr_type, "");
4345      addr = LLVMBuildGEP2(ctx->ac.builder, ctx->ac.i8, addr, &offset, 1, "");
4346      addr = LLVMBuildBitCast(ctx->ac.builder, addr, ptr_type, "");
4347
4348      LLVMSetMetadata(addr, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
4349      result = LLVMBuildLoad2(ctx->ac.builder, result_type, addr, "");
4350      LLVMSetMetadata(result, ctx->ac.invariant_load_md_kind, ctx->ac.empty_md);
4351      break;
4352   }
4353   default:
4354      fprintf(stderr, "Unknown intrinsic: ");
4355      nir_print_instr(&instr->instr, stderr);
4356      fprintf(stderr, "\n");
4357      abort();
4358      break;
4359   }
4360   if (result) {
4361      ctx->ssa_defs[instr->dest.ssa.index] = result;
4362   }
4363}
4364
4365static LLVMValueRef get_bindless_index_from_uniform(struct ac_nir_context *ctx, unsigned base_index,
4366                                                    unsigned constant_index,
4367                                                    LLVMValueRef dynamic_index)
4368{
4369   LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, base_index * 4, 0);
4370   LLVMValueRef index = LLVMBuildAdd(ctx->ac.builder, dynamic_index,
4371                                     LLVMConstInt(ctx->ac.i32, constant_index, 0), "");
4372
4373   /* Bindless uniforms are 64bit so multiple index by 8 */
4374   index = LLVMBuildMul(ctx->ac.builder, index, LLVMConstInt(ctx->ac.i32, 8, 0), "");
4375   offset = LLVMBuildAdd(ctx->ac.builder, offset, index, "");
4376
4377   LLVMValueRef ubo_index = ctx->abi->load_ubo(ctx->abi, ctx->ac.i32_0);
4378
4379   LLVMValueRef ret =
4380      ac_build_buffer_load(&ctx->ac, ubo_index, 1, NULL, offset, NULL, ctx->ac.f32, 0, true, true);
4381
4382   return LLVMBuildBitCast(ctx->ac.builder, ret, ctx->ac.i32, "");
4383}
4384
4385struct sampler_desc_address {
4386   unsigned descriptor_set;
4387   unsigned base_index; /* binding in vulkan */
4388   unsigned constant_index;
4389   LLVMValueRef dynamic_index;
4390   bool image;
4391   bool bindless;
4392};
4393
4394static struct sampler_desc_address get_sampler_desc_internal(struct ac_nir_context *ctx,
4395                                                             nir_deref_instr *deref_instr,
4396                                                             const nir_instr *instr, bool image)
4397{
4398   LLVMValueRef index = NULL;
4399   unsigned constant_index = 0;
4400   unsigned descriptor_set;
4401   unsigned base_index;
4402   bool bindless = false;
4403
4404   if (!deref_instr) {
4405      descriptor_set = 0;
4406      if (image) {
4407         nir_intrinsic_instr *img_instr = nir_instr_as_intrinsic(instr);
4408         base_index = 0;
4409         bindless = true;
4410         index = get_src(ctx, img_instr->src[0]);
4411      } else {
4412         nir_tex_instr *tex_instr = nir_instr_as_tex(instr);
4413         int sampSrcIdx = nir_tex_instr_src_index(tex_instr, nir_tex_src_sampler_handle);
4414         if (sampSrcIdx != -1) {
4415            base_index = 0;
4416            bindless = true;
4417            index = get_src(ctx, tex_instr->src[sampSrcIdx].src);
4418         } else {
4419            assert(tex_instr && !image);
4420            base_index = tex_instr->sampler_index;
4421         }
4422      }
4423   } else {
4424      while (deref_instr->deref_type != nir_deref_type_var) {
4425         if (deref_instr->deref_type == nir_deref_type_array) {
4426            unsigned array_size = glsl_get_aoa_size(deref_instr->type);
4427            if (!array_size)
4428               array_size = 1;
4429
4430            if (nir_src_is_const(deref_instr->arr.index)) {
4431               constant_index += array_size * nir_src_as_uint(deref_instr->arr.index);
4432            } else {
4433               LLVMValueRef indirect = get_src(ctx, deref_instr->arr.index);
4434
4435               indirect = LLVMBuildMul(ctx->ac.builder, indirect,
4436                                       LLVMConstInt(ctx->ac.i32, array_size, false), "");
4437
4438               if (!index)
4439                  index = indirect;
4440               else
4441                  index = LLVMBuildAdd(ctx->ac.builder, index, indirect, "");
4442            }
4443
4444            deref_instr = nir_src_as_deref(deref_instr->parent);
4445         } else if (deref_instr->deref_type == nir_deref_type_struct) {
4446            unsigned sidx = deref_instr->strct.index;
4447            deref_instr = nir_src_as_deref(deref_instr->parent);
4448            constant_index += glsl_get_struct_location_offset(deref_instr->type, sidx);
4449         } else {
4450            unreachable("Unsupported deref type");
4451         }
4452      }
4453      descriptor_set = deref_instr->var->data.descriptor_set;
4454
4455      if (deref_instr->var->data.bindless) {
4456         /* For now just assert on unhandled variable types */
4457         assert(deref_instr->var->data.mode == nir_var_uniform);
4458
4459         base_index = deref_instr->var->data.driver_location;
4460         bindless = true;
4461
4462         index = index ? index : ctx->ac.i32_0;
4463         index = get_bindless_index_from_uniform(ctx, base_index, constant_index, index);
4464      } else
4465         base_index = deref_instr->var->data.binding;
4466   }
4467   return (struct sampler_desc_address){
4468      .descriptor_set = descriptor_set,
4469      .base_index = base_index,
4470      .constant_index = constant_index,
4471      .dynamic_index = index,
4472      .image = image,
4473      .bindless = bindless,
4474   };
4475}
4476
4477/* Extract any possibly divergent index into a separate value that can be fed
4478 * into get_sampler_desc with the same arguments. */
4479static LLVMValueRef get_sampler_desc_index(struct ac_nir_context *ctx, nir_deref_instr *deref_instr,
4480                                           const nir_instr *instr, bool image)
4481{
4482   struct sampler_desc_address addr = get_sampler_desc_internal(ctx, deref_instr, instr, image);
4483   return addr.dynamic_index;
4484}
4485
4486static LLVMValueRef get_sampler_desc(struct ac_nir_context *ctx, nir_deref_instr *deref_instr,
4487                                     enum ac_descriptor_type desc_type, const nir_instr *instr,
4488                                     LLVMValueRef index, bool image, bool write)
4489{
4490   struct sampler_desc_address addr = get_sampler_desc_internal(ctx, deref_instr, instr, image);
4491   return ctx->abi->load_sampler_desc(ctx->abi, addr.descriptor_set, addr.base_index,
4492                                      addr.constant_index, index, desc_type, addr.image, write,
4493                                      addr.bindless);
4494}
4495
4496/* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
4497 *
4498 * GFX6-GFX7:
4499 *   If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
4500 *   filtering manually. The driver sets img7 to a mask clearing
4501 *   MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
4502 *     s_and_b32 samp0, samp0, img7
4503 *
4504 * GFX8:
4505 *   The ANISO_OVERRIDE sampler field enables this fix in TA.
