1/*
2 * Copyright (C) 2021 Collabora, Ltd.
3 * Copyright (C) 2021 Alyssa Rosenzweig <alyssa@rosenzweig.io>
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24
25#include "compiler.h"
26#include "bi_builder.h"
27
28/*
29 * Due to a Bifrost encoding restriction, some instructions cannot have an abs
30 * modifier on both sources. Check if adding a fabs modifier to a given source
31 * of a binary instruction would cause this restriction to be hit.
32 */
33static bool
34bi_would_impact_abs(unsigned arch, bi_instr *I, bi_index repl, unsigned s)
35{
36        return (arch <= 8) && I->src[1 - s].abs &&
37               bi_is_word_equiv(I->src[1 - s], repl);
38}
39
40static bool
41bi_takes_fabs(unsigned arch, bi_instr *I, bi_index repl, unsigned s)
42{
43        switch (I->op) {
44        case BI_OPCODE_FCMP_V2F16:
45        case BI_OPCODE_FMAX_V2F16:
46        case BI_OPCODE_FMIN_V2F16:
47                return !bi_would_impact_abs(arch, I, repl, s);
48        case BI_OPCODE_FADD_V2F16:
49                /*
50                 * For FADD.v2f16, the FMA pipe has the abs encoding hazard,
51                 * while the FADD pipe cannot encode a clamp. Either case in
52                 * isolation can be worked around in the scheduler, but both
53                 * together is impossible to encode. Avoid the hazard.
54                 */
55                return !(I->clamp && bi_would_impact_abs(arch, I, repl, s));
56        case BI_OPCODE_V2F32_TO_V2F16:
57                /* TODO: Needs both match or lower */
58                return false;
59        case BI_OPCODE_FLOG_TABLE_F32:
60                /* TODO: Need to check mode */
61                return false;
62        default:
63                return bi_opcode_props[I->op].abs & BITFIELD_BIT(s);
64        }
65}
66
67static bool
68bi_takes_fneg(unsigned arch, bi_instr *I, unsigned s)
69{
70        switch (I->op) {
71        case BI_OPCODE_CUBE_SSEL:
72        case BI_OPCODE_CUBE_TSEL:
73        case BI_OPCODE_CUBEFACE:
74                /* TODO: Bifrost encoding restriction: need to match or lower */
75                return arch >= 9;
76        case BI_OPCODE_FREXPE_F32:
77        case BI_OPCODE_FREXPE_V2F16:
78        case BI_OPCODE_FLOG_TABLE_F32:
79                /* TODO: Need to check mode */
80                return false;
81        default:
82                return bi_opcode_props[I->op].neg & BITFIELD_BIT(s);
83        }
84}
85
86static bool
87bi_is_fabsneg(enum bi_opcode op, enum bi_size size)
88{
89        return (size == BI_SIZE_32 && op == BI_OPCODE_FABSNEG_F32) ||
90               (size == BI_SIZE_16 && op == BI_OPCODE_FABSNEG_V2F16);
91}
92
93static enum bi_swizzle
94bi_compose_swizzle_16(enum bi_swizzle a, enum bi_swizzle b)
95{
96        assert(a <= BI_SWIZZLE_H11);
97        assert(b <= BI_SWIZZLE_H11);
98
99        bool al = (a & BI_SWIZZLE_H10);
100        bool ar = (a & BI_SWIZZLE_H01);
101        bool bl = (b & BI_SWIZZLE_H10);
102        bool br = (b & BI_SWIZZLE_H01);
103
104        return ((al ? br : bl) ? BI_SWIZZLE_H10 : 0) |
105               ((ar ? br : bl) ? BI_SWIZZLE_H01 : 0);
106}
107
108/* Like bi_replace_index, but composes instead of overwrites */
109
110static inline bi_index
111bi_compose_float_index(bi_index old, bi_index repl)
112{
113        /* abs(-x) = abs(+x) so ignore repl.neg if old.abs is set, otherwise
114         * -(-x) = x but -(+x) = +(-x) so need to exclusive-or the negates */
115        repl.neg = old.neg ^ (repl.neg && !old.abs);
116
117        /* +/- abs(+/- abs(x)) = +/- abs(x), etc so just or the two */
