1 /*
2  * Copyright © 2016 Intel Corporation
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 "nir.h"
25 #include "nir/nir_builder.h"
26 #include "nir_constant_expressions.h"
27 #include "nir_control_flow.h"
28 #include "nir_loop_analyze.h"
29 
30 static nir_ssa_def *clone_alu_and_replace_src_defs(nir_builder *b,
31                                                    const nir_alu_instr *alu,
32                                                    nir_ssa_def **src_defs);
33 
34 /**
35  * Gets the single block that jumps back to the loop header. Already assumes
36  * there is exactly one such block.
37  */
38 static nir_block*
find_continue_block(nir_loop *loop)39 find_continue_block(nir_loop *loop)
40 {
41    nir_block *header_block = nir_loop_first_block(loop);
42    nir_block *prev_block =
43       nir_cf_node_as_block(nir_cf_node_prev(&loop->cf_node));
44 
45    assert(header_block->predecessors->entries == 2);
46 
47    set_foreach(header_block->predecessors, pred_entry) {
48       if (pred_entry->key != prev_block)
49          return (nir_block*)pred_entry->key;
50    }
51 
52    unreachable("Continue block not found!");
53 }
54 
55 /**
56  * Does a phi have one constant value from outside a loop and one from inside?
57  */
58 static bool
phi_has_constant_from_outside_and_one_from_inside_loop(nir_phi_instr *phi, const nir_block *entry_block, bool *entry_val, bool *continue_val)59 phi_has_constant_from_outside_and_one_from_inside_loop(nir_phi_instr *phi,
60                                                        const nir_block *entry_block,
61                                                        bool *entry_val,
62                                                        bool *continue_val)
63 {
64    /* We already know we have exactly one continue */
65    assert(exec_list_length(&phi->srcs) == 2);
66 
67    *entry_val = false;
68    *continue_val = false;
69 
70     nir_foreach_phi_src(src, phi) {
71        if (!nir_src_is_const(src->src))
72          return false;
73 
74        if (src->pred != entry_block) {
75           *continue_val = nir_src_as_bool(src->src);
76        } else {
77           *entry_val = nir_src_as_bool(src->src);
78        }
79     }
80 
81     return true;
82 }
83 
84 /**
85  * This optimization detects if statements at the tops of loops where the
86  * condition is a phi node of two constants and moves half of the if to above
87  * the loop and the other half of the if to the end of the loop.  A simple for
88  * loop "for (int i = 0; i < 4; i++)", when run through the SPIR-V front-end,
89  * ends up looking something like this:
90  *
91  * vec1 32 ssa_0 = load_const (0x00000000)
92  * vec1 32 ssa_1 = load_const (0xffffffff)
93  * loop {
94  *    block block_1:
95  *    vec1 32 ssa_2 = phi block_0: ssa_0, block_7: ssa_5
96  *    vec1 32 ssa_3 = phi block_0: ssa_0, block_7: ssa_1
97  *    if ssa_3 {
98  *       block block_2:
99  *       vec1 32 ssa_4 = load_const (0x00000001)
100  *       vec1 32 ssa_5 = iadd ssa_2, ssa_4
101  *    } else {
102  *       block block_3:
103  *    }
104  *    block block_4:
105  *    vec1 32 ssa_6 = load_const (0x00000004)
106  *    vec1 32 ssa_7 = ilt ssa_5, ssa_6
107  *    if ssa_7 {
108  *       block block_5:
109  *    } else {
110  *       block block_6:
111  *       break
112  *    }
113  *    block block_7:
114  * }
115  *
116  * This turns it into something like this:
117  *
118  * // Stuff from block 1
119  * // Stuff from block 3
120  * loop {
121  *    block block_1:
122  *    vec1 32 ssa_2 = phi block_0: ssa_0, block_7: ssa_5
123  *    vec1 32 ssa_6 = load_const (0x00000004)
124  *    vec1 32 ssa_7 = ilt ssa_2, ssa_6
125  *    if ssa_7 {
126  *       block block_5:
127  *    } else {
128  *       block block_6:
129  *       break
130  *    }
131  *    block block_7:
132  *    // Stuff from block 1
133  *    // Stuff from block 2
134  *    vec1 32 ssa_4 = load_const (0x00000001)
135  *    vec1 32 ssa_5 = iadd ssa_2, ssa_4
136  * }
137  */
138 static bool
opt_peel_loop_initial_if(nir_loop *loop)139 opt_peel_loop_initial_if(nir_loop *loop)
140 {
141    nir_block *header_block = nir_loop_first_block(loop);
142    nir_block *const prev_block =
143       nir_cf_node_as_block(nir_cf_node_prev(&loop->cf_node));
144 
145    /* It would be insane if this were not true */
146    assert(_mesa_set_search(header_block->predecessors, prev_block));
147 
148    /* The loop must have exactly one continue block which could be a block
149     * ending in a continue instruction or the "natural" continue from the
150     * last block in the loop back to the top.
151     */
152    if (header_block->predecessors->entries != 2)
153       return false;
154 
155    nir_cf_node *if_node = nir_cf_node_next(&header_block->cf_node);
156    if (!if_node || if_node->type != nir_cf_node_if)
157       return false;
158 
159    nir_if *nif = nir_cf_node_as_if(if_node);
160    if (!nif->condition.is_ssa)
161       return false;
162 
163    nir_ssa_def *cond = nif->condition.ssa;
164    if (cond->parent_instr->type != nir_instr_type_phi)
165       return false;
166 
167    nir_phi_instr *cond_phi = nir_instr_as_phi(cond->parent_instr);
168    if (cond->parent_instr->block != header_block)
169       return false;
170 
171    bool entry_val = false, continue_val = false;
172    if (!phi_has_constant_from_outside_and_one_from_inside_loop(cond_phi,
173                                                                prev_block,
174                                                                &entry_val,
175                                                                &continue_val))
176       return false;
177 
178    /* If they both execute or both don't execute, this is a job for
179     * nir_dead_cf, not this pass.
180     */
181    if ((entry_val && continue_val) || (!entry_val && !continue_val))
182       return false;
183 
184    struct exec_list *continue_list, *entry_list;
185    if (continue_val) {
186       continue_list = &nif->then_list;
187       entry_list = &nif->else_list;
188    } else {
189       continue_list = &nif->else_list;
190       entry_list = &nif->then_list;
191    }
192 
193    /* We want to be moving the contents of entry_list to above the loop so it
194     * can't contain any break or continue instructions.
195     */
196    foreach_list_typed(nir_cf_node, cf_node, node, entry_list) {
197       nir_foreach_block_in_cf_node(block, cf_node) {
198          nir_instr *last_instr = nir_block_last_instr(block);
199          if (last_instr && last_instr->type == nir_instr_type_jump)
200             return false;
201       }
202    }
203 
204    /* We're about to re-arrange a bunch of blocks so make sure that we don't
205     * have deref uses which cross block boundaries.  We don't want a deref
206     * accidentally ending up in a phi.
207     */
208    nir_rematerialize_derefs_in_use_blocks_impl(
209       nir_cf_node_get_function(&loop->cf_node));
210 
211    /* Before we do anything, convert the loop to LCSSA.  We're about to
212     * replace a bunch of SSA defs with registers and this will prevent any of
213     * it from leaking outside the loop.