4506 */
4507static LLVMValueRef sici_fix_sampler_aniso(struct ac_nir_context *ctx, LLVMValueRef res,
4508                                           LLVMValueRef samp)
4509{
4510   LLVMBuilderRef builder = ctx->ac.builder;
4511   LLVMValueRef img7, samp0;
4512
4513   if (ctx->ac.gfx_level >= GFX8)
4514      return samp;
4515
4516   img7 = LLVMBuildExtractElement(builder, res, LLVMConstInt(ctx->ac.i32, 7, 0), "");
4517   samp0 = LLVMBuildExtractElement(builder, samp, LLVMConstInt(ctx->ac.i32, 0, 0), "");
4518   samp0 = LLVMBuildAnd(builder, samp0, img7, "");
4519   return LLVMBuildInsertElement(builder, samp, samp0, LLVMConstInt(ctx->ac.i32, 0, 0), "");
4520}
4521
4522static void tex_fetch_ptrs(struct ac_nir_context *ctx, nir_tex_instr *instr,
4523                           struct waterfall_context *wctx, LLVMValueRef *res_ptr,
4524                           LLVMValueRef *samp_ptr, LLVMValueRef *fmask_ptr)
4525{
4526   LLVMValueRef texture_dynamic_handle = NULL;
4527   LLVMValueRef sampler_dynamic_handle = NULL;
4528   nir_deref_instr *texture_deref_instr = NULL;
4529   nir_deref_instr *sampler_deref_instr = NULL;
4530   int plane = -1;
4531
4532   *res_ptr = NULL;
4533   *samp_ptr = NULL;
4534   *fmask_ptr = NULL;
4535   for (unsigned i = 0; i < instr->num_srcs; i++) {
4536      switch (instr->src[i].src_type) {
4537      case nir_tex_src_texture_deref:
4538         texture_deref_instr = nir_src_as_deref(instr->src[i].src);
4539         break;
4540      case nir_tex_src_sampler_deref:
4541         sampler_deref_instr = nir_src_as_deref(instr->src[i].src);
4542         break;
4543      case nir_tex_src_texture_handle:
4544      case nir_tex_src_sampler_handle: {
4545         LLVMValueRef val = get_src(ctx, instr->src[i].src);
4546         if (LLVMGetTypeKind(LLVMTypeOf(val)) == LLVMVectorTypeKind) {
4547            if (instr->src[i].src_type == nir_tex_src_texture_handle)
4548               *res_ptr = val;
4549            else
4550               *samp_ptr = val;
4551         } else {
4552            if (instr->src[i].src_type == nir_tex_src_texture_handle)
4553               texture_dynamic_handle = val;
4554            else
4555               sampler_dynamic_handle = val;
4556         }
4557         break;
4558      }
4559      case nir_tex_src_plane:
4560         plane = nir_src_as_int(instr->src[i].src);
4561         break;
4562      default:
4563         break;
4564      }
4565   }
4566
4567   if (*res_ptr) {
4568      /* descriptors given through nir_tex_src_{texture,sampler}_handle */
4569      return;
4570   }
4571
4572   enum ac_descriptor_type main_descriptor =
4573      instr->sampler_dim == GLSL_SAMPLER_DIM_BUF ? AC_DESC_BUFFER : AC_DESC_IMAGE;
4574
4575   if (plane >= 0) {
4576      assert(instr->op != nir_texop_txf_ms && instr->op != nir_texop_samples_identical);
4577      assert(instr->sampler_dim != GLSL_SAMPLER_DIM_BUF);
4578
4579      main_descriptor = AC_DESC_PLANE_0 + plane;
4580   }
4581
4582   if (instr->op == nir_texop_fragment_mask_fetch_amd || instr->op == nir_texop_samples_identical) {
4583      /* The fragment mask is fetched from the compressed
4584       * multisampled surface.
4585       */
4586      assert(ctx->ac.gfx_level < GFX11);
4587      main_descriptor = AC_DESC_FMASK;
4588   }
4589
4590   if (texture_dynamic_handle) {
4591      /* descriptor handles given through nir_tex_src_{texture,sampler}_handle */
4592      if (instr->texture_non_uniform)
4593         texture_dynamic_handle = enter_waterfall(ctx, &wctx[0], texture_dynamic_handle, true);
4594
4595      if (instr->sampler_non_uniform)
4596         sampler_dynamic_handle = enter_waterfall(ctx, &wctx[1], sampler_dynamic_handle, true);
4597
4598      *res_ptr = ctx->abi->load_sampler_desc(ctx->abi, 0, 0, 0, texture_dynamic_handle,
4599                                             main_descriptor, false, false, true);
4600
4601      if (samp_ptr)
4602         *samp_ptr = ctx->abi->load_sampler_desc(ctx->abi, 0, 0, 0, sampler_dynamic_handle,
4603                                                 AC_DESC_SAMPLER, false, false, true);
4604      return;
4605   }
4606
4607   LLVMValueRef texture_dynamic_index =
4608      get_sampler_desc_index(ctx, texture_deref_instr, &instr->instr, false);
4609   if (!sampler_deref_instr)
4610      sampler_deref_instr = texture_deref_instr;
4611
4612   LLVMValueRef sampler_dynamic_index =
4613      get_sampler_desc_index(ctx, sampler_deref_instr, &instr->instr, false);
4614
4615   /* instr->sampler_non_uniform and texture_non_uniform are always false in GLSL,
4616    * but this can lead to unexpected behavior if texture/sampler index come from
4617    * a vertex attribute.
4618    * For instance, 2 consecutive draws using 2 different index values,
4619    * could be squashed together by the hw - producing a single draw with
4620    * non-dynamically uniform index.
4621    * To avoid this, detect divergent indexing, and use enter_waterfall.
4622    * See https://gitlab.freedesktop.org/mesa/mesa/-/issues/2253.
