• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2018 Valve 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 
25 #include "aco_ir.h"
26 
27 #include <map>
28 #include <unordered_map>
29 #include <vector>
30 
31 /*
32  * Implements the algorithm for dominator-tree value numbering
33  * from "Value Numbering" by Briggs, Cooper, and Simpson.
34  */
35 
36 namespace aco {
37 namespace {
38 
39 inline uint32_t
murmur_32_scramble(uint32_t h,uint32_t k)40 murmur_32_scramble(uint32_t h, uint32_t k)
41 {
42    k *= 0xcc9e2d51;
43    k = (k << 15) | (k >> 17);
44    h ^= k * 0x1b873593;
45    h = (h << 13) | (h >> 19);
46    h = h * 5 + 0xe6546b64;
47    return h;
48 }
49 
50 template <typename T>
51 uint32_t
hash_murmur_32(Instruction * instr)52 hash_murmur_32(Instruction* instr)
53 {
54    uint32_t hash = uint32_t(instr->format) << 16 | uint32_t(instr->opcode);
55 
56    for (const Operand& op : instr->operands)
57       hash = murmur_32_scramble(hash, op.constantValue());
58 
59    /* skip format, opcode and pass_flags */
60    for (unsigned i = 2; i < (sizeof(T) >> 2); i++) {
61       uint32_t u;
62       /* Accesses it though a byte array, so doesn't violate the strict aliasing rule */
63       memcpy(&u, reinterpret_cast<uint8_t*>(instr) + i * 4, 4);
64       hash = murmur_32_scramble(hash, u);
65    }
66 
67    /* Finalize. */
68    uint32_t len = instr->operands.size() + instr->definitions.size() + sizeof(T);
69    hash ^= len;
70    hash ^= hash >> 16;
71    hash *= 0x85ebca6b;
72    hash ^= hash >> 13;
73    hash *= 0xc2b2ae35;
74    hash ^= hash >> 16;
75    return hash;
76 }
77 
78 struct InstrHash {
79    /* This hash function uses the Murmur3 algorithm written by Austin Appleby
80     * https://github.com/aappleby/smhasher/blob/master/src/MurmurHash3.cpp
81     *
82     * In order to calculate the expression set, only the right-hand-side of an
83     * instruction is used for the hash, i.e. everything except the definitions.
84     */
operator ()aco::__anon9e7ccdea0111::InstrHash85    std::size_t operator()(Instruction* instr) const
86    {
87       if (instr->isVOP3())
88          return hash_murmur_32<VOP3_instruction>(instr);
89 
90       if (instr->isDPP())
91          return hash_murmur_32<DPP_instruction>(instr);
92 
93       if (instr->isSDWA())
94          return hash_murmur_32<SDWA_instruction>(instr);
95 
96       switch (instr->format) {
97       case Format::SMEM: return hash_murmur_32<SMEM_instruction>(instr);
98       case Format::VINTRP: return hash_murmur_32<Interp_instruction>(instr);
99       case Format::DS: return hash_murmur_32<DS_instruction>(instr);
100       case Format::SOPP: return hash_murmur_32<SOPP_instruction>(instr);
101       case Format::SOPK: return hash_murmur_32<SOPK_instruction>(instr);
102       case Format::EXP: return hash_murmur_32<Export_instruction>(instr);
103       case Format::MUBUF: return hash_murmur_32<MUBUF_instruction>(instr);
104       case Format::MIMG: return hash_murmur_32<MIMG_instruction>(instr);
105       case Format::MTBUF: return hash_murmur_32<MTBUF_instruction>(instr);
106       case Format::FLAT: return hash_murmur_32<FLAT_instruction>(instr);
107       case Format::PSEUDO_BRANCH: return hash_murmur_32<Pseudo_branch_instruction>(instr);
108       case Format::PSEUDO_REDUCTION: return hash_murmur_32<Pseudo_reduction_instruction>(instr);
109       default: return hash_murmur_32<Instruction>(instr);
110       }
111    }
112 };
113 
114 struct InstrPred {
operator ()aco::__anon9e7ccdea0111::InstrPred115    bool operator()(Instruction* a, Instruction* b) const
116    {
117       if (a->format != b->format)
118          return false;
119       if (a->opcode != b->opcode)
120          return false;
121       if (a->operands.size() != b->operands.size() ||
122           a->definitions.size() != b->definitions.size())