214     */
215    nir_convert_loop_to_lcssa(loop);
216 
217    nir_block *after_if_block =
218       nir_cf_node_as_block(nir_cf_node_next(&nif->cf_node));
219 
220    /* Get rid of phis in the header block since we will be duplicating it */
221    nir_lower_phis_to_regs_block(header_block);
222    /* Get rid of phis after the if since dominance will change */
223    nir_lower_phis_to_regs_block(after_if_block);
224 
225    /* Get rid of SSA defs in the pieces we're about to move around */
226    nir_lower_ssa_defs_to_regs_block(header_block);
227    nir_foreach_block_in_cf_node(block, &nif->cf_node)
228       nir_lower_ssa_defs_to_regs_block(block);
229 
230    nir_cf_list header, tmp;
231    nir_cf_extract(&header, nir_before_block(header_block),
232                            nir_after_block(header_block));
233 
234    nir_cf_list_clone(&tmp, &header, &loop->cf_node, NULL);
235    nir_cf_reinsert(&tmp, nir_before_cf_node(&loop->cf_node));
236    nir_cf_extract(&tmp, nir_before_cf_list(entry_list),
237                         nir_after_cf_list(entry_list));
238    nir_cf_reinsert(&tmp, nir_before_cf_node(&loop->cf_node));
239 
240    nir_cf_reinsert(&header,
241                    nir_after_block_before_jump(find_continue_block(loop)));
242 
243    bool continue_list_jumps =
244       nir_block_ends_in_jump(exec_node_data(nir_block,
245                                             exec_list_get_tail(continue_list),
246                                             cf_node.node));
247 
248    nir_cf_extract(&tmp, nir_before_cf_list(continue_list),
249                         nir_after_cf_list(continue_list));
250 
251    /* Get continue block again as the previous reinsert might have removed the
252     * block.  Also, if both the continue list and the continue block ends in
253     * jump instructions, removes the jump from the latter, as it will not be
254     * executed if we insert the continue list before it. */
255 
256    nir_block *continue_block = find_continue_block(loop);
257 
258    if (continue_list_jumps) {
259       nir_instr *last_instr = nir_block_last_instr(continue_block);
260       if (last_instr && last_instr->type == nir_instr_type_jump)
261          nir_instr_remove(last_instr);
262    }
263 
264    nir_cf_reinsert(&tmp,
265                    nir_after_block_before_jump(continue_block));
266 
267    nir_cf_node_remove(&nif->cf_node);
268 
269    return true;
270 }
271 
272 static bool
alu_instr_is_comparison(const nir_alu_instr *alu)273 alu_instr_is_comparison(const nir_alu_instr *alu)
274 {
275    switch (alu->op) {
276    case nir_op_flt32:
277    case nir_op_fge32:
278    case nir_op_feq32:
279    case nir_op_fneu32:
280    case nir_op_ilt32:
281    case nir_op_ult32:
282    case nir_op_ige32:
283    case nir_op_uge32:
284    case nir_op_ieq32:
285    case nir_op_ine32:
286       return true;
287    default:
288       return nir_alu_instr_is_comparison(alu);
289    }
290 }
291 
292 static bool
alu_instr_is_type_conversion(const nir_alu_instr *alu)293 alu_instr_is_type_conversion(const nir_alu_instr *alu)
294 {
295    return nir_op_infos[alu->op].num_inputs == 1 &&
296           nir_op_infos[alu->op].output_type != nir_op_infos[alu->op].input_types[0];
297 }
298 
299 static bool
is_trivial_bcsel(const nir_instr *instr, bool allow_non_phi_src)300 is_trivial_bcsel(const nir_instr *instr, bool allow_non_phi_src)
301 {
302    if (instr->type != nir_instr_type_alu)
303       return false;
304 
305    nir_alu_instr *const bcsel = nir_instr_as_alu(instr);
306    if (!nir_op_is_selection(bcsel->op))
307       return false;
308 
309    for (unsigned i = 0; i < 3; i++) {
310       if (!nir_alu_src_is_trivial_ssa(bcsel, i) ||
311           bcsel->src[i].src.ssa->parent_instr->block != instr->block)
312          return false;
313 
314       if (bcsel->src[i].src.ssa->parent_instr->type != nir_instr_type_phi) {
315          /* opt_split_alu_of_phi() is able to peel that src from the loop */
316          if (i == 0 || !allow_non_phi_src)
317             return false;
318          allow_non_phi_src = false;
319       }
320    }
321 
322    nir_foreach_phi_src(src, nir_instr_as_phi(bcsel->src[0].src.ssa->parent_instr)) {
323       if (!nir_src_is_const(src->src))
324          return false;
325    }
326 
327    return true;
328 }
329 
330 /**
331  * Splits ALU instructions that have a source that is a phi node
332  *
333  * ALU instructions in the header block of a loop that meet the following
334  * criteria can be split.
335  *
336  * - The loop has no continue instructions other than the "natural" continue
337  *   at the bottom of the loop.
338  *
339  * - At least one source of the instruction is a phi node from the header block.
340  *
341  * - Any non-phi sources of the ALU instruction come from a block that
342  *   dominates the block before the loop.  The most common failure mode for
343  *   this check is sources that are generated in the loop header block.
344  *
345  * - The phi node selects a constant or undef from the block before the loop or
346  *   the only ALU user is a trivial bcsel that gets removed by peeling the ALU
347  *
348  * The split process splits the original ALU instruction into two, one at the
349  * bottom of the loop and one at the block before the loop. The instruction
350  * before the loop computes the value on the first iteration, and the
351  * instruction at the bottom computes the value on the second, third, and so
352  * on. A new phi node is added to the header block that selects either the
353  * instruction before the loop or the one at the end, and uses of the original
354  * instruction are replaced by this phi.
355  *
356  * The splitting transforms a loop like:
357  *
358  *    vec1 32 ssa_8 = load_const (0x00000001)
359  *    vec1 32 ssa_10 = load_const (0x00000000)
360  *    // succs: block_1
361  *    loop {
362  *            block block_1:
363  *            // preds: block_0 block_4
364  *            vec1 32 ssa_11 = phi block_0: ssa_10, block_4: ssa_15
365  *            vec1 32 ssa_12 = phi block_0: ssa_1, block_4: ssa_15
366  *            vec1 32 ssa_13 = phi block_0: ssa_10, block_4: ssa_16
367  *            vec1 32 ssa_14 = iadd ssa_11, ssa_8
368  *            vec1 32 ssa_15 = b32csel ssa_13, ssa_14, ssa_12
369  *            ...
370  *            // succs: block_1
371  *    }
372  *
373  * into:
374  *
375  *    vec1 32 ssa_8 = load_const (0x00000001)
376  *    vec1 32 ssa_10 = load_const (0x00000000)
377  *    vec1 32 ssa_22 = iadd ssa_10, ssa_8
378  *    // succs: block_1
379  *    loop {
380  *            block block_1:
381  *            // preds: block_0 block_4
382  *            vec1 32 ssa_11 = phi block_0: ssa_10, block_4: ssa_15
383  *            vec1 32 ssa_12 = phi block_0: ssa_1, block_4: ssa_15
384  *            vec1 32 ssa_13 = phi block_0: ssa_10, block_4: ssa_16
385  *            vec1 32 ssa_21 = phi block_0: ssa_22, block_4: ssa_20
386  *            vec1 32 ssa_15 = b32csel ssa_13, ssa_21, ssa_12
387  *            ...
388  *            vec1 32 ssa_20 = iadd ssa_15, ssa_8
389  *            // succs: block_1
390  *    }
391  */
392 static bool
opt_split_alu_of_phi(nir_builder *b, nir_loop *loop)393 opt_split_alu_of_phi(nir_builder *b, nir_loop *loop)
394 {
395    bool progress = false;
396    nir_block *header_block = nir_loop_first_block(loop);
397    nir_block *const prev_block =
398       nir_cf_node_as_block(nir_cf_node_prev(&loop->cf_node));
399 
400    /* It would be insane if this were not true */
401    assert(_mesa_set_search(header_block->predecessors, prev_block));
402 
403    /* The loop must have exactly one continue block which could be a block
404     * ending in a continue instruction or the "natural" continue from the
405     * last block in the loop back to the top.
406     */
407    if (header_block->predecessors->entries != 2)
408       return false;
409 
410    nir_block *continue_block = find_continue_block(loop);
411    if (continue_block == header_block)
412       return false;
413 
414    nir_foreach_instr_safe(instr, header_block) {
415       if (instr->type != nir_instr_type_alu)
416          continue;
417 
418       nir_alu_instr *const alu = nir_instr_as_alu(instr);
419 
420       /* nir_op_vec{2,3,4} and nir_op_mov are excluded because they can easily
421        * lead to infinite optimization loops. Splitting comparisons can lead
422        * to loop unrolling not recognizing loop termintators, and type
423        * conversions also lead to regressions.
424        */
425       if (nir_op_is_vec(alu->op) ||
426           alu_instr_is_comparison(alu) ||
427           alu_instr_is_type_conversion(alu))
428          continue;
429 
430       bool has_phi_src_from_prev_block = false;
431       bool all_non_phi_exist_in_prev_block = true;
432       bool is_prev_result_undef = true;
433       bool is_prev_result_const = true;
434       nir_ssa_def *prev_srcs[8];     // FINISHME: Array size?
435       nir_ssa_def *continue_srcs[8]; // FINISHME: Array size?
436 
437       for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
438          nir_instr *const src_instr = alu->src[i].src.ssa->parent_instr;
439 
440          /* If the source is a phi in the loop header block, then the
441           * prev_srcs and continue_srcs will come from the different sources
442           * of the phi.
443           */
444          if (src_instr->type == nir_instr_type_phi &&
445              src_instr->block == header_block) {
446             nir_phi_instr *const phi = nir_instr_as_phi(src_instr);
447 
448             /* Only strictly need to NULL out the pointers when the assertions
449              * (below) are compiled in.  Debugging a NULL pointer deref in the
450              * wild is easier than debugging a random pointer deref, so set
451              * NULL unconditionally just to be safe.