4623    */
4624   if (instr->texture_non_uniform ||
4625       (ctx->abi->use_waterfall_for_divergent_tex_samplers && texture_deref_instr->dest.ssa.divergent))
4626      texture_dynamic_index = enter_waterfall(ctx, wctx + 0, texture_dynamic_index, true);
4627
4628   if (instr->sampler_non_uniform ||
4629       (ctx->abi->use_waterfall_for_divergent_tex_samplers && sampler_deref_instr->dest.ssa.divergent))
4630      sampler_dynamic_index = enter_waterfall(ctx, wctx + 1, sampler_dynamic_index, true);
4631
4632   *res_ptr = get_sampler_desc(ctx, texture_deref_instr, main_descriptor, &instr->instr,
4633                               texture_dynamic_index, false, false);
4634
4635   if (samp_ptr) {
4636      *samp_ptr = get_sampler_desc(ctx, sampler_deref_instr, AC_DESC_SAMPLER, &instr->instr,
4637                                   sampler_dynamic_index, false, false);
4638      if (instr->sampler_dim < GLSL_SAMPLER_DIM_RECT)
4639         *samp_ptr = sici_fix_sampler_aniso(ctx, *res_ptr, *samp_ptr);
4640   }
4641   if (ctx->ac.gfx_level < GFX11 &&
4642       fmask_ptr && (instr->op == nir_texop_txf_ms || instr->op == nir_texop_samples_identical))
4643      *fmask_ptr = get_sampler_desc(ctx, texture_deref_instr, AC_DESC_FMASK, &instr->instr,
4644                                    texture_dynamic_index, false, false);
4645}
4646
4647static LLVMValueRef apply_round_slice(struct ac_llvm_context *ctx, LLVMValueRef coord)
4648{
4649   coord = ac_to_float(ctx, coord);
4650   coord = ac_build_round(ctx, coord);
4651   coord = ac_to_integer(ctx, coord);
4652   return coord;
4653}
4654
4655static void visit_tex(struct ac_nir_context *ctx, nir_tex_instr *instr)
4656{
4657   LLVMValueRef result = NULL;
4658   struct ac_image_args args = {0};
4659   LLVMValueRef fmask_ptr = NULL, sample_index = NULL;
4660   LLVMValueRef ddx = NULL, ddy = NULL;
4661   unsigned offset_src = 0;
4662   struct waterfall_context wctx[2] = {{{0}}};
4663
4664   tex_fetch_ptrs(ctx, instr, wctx, &args.resource, &args.sampler, &fmask_ptr);
4665
4666   for (unsigned i = 0; i < instr->num_srcs; i++) {
4667      switch (instr->src[i].src_type) {
4668      case nir_tex_src_coord: {
4669         LLVMValueRef coord = get_src(ctx, instr->src[i].src);
4670         args.a16 = instr->src[i].src.ssa->bit_size == 16;
4671         for (unsigned chan = 0; chan < instr->coord_components; ++chan)
4672            args.coords[chan] = ac_llvm_extract_elem(&ctx->ac, coord, chan);
4673         break;
4674      }
4675      case nir_tex_src_projector:
4676         break;
4677      case nir_tex_src_comparator:
4678         if (instr->is_shadow) {
4679            args.compare = get_src(ctx, instr->src[i].src);
4680            args.compare = ac_to_float(&ctx->ac, args.compare);
4681            assert(instr->src[i].src.ssa->bit_size == 32);
4682         }
4683         break;
4684      case nir_tex_src_offset:
4685         args.offset = get_src(ctx, instr->src[i].src);
4686         offset_src = i;
4687         /* We pack it with bit shifts, so we need it to be 32-bit. */
4688         assert(ac_get_elem_bits(&ctx->ac, LLVMTypeOf(args.offset)) == 32);
4689         break;
4690      case nir_tex_src_bias:
4691         args.bias = get_src(ctx, instr->src[i].src);
4692         assert(ac_get_elem_bits(&ctx->ac, LLVMTypeOf(args.bias)) == 32);
4693         break;
4694      case nir_tex_src_lod:
4695         if (nir_src_is_const(instr->src[i].src) && nir_src_as_uint(instr->src[i].src) == 0)
4696            args.level_zero = true;
4697         else
4698            args.lod = get_src(ctx, instr->src[i].src);
4699         break;
4700      case nir_tex_src_ms_index:
4701         sample_index = get_src(ctx, instr->src[i].src);
4702         break;
4703      case nir_tex_src_ddx:
4704         ddx = get_src(ctx, instr->src[i].src);
4705         args.g16 = instr->src[i].src.ssa->bit_size == 16;
4706         break;
4707      case nir_tex_src_ddy:
4708         ddy = get_src(ctx, instr->src[i].src);
4709         assert(LLVMTypeOf(ddy) == LLVMTypeOf(ddx));
4710         break;
4711      case nir_tex_src_min_lod:
4712         args.min_lod = get_src(ctx, instr->src[i].src);
4713         break;
4714      case nir_tex_src_texture_offset:
4715      case nir_tex_src_sampler_offset:
4716      case nir_tex_src_plane:
4717      default:
4718         break;
4719      }
4720   }
4721
4722   if (instr->op == nir_texop_txs && instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
4723      result = get_buffer_size(ctx, args.resource, true);
4724      goto write_result;
4725   }
4726
4727   if (instr->op == nir_texop_texture_samples) {
4728      LLVMValueRef res, samples, is_msaa;
4729      LLVMValueRef default_sample;
4730
4731      res = LLVMBuildBitCast(ctx->ac.builder, args.resource, ctx->ac.v8i32, "");
4732      samples =
4733         LLVMBuildExtractElement(ctx->ac.builder, res, LLVMConstInt(ctx->ac.i32, 3, false), "");
4734      is_msaa = LLVMBuildLShr(ctx->ac.builder, samples, LLVMConstInt(ctx->ac.i32, 28, false), "");
4735      is_msaa = LLVMBuildAnd(ctx->ac.builder, is_msaa, LLVMConstInt(ctx->ac.i32, 0xe, false), "");
4736      is_msaa = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, is_msaa,
4737                              LLVMConstInt(ctx->ac.i32, 0xe, false), "");
4738
4739      samples = LLVMBuildLShr(ctx->ac.builder, samples, LLVMConstInt(ctx->ac.i32, 16, false), "");
4740      samples = LLVMBuildAnd(ctx->ac.builder, samples, LLVMConstInt(ctx->ac.i32, 0xf, false), "");
4741      samples = LLVMBuildShl(ctx->ac.builder, ctx->ac.i32_1, samples, "");
4742
4743      if (ctx->abi->robust_buffer_access) {
4744         LLVMValueRef dword1, is_null_descriptor;
4745
4746         /* Extract the second dword of the descriptor, if it's
4747          * all zero, then it's a null descriptor.
4748          */
4749         dword1 =
4750            LLVMBuildExtractElement(ctx->ac.builder, res, LLVMConstInt(ctx->ac.i32, 1, false), "");
4751         is_null_descriptor = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, dword1,
4752                                            LLVMConstInt(ctx->ac.i32, 0, false), "");
4753         default_sample =
4754            LLVMBuildSelect(ctx->ac.builder, is_null_descriptor, ctx->ac.i32_0, ctx->ac.i32_1, "");
4755      } else {
4756         default_sample = ctx->ac.i32_1;
4757      }
4758
4759      samples = LLVMBuildSelect(ctx->ac.builder, is_msaa, samples, default_sample, "");
4760      result = samples;
4761      goto write_result;
4762   }
4763
4764   if (args.offset && instr->op != nir_texop_txf && instr->op != nir_texop_txf_ms) {
4765      LLVMValueRef offset[3], pack;
4766      for (unsigned chan = 0; chan < 3; ++chan)
4767         offset[chan] = ctx->ac.i32_0;
4768
4769      unsigned num_components = ac_get_llvm_num_components(args.offset);
4770      for (unsigned chan = 0; chan < num_components; chan++) {
4771         offset[chan] = ac_llvm_extract_elem(&ctx->ac, args.offset, chan);
4772         offset[chan] =
4773            LLVMBuildAnd(ctx->ac.builder, offset[chan], LLVMConstInt(ctx->ac.i32, 0x3f, false), "");
4774         if (chan)
4775            offset[chan] = LLVMBuildShl(ctx->ac.builder, offset[chan],
4776                                        LLVMConstInt(ctx->ac.i32, chan * 8, false), "");
4777      }
4778      pack = LLVMBuildOr(ctx->ac.builder, offset[0], offset[1], "");
4779      pack = LLVMBuildOr(ctx->ac.builder, pack, offset[2], "");
4780      args.offset = pack;
4781   }
4782
4783   /* Section 8.23.1 (Depth Texture Comparison Mode) of the
4784    * OpenGL 4.5 spec says:
4785    *
4786    *    "If the texture’s internal format indicates a fixed-point
4787    *     depth texture, then D_t and D_ref are clamped to the
4788    *     range [0, 1]; otherwise no clamping is performed."