123          return false; /* possible with pseudo-instructions */
124       for (unsigned i = 0; i < a->operands.size(); i++) {
125          if (a->operands[i].isConstant()) {
126             if (!b->operands[i].isConstant())
127                return false;
128             if (a->operands[i].constantValue() != b->operands[i].constantValue())
129                return false;
130          } else if (a->operands[i].isTemp()) {
131             if (!b->operands[i].isTemp())
132                return false;
133             if (a->operands[i].tempId() != b->operands[i].tempId())
134                return false;
135          } else if (a->operands[i].isUndefined() ^ b->operands[i].isUndefined())
136             return false;
137          if (a->operands[i].isFixed()) {
138             if (!b->operands[i].isFixed())
139                return false;
140             if (a->operands[i].physReg() != b->operands[i].physReg())
141                return false;
142             if (a->operands[i].physReg() == exec && a->pass_flags != b->pass_flags)
143                return false;
144          }
145       }
146       for (unsigned i = 0; i < a->definitions.size(); i++) {
147          if (a->definitions[i].isTemp()) {
148             if (!b->definitions[i].isTemp())
149                return false;
150             if (a->definitions[i].regClass() != b->definitions[i].regClass())
151                return false;
152          }
153          if (a->definitions[i].isFixed()) {
154             if (!b->definitions[i].isFixed())
155                return false;
156             if (a->definitions[i].physReg() != b->definitions[i].physReg())
157                return false;
158             if (a->definitions[i].physReg() == exec)
159                return false;
160          }
161       }
162 
163       if (a->opcode == aco_opcode::v_readfirstlane_b32)
164          return a->pass_flags == b->pass_flags;
165 
166       if (a->isVOP3()) {
167          VOP3_instruction& a3 = a->vop3();
168          VOP3_instruction& b3 = b->vop3();
169          for (unsigned i = 0; i < 3; i++) {
170             if (a3.abs[i] != b3.abs[i] || a3.neg[i] != b3.neg[i])
171                return false;
172          }
173          return a3.clamp == b3.clamp && a3.omod == b3.omod && a3.opsel == b3.opsel;
174       }
175       if (a->isDPP()) {
176          DPP_instruction& aDPP = a->dpp();
177          DPP_instruction& bDPP = b->dpp();
178          return aDPP.pass_flags == bDPP.pass_flags && aDPP.dpp_ctrl == bDPP.dpp_ctrl &&
179                 aDPP.bank_mask == bDPP.bank_mask && aDPP.row_mask == bDPP.row_mask &&
180                 aDPP.bound_ctrl == bDPP.bound_ctrl && aDPP.abs[0] == bDPP.abs[0] &&
181                 aDPP.abs[1] == bDPP.abs[1] && aDPP.neg[0] == bDPP.neg[0] &&
182                 aDPP.neg[1] == bDPP.neg[1];
183       }
184       if (a->isSDWA()) {
185          SDWA_instruction& aSDWA = a->sdwa();
186          SDWA_instruction& bSDWA = b->sdwa();
187          return aSDWA.sel[0] == bSDWA.sel[0] && aSDWA.sel[1] == bSDWA.sel[1] &&
188                 aSDWA.dst_sel == bSDWA.dst_sel && aSDWA.abs[0] == bSDWA.abs[0] &&
189                 aSDWA.abs[1] == bSDWA.abs[1] && aSDWA.neg[0] == bSDWA.neg[0] &&
190                 aSDWA.neg[1] == bSDWA.neg[1] && aSDWA.clamp == bSDWA.clamp &&
191                 aSDWA.omod == bSDWA.omod;
192       }
193 
194       switch (a->format) {
195       case Format::SOPK: {
196          if (a->opcode == aco_opcode::s_getreg_b32)
197             return false;
198          SOPK_instruction& aK = a->sopk();
199          SOPK_instruction& bK = b->sopk();
200          return aK.imm == bK.imm;
201       }
202       case Format::SMEM: {
203          SMEM_instruction& aS = a->smem();
204          SMEM_instruction& bS = b->smem();
205          /* isel shouldn't be creating situations where this assertion fails */