452              */
453             prev_srcs[i] = NULL;
454             continue_srcs[i] = NULL;
455 
456             nir_foreach_phi_src(src_of_phi, phi) {
457                if (src_of_phi->pred == prev_block) {
458                   if (src_of_phi->src.ssa->parent_instr->type !=
459                       nir_instr_type_ssa_undef) {
460                      is_prev_result_undef = false;
461                   }
462 
463                   if (src_of_phi->src.ssa->parent_instr->type !=
464                       nir_instr_type_load_const) {
465                      is_prev_result_const = false;
466                   }
467 
468                   prev_srcs[i] = src_of_phi->src.ssa;
469                   has_phi_src_from_prev_block = true;
470                } else
471                   continue_srcs[i] = src_of_phi->src.ssa;
472             }
473 
474             assert(prev_srcs[i] != NULL);
475             assert(continue_srcs[i] != NULL);
476          } else {
477             /* If the source is not a phi (or a phi in a block other than the
478              * loop header), then the value must exist in prev_block.
479              */
480             if (!nir_block_dominates(src_instr->block, prev_block)) {
481                all_non_phi_exist_in_prev_block = false;
482                break;
483             }
484 
485             prev_srcs[i] = alu->src[i].src.ssa;
486             continue_srcs[i] = alu->src[i].src.ssa;
487          }
488       }
489 
490       if (!has_phi_src_from_prev_block || !all_non_phi_exist_in_prev_block)
491          continue;
492 
493       if (!is_prev_result_undef && !is_prev_result_const) {
494          /* check if the only user is a trivial bcsel */
495          if (!list_is_empty(&alu->dest.dest.ssa.if_uses) ||
496              !list_is_singular(&alu->dest.dest.ssa.uses))
497             continue;
498 
499          nir_src *use = list_first_entry(&alu->dest.dest.ssa.uses, nir_src, use_link);
500          if (!is_trivial_bcsel(use->parent_instr, true))
501             continue;
502       }
503 
504       /* Split ALU of Phi */
505       b->cursor = nir_after_block(prev_block);
506       nir_ssa_def *prev_value = clone_alu_and_replace_src_defs(b, alu, prev_srcs);
507 
508       /* Make a copy of the original ALU instruction.  Replace the sources
509        * of the new instruction that read a phi with an undef source from
510        * prev_block with the non-undef source of that phi.
511        *
512        * Insert the new instruction at the end of the continue block.
513        */
514       b->cursor = nir_after_block_before_jump(continue_block);
515 
516       nir_ssa_def *const alu_copy =
517          clone_alu_and_replace_src_defs(b, alu, continue_srcs);
518 
519       /* Make a new phi node that selects a value from prev_block and the
520        * result of the new instruction from continue_block.
521        */
522       nir_phi_instr *const phi = nir_phi_instr_create(b->shader);
523       nir_phi_instr_add_src(phi, prev_block, nir_src_for_ssa(prev_value));
524       nir_phi_instr_add_src(phi, continue_block, nir_src_for_ssa(alu_copy));
525 
526       nir_ssa_dest_init(&phi->instr, &phi->dest,
527                         alu_copy->num_components, alu_copy->bit_size, NULL);
528 
529       b->cursor = nir_after_phis(header_block);
530       nir_builder_instr_insert(b, &phi->instr);
531 
532       /* Modify all readers of the original ALU instruction to read the
533        * result of the phi.
534        */
535       nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa,
536                                &phi->dest.ssa);
537 
538       /* Since the original ALU instruction no longer has any readers, just
539        * remove it.
540        */
541       nir_instr_remove_v(&alu->instr);
542       nir_instr_free(&alu->instr);
543 
544       progress = true;
545    }
546 
547    return progress;
548 }
549 
550 /**
551  * Simplify a bcsel whose sources are all phi nodes from the loop header block
552  *
553  * bcsel instructions in a loop that meet the following criteria can be
554  * converted to phi nodes:
555  *
556  * - The loop has no continue instructions other than the "natural" continue
557  *   at the bottom of the loop.
558  *
559  * - All of the sources of the bcsel are phi nodes in the header block of the
560  *   loop.
561  *
562  * - The phi node representing the condition of the bcsel instruction chooses
563  *   only constant values.
564  *
565  * The contant value from the condition will select one of the other sources
566  * when entered from outside the loop and the remaining source when entered
567  * from the continue block.  Since each of these sources is also a phi node in
568  * the header block, the value of the phi node can be "evaluated."  These
569  * evaluated phi nodes provide the sources for a new phi node.  All users of
570  * the bcsel result are updated to use the phi node result.
571  *
572  * The replacement transforms loops like:
573  *
574  *    vec1 32 ssa_7 = undefined
575  *    vec1 32 ssa_8 = load_const (0x00000001)
576  *    vec1 32 ssa_9 = load_const (0x000000c8)
577  *    vec1 32 ssa_10 = load_const (0x00000000)
578  *    // succs: block_1
579  *    loop {
580  *            block block_1:
581  *            // preds: block_0 block_4
582  *            vec1 32 ssa_11 = phi block_0: ssa_1, block_4: ssa_14
583  *            vec1 32 ssa_12 = phi block_0: ssa_10, block_4: ssa_15
584  *            vec1 32 ssa_13 = phi block_0: ssa_7, block_4: ssa_25
585  *            vec1 32 ssa_14 = b32csel ssa_12, ssa_13, ssa_11
586  *            vec1 32 ssa_16 = ige32 ssa_14, ssa_9
587  *            ...
588  *            vec1 32 ssa_15 = load_const (0xffffffff)
589  *            ...
590  *            vec1 32 ssa_25 = iadd ssa_14, ssa_8
591  *            // succs: block_1
592  *    }
593  *
594  * into:
595  *
596  *    vec1 32 ssa_7 = undefined
597  *    vec1 32 ssa_8 = load_const (0x00000001)
598  *    vec1 32 ssa_9 = load_const (0x000000c8)
599  *    vec1 32 ssa_10 = load_const (0x00000000)
600  *    // succs: block_1
601  *    loop {
602  *            block block_1:
603  *            // preds: block_0 block_4
604  *            vec1 32 ssa_11 = phi block_0: ssa_1, block_4: ssa_14
605  *            vec1 32 ssa_12 = phi block_0: ssa_10, block_4: ssa_15
606  *            vec1 32 ssa_13 = phi block_0: ssa_7, block_4: ssa_25
607  *            vec1 32 sss_26 = phi block_0: ssa_1, block_4: ssa_25
608  *            vec1 32 ssa_16 = ige32 ssa_26, ssa_9
609  *            ...
610  *            vec1 32 ssa_15 = load_const (0xffffffff)
611  *            ...
612  *            vec1 32 ssa_25 = iadd ssa_26, ssa_8
613  *            // succs: block_1
614  *    }
615  *
616  * \note
617  * It may be possible modify this function to not require a phi node as the
618  * source of the bcsel that is selected when entering from outside the loop.
619  * The only restriction is that the source must be geneated outside the loop
620  * (since it will become the source of a phi node in the header block of the
621  * loop).
622  */
623 static bool
opt_simplify_bcsel_of_phi(nir_builder *b, nir_loop *loop)624 opt_simplify_bcsel_of_phi(nir_builder *b, nir_loop *loop)
625 {
626    bool progress = false;
627    nir_block *header_block = nir_loop_first_block(loop);
628    nir_block *const prev_block =
629       nir_cf_node_as_block(nir_cf_node_prev(&loop->cf_node));
630 
631    /* It would be insane if this were not true */
632    assert(_mesa_set_search(header_block->predecessors, prev_block));
633 
634    /* The loop must have exactly one continue block which could be a block
635     * ending in a continue instruction or the "natural" continue from the
636     * last block in the loop back to the top.
637     */
638    if (header_block->predecessors->entries != 2)
639       return false;
640 
641    /* We can move any bcsel that can guaranteed to execut on every iteration
642     * of a loop.  For now this is accomplished by only taking bcsels from the
643     * header_block.  In the future, this could be expanced to include any
644     * bcsel that must come before any break.
645     *
646     * For more details, see
647     * https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/170#note_110305
648     */
649    nir_foreach_instr_safe(instr, header_block) {
650       if (!is_trivial_bcsel(instr, false))
651          continue;
652 
653       nir_alu_instr *const bcsel = nir_instr_as_alu(instr);
654       nir_phi_instr *const cond_phi =
655          nir_instr_as_phi(bcsel->src[0].src.ssa->parent_instr);
656 
657       bool entry_val = false, continue_val = false;
658       if (!phi_has_constant_from_outside_and_one_from_inside_loop(cond_phi,
659                                                                   prev_block,
660                                                                   &entry_val,
661                                                                   &continue_val))
662          continue;
663 
664       /* If they both execute or both don't execute, this is a job for
665        * nir_dead_cf, not this pass.