4789    *
4790    * TC-compatible HTILE promotes Z16 and Z24 to Z32_FLOAT,
4791    * so the depth comparison value isn't clamped for Z16 and
4792    * Z24 anymore. Do it manually here for GFX8-9; GFX10 has
4793    * an explicitly clamped 32-bit float format.
4794    */
4795   if (args.compare && ctx->ac.gfx_level >= GFX8 && ctx->ac.gfx_level <= GFX9 &&
4796       ctx->abi->clamp_shadow_reference) {
4797      LLVMValueRef upgraded, clamped;
4798
4799      upgraded = LLVMBuildExtractElement(ctx->ac.builder, args.sampler,
4800                                         LLVMConstInt(ctx->ac.i32, 3, false), "");
4801      upgraded = LLVMBuildLShr(ctx->ac.builder, upgraded, LLVMConstInt(ctx->ac.i32, 29, false), "");
4802      upgraded = LLVMBuildTrunc(ctx->ac.builder, upgraded, ctx->ac.i1, "");
4803      clamped = ac_build_clamp(&ctx->ac, args.compare);
4804      args.compare = LLVMBuildSelect(ctx->ac.builder, upgraded, clamped, args.compare, "");
4805   }
4806
4807   /* pack derivatives */
4808   if (ddx || ddy) {
4809      int num_src_deriv_channels, num_dest_deriv_channels;
4810      switch (instr->sampler_dim) {
4811      case GLSL_SAMPLER_DIM_3D:
4812      case GLSL_SAMPLER_DIM_CUBE:
4813         num_src_deriv_channels = 3;
4814         num_dest_deriv_channels = 3;
4815         break;
4816      case GLSL_SAMPLER_DIM_2D:
4817      default:
4818         num_src_deriv_channels = 2;
4819         num_dest_deriv_channels = 2;
4820         break;
4821      case GLSL_SAMPLER_DIM_1D:
4822         num_src_deriv_channels = 1;
4823         if (ctx->ac.gfx_level == GFX9) {
4824            num_dest_deriv_channels = 2;
4825         } else {
4826            num_dest_deriv_channels = 1;
4827         }
4828         break;
4829      }
4830
4831      for (unsigned i = 0; i < num_src_deriv_channels; i++) {
4832         args.derivs[i] = ac_to_float(&ctx->ac, ac_llvm_extract_elem(&ctx->ac, ddx, i));
4833         args.derivs[num_dest_deriv_channels + i] =
4834            ac_to_float(&ctx->ac, ac_llvm_extract_elem(&ctx->ac, ddy, i));
4835      }
4836      for (unsigned i = num_src_deriv_channels; i < num_dest_deriv_channels; i++) {
4837         LLVMValueRef zero = args.g16 ? ctx->ac.f16_0 : ctx->ac.f32_0;
4838         args.derivs[i] = zero;
4839         args.derivs[num_dest_deriv_channels + i] = zero;
4840      }
4841   }
4842
4843   if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE && args.coords[0]) {
4844      for (unsigned chan = 0; chan < instr->coord_components; chan++)
4845         args.coords[chan] = ac_to_float(&ctx->ac, args.coords[chan]);
4846      if (instr->coord_components == 3)
4847         args.coords[3] = LLVMGetUndef(args.a16 ? ctx->ac.f16 : ctx->ac.f32);
4848      ac_prepare_cube_coords(&ctx->ac, instr->op == nir_texop_txd, instr->is_array,
4849                             instr->op == nir_texop_lod, args.coords, args.derivs);
4850   }
4851
4852   /* Texture coordinates fixups */
4853   if (instr->coord_components > 1 && instr->sampler_dim == GLSL_SAMPLER_DIM_1D &&
4854       instr->is_array && instr->op != nir_texop_txf) {
4855      args.coords[1] = apply_round_slice(&ctx->ac, args.coords[1]);
4856   }
4857
4858   if (instr->coord_components > 2 &&
4859       (instr->sampler_dim == GLSL_SAMPLER_DIM_2D || instr->sampler_dim == GLSL_SAMPLER_DIM_MS ||
4860        instr->sampler_dim == GLSL_SAMPLER_DIM_SUBPASS ||
4861        instr->sampler_dim == GLSL_SAMPLER_DIM_SUBPASS_MS) &&
4862       instr->is_array && instr->op != nir_texop_txf && instr->op != nir_texop_txf_ms &&
4863       instr->op != nir_texop_fragment_fetch_amd && instr->op != nir_texop_fragment_mask_fetch_amd) {
4864      args.coords[2] = apply_round_slice(&ctx->ac, args.coords[2]);
4865   }
4866
4867   if (ctx->ac.gfx_level == GFX9 && instr->sampler_dim == GLSL_SAMPLER_DIM_1D &&
4868       instr->op != nir_texop_lod) {
4869      LLVMValueRef filler;
4870      if (instr->op == nir_texop_txf)
4871         filler = args.a16 ? ctx->ac.i16_0 : ctx->ac.i32_0;
4872      else
4873         filler = LLVMConstReal(args.a16 ? ctx->ac.f16 : ctx->ac.f32, 0.5);
4874
4875      if (instr->is_array)
4876         args.coords[2] = args.coords[1];
4877      args.coords[1] = filler;
4878   }
4879
4880   /* Pack sample index */
4881   if (sample_index && (instr->op == nir_texop_txf_ms || instr->op == nir_texop_fragment_fetch_amd))
4882      args.coords[instr->coord_components] = sample_index;
4883
4884   if (instr->op == nir_texop_samples_identical) {
4885      assert(ctx->ac.gfx_level < GFX11);
4886      struct ac_image_args txf_args = {0};
4887      memcpy(txf_args.coords, args.coords, sizeof(txf_args.coords));
4888
4889      txf_args.dmask = 0xf;
4890      txf_args.resource = args.resource;
4891      txf_args.dim = instr->is_array ? ac_image_2darray : ac_image_2d;
4892      result = build_tex_intrinsic(ctx, instr, &txf_args);
4893
4894      result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
4895      result = emit_int_cmp(&ctx->ac, LLVMIntEQ, result, ctx->ac.i32_0);
4896      goto write_result;
4897   }
4898
4899   if (ctx->ac.gfx_level < GFX11 &&
4900       (instr->sampler_dim == GLSL_SAMPLER_DIM_SUBPASS_MS ||
4901        instr->sampler_dim == GLSL_SAMPLER_DIM_MS) &&
4902       instr->op != nir_texop_txs && instr->op != nir_texop_fragment_fetch_amd &&
4903       instr->op != nir_texop_fragment_mask_fetch_amd) {
4904      unsigned sample_chan = instr->is_array ? 3 : 2;
4905      args.coords[sample_chan] = adjust_sample_index_using_fmask(
4906         &ctx->ac, args.coords[0], args.coords[1], instr->is_array ? args.coords[2] : NULL,
4907         args.coords[sample_chan], fmask_ptr);
4908   }
4909
4910   if (args.offset && (instr->op == nir_texop_txf || instr->op == nir_texop_txf_ms)) {
4911      int num_offsets = instr->src[offset_src].src.ssa->num_components;
4912      num_offsets = MIN2(num_offsets, instr->coord_components);
4913      for (unsigned i = 0; i < num_offsets; ++i) {
4914         LLVMValueRef off = ac_llvm_extract_elem(&ctx->ac, args.offset, i);
4915         if (args.a16)
4916            off = LLVMBuildTrunc(ctx->ac.builder, off, ctx->ac.i16, "");
4917         args.coords[i] = LLVMBuildAdd(ctx->ac.builder, args.coords[i], off, "");
4918      }
4919      args.offset = NULL;
4920   }
4921
4922   /* DMASK was repurposed for GATHER4. 4 components are always
4923    * returned and DMASK works like a swizzle - it selects
4924    * the component to fetch. The only valid DMASK values are
4925    * 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns
4926    * (red,red,red,red) etc.) The ISA document doesn't mention
4927    * this.