118        repl.abs |= old.abs;
119
120        /* Use the old swizzle to select from the replacement swizzle */
121        repl.swizzle = bi_compose_swizzle_16(old.swizzle, repl.swizzle);
122
123        return repl;
124}
125
126/* DISCARD.b32(FCMP.f(x, y)) --> DISCARD.f(x, y) */
127
128static inline void
129bi_fuse_discard_fcmp(bi_instr *I, bi_instr *mod, unsigned arch)
130{
131        if (I->op != BI_OPCODE_DISCARD_B32) return;
132        if (mod->op != BI_OPCODE_FCMP_F32 && mod->op != BI_OPCODE_FCMP_V2F16) return;
133        if (mod->cmpf >= BI_CMPF_GTLT) return;
134
135        /* .abs and .neg modifiers allowed on Valhall DISCARD but not Bifrost */
136        bool absneg = mod->src[0].neg || mod->src[0].abs;
137        absneg     |= mod->src[1].neg || mod->src[1].abs;
138
139        if (arch <= 8 && absneg) return;
140
141        enum bi_swizzle r = I->src[0].swizzle;
142
143        /* result_type doesn't matter */
144        I->op = BI_OPCODE_DISCARD_F32;
145        I->cmpf = mod->cmpf;
146        I->src[0] = mod->src[0];
147        I->src[1] = mod->src[1];
148
149        if (mod->op == BI_OPCODE_FCMP_V2F16) {
150                I->src[0].swizzle = bi_compose_swizzle_16(r, I->src[0].swizzle);
151                I->src[1].swizzle = bi_compose_swizzle_16(r, I->src[1].swizzle);
152        }
153}
154
155void
156bi_opt_mod_prop_forward(bi_context *ctx)
157{
158        bi_instr **lut = calloc(sizeof(bi_instr *), ctx->ssa_alloc);
159
160        bi_foreach_instr_global_safe(ctx, I) {
161                if (bi_is_ssa(I->dest[0]))
162                        lut[I->dest[0].value] = I;
163
164                bi_foreach_src(I, s) {
165                        if (!bi_is_ssa(I->src[s]))
166                                continue;
167
168                        bi_instr *mod = lut[I->src[s].value];
169
170                        if (!mod)
171                                continue;
172
173                        unsigned size = bi_opcode_props[I->op].size;
174
175                        bi_fuse_discard_fcmp(I, mod, ctx->arch);
176
177                        if (bi_is_fabsneg(mod->op, size)) {
178                                if (mod->src[0].abs && !bi_takes_fabs(ctx->arch, I, mod->src[0], s))
179                                        continue;
180
181                                if (mod->src[0].neg && !bi_takes_fneg(ctx->arch, I, s))
182                                        continue;
183
184                                I->src[s] = bi_compose_float_index(I->src[s], mod->src[0]);
185                        }
186                }
187        }
188
189        free(lut);
190}
191
192/* RSCALE has restrictions on how the clamp may be used, only used for
193 * specialized transcendental sequences that set the clamp explicitly anyway */
194
195static bool
196bi_takes_clamp(bi_instr *I)
197{
198        switch (I->op) {
199        case BI_OPCODE_FMA_RSCALE_F32:
200        case BI_OPCODE_FMA_RSCALE_V2F16:
201        case BI_OPCODE_FADD_RSCALE_F32:
202                return false;
203        case BI_OPCODE_FADD_V2F16:
204                /* Encoding restriction */
205                return !(I->src[0].abs && I->src[1].abs &&
206                         bi_is_word_equiv(I->src[0], I->src[1]));
207        default:
208                return bi_opcode_props[I->op].clamp;
209        }
210}
211
212static bool
213bi_is_fclamp(enum bi_opcode op, enum bi_size size)
214{
215        return (size == BI_SIZE_32 && op == BI_OPCODE_FCLAMP_F32) ||