206          assert(aS.prevent_overflow == bS.prevent_overflow);
207          return aS.sync == bS.sync && aS.glc == bS.glc && aS.dlc == bS.dlc && aS.nv == bS.nv &&
208                 aS.disable_wqm == bS.disable_wqm && aS.prevent_overflow == bS.prevent_overflow;
209       }
210       case Format::VINTRP: {
211          Interp_instruction& aI = a->vintrp();
212          Interp_instruction& bI = b->vintrp();
213          if (aI.attribute != bI.attribute)
214             return false;
215          if (aI.component != bI.component)
216             return false;
217          return true;
218       }
219       case Format::VOP3P: {
220          VOP3P_instruction& a3P = a->vop3p();
221          VOP3P_instruction& b3P = b->vop3p();
222          for (unsigned i = 0; i < 3; i++) {
223             if (a3P.neg_lo[i] != b3P.neg_lo[i] || a3P.neg_hi[i] != b3P.neg_hi[i])
224                return false;
225          }
226          return a3P.opsel_lo == b3P.opsel_lo && a3P.opsel_hi == b3P.opsel_hi &&
227                 a3P.clamp == b3P.clamp;
228       }
229       case Format::PSEUDO_REDUCTION: {
230          Pseudo_reduction_instruction& aR = a->reduction();
231          Pseudo_reduction_instruction& bR = b->reduction();
232          return aR.pass_flags == bR.pass_flags && aR.reduce_op == bR.reduce_op &&
233                 aR.cluster_size == bR.cluster_size;
234       }
235       case Format::DS: {
236          assert(a->opcode == aco_opcode::ds_bpermute_b32 ||
237                 a->opcode == aco_opcode::ds_permute_b32 || a->opcode == aco_opcode::ds_swizzle_b32);
238          DS_instruction& aD = a->ds();
239          DS_instruction& bD = b->ds();
240          return aD.sync == bD.sync && aD.pass_flags == bD.pass_flags && aD.gds == bD.gds &&
241                 aD.offset0 == bD.offset0 && aD.offset1 == bD.offset1;
242       }
243       case Format::MTBUF: {
244          MTBUF_instruction& aM = a->mtbuf();
245          MTBUF_instruction& bM = b->mtbuf();
246          return aM.sync == bM.sync && aM.dfmt == bM.dfmt && aM.nfmt == bM.nfmt &&
247                 aM.offset == bM.offset && aM.offen == bM.offen && aM.idxen == bM.idxen &&
248                 aM.glc == bM.glc && aM.dlc == bM.dlc && aM.slc == bM.slc && aM.tfe == bM.tfe &&
249                 aM.disable_wqm == bM.disable_wqm;
250       }
251       case Format::MUBUF: {
252          MUBUF_instruction& aM = a->mubuf();
253          MUBUF_instruction& bM = b->mubuf();
254          return aM.sync == bM.sync && aM.offset == bM.offset && aM.offen == bM.offen &&
255                 aM.idxen == bM.idxen && aM.glc == bM.glc && aM.dlc == bM.dlc && aM.slc == bM.slc &&
256                 aM.tfe == bM.tfe && aM.lds == bM.lds && aM.disable_wqm == bM.disable_wqm;
257       }
258       case Format::MIMG: {
259          MIMG_instruction& aM = a->mimg();
260          MIMG_instruction& bM = b->mimg();
261          return aM.sync == bM.sync && aM.dmask == bM.dmask && aM.unrm == bM.unrm &&
262                 aM.glc == bM.glc && aM.slc == bM.slc && aM.tfe == bM.tfe && aM.da == bM.da &&
263                 aM.lwe == bM.lwe && aM.r128 == bM.r128 && aM.a16 == bM.a16 && aM.d16 == bM.d16 &&
264                 aM.disable_wqm == bM.disable_wqm;
265       }
266       case Format::FLAT:
267       case Format::GLOBAL:
268       case Format::SCRATCH:
269       case Format::EXP:
270       case Format::SOPP:
271       case Format::PSEUDO_BRANCH:
272       case Format::PSEUDO_BARRIER: assert(false);
273       default: return true;
274       }
275    }
276 };
277 
278 using expr_set = std::unordered_map<Instruction*, uint32_t, InstrHash, InstrPred>;
279 
280 struct vn_ctx {
281    Program* program;
282    expr_set expr_values;
283    std::map<uint32_t, Temp> renames;
284 
285    /* The exec id should be the same on the same level of control flow depth.