4928    */
4929   args.dmask = 0xf;
4930   if (instr->op == nir_texop_tg4) {
4931      if (instr->is_shadow)
4932         args.dmask = 1;
4933      else
4934         args.dmask = 1 << instr->component;
4935   }
4936
4937   if (instr->sampler_dim != GLSL_SAMPLER_DIM_BUF) {
4938      args.dim = ac_get_sampler_dim(ctx->ac.gfx_level, instr->sampler_dim, instr->is_array);
4939      args.unorm = instr->sampler_dim == GLSL_SAMPLER_DIM_RECT;
4940   }
4941
4942   /* Adjust the number of coordinates because we only need (x,y) for 2D
4943    * multisampled images and (x,y,layer) for 2D multisampled layered
4944    * images or for multisampled input attachments.
4945    */
4946   if (instr->op == nir_texop_fragment_mask_fetch_amd) {
4947      if (args.dim == ac_image_2dmsaa) {
4948         args.dim = ac_image_2d;
4949      } else {
4950         assert(args.dim == ac_image_2darraymsaa);
4951         args.dim = ac_image_2darray;
4952      }
4953   }
4954
4955   /* Set TRUNC_COORD=0 for textureGather(). */
4956   if (instr->op == nir_texop_tg4) {
4957      LLVMValueRef dword0 = LLVMBuildExtractElement(ctx->ac.builder, args.sampler, ctx->ac.i32_0, "");
4958      dword0 = LLVMBuildAnd(ctx->ac.builder, dword0, LLVMConstInt(ctx->ac.i32, C_008F30_TRUNC_COORD, 0), "");
4959      args.sampler = LLVMBuildInsertElement(ctx->ac.builder, args.sampler, dword0, ctx->ac.i32_0, "");
4960   }
4961
4962   assert(instr->dest.is_ssa);
4963   args.d16 = instr->dest.ssa.bit_size == 16;
4964   args.tfe = instr->is_sparse;
4965
4966   result = build_tex_intrinsic(ctx, instr, &args);
4967
4968   LLVMValueRef code = NULL;
4969   if (instr->is_sparse) {
4970      code = ac_llvm_extract_elem(&ctx->ac, result, 4);
4971      result = ac_trim_vector(&ctx->ac, result, 4);
4972   }
4973
4974   if (instr->op == nir_texop_query_levels)
4975      result =
4976         LLVMBuildExtractElement(ctx->ac.builder, result, LLVMConstInt(ctx->ac.i32, 3, false), "");
4977   else if (instr->is_shadow && instr->is_new_style_shadow && instr->op != nir_texop_txs &&
4978            instr->op != nir_texop_lod && instr->op != nir_texop_tg4)
4979      result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
4980   else if (ctx->ac.gfx_level == GFX9 && instr->op == nir_texop_txs &&
4981              instr->sampler_dim == GLSL_SAMPLER_DIM_1D && instr->is_array) {
4982      LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
4983      LLVMValueRef layers = LLVMBuildExtractElement(ctx->ac.builder, result, two, "");
4984      result = LLVMBuildInsertElement(ctx->ac.builder, result, layers, ctx->ac.i32_1, "");
4985   } else if (instr->op == nir_texop_fragment_mask_fetch_amd) {
4986      /* Use 0x76543210 if the image doesn't have FMASK. */
4987      LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, args.resource, ctx->ac.v8i32, "");
4988      tmp = LLVMBuildExtractElement(ctx->ac.builder, tmp, ctx->ac.i32_1, "");
4989      tmp = LLVMBuildICmp(ctx->ac.builder, LLVMIntNE, tmp, ctx->ac.i32_0, "");
4990      result = LLVMBuildSelect(ctx->ac.builder, tmp,
4991                               LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, ""),
4992                               LLVMConstInt(ctx->ac.i32, 0x76543210, false), "");
4993   } else if (nir_tex_instr_result_size(instr) != 4)
4994      result = ac_trim_vector(&ctx->ac, result, instr->dest.ssa.num_components);
4995
4996   if (instr->is_sparse)
4997      result = ac_build_concat(&ctx->ac, result, code);
4998
4999write_result:
5000   if (result) {
5001      assert(instr->dest.is_ssa);
5002      result = ac_to_integer(&ctx->ac, result);
5003
5004      for (int i = ARRAY_SIZE(wctx); --i >= 0;) {
5005         result = exit_waterfall(ctx, wctx + i, result);
5006      }
5007
5008      ctx->ssa_defs[instr->dest.ssa.index] = result;
5009   }
5010}
5011
5012static void visit_phi(struct ac_nir_context *ctx, nir_phi_instr *instr)
5013{
5014   LLVMTypeRef type = get_def_type(ctx, &instr->dest.ssa);
5015   LLVMValueRef result = LLVMBuildPhi(ctx->ac.builder, type, "");
5016
5017   ctx->ssa_defs[instr->dest.ssa.index] = result;
5018   _mesa_hash_table_insert(ctx->phis, instr, result);
5019}
5020
5021static void visit_post_phi(struct ac_nir_context *ctx, nir_phi_instr *instr, LLVMValueRef llvm_phi)
5022{
5023   nir_foreach_phi_src (src, instr) {
5024      LLVMBasicBlockRef block = get_block(ctx, src->pred);
5025      LLVMValueRef llvm_src = get_src(ctx, src->src);
5026
5027      LLVMAddIncoming(llvm_phi, &llvm_src, &block, 1);
5028   }
5029}
5030
5031static void phi_post_pass(struct ac_nir_context *ctx)
5032{
5033   hash_table_foreach(ctx->phis, entry)
5034   {
5035      visit_post_phi(ctx, (nir_phi_instr *)entry->key, (LLVMValueRef)entry->data);
5036   }
5037}
5038
5039static bool is_def_used_in_an_export(const nir_ssa_def *def)
5040{
5041   nir_foreach_use (use_src, def) {
5042      if (use_src->parent_instr->type == nir_instr_type_intrinsic) {
5043         nir_intrinsic_instr *instr = nir_instr_as_intrinsic(use_src->parent_instr);
5044         if (instr->intrinsic == nir_intrinsic_store_deref)
5045            return true;
5046      } else if (use_src->parent_instr->type == nir_instr_type_alu) {
5047         nir_alu_instr *instr = nir_instr_as_alu(use_src->parent_instr);
5048         if (instr->op == nir_op_vec4 && is_def_used_in_an_export(&instr->dest.dest.ssa)) {
5049            return true;
5050         }
5051      }
5052   }
5053   return false;
5054}
5055
5056static void visit_ssa_undef(struct ac_nir_context *ctx, const nir_ssa_undef_instr *instr)
5057{
5058   unsigned num_components = instr->def.num_components;
5059   LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
5060
5061   if (!ctx->abi->convert_undef_to_zero || is_def_used_in_an_export(&instr->def)) {