216               (size == BI_SIZE_16 && op == BI_OPCODE_FCLAMP_V2F16);
217}
218
219static bool
220bi_optimizer_clamp(bi_instr *I, bi_instr *use)
221{
222        if (!bi_is_fclamp(use->op, bi_opcode_props[I->op].size)) return false;
223        if (!bi_takes_clamp(I)) return false;
224
225        /* Clamps are bitfields (clamp_m1_1/clamp_0_inf) so composition is OR */
226        I->clamp |= use->clamp;
227        I->dest[0] = use->dest[0];
228        return true;
229}
230
231static enum bi_opcode
232bi_sized_mux_op(unsigned size)
233{
234        switch (size) {
235        case  8: return BI_OPCODE_MUX_V4I8;
236        case 16: return BI_OPCODE_MUX_V2I16;
237        case 32: return BI_OPCODE_MUX_I32;
238        default: unreachable("invalid size");
239        }
240}
241
242static bool
243bi_is_fixed_mux(bi_instr *I, unsigned size, bi_index v1)
244{
245        return I->op == bi_sized_mux_op(size) &&
246               bi_is_value_equiv(I->src[0], bi_zero()) &&
247               bi_is_value_equiv(I->src[1], v1);
248}
249
250static bool
251bi_takes_int_result_type(enum bi_opcode op)
252{
253        switch (op) {
254        case BI_OPCODE_ICMP_I32:
255        case BI_OPCODE_ICMP_S32:
256        case BI_OPCODE_ICMP_U32:
257        case BI_OPCODE_ICMP_V2I16:
258        case BI_OPCODE_ICMP_V2S16:
259        case BI_OPCODE_ICMP_V2U16:
260        case BI_OPCODE_ICMP_V4I8:
261        case BI_OPCODE_ICMP_V4S8:
262        case BI_OPCODE_ICMP_V4U8:
263        case BI_OPCODE_FCMP_F32:
264        case BI_OPCODE_FCMP_V2F16:
265                return true;
266        default:
267                return false;
268        }
269}
270
271static bool
272bi_takes_float_result_type(enum bi_opcode op)
273{
274        return (op == BI_OPCODE_FCMP_F32) ||
275               (op == BI_OPCODE_FCMP_V2F16);
276}
277
278/* CMP+MUX -> CMP with result type */
279static bool
280bi_optimizer_result_type(bi_instr *I, bi_instr *mux)
281{
282        if (bi_opcode_props[I->op].size != bi_opcode_props[mux->op].size)
283                return false;
284
285        if (bi_is_fixed_mux(mux, 32, bi_imm_f32(1.0)) ||
286            bi_is_fixed_mux(mux, 16, bi_imm_f16(1.0))) {
287
288                if (!bi_takes_float_result_type(I->op))
289                        return false;
290
291                I->result_type = BI_RESULT_TYPE_F1;
292        } else if (bi_is_fixed_mux(mux, 32, bi_imm_u32(1)) ||
293                   bi_is_fixed_mux(mux, 16, bi_imm_u16(1)) ||
294                   bi_is_fixed_mux(mux,  8, bi_imm_u8(1))) {
295
296                if (!bi_takes_int_result_type(I->op))
297                        return false;
298
299                I->result_type = BI_RESULT_TYPE_I1;
300        } else {
301                return false;
302        }
303
304        I->dest[0] = mux->dest[0];
305        return true;
306}
307
308static bool
309bi_is_var_tex(bi_instr *var, bi_instr *tex)
310{
311        return (var->op == BI_OPCODE_LD_VAR_IMM) &&
312                (tex->op == BI_OPCODE_TEXS_2D_F16 || tex->op == BI_OPCODE_TEXS_2D_F32) &&
313                (var->register_format == BI_REGISTER_FORMAT_F32) &&
314                ((var->sample == BI_SAMPLE_CENTER && var->update == BI_UPDATE_STORE) ||
315                 (var->sample == BI_SAMPLE_NONE && var->update == BI_UPDATE_RETRIEVE)) &&
316                (tex->texture_index == tex->sampler_index) &&
317                (tex->texture_index < 4) &&
318                (var->index < 8);
319}
320
321static bool
322bi_optimizer_var_tex(bi_context *ctx, bi_instr *var, bi_instr *tex)
323{
324        if (!bi_is_var_tex(var, tex)) return false;