666        */
667       if ((entry_val && continue_val) || (!entry_val && !continue_val))
668          continue;
669 
670       const unsigned entry_src = entry_val ? 1 : 2;
671       const unsigned continue_src = entry_val ? 2 : 1;
672 
673       /* Create a new phi node that selects the value for prev_block from
674        * the bcsel source that is selected by entry_val and the value for
675        * continue_block from the other bcsel source.  Both sources have
676        * already been verified to be phi nodes.
677        */
678       nir_block *continue_block = find_continue_block(loop);
679       nir_phi_instr *const phi = nir_phi_instr_create(b->shader);
680       nir_phi_instr_add_src(phi, prev_block,
681                             nir_phi_get_src_from_block(nir_instr_as_phi(bcsel->src[entry_src].src.ssa->parent_instr),
682                                                        prev_block)->src);
683 
684       nir_phi_instr_add_src(phi, continue_block,
685                             nir_phi_get_src_from_block(nir_instr_as_phi(bcsel->src[continue_src].src.ssa->parent_instr),
686                                     continue_block)->src);
687 
688       nir_ssa_dest_init(&phi->instr,
689                         &phi->dest,
690                         nir_dest_num_components(bcsel->dest.dest),
691                         nir_dest_bit_size(bcsel->dest.dest),
692                         NULL);
693 
694       b->cursor = nir_after_phis(header_block);
695       nir_builder_instr_insert(b, &phi->instr);
696 
697       /* Modify all readers of the bcsel instruction to read the result of
698        * the phi.
699        */
700       nir_ssa_def_rewrite_uses(&bcsel->dest.dest.ssa,
701                                &phi->dest.ssa);
702 
703       /* Since the original bcsel instruction no longer has any readers,
704        * just remove it.
705        */
706       nir_instr_remove_v(&bcsel->instr);
707       nir_instr_free(&bcsel->instr);
708 
709       progress = true;
710    }
711 
712    return progress;
713 }
714 
715 static bool
is_block_empty(nir_block *block)716 is_block_empty(nir_block *block)
717 {
718    return nir_cf_node_is_last(&block->cf_node) &&
719           exec_list_is_empty(&block->instr_list);
720 }
721 
722 static bool
nir_block_ends_in_continue(nir_block *block)723 nir_block_ends_in_continue(nir_block *block)
724 {
725    if (exec_list_is_empty(&block->instr_list))
726       return false;
727 
728    nir_instr *instr = nir_block_last_instr(block);
729    return instr->type == nir_instr_type_jump &&
730       nir_instr_as_jump(instr)->type == nir_jump_continue;
731 }
732 
733 /**
734  * This optimization turns:
735  *
736  *     loop {
737  *        ...
738  *        if (cond) {
739  *           do_work_1();
740  *           continue;
741  *        } else {
742  *        }
743  *        do_work_2();
744  *     }
745  *
746  * into:
747  *
748  *     loop {
749  *        ...
750  *        if (cond) {
751  *           do_work_1();
752  *           continue;
753  *        } else {
754  *           do_work_2();
755  *        }
756  *     }
757  *
758  * The continue should then be removed by nir_opt_trivial_continues() and the
759  * loop can potentially be unrolled.
760  *
761  * Note: Unless the function param aggressive_last_continue==true do_work_2()
762  * is only ever blocks and nested loops. We avoid nesting other if-statments
763  * in the branch as this can result in increased register pressure, and in
764  * the i965 driver it causes a large amount of spilling in shader-db.
765  * For RADV however nesting these if-statements allows further continues to be
766  * remove and provides a significant FPS boost in Doom, which is why we have
767  * opted for this special bool to enable more aggresive optimisations.
768  * TODO: The GCM pass solves most of the spilling regressions in i965, if it
769  * is ever enabled we should consider removing the aggressive_last_continue
770  * param.
771  */
772 static bool
opt_if_loop_last_continue(nir_loop *loop, bool aggressive_last_continue)773 opt_if_loop_last_continue(nir_loop *loop, bool aggressive_last_continue)
774 {
775    nir_if *nif = NULL;
776    bool then_ends_in_continue = false;
777    bool else_ends_in_continue = false;
778 
779    /* Scan the control flow of the loop from the last to the first node
780     * looking for an if-statement we can optimise.
781     */
782    nir_block *last_block = nir_loop_last_block(loop);
783    nir_cf_node *if_node = nir_cf_node_prev(&last_block->cf_node);
784    while (if_node) {
785       if (if_node->type == nir_cf_node_if) {
786          nif = nir_cf_node_as_if(if_node);
787          nir_block *then_block = nir_if_last_then_block(nif);
788          nir_block *else_block = nir_if_last_else_block(nif);
789 
790          then_ends_in_continue = nir_block_ends_in_continue(then_block);
791          else_ends_in_continue = nir_block_ends_in_continue(else_block);
792 
793          /* If both branches end in a jump do nothing, this should be handled
794           * by nir_opt_dead_cf().
795           */
796          if ((then_ends_in_continue || nir_block_ends_in_break(then_block)) &&
797              (else_ends_in_continue || nir_block_ends_in_break(else_block)))
798             return false;
799 
800          /* If continue found stop scanning and attempt optimisation, or
801           */
802          if (then_ends_in_continue || else_ends_in_continue ||
803              !aggressive_last_continue)
804             break;
805       }
806 
807       if_node = nir_cf_node_prev(if_node);
808    }
809 
810    /* If we didn't find an if to optimise return */
811    if (!nif || (!then_ends_in_continue && !else_ends_in_continue))
812       return false;
813 
814    /* If there is nothing after the if-statement we bail */
815    if (&nif->cf_node == nir_cf_node_prev(&last_block->cf_node) &&
816        exec_list_is_empty(&last_block->instr_list))
817       return false;
818 
819    /* If there are single-source phis in the last block,
820     * get rid of them first
821     */
822    nir_opt_remove_phis_block(last_block);
823 
824    /* Move the last block of the loop inside the last if-statement */
825    nir_cf_list tmp;
826    nir_cf_extract(&tmp, nir_after_cf_node(if_node),
827                         nir_after_block(last_block));
828    if (then_ends_in_continue)
829       nir_cf_reinsert(&tmp, nir_after_cf_list(&nif->else_list));
830    else
831       nir_cf_reinsert(&tmp, nir_after_cf_list(&nif->then_list));
832 
833    /* In order to avoid running nir_lower_regs_to_ssa_impl() every time an if
834     * opt makes progress we leave nir_opt_trivial_continues() to remove the
835     * continue now that the end of the loop has been simplified.
836     */
837 
838    return true;
839 }
840 
841 /* Walk all the phis in the block immediately following the if statement and
842  * swap the blocks.
843  */
844 static void
rewrite_phi_predecessor_blocks(nir_if *nif, nir_block *old_then_block, nir_block *old_else_block, nir_block *new_then_block, nir_block *new_else_block)845 rewrite_phi_predecessor_blocks(nir_if *nif,
846                                nir_block *old_then_block,
847                                nir_block *old_else_block,
848                                nir_block *new_then_block,
849                                nir_block *new_else_block)
850 {
851    nir_block *after_if_block =
852       nir_cf_node_as_block(nir_cf_node_next(&nif->cf_node));
853 
854    nir_foreach_instr(instr, after_if_block) {
855       if (instr->type != nir_instr_type_phi)
856          continue;
857 
858       nir_phi_instr *phi = nir_instr_as_phi(instr);
859 
860       nir_foreach_phi_src(src, phi) {
861          if (src->pred == old_then_block) {
862             src->pred = new_then_block;
863          } else if (src->pred == old_else_block) {
864             src->pred = new_else_block;
865          }
866       }
867    }
868 }
869 
870 /**
871  * This optimization turns:
872  *
873  *     if (cond) {
874  *     } else {
875  *         do_work();
876  *     }
877  *
878  * into:
879  *
880  *     if (!cond) {
881  *         do_work();
882  *     } else {
883  *     }
884  */
885 static bool
opt_if_simplification(nir_builder *b, nir_if *nif)886 opt_if_simplification(nir_builder *b, nir_if *nif)
887 {
888    /* Only simplify if the then block is empty and the else block is not. */
889    if (!is_block_empty(nir_if_first_then_block(nif)) ||
890        is_block_empty(nir_if_first_else_block(nif)))
891       return false;
892 
893    /* Make sure the condition is a comparison operation. */
894    nir_instr *src_instr = nif->condition.ssa->parent_instr;
895    if (src_instr->type != nir_instr_type_alu)
896       return false;
897 
898    nir_alu_instr *alu_instr = nir_instr_as_alu(src_instr);
899    if (!nir_alu_instr_is_comparison(alu_instr))
900       return false;
901 
902    /* Insert the inverted instruction and rewrite the condition. */
903    b->cursor = nir_after_instr(&alu_instr->instr);
904 
905    nir_ssa_def *new_condition =
906       nir_inot(b, &alu_instr->dest.dest.ssa);
907 
908    nir_if_rewrite_condition(nif, nir_src_for_ssa(new_condition));
909 
910    /* Grab pointers to the last then/else blocks for fixing up the phis. */
911    nir_block *then_block = nir_if_last_then_block(nif);
912    nir_block *else_block = nir_if_last_else_block(nif);
913 
914    if (nir_block_ends_in_jump(else_block)) {
915       /* Even though this if statement has a jump on one side, we may still have
916        * phis afterwards.  Single-source phis can be produced by loop unrolling
917        * or dead control-flow passes and are perfectly legal.  Run a quick phi
918        * removal on the block after the if to clean up any such phis.