5062      LLVMValueRef undef;
5063
5064      if (num_components == 1)
5065         undef = LLVMGetUndef(type);
5066      else {
5067         undef = LLVMGetUndef(LLVMVectorType(type, num_components));
5068      }
5069      ctx->ssa_defs[instr->def.index] = undef;
5070   } else {
5071      LLVMValueRef zero = LLVMConstInt(type, 0, false);
5072      if (num_components > 1) {
5073         zero = ac_build_gather_values_extended(&ctx->ac, &zero, num_components, 0, false);
5074      }
5075      ctx->ssa_defs[instr->def.index] = zero;
5076   }
5077}
5078
5079static void visit_jump(struct ac_llvm_context *ctx, const nir_jump_instr *instr)
5080{
5081   switch (instr->type) {
5082   case nir_jump_break:
5083      ac_build_break(ctx);
5084      break;
5085   case nir_jump_continue:
5086      ac_build_continue(ctx);
5087      break;
5088   default:
5089      fprintf(stderr, "Unknown NIR jump instr: ");
5090      nir_print_instr(&instr->instr, stderr);
5091      fprintf(stderr, "\n");
5092      abort();
5093   }
5094}
5095
5096static LLVMTypeRef glsl_base_to_llvm_type(struct ac_llvm_context *ac, enum glsl_base_type type)
5097{
5098   switch (type) {
5099   case GLSL_TYPE_INT:
5100   case GLSL_TYPE_UINT:
5101   case GLSL_TYPE_BOOL:
5102   case GLSL_TYPE_SUBROUTINE:
5103      return ac->i32;
5104   case GLSL_TYPE_INT8:
5105   case GLSL_TYPE_UINT8:
5106      return ac->i8;
5107   case GLSL_TYPE_INT16:
5108   case GLSL_TYPE_UINT16:
5109      return ac->i16;
5110   case GLSL_TYPE_FLOAT:
5111      return ac->f32;
5112   case GLSL_TYPE_FLOAT16:
5113      return ac->f16;
5114   case GLSL_TYPE_INT64:
5115   case GLSL_TYPE_UINT64:
5116      return ac->i64;
5117   case GLSL_TYPE_DOUBLE:
5118      return ac->f64;
5119   default:
5120      unreachable("unknown GLSL type");
5121   }
5122}
5123
5124static LLVMTypeRef glsl_to_llvm_type(struct ac_llvm_context *ac, const struct glsl_type *type)
5125{
5126   if (glsl_type_is_scalar(type)) {
5127      return glsl_base_to_llvm_type(ac, glsl_get_base_type(type));
5128   }
5129
5130   if (glsl_type_is_vector(type)) {
5131      return LLVMVectorType(glsl_base_to_llvm_type(ac, glsl_get_base_type(type)),
5132                            glsl_get_vector_elements(type));
5133   }
5134
5135   if (glsl_type_is_matrix(type)) {
5136      return LLVMArrayType(glsl_to_llvm_type(ac, glsl_get_column_type(type)),
5137                           glsl_get_matrix_columns(type));
5138   }
5139
5140   if (glsl_type_is_array(type)) {
5141      return LLVMArrayType(glsl_to_llvm_type(ac, glsl_get_array_element(type)),
5142                           glsl_get_length(type));
5143   }
5144
5145   assert(glsl_type_is_struct_or_ifc(type));
5146
5147   LLVMTypeRef *const member_types = alloca(glsl_get_length(type) * sizeof(LLVMTypeRef));
5148
5149   for (unsigned i = 0; i < glsl_get_length(type); i++) {
5150      member_types[i] = glsl_to_llvm_type(ac, glsl_get_struct_field(type, i));
5151   }
5152
5153   return LLVMStructTypeInContext(ac->context, member_types, glsl_get_length(type), false);
5154}
5155
5156static void visit_deref(struct ac_nir_context *ctx, nir_deref_instr *instr)
5157{
5158   if (!nir_deref_mode_is_one_of(instr, nir_var_mem_shared | nir_var_mem_global))
5159      return;
5160
5161   LLVMValueRef result = NULL;
5162   switch (instr->deref_type) {
5163   case nir_deref_type_var: {
5164      struct hash_entry *entry = _mesa_hash_table_search(ctx->vars, instr->var);
5165      result = entry->data;
5166      break;
5167   }
5168   case nir_deref_type_struct:
5169      if (nir_deref_mode_is(instr, nir_var_mem_global)) {
5170         nir_deref_instr *parent = nir_deref_instr_parent(instr);
5171         uint64_t offset = glsl_get_struct_field_offset(parent->type, instr->strct.index);
5172         result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent),
5173                                   LLVMConstInt(ctx->ac.i32, offset, 0));
5174      } else {
5175         result = ac_build_gep0(&ctx->ac, get_src(ctx, instr->parent),
5176                                LLVMConstInt(ctx->ac.i32, instr->strct.index, 0));
5177      }
5178      break;
5179   case nir_deref_type_array:
5180      if (nir_deref_mode_is(instr, nir_var_mem_global)) {
5181         nir_deref_instr *parent = nir_deref_instr_parent(instr);
5182         unsigned stride = glsl_get_explicit_stride(parent->type);
5183
5184         if ((glsl_type_is_matrix(parent->type) && glsl_matrix_type_is_row_major(parent->type)) ||
5185             (glsl_type_is_vector(parent->type) && stride == 0))
5186            stride = type_scalar_size_bytes(parent->type);
5187
5188         assert(stride > 0);
5189         LLVMValueRef index = get_src(ctx, instr->arr.index);
5190         if (LLVMTypeOf(index) != ctx->ac.i64)
5191            index = LLVMBuildZExt(ctx->ac.builder, index, ctx->ac.i64, "");
5192
5193         LLVMValueRef offset =
5194            LLVMBuildMul(ctx->ac.builder, index, LLVMConstInt(ctx->ac.i64, stride, 0), "");
5195
5196         result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent), offset);
5197      } else {
5198         result =
5199            ac_build_gep0(&ctx->ac, get_src(ctx, instr->parent), get_src(ctx, instr->arr.index));
5200      }
5201      break;
5202   case nir_deref_type_ptr_as_array:
5203      if (nir_deref_mode_is(instr, nir_var_mem_global)) {
5204         unsigned stride = nir_deref_instr_array_stride(instr);
5205
5206         LLVMValueRef index = get_src(ctx, instr->arr.index);
5207         if (LLVMTypeOf(index) != ctx->ac.i64)
5208            index = LLVMBuildZExt(ctx->ac.builder, index, ctx->ac.i64, "");
5209
5210         LLVMValueRef offset =
5211            LLVMBuildMul(ctx->ac.builder, index, LLVMConstInt(ctx->ac.i64, stride, 0), "");