286     * Together with the check for dominator relations, it is safe to assume
287     * that the same exec_id also means the same execution mask.
288     * Discards increment the exec_id, so that it won't return to the previous value.
289     */
290    uint32_t exec_id = 1;
291 
vn_ctxaco::__anon9e7ccdea0111::vn_ctx292    vn_ctx(Program* program_) : program(program_)
293    {
294       static_assert(sizeof(Temp) == 4, "Temp must fit in 32bits");
295       unsigned size = 0;
296       for (Block& block : program->blocks)
297          size += block.instructions.size();
298       expr_values.reserve(size);
299    }
300 };
301 
302 /* dominates() returns true if the parent block dominates the child block and
303  * if the parent block is part of the same loop or has a smaller loop nest depth.
304  */
305 bool
dominates(vn_ctx & ctx,uint32_t parent,uint32_t child)306 dominates(vn_ctx& ctx, uint32_t parent, uint32_t child)
307 {
308    unsigned parent_loop_nest_depth = ctx.program->blocks[parent].loop_nest_depth;
309    while (parent < child && parent_loop_nest_depth <= ctx.program->blocks[child].loop_nest_depth)
310       child = ctx.program->blocks[child].logical_idom;
311 
312    return parent == child;
313 }
314 
315 /** Returns whether this instruction can safely be removed
316  *  and replaced by an equal expression.
317  *  This is in particular true for ALU instructions and
318  *  read-only memory instructions.
319  *
320  *  Note that expr_set must not be used with instructions
321  *  which cannot be eliminated.
322  */
323 bool
can_eliminate(aco_ptr<Instruction> & instr)324 can_eliminate(aco_ptr<Instruction>& instr)
325 {
326    switch (instr->format) {
327    case Format::FLAT:
328    case Format::GLOBAL:
329    case Format::SCRATCH:
330    case Format::EXP:
331    case Format::SOPP:
332    case Format::PSEUDO_BRANCH:
333    case Format::PSEUDO_BARRIER: return false;
334    case Format::DS:
335       return instr->opcode == aco_opcode::ds_bpermute_b32 ||
336              instr->opcode == aco_opcode::ds_permute_b32 ||
337              instr->opcode == aco_opcode::ds_swizzle_b32;
338    case Format::SMEM:
339    case Format::MUBUF:
340    case Format::MIMG:
341    case Format::MTBUF:
342       if (!get_sync_info(instr.get()).can_reorder())
343          return false;
344       break;
345    default: break;
346    }
347 
348    if (instr->definitions.empty() || instr->opcode == aco_opcode::p_phi ||
349        instr->opcode == aco_opcode::p_linear_phi || instr->definitions[0].isNoCSE())
350       return false;
351 
352    return true;
353 }
354 
355 void
process_block(vn_ctx & ctx,Block & block)356 process_block(vn_ctx& ctx, Block& block)
357 {
358    std::vector<aco_ptr<Instruction>> new_instructions;
359    new_instructions.reserve(block.instructions.size());
360 
361    for (aco_ptr<Instruction>& instr : block.instructions) {
362       /* first, rename operands */
363       for (Operand& op : instr->operands) {
364          if (!op.isTemp())
365             continue;
366          auto it = ctx.renames.find(op.tempId());
367          if (it != ctx.renames.end())
368             op.setTemp(it->second);
369       }
370 
371       if (instr->opcode == aco_opcode::p_discard_if ||
372           instr->opcode == aco_opcode::p_demote_to_helper)
373          ctx.exec_id++;
374 
375       if (!can_eliminate(instr)) {
376          new_instructions.emplace_back(std::move(instr));
377          continue;
378       }
379 
380       /* simple copy-propagation through renaming */
381       bool copy_instr =
382          instr->opcode == aco_opcode::p_parallelcopy ||
383          (instr->opcode == aco_opcode::p_create_vector && instr->operands.size() == 1);
384       if (copy_instr && !instr->definitions[0].isFixed() && instr->operands[0].isTemp() &&
385           instr->operands[0].regClass() == instr->definitions[0].regClass()) {
386          ctx.renames[instr->definitions[0].tempId()] = instr->operands[0].getTemp();
387          continue;
388       }
389 