919        */
920       nir_block *const next_block =
921          nir_cf_node_as_block(nir_cf_node_next(&nif->cf_node));
922       nir_opt_remove_phis_block(next_block);
923    }
924 
925    rewrite_phi_predecessor_blocks(nif, then_block, else_block, else_block,
926                                   then_block);
927 
928    /* Finally, move the else block to the then block. */
929    nir_cf_list tmp;
930    nir_cf_extract(&tmp, nir_before_cf_list(&nif->else_list),
931                         nir_after_cf_list(&nif->else_list));
932    nir_cf_reinsert(&tmp, nir_before_cf_list(&nif->then_list));
933 
934    return true;
935 }
936 
937 /* Find phi statements after an if that choose between true and false, and
938  * replace them with the if statement's condition (or an inot of it).
939  */
940 static bool
opt_if_phi_is_condition(nir_builder *b, nir_if *nif)941 opt_if_phi_is_condition(nir_builder *b, nir_if *nif)
942 {
943    /* Grab pointers to the last then/else blocks for looking in the phis. */
944    nir_block *then_block = nir_if_last_then_block(nif);
945    nir_block *else_block = nir_if_last_else_block(nif);
946    nir_ssa_def *cond = nif->condition.ssa;
947    bool progress = false;
948 
949    nir_block *after_if_block = nir_cf_node_as_block(nir_cf_node_next(&nif->cf_node));
950    nir_foreach_instr_safe(instr, after_if_block) {
951       if (instr->type != nir_instr_type_phi)
952          break;
953 
954       nir_phi_instr *phi = nir_instr_as_phi(instr);
955       if (phi->dest.ssa.bit_size != cond->bit_size ||
956           phi->dest.ssa.num_components != 1)
957          continue;
958 
959       enum opt_bool {
960          T, F, UNKNOWN
961       } then_val = UNKNOWN, else_val = UNKNOWN;
962 
963       nir_foreach_phi_src(src, phi) {
964          assert(src->pred == then_block || src->pred == else_block);
965          enum opt_bool *pred_val = src->pred == then_block ? &then_val : &else_val;
966 
967          nir_ssa_scalar val = nir_ssa_scalar_resolved(src->src.ssa, 0);
968          if (!nir_ssa_scalar_is_const(val))
969             break;
970 
971          if (nir_ssa_scalar_as_int(val) == -1)
972             *pred_val = T;
973          else if (nir_ssa_scalar_as_uint(val) == 0)
974             *pred_val = F;
975          else
976             break;
977       }
978       if (then_val == T && else_val == F) {
979          nir_ssa_def_rewrite_uses(&phi->dest.ssa, cond);
980          progress = true;
981       } else if (then_val == F && else_val == T) {
982          b->cursor = nir_before_cf_node(&nif->cf_node);
983          nir_ssa_def_rewrite_uses(&phi->dest.ssa, nir_inot(b, cond));
984          progress = true;
985       }
986    }
987 
988    return progress;
989 }
990 
991 /**
992  * This optimization tries to merge two break statements into a single break.
993  * For this purpose, it checks if both branch legs end in a break or
994  * if one branch leg ends in a break, and the other one does so after the
995  * branch.
996  *
997  * This optimization turns
998  *
999  *     loop {
1000  *        ...
1001  *        if (cond) {
1002  *           do_work_1();
1003  *           break;
1004  *        } else {
1005  *           do_work_2();
1006  *           break;
1007  *        }
1008  *     }
1009  *
1010  * into:
1011  *
1012  *     loop {
1013  *        ...
1014  *        if (cond) {
1015  *           do_work_1();
1016  *        } else {
1017  *           do_work_2();
1018  *        }
1019  *        break;
1020  *     }
1021  *
1022  * but also situations like
1023  *
1024  *     loop {
1025  *        ...
1026  *        if (cond1) {
1027  *           if (cond2) {
1028  *              do_work_1();
1029  *              break;
1030  *           } else {
1031  *              do_work_2();
1032  *           }
1033  *           do_work_3();
1034  *           break;
1035  *        } else {
1036  *           ...
1037  *        }
1038  *     }
1039  *
1040  *  into:
1041  *
1042  *     loop {
1043  *        ...
1044  *        if (cond1) {
1045  *           if (cond2) {
1046  *              do_work_1();
1047  *           } else {
1048  *              do_work_2();
1049  *              do_work_3();
1050  *           }
1051  *           break;
1052  *        } else {
1053  *           ...
1054  *        }
1055  *     }
1056  */
1057 static bool
opt_merge_breaks(nir_if *nif)1058 opt_merge_breaks(nir_if *nif)
1059 {
1060    nir_block *last_then = nir_if_last_then_block(nif);
1061    nir_block *last_else = nir_if_last_else_block(nif);
1062    bool then_break = nir_block_ends_in_break(last_then);
1063    bool else_break = nir_block_ends_in_break(last_else);
1064 
1065    /* If both branch legs end in a break, merge the break after the branch */
1066    if (then_break && else_break) {
1067       nir_block *after_if = nir_cf_node_cf_tree_next(&nif->cf_node);
1068       /* Make sure that the successor is empty.
1069        * If not we let nir_opt_dead_cf() clean it up first.
1070        */
1071       if (!is_block_empty(after_if))
1072          return false;
1073 
1074       nir_lower_phis_to_regs_block(last_then->successors[0]);
1075       nir_instr_remove_v(nir_block_last_instr(last_then));
1076       nir_instr *jump = nir_block_last_instr(last_else);
1077       nir_instr_remove_v(jump);
1078       nir_instr_insert(nir_after_block(after_if), jump);
1079       return true;
1080     }
1081 
1082    /* Single break: If there's a break after the branch and the non-breaking
1083     * side of the if falls through to it, then hoist that code after up into
1084     * the if and leave just a single break there.
1085     */
1086    if (then_break || else_break) {
1087 
1088       /* At least one branch leg must fall-through */
1089       if (nir_block_ends_in_jump(last_then) && nir_block_ends_in_jump(last_else))
1090          return false;
1091 
1092       /* Check if there is a single break after the IF */
1093       nir_cf_node *first = nir_cf_node_next(&nif->cf_node);
1094       nir_cf_node *last = first;
1095       while (!nir_cf_node_is_last(last)) {
1096          if (contains_other_jump (last, NULL))
1097             return false;
1098          last = nir_cf_node_next(last);
1099       }
1100 
1101       assert(last->type == nir_cf_node_block);
1102       if (!nir_block_ends_in_break(nir_cf_node_as_block(last)))
1103          return false;
1104 
1105       /* Hoist the code from after the IF into the falling-through branch leg */
1106       nir_opt_remove_phis_block(nir_cf_node_as_block(first));
1107       nir_block *break_block = then_break ? last_then : last_else;
1108       nir_lower_phis_to_regs_block(break_block->successors[0]);
1109 
1110       nir_cf_list tmp;
1111       nir_cf_extract(&tmp, nir_before_cf_node(first),
1112                            nir_after_block_before_jump(nir_cf_node_as_block(last)));
1113       if (then_break)
1114          nir_cf_reinsert(&tmp, nir_after_block(last_else));
1115       else
1116          nir_cf_reinsert(&tmp, nir_after_block(last_then));
1117 
1118       nir_instr_remove_v(nir_block_last_instr(break_block));
1119       return true;
1120    }
1121 
1122    return false;
1123 }
1124 
1125 /**
1126  * This optimization simplifies potential loop terminators which then allows
1127  * other passes such as opt_if_simplification() and loop unrolling to progress
1128  * further:
1129  *
1130  *     if (cond) {
1131  *        ... then block instructions ...