5212
5213         result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent), offset);
5214      } else {
5215         result =
5216            ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent), get_src(ctx, instr->arr.index));
5217      }
5218      break;
5219   case nir_deref_type_cast: {
5220      result = get_src(ctx, instr->parent);
5221
5222      /* We can't use the structs from LLVM because the shader
5223       * specifies its own offsets. */
5224      LLVMTypeRef pointee_type = ctx->ac.i8;
5225      if (nir_deref_mode_is(instr, nir_var_mem_shared))
5226         pointee_type = glsl_to_llvm_type(&ctx->ac, instr->type);
5227
5228      unsigned address_space;
5229
5230      switch (instr->modes) {
5231      case nir_var_mem_shared:
5232         address_space = AC_ADDR_SPACE_LDS;
5233         break;
5234      case nir_var_mem_global:
5235         address_space = AC_ADDR_SPACE_GLOBAL;
5236         break;
5237      default:
5238         unreachable("Unhandled address space");
5239      }
5240
5241      LLVMTypeRef type = LLVMPointerType(pointee_type, address_space);
5242
5243      if (LLVMTypeOf(result) != type) {
5244         if (LLVMGetTypeKind(LLVMTypeOf(result)) == LLVMVectorTypeKind) {
5245            result = LLVMBuildBitCast(ctx->ac.builder, result, type, "");
5246         } else {
5247            result = LLVMBuildIntToPtr(ctx->ac.builder, result, type, "");
5248         }
5249      }
5250      break;
5251   }
5252   default:
5253      unreachable("Unhandled deref_instr deref type");
5254   }
5255
5256   ctx->ssa_defs[instr->dest.ssa.index] = result;
5257}
5258
5259static void visit_cf_list(struct ac_nir_context *ctx, struct exec_list *list);
5260
5261static void visit_block(struct ac_nir_context *ctx, nir_block *block)
5262{
5263   LLVMBasicBlockRef blockref = LLVMGetInsertBlock(ctx->ac.builder);
5264   LLVMValueRef first = LLVMGetFirstInstruction(blockref);
5265   if (first) {
5266      /* ac_branch_exited() might have already inserted non-phis */
5267      LLVMPositionBuilderBefore(ctx->ac.builder, LLVMGetFirstInstruction(blockref));
5268   }
5269
5270   nir_foreach_instr(instr, block) {
5271      if (instr->type != nir_instr_type_phi)
5272         break;
5273      visit_phi(ctx, nir_instr_as_phi(instr));
5274   }
5275
5276   LLVMPositionBuilderAtEnd(ctx->ac.builder, blockref);
5277
5278   nir_foreach_instr (instr, block) {
5279      switch (instr->type) {
5280      case nir_instr_type_alu:
5281         visit_alu(ctx, nir_instr_as_alu(instr));
5282         break;
5283      case nir_instr_type_load_const:
5284         visit_load_const(ctx, nir_instr_as_load_const(instr));
5285         break;
5286      case nir_instr_type_intrinsic:
5287         visit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
5288         break;
5289      case nir_instr_type_tex:
5290         visit_tex(ctx, nir_instr_as_tex(instr));
5291         break;
5292      case nir_instr_type_phi:
5293         break;
5294      case nir_instr_type_ssa_undef:
5295         visit_ssa_undef(ctx, nir_instr_as_ssa_undef(instr));
5296         break;
5297      case nir_instr_type_jump:
5298         visit_jump(&ctx->ac, nir_instr_as_jump(instr));
5299         break;
5300      case nir_instr_type_deref:
5301         visit_deref(ctx, nir_instr_as_deref(instr));
5302         break;
5303      default:
5304         fprintf(stderr, "Unknown NIR instr type: ");
5305         nir_print_instr(instr, stderr);
5306         fprintf(stderr, "\n");
5307         abort();
5308      }
5309   }
5310
5311   _mesa_hash_table_insert(ctx->defs, block, LLVMGetInsertBlock(ctx->ac.builder));
5312}
5313
5314static void visit_if(struct ac_nir_context *ctx, nir_if *if_stmt)
5315{
5316   LLVMValueRef value = get_src(ctx, if_stmt->condition);
5317
5318   nir_block *then_block = (nir_block *)exec_list_get_head(&if_stmt->then_list);
5319
5320   ac_build_ifcc(&ctx->ac, value, then_block->index);
5321
5322   visit_cf_list(ctx, &if_stmt->then_list);
5323
5324   if (!exec_list_is_empty(&if_stmt->else_list)) {
5325      nir_block *else_block = (nir_block *)exec_list_get_head(&if_stmt->else_list);
5326
5327      ac_build_else(&ctx->ac, else_block->index);
5328      visit_cf_list(ctx, &if_stmt->else_list);
5329   }
5330
5331   ac_build_endif(&ctx->ac, then_block->index);
5332}
5333
5334static void visit_loop(struct ac_nir_context *ctx, nir_loop *loop)
5335{
5336   nir_block *first_loop_block = (nir_block *)exec_list_get_head(&loop->body);
5337
5338   ac_build_bgnloop(&ctx->ac, first_loop_block->index);
5339
5340   visit_cf_list(ctx, &loop->body);
5341
5342   ac_build_endloop(&ctx->ac, first_loop_block->index);
5343}
5344
5345static void visit_cf_list(struct ac_nir_context *ctx, struct exec_list *list)
5346{
5347   foreach_list_typed(nir_cf_node, node, node, list)
5348   {
5349      switch (node->type) {
5350      case nir_cf_node_block:
5351         visit_block(ctx, nir_cf_node_as_block(node));
5352         break;
5353
5354      case nir_cf_node_if:
5355         visit_if(ctx, nir_cf_node_as_if(node));
5356         break;
5357
5358      case nir_cf_node_loop:
5359         visit_loop(ctx, nir_cf_node_as_loop(node));
5360         break;
5361
5362      default:
5363         assert(0);
5364      }
5365   }
5366}
5367
5368void ac_handle_shader_output_decl(struct ac_llvm_context *ctx, struct ac_shader_abi *abi,
5369                                  struct nir_shader *nir, struct nir_variable *variable,
5370                                  gl_shader_stage stage)
5371{
5372   unsigned output_loc = variable->data.driver_location;
5373   unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
5374
5375   /* tess ctrl has it's own load/store paths for outputs */
5376   if (stage == MESA_SHADER_TESS_CTRL)