325
326        /* Construct the corresponding VAR_TEX intruction */
327        bi_builder b = bi_init_builder(ctx, bi_after_instr(var));
328
329        bi_instr *I = bi_var_tex_f32_to(&b, tex->dest[0], tex->lod_mode,
330                        var->sample, var->update, tex->texture_index, var->index);
331        I->skip = tex->skip;
332
333        if (tex->op == BI_OPCODE_TEXS_2D_F16)
334                I->op = BI_OPCODE_VAR_TEX_F16;
335
336        /* Dead code elimination will clean up for us */
337        return true;
338}
339
340void
341bi_opt_mod_prop_backward(bi_context *ctx)
342{
343        unsigned count = ctx->ssa_alloc;
344        bi_instr **uses = calloc(count, sizeof(*uses));
345        BITSET_WORD *multiple = calloc(BITSET_WORDS(count), sizeof(*multiple));
346
347        bi_foreach_instr_global_rev(ctx, I) {
348                bi_foreach_src(I, s) {
349                        if (bi_is_ssa(I->src[s])) {
350                                unsigned v = I->src[s].value;
351
352                                if (uses[v] && uses[v] != I)
353                                        BITSET_SET(multiple, v);
354                                else
355                                        uses[v] = I;
356                        }
357                }
358
359                if (!bi_is_ssa(I->dest[0]))
360                        continue;
361
362                bi_instr *use = uses[I->dest[0].value];
363
364                if (!use || BITSET_TEST(multiple, I->dest[0].value))
365                        continue;
366
367                /* Destination has a single use, try to propagate */
368                bool propagated =
369                        bi_optimizer_clamp(I, use) ||
370                        bi_optimizer_result_type(I, use);
371
372                if (!propagated && I->op == BI_OPCODE_LD_VAR_IMM && use->op == BI_OPCODE_SPLIT_I32) {
373                        /* Need to see through the split in a
374                         * ld_var_imm/split/var_tex  sequence
375                         */
376                        assert(bi_is_ssa(use->dest[0]));
377                        bi_instr *tex = uses[use->dest[0].value];
378
379                        if (!tex || BITSET_TEST(multiple, use->dest[0].value))
380                                continue;
381
382                        use = tex;
383                        propagated = bi_optimizer_var_tex(ctx, I, use);
384                }
385
386                if (propagated) {
387                        bi_remove_instruction(use);
388                        continue;
389                }
390        }
391
392        free(uses);
393        free(multiple);
394}
395
396/** Lower pseudo instructions that exist to simplify the optimizer */
397
398void
399bi_lower_opt_instruction(bi_instr *I)
400{
401        switch (I->op) {
402        case BI_OPCODE_FABSNEG_F32:
403        case BI_OPCODE_FABSNEG_V2F16:
404        case BI_OPCODE_FCLAMP_F32:
405        case BI_OPCODE_FCLAMP_V2F16:
406                I->op = (bi_opcode_props[I->op].size == BI_SIZE_32) ?
407                        BI_OPCODE_FADD_F32 : BI_OPCODE_FADD_V2F16;
408
409                I->round = BI_ROUND_NONE;
410                I->src[1] = bi_negzero();
411                break;
412
413        case BI_OPCODE_DISCARD_B32:
414                I->op = BI_OPCODE_DISCARD_F32;
415                I->src[1] = bi_imm_u32(0);
416                I->cmpf = BI_CMPF_NE;
417                break;
418
419        default:
420                break;
421        }
422}
423