1132  *     } else {
1133  *         ...
1134  *        break;
1135  *     }
1136  *
1137  * into:
1138  *
1139  *     if (cond) {
1140  *     } else {
1141  *         ...
1142  *        break;
1143  *     }
1144  *     ... then block instructions ...
1145  */
1146 static bool
opt_if_loop_terminator(nir_if *nif)1147 opt_if_loop_terminator(nir_if *nif)
1148 {
1149    nir_block *break_blk = NULL;
1150    nir_block *continue_from_blk = NULL;
1151    bool continue_from_then = true;
1152 
1153    nir_block *last_then = nir_if_last_then_block(nif);
1154    nir_block *last_else = nir_if_last_else_block(nif);
1155 
1156    if (nir_block_ends_in_break(last_then)) {
1157       break_blk = last_then;
1158       continue_from_blk = last_else;
1159       continue_from_then = false;
1160    } else if (nir_block_ends_in_break(last_else)) {
1161       break_blk = last_else;
1162       continue_from_blk = last_then;
1163    }
1164 
1165    /* Continue if the if-statement contained no jumps at all */
1166    if (!break_blk)
1167       return false;
1168 
1169    /* If the continue from block is empty then return as there is nothing to
1170     * move.
1171     */
1172    nir_block *first_continue_from_blk = continue_from_then ?
1173       nir_if_first_then_block(nif) :
1174       nir_if_first_else_block(nif);
1175    if (is_block_empty(first_continue_from_blk))
1176       return false;
1177 
1178    if (nir_block_ends_in_jump(continue_from_blk))
1179       return false;
1180 
1181    /* Even though this if statement has a jump on one side, we may still have
1182     * phis afterwards.  Single-source phis can be produced by loop unrolling
1183     * or dead control-flow passes and are perfectly legal.  Run a quick phi
1184     * removal on the block after the if to clean up any such phis.
1185     */
1186    nir_opt_remove_phis_block(nir_cf_node_as_block(nir_cf_node_next(&nif->cf_node)));
1187 
1188    /* Finally, move the continue from branch after the if-statement. */
1189    nir_cf_list tmp;
1190    nir_cf_extract(&tmp, nir_before_block(first_continue_from_blk),
1191                         nir_after_block(continue_from_blk));
1192    nir_cf_reinsert(&tmp, nir_after_cf_node(&nif->cf_node));
1193 
1194    return true;
1195 }
1196 
1197 static bool
evaluate_if_condition(nir_if *nif, nir_cursor cursor, bool *value)1198 evaluate_if_condition(nir_if *nif, nir_cursor cursor, bool *value)
1199 {
1200    nir_block *use_block = nir_cursor_current_block(cursor);
1201    if (nir_block_dominates(nir_if_first_then_block(nif), use_block)) {
1202       *value = true;
1203       return true;
1204    } else if (nir_block_dominates(nir_if_first_else_block(nif), use_block)) {
1205       *value = false;
1206       return true;
1207    } else {
1208       return false;
1209    }
1210 }
1211 
1212 static nir_ssa_def *
clone_alu_and_replace_src_defs(nir_builder *b, const nir_alu_instr *alu, nir_ssa_def **src_defs)1213 clone_alu_and_replace_src_defs(nir_builder *b, const nir_alu_instr *alu,
1214                                nir_ssa_def **src_defs)
1215 {
1216    nir_alu_instr *nalu = nir_alu_instr_create(b->shader, alu->op);
1217    nalu->exact = alu->exact;
1218 
1219    nir_ssa_dest_init(&nalu->instr, &nalu->dest.dest,
1220                      alu->dest.dest.ssa.num_components,
1221                      alu->dest.dest.ssa.bit_size, NULL);
1222 
1223    nalu->dest.saturate = alu->dest.saturate;
1224    nalu->dest.write_mask = alu->dest.write_mask;
1225 
1226    for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
1227       assert(alu->src[i].src.is_ssa);
1228       nalu->src[i].src = nir_src_for_ssa(src_defs[i]);
1229       nalu->src[i].negate = alu->src[i].negate;
1230       nalu->src[i].abs = alu->src[i].abs;
1231       memcpy(nalu->src[i].swizzle, alu->src[i].swizzle,
1232              sizeof(nalu->src[i].swizzle));
1233    }
1234 
1235    nir_builder_instr_insert(b, &nalu->instr);
1236 
1237    return &nalu->dest.dest.ssa;;
1238 }
1239 
1240 /*
1241  * This propagates if condition evaluation down the chain of some alu
1242  * instructions. For example by checking the use of some of the following alu
1243  * instruction we can eventually replace ssa_107 with NIR_TRUE.
1244  *
1245  *   loop {
1246  *      block block_1:
1247  *      vec1 32 ssa_85 = load_const (0x00000002)
1248  *      vec1 32 ssa_86 = ieq ssa_48, ssa_85
1249  *      vec1 32 ssa_87 = load_const (0x00000001)
1250  *      vec1 32 ssa_88 = ieq ssa_48, ssa_87
1251  *      vec1 32 ssa_89 = ior ssa_86, ssa_88
1252  *      vec1 32 ssa_90 = ieq ssa_48, ssa_0
1253  *      vec1 32 ssa_91 = ior ssa_89, ssa_90
1254  *      if ssa_86 {
1255  *         block block_2:
1256  *             ...
1257  *            break
1258  *      } else {
1259  *            block block_3:
1260  *      }
1261  *      block block_4:
1262  *      if ssa_88 {
1263  *            block block_5:
1264  *             ...
1265  *            break
1266  *      } else {
1267  *            block block_6:
1268  *      }
1269  *      block block_7:
1270  *      if ssa_90 {
1271  *            block block_8:
1272  *             ...
1273  *            break
1274  *      } else {
1275  *            block block_9:
1276  *      }
1277  *      block block_10:
1278  *      vec1 32 ssa_107 = inot ssa_91
1279  *      if ssa_107 {
1280  *            block block_11:
1281  *            break
1282  *      } else {
1283  *            block block_12:
1284  *      }
1285  *   }
1286  */
1287 static bool
propagate_condition_eval(nir_builder *b, nir_if *nif, nir_src *use_src, nir_src *alu_use, nir_alu_instr *alu, bool is_if_condition)1288 propagate_condition_eval(nir_builder *b, nir_if *nif, nir_src *use_src,
1289                          nir_src *alu_use, nir_alu_instr *alu,
1290                          bool is_if_condition)
1291 {
1292    bool bool_value;
1293    b->cursor = nir_before_src(alu_use, is_if_condition);
1294    if (!evaluate_if_condition(nif, b->cursor, &bool_value))
1295       return false;
1296 
1297    nir_ssa_def *def[NIR_MAX_VEC_COMPONENTS] = {0};
1298    for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
1299       if (alu->src[i].src.ssa == use_src->ssa) {
1300          def[i] = nir_imm_bool(b, bool_value);
1301       } else {
1302          def[i] = alu->src[i].src.ssa;
1303       }
1304    }
1305 
1306    nir_ssa_def *nalu = clone_alu_and_replace_src_defs(b, alu, def);
1307 
1308    /* Rewrite use to use new alu instruction */
1309    nir_src new_src = nir_src_for_ssa(nalu);
1310 
1311    if (is_if_condition)
1312       nir_if_rewrite_condition(alu_use->parent_if, new_src);
1313    else
1314       nir_instr_rewrite_src(alu_use->parent_instr, alu_use, new_src);
1315 
1316    return true;
1317 }
1318 
1319 static bool
can_propagate_through_alu(nir_src *src)1320 can_propagate_through_alu(nir_src *src)
1321 {
1322    if (src->parent_instr->type != nir_instr_type_alu)
1323       return false;
1324 
1325    nir_alu_instr *alu = nir_instr_as_alu(src->parent_instr);
1326    switch (alu->op) {
1327       case nir_op_ior:
1328       case nir_op_iand:
1329       case nir_op_inot:
1330       case nir_op_b2i32:
1331          return true;
1332       case nir_op_bcsel:
1333          return src == &alu->src[0].src;
1334       default:
1335          return false;
1336    }
1337 }
1338 
1339 static bool
evaluate_condition_use(nir_builder *b, nir_if *nif, nir_src *use_src, bool is_if_condition)1340 evaluate_condition_use(nir_builder *b, nir_if *nif, nir_src *use_src,
1341                        bool is_if_condition)
1342 {
1343    bool progress = false;
1344 
1345    b->cursor = nir_before_src(use_src, is_if_condition);
1346 
1347    bool bool_value;