5377      return;
5378
5379   if (stage == MESA_SHADER_VERTEX || stage == MESA_SHADER_TESS_EVAL ||
5380       stage == MESA_SHADER_GEOMETRY) {
5381      int idx = variable->data.location + variable->data.index;
5382      if (idx == VARYING_SLOT_CLIP_DIST0) {
5383         int length = nir->info.clip_distance_array_size + nir->info.cull_distance_array_size;
5384
5385         if (length > 4)
5386            attrib_count = 2;
5387         else
5388            attrib_count = 1;
5389      }
5390   }
5391
5392   bool is_16bit = glsl_type_is_16bit(glsl_without_array(variable->type));
5393   LLVMTypeRef type = is_16bit ? ctx->f16 : ctx->f32;
5394   for (unsigned i = 0; i < attrib_count; ++i) {
5395      for (unsigned chan = 0; chan < 4; chan++) {
5396         int idx = ac_llvm_reg_index_soa(output_loc + i, chan);
5397         abi->outputs[idx] = ac_build_alloca_undef(ctx, type, "");
5398         abi->is_16bit[idx] = is_16bit;
5399      }
5400   }
5401}
5402
5403static void setup_scratch(struct ac_nir_context *ctx, struct nir_shader *shader)
5404{
5405   if (shader->scratch_size == 0)
5406      return;
5407
5408   ctx->scratch =
5409      ac_build_alloca_undef(&ctx->ac, LLVMArrayType(ctx->ac.i8, shader->scratch_size), "scratch");
5410}
5411
5412static void setup_constant_data(struct ac_nir_context *ctx, struct nir_shader *shader)
5413{
5414   if (!shader->constant_data)
5415      return;
5416
5417   LLVMValueRef data = LLVMConstStringInContext(ctx->ac.context, shader->constant_data,
5418                                                shader->constant_data_size, true);
5419   LLVMTypeRef type = LLVMArrayType(ctx->ac.i8, shader->constant_data_size);
5420   LLVMValueRef global =
5421      LLVMAddGlobalInAddressSpace(ctx->ac.module, type, "const_data", AC_ADDR_SPACE_CONST);
5422
5423   LLVMSetInitializer(global, data);
5424   LLVMSetGlobalConstant(global, true);
5425   LLVMSetVisibility(global, LLVMHiddenVisibility);
5426   ctx->constant_data = global;
5427}
5428
5429static void setup_shared(struct ac_nir_context *ctx, struct nir_shader *nir)
5430{
5431   if (ctx->ac.lds)
5432      return;
5433
5434   LLVMTypeRef type = LLVMArrayType(ctx->ac.i8, nir->info.shared_size);
5435
5436   LLVMValueRef lds =
5437      LLVMAddGlobalInAddressSpace(ctx->ac.module, type, "compute_lds", AC_ADDR_SPACE_LDS);
5438   LLVMSetAlignment(lds, 64 * 1024);
5439
5440   ctx->ac.lds =
5441      LLVMBuildBitCast(ctx->ac.builder, lds, LLVMPointerType(ctx->ac.i8, AC_ADDR_SPACE_LDS), "");
5442}
5443
5444static void setup_gds(struct ac_nir_context *ctx, nir_function_impl *impl)
5445{
5446   bool has_gds_atomic = false;
5447
5448   if (ctx->ac.gfx_level >= GFX10 &&
5449       (ctx->stage == MESA_SHADER_VERTEX ||
5450        ctx->stage == MESA_SHADER_TESS_EVAL ||
5451        ctx->stage == MESA_SHADER_GEOMETRY)) {
5452
5453      nir_foreach_block(block, impl) {
5454         nir_foreach_instr(instr, block) {
5455            if (instr->type != nir_instr_type_intrinsic)
5456               continue;
5457
5458            nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
5459            has_gds_atomic |= intrin->intrinsic == nir_intrinsic_gds_atomic_add_amd;
5460         }
5461      }
5462   }
5463
5464   unsigned gds_size = has_gds_atomic ? 0x100 : 0;
5465
5466   if (gds_size)
5467      ac_llvm_add_target_dep_function_attr(ctx->main_function, "amdgpu-gds-size", gds_size);
5468}
5469
5470void ac_nir_translate(struct ac_llvm_context *ac, struct ac_shader_abi *abi,
5471                      const struct ac_shader_args *args, struct nir_shader *nir)
5472{
5473   struct ac_nir_context ctx = {0};
5474   struct nir_function *func;
5475
5476   ctx.ac = *ac;
5477   ctx.abi = abi;
5478   ctx.args = args;
5479
5480   ctx.stage = nir->info.stage;
5481   ctx.info = &nir->info;
5482
5483   ctx.main_function = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx.ac.builder));
5484
5485   /* TODO: remove this after RADV switches to lowered IO */
5486   if (!nir->info.io_lowered) {
5487      nir_foreach_shader_out_variable(variable, nir)
5488      {
5489         ac_handle_shader_output_decl(&ctx.ac, ctx.abi, nir, variable, ctx.stage);
5490      }
5491   }
5492
5493   ctx.defs = _mesa_hash_table_create(NULL, _mesa_hash_pointer, _mesa_key_pointer_equal);
5494   ctx.phis = _mesa_hash_table_create(NULL, _mesa_hash_pointer, _mesa_key_pointer_equal);
5495   ctx.vars = _mesa_hash_table_create(NULL, _mesa_hash_pointer, _mesa_key_pointer_equal);
5496
5497   if (ctx.abi->kill_ps_if_inf_interp)
5498      ctx.verified_interp =
5499         _mesa_hash_table_create(NULL, _mesa_hash_pointer, _mesa_key_pointer_equal);
5500
5501   func = (struct nir_function *)exec_list_get_head(&nir->functions);
5502
5503   nir_index_ssa_defs(func->impl);
5504   ctx.ssa_defs = calloc(func->impl->ssa_alloc, sizeof(LLVMValueRef));
5505
5506   setup_scratch(&ctx, nir);
5507   setup_constant_data(&ctx, nir);
5508   setup_gds(&ctx, func->impl);
5509
5510   if (gl_shader_stage_is_compute(nir->info.stage))
5511      setup_shared(&ctx, nir);
5512
5513   if (nir->info.stage == MESA_SHADER_FRAGMENT && nir->info.fs.uses_demote &&
5514       LLVM_VERSION_MAJOR < 13) {
5515      /* true = don't kill. */
5516      ctx.ac.postponed_kill = ac_build_alloca_init(&ctx.ac, ctx.ac.i1true, "");
5517   }
5518
5519   visit_cf_list(&ctx, &func->impl->body);
5520   phi_post_pass(&ctx);
5521
5522   if (ctx.ac.postponed_kill)
5523      ac_build_kill_if_false(&ctx.ac, LLVMBuildLoad2(ctx.ac.builder, ctx.ac.i1, ctx.ac.postponed_kill, ""));
5524
5525   free(ctx.ssa_defs);
5526   ralloc_free(ctx.defs);
5527   ralloc_free(ctx.phis);
5528   ralloc_free(ctx.vars);
5529   if (ctx.abi->kill_ps_if_inf_interp)
5530      ralloc_free(ctx.verified_interp);
5531}
5532