1348    if (evaluate_if_condition(nif, b->cursor, &bool_value)) {
1349       /* Rewrite use to use const */
1350       nir_src imm_src = nir_src_for_ssa(nir_imm_bool(b, bool_value));
1351       if (is_if_condition)
1352          nir_if_rewrite_condition(use_src->parent_if, imm_src);
1353       else
1354          nir_instr_rewrite_src(use_src->parent_instr, use_src, imm_src);
1355 
1356       progress = true;
1357    }
1358 
1359    if (!is_if_condition && can_propagate_through_alu(use_src)) {
1360       nir_alu_instr *alu = nir_instr_as_alu(use_src->parent_instr);
1361 
1362       nir_foreach_use_safe(alu_use, &alu->dest.dest.ssa) {
1363          progress |= propagate_condition_eval(b, nif, use_src, alu_use, alu,
1364                                               false);
1365       }
1366 
1367       nir_foreach_if_use_safe(alu_use, &alu->dest.dest.ssa) {
1368          progress |= propagate_condition_eval(b, nif, use_src, alu_use, alu,
1369                                               true);
1370       }
1371    }
1372 
1373    return progress;
1374 }
1375 
1376 static bool
opt_if_evaluate_condition_use(nir_builder *b, nir_if *nif)1377 opt_if_evaluate_condition_use(nir_builder *b, nir_if *nif)
1378 {
1379    bool progress = false;
1380 
1381    /* Evaluate any uses of the if condition inside the if branches */
1382    assert(nif->condition.is_ssa);
1383    nir_foreach_use_safe(use_src, nif->condition.ssa) {
1384       progress |= evaluate_condition_use(b, nif, use_src, false);
1385    }
1386 
1387    nir_foreach_if_use_safe(use_src, nif->condition.ssa) {
1388       if (use_src->parent_if != nif)
1389          progress |= evaluate_condition_use(b, nif, use_src, true);
1390    }
1391 
1392    return progress;
1393 }
1394 
1395 static bool
rewrite_comp_uses_within_if(nir_builder *b, nir_if *nif, bool invert, nir_ssa_scalar scalar, nir_ssa_scalar new_scalar)1396 rewrite_comp_uses_within_if(nir_builder *b, nir_if *nif, bool invert,
1397                             nir_ssa_scalar scalar, nir_ssa_scalar new_scalar)
1398 {
1399    bool progress = false;
1400 
1401    nir_block *first = invert ? nir_if_first_else_block(nif) : nir_if_first_then_block(nif);
1402    nir_block *last = invert ? nir_if_last_else_block(nif) : nir_if_last_then_block(nif);
1403 
1404    nir_ssa_def *new_ssa = NULL;
1405    nir_foreach_use_safe(use, scalar.def) {
1406       if (use->parent_instr->block->index < first->index ||
1407           use->parent_instr->block->index > last->index)
1408          continue;
1409 
1410       /* Only rewrite users which use only the new component. This is to avoid a
1411        * situation where copy propagation will undo the rewrite and we risk an infinite
1412        * loop.
1413        *
1414        * We could rewrite users which use a mix of the old and new components, but if
1415        * nir_src_components_read() is incomplete, then we risk the new component actually being
1416        * unused and some optimization later undoing the rewrite.
1417        */
1418       if (nir_src_components_read(use) != BITFIELD64_BIT(scalar.comp))
1419          continue;
1420 
1421       if (!new_ssa) {
1422          b->cursor = nir_before_cf_node(&nif->cf_node);
1423          new_ssa = nir_channel(b, new_scalar.def, new_scalar.comp);
1424          if (scalar.def->num_components > 1) {
1425             nir_ssa_def *vec = nir_ssa_undef(b, scalar.def->num_components, scalar.def->bit_size);
1426             new_ssa = nir_vector_insert_imm(b, vec, new_ssa, scalar.comp);
1427          }
1428       }
1429 
1430       nir_instr_rewrite_src_ssa(use->parent_instr, use, new_ssa);
1431       progress = true;
1432    }
1433 
1434    return progress;
1435 }
1436 
1437 /*
1438  * This optimization turns:
1439  *
1440  *     if (a == (b=readfirstlane(a)))
1441  *        use(a)
1442  *     if (c == (d=load_const))
1443  *        use(c)
1444  *
1445  * into:
1446  *
1447  *     if (a == (b=readfirstlane(a)))
1448  *        use(b)
1449  *     if (c == (d=load_const))
1450  *        use(d)
1451 */
1452 static bool
opt_if_rewrite_uniform_uses(nir_builder *b, nir_if *nif, nir_ssa_scalar cond, bool accept_ine)1453 opt_if_rewrite_uniform_uses(nir_builder *b, nir_if *nif, nir_ssa_scalar cond, bool accept_ine)
1454 {
1455    bool progress = false;
1456 
1457    if (!nir_ssa_scalar_is_alu(cond))
1458       return false;
1459 
1460    nir_op op = nir_ssa_scalar_alu_op(cond);
1461    if (op == nir_op_iand) {
1462       progress |= opt_if_rewrite_uniform_uses(b, nif, nir_ssa_scalar_chase_alu_src(cond, 0), false);
1463       progress |= opt_if_rewrite_uniform_uses(b, nif, nir_ssa_scalar_chase_alu_src(cond, 1), false);
1464       return progress;
1465    }
1466 
1467    if (op != nir_op_ieq && (op != nir_op_ine || !accept_ine))
1468       return false;
1469 
1470    for (unsigned i = 0; i < 2; i++) {
1471       nir_ssa_scalar src_uni = nir_ssa_scalar_chase_alu_src(cond, i);
1472       nir_ssa_scalar src_div = nir_ssa_scalar_chase_alu_src(cond, !i);
1473 
1474       if (src_uni.def->parent_instr->type == nir_instr_type_load_const && src_div.def != src_uni.def)
1475          return rewrite_comp_uses_within_if(b, nif, op == nir_op_ine, src_div, src_uni);
1476 
1477       if (src_uni.def->parent_instr->type != nir_instr_type_intrinsic)
1478          continue;
1479       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(src_uni.def->parent_instr);
1480       if (intrin->intrinsic != nir_intrinsic_read_first_invocation &&
1481           (intrin->intrinsic != nir_intrinsic_reduce || nir_intrinsic_cluster_size(intrin)))
1482          continue;
1483 
1484       nir_ssa_scalar intrin_src = {intrin->src[0].ssa, src_uni.comp};
1485       nir_ssa_scalar resolved_intrin_src = nir_ssa_scalar_resolved(intrin_src.def, intrin_src.comp);
1486 
1487       if (resolved_intrin_src.comp != src_div.comp || resolved_intrin_src.def != src_div.def)
1488          continue;
1489 
1490       progress |= rewrite_comp_uses_within_if(b, nif, op == nir_op_ine, resolved_intrin_src, src_uni);
1491       if (intrin_src.comp != resolved_intrin_src.comp || intrin_src.def != resolved_intrin_src.def)
1492          progress |= rewrite_comp_uses_within_if(b, nif, op == nir_op_ine, intrin_src, src_uni);
1493 
1494       return progress;
1495    }
1496 
1497    return false;
1498 }
1499 
1500 static void
simple_merge_if(nir_if *dest_if, nir_if *src_if, bool dest_if_then, bool src_if_then)1501 simple_merge_if(nir_if *dest_if, nir_if *src_if, bool dest_if_then,
1502                 bool src_if_then)
1503 {
1504    /* Now merge the if branch */
1505    nir_block *dest_blk = dest_if_then ? nir_if_last_then_block(dest_if)
1506                                       : nir_if_last_else_block(dest_if);
1507 
1508    struct exec_list *list = src_if_then ? &src_if->then_list
1509                                         : &src_if->else_list;
1510 
1511    nir_cf_list if_cf_list;
1512    nir_cf_extract(&if_cf_list, nir_before_cf_list(list),
1513                   nir_after_cf_list(list));
1514    nir_cf_reinsert(&if_cf_list, nir_after_block(dest_blk));
1515 }
1516 
1517 static bool
opt_if_merge(nir_if *nif)1518 opt_if_merge(nir_if *nif)
1519 {
1520    bool progress = false;
1521 
1522    nir_block *next_blk = nir_cf_node_cf_tree_next(&nif->cf_node);
1523    if (!next_blk || !nif->condition.is_ssa)
1524       return false;
1525 
1526    nir_if *next_if = nir_block_get_following_if(next_blk);
1527    if (!next_if || !next_if->condition.is_ssa)
1528       return false;
1529 
1530    /* Here we merge two consecutive ifs that have the same condition e.g:
1531     *
1532     *   if ssa_12 {
1533     *      ...
1534     *   } else {
1535     *      ...
1536     *   }
1537     *   if ssa_12 {
1538     *      ...
1539     *   } else {
1540     *      ...
1541     *   }
1542     *
1543     * Note: This only merges if-statements when the block between them is
1544     * empty. The reason we don't try to merge ifs that just have phis between
1545     * them is because this can result in increased register pressure. For
1546     * example when merging if ladders created by indirect indexing.
1547     */
1548    if (nif->condition.ssa == next_if->condition.ssa &&
1549        exec_list_is_empty(&next_blk->instr_list)) {
1550 
1551       /* This optimization isn't made to work in this case and
1552        * opt_if_evaluate_condition_use will optimize it later.
1553        */
1554       if (nir_block_ends_in_jump(nir_if_last_then_block(nif)) ||
1555           nir_block_ends_in_jump(nir_if_last_else_block(nif)))
1556          return false;
1557 
1558       simple_merge_if(nif, next_if, true, true);
1559       simple_merge_if(nif, next_if, false, false);
1560 
1561       nir_block *new_then_block = nir_if_last_then_block(nif);
1562       nir_block *new_else_block = nir_if_last_else_block(nif);
1563 
1564       nir_block *old_then_block = nir_if_last_then_block(next_if);
1565       nir_block *old_else_block = nir_if_last_else_block(next_if);
1566 
1567       /* Rewrite the predecessor block for any phis following the second
1568        * if-statement.
1569        */
1570       rewrite_phi_predecessor_blocks(next_if, old_then_block,
1571                                      old_else_block,
1572                                      new_then_block,
1573                                      new_else_block);
1574 
1575       /* Move phis after merged if to avoid them being deleted when we remove
1576        * the merged if-statement.
1577        */
1578       nir_block *after_next_if_block =
1579          nir_cf_node_as_block(nir_cf_node_next(&next_if->cf_node));
1580 
1581       nir_foreach_instr_safe(instr, after_next_if_block) {
1582          if (instr->type != nir_instr_type_phi)
1583             break;
1584 
1585          exec_node_remove(&instr->node);
1586          exec_list_push_tail(&next_blk->instr_list, &instr->node);
1587          instr->block = next_blk;
1588       }
1589 
1590       nir_cf_node_remove(&next_if->cf_node);
1591 
1592       progress = true;
1593    }
1594 
1595    return progress;
1596 }
1597 
1598 static bool
opt_if_cf_list(nir_builder *b, struct exec_list *cf_list, nir_opt_if_options options)1599 opt_if_cf_list(nir_builder *b, struct exec_list *cf_list,
1600                nir_opt_if_options options)
1601 {
1602    bool progress = false;
1603    foreach_list_typed(nir_cf_node, cf_node, node, cf_list) {
1604       switch (cf_node->type) {
1605       case nir_cf_node_block:
1606          break;
1607 
1608       case nir_cf_node_if: {
1609          nir_if *nif = nir_cf_node_as_if(cf_node);
1610          progress |= opt_if_cf_list(b, &nif->then_list,
1611                                     options);
1612          progress |= opt_if_cf_list(b, &nif->else_list,
1613                                     options);
1614          progress |= opt_if_loop_terminator(nif);
1615          progress |= opt_if_merge(nif);
1616          progress |= opt_if_simplification(b, nif);
1617          if (options & nir_opt_if_optimize_phi_true_false)
1618             progress |= opt_if_phi_is_condition(b, nif);
1619          break;
1620       }
1621 
1622       case nir_cf_node_loop: {
1623          nir_loop *loop = nir_cf_node_as_loop(cf_node);
1624          progress |= opt_if_cf_list(b, &loop->body,
1625                                     options);
1626          progress |= opt_simplify_bcsel_of_phi(b, loop);
1627          progress |= opt_if_loop_last_continue(loop,
1628                                                options & nir_opt_if_aggressive_last_continue);
1629          break;
1630       }
1631 
1632       case nir_cf_node_function:
1633          unreachable("Invalid cf type");
1634       }
1635    }
1636 
1637    return progress;
1638 }
1639 
1640 /**
1641  * Optimizations which can create registers are done after other optimizations
1642  * which require SSA.
1643  */
1644 static bool
opt_if_regs_cf_list(struct exec_list *cf_list)1645 opt_if_regs_cf_list(struct exec_list *cf_list)
1646 {
1647    bool progress = false;
1648    foreach_list_typed(nir_cf_node, cf_node, node, cf_list) {
1649       switch (cf_node->type) {
1650       case nir_cf_node_block:
1651          break;
1652 
1653       case nir_cf_node_if: {
1654          nir_if *nif = nir_cf_node_as_if(cf_node);
1655          progress |= opt_if_regs_cf_list(&nif->then_list);
1656          progress |= opt_if_regs_cf_list(&nif->else_list);
1657          if (opt_merge_breaks(nif)) {
1658             /* This optimization might move blocks
1659              * from after the NIF into the NIF */
1660             progress = true;
1661             opt_if_regs_cf_list(&nif->then_list);
1662             opt_if_regs_cf_list(&nif->else_list);
1663          }
1664          break;
1665       }
1666 
1667       case nir_cf_node_loop: {
1668          nir_loop *loop = nir_cf_node_as_loop(cf_node);
1669          progress |= opt_if_regs_cf_list(&loop->body);
1670          progress |= opt_peel_loop_initial_if(loop);
1671          break;
1672       }
1673 
1674       case nir_cf_node_function:
1675          unreachable("Invalid cf type");
1676       }
1677    }
1678 
1679    return progress;
1680 }
1681 
1682 /**
1683  * These optimisations depend on nir_metadata_block_index and therefore must
1684  * not do anything to cause the metadata to become invalid.
1685  */
1686 static bool
opt_if_safe_cf_list(nir_builder *b, struct exec_list *cf_list)1687 opt_if_safe_cf_list(nir_builder *b, struct exec_list *cf_list)
1688 {
1689    bool progress = false;
1690    foreach_list_typed(nir_cf_node, cf_node, node, cf_list) {
1691       switch (cf_node->type) {
1692       case nir_cf_node_block:
1693          break;
1694 
1695       case nir_cf_node_if: {
1696          nir_if *nif = nir_cf_node_as_if(cf_node);
1697          progress |= opt_if_safe_cf_list(b, &nif->then_list);
1698          progress |= opt_if_safe_cf_list(b, &nif->else_list);
1699          progress |= opt_if_evaluate_condition_use(b, nif);
1700          nir_ssa_scalar cond = nir_ssa_scalar_resolved(nif->condition.ssa, 0);
1701          progress |= opt_if_rewrite_uniform_uses(b, nif, cond, true);
1702          break;
1703       }
1704 
1705       case nir_cf_node_loop: {
1706          nir_loop *loop = nir_cf_node_as_loop(cf_node);
1707          progress |= opt_if_safe_cf_list(b, &loop->body);
1708          progress |= opt_split_alu_of_phi(b, loop);
1709          break;
1710       }
1711 
1712       case nir_cf_node_function:
1713          unreachable("Invalid cf type");
1714       }
1715    }
1716 
1717    return progress;
1718 }
1719 
1720 bool
nir_opt_if(nir_shader *shader, nir_opt_if_options options)1721 nir_opt_if(nir_shader *shader, nir_opt_if_options options)
1722 {
1723    bool progress = false;
1724 
1725    nir_foreach_function(function, shader) {
1726       if (function->impl == NULL)
1727          continue;
1728 
1729       nir_builder b;
1730       nir_builder_init(&b, function->impl);
1731 
1732       nir_metadata_require(function->impl, nir_metadata_block_index |
1733                            nir_metadata_dominance);
1734       progress = opt_if_safe_cf_list(&b, &function->impl->body);
1735       nir_metadata_preserve(function->impl, nir_metadata_block_index |
1736                             nir_metadata_dominance);
1737 
1738       bool preserve = true;
1739 
1740       if (opt_if_cf_list(&b, &function->impl->body, options)) {
1741          preserve = false;
1742          progress = true;
1743       }
1744 
1745       if (opt_if_regs_cf_list(&function->impl->body)) {
1746          preserve = false;
1747          progress = true;
1748 
1749          /* If that made progress, we're no longer really in SSA form.  We
1750           * need to convert registers back into SSA defs and clean up SSA defs
1751           * that don't dominate their uses.
1752           */
1753          nir_lower_regs_to_ssa_impl(function->impl);
1754       }
1755 
1756       if (preserve) {
1757          nir_metadata_preserve(function->impl, nir_metadata_none);
1758       } else {
1759          nir_metadata_preserve(function->impl, nir_metadata_all);
1760       }
1761    }
1762 
1763    return progress;
1764 }
1765