390       instr->pass_flags = ctx.exec_id;
391       std::pair<expr_set::iterator, bool> res = ctx.expr_values.emplace(instr.get(), block.index);
392 
393       /* if there was already an expression with the same value number */
394       if (!res.second) {
395          Instruction* orig_instr = res.first->first;
396          assert(instr->definitions.size() == orig_instr->definitions.size());
397          /* check if the original instruction dominates the current one */
398          if (dominates(ctx, res.first->second, block.index) &&
399              ctx.program->blocks[res.first->second].fp_mode.canReplace(block.fp_mode)) {
400             for (unsigned i = 0; i < instr->definitions.size(); i++) {
401                assert(instr->definitions[i].regClass() == orig_instr->definitions[i].regClass());
402                assert(instr->definitions[i].isTemp());
403                ctx.renames[instr->definitions[i].tempId()] = orig_instr->definitions[i].getTemp();
404                if (instr->definitions[i].isPrecise())
405                   orig_instr->definitions[i].setPrecise(true);
406                /* SPIR_V spec says that an instruction marked with NUW wrapping
407                 * around is undefined behaviour, so we can break additions in
408                 * other contexts.
409                 */
410                if (instr->definitions[i].isNUW())
411                   orig_instr->definitions[i].setNUW(true);
412             }
413          } else {
414             ctx.expr_values.erase(res.first);
415             ctx.expr_values.emplace(instr.get(), block.index);
416             new_instructions.emplace_back(std::move(instr));
417          }
418       } else {
419          new_instructions.emplace_back(std::move(instr));
420       }
421    }
422 
423    block.instructions = std::move(new_instructions);
424 }
425 
426 void
rename_phi_operands(Block & block,std::map<uint32_t,Temp> & renames)427 rename_phi_operands(Block& block, std::map<uint32_t, Temp>& renames)
428 {
429    for (aco_ptr<Instruction>& phi : block.instructions) {
430       if (phi->opcode != aco_opcode::p_phi && phi->opcode != aco_opcode::p_linear_phi)
431          break;
432 
433       for (Operand& op : phi->operands) {
434          if (!op.isTemp())
435             continue;
436          auto it = renames.find(op.tempId());
437          if (it != renames.end())
438             op.setTemp(it->second);
439       }
440    }
441 }
442 } /* end namespace */
443 
444 void
value_numbering(Program * program)445 value_numbering(Program* program)
446 {
447    vn_ctx ctx(program);
448    std::vector<unsigned> loop_headers;
449 
450    for (Block& block : program->blocks) {
451       assert(ctx.exec_id > 0);
452       /* decrement exec_id when leaving nested control flow */
453       if (block.kind & block_kind_loop_header)
454          loop_headers.push_back(block.index);
455       if (block.kind & block_kind_merge) {
456          ctx.exec_id--;
457       } else if (block.kind & block_kind_loop_exit) {
458          ctx.exec_id -= program->blocks[loop_headers.back()].linear_preds.size();
459          ctx.exec_id -= block.linear_preds.size();
460          loop_headers.pop_back();
461       }
462 
463       if (block.logical_idom != -1)
464          process_block(ctx, block);
465       else
466          rename_phi_operands(block, ctx.renames);
467 
468       /* increment exec_id when entering nested control flow */
469       if (block.kind & block_kind_branch || block.kind & block_kind_loop_preheader ||
470           block.kind & block_kind_break || block.kind & block_kind_continue ||
471           block.kind & block_kind_discard)
472          ctx.exec_id++;
473       else if (block.kind & block_kind_continue_or_break)
474          ctx.exec_id += 2;
475    }
476 
477    /* rename loop header phi operands */
478    for (Block& block : program->blocks) {
479       if (block.kind & block_kind_loop_header)
480          rename_phi_operands(block, ctx.renames);
481    }
482 }
483 
484 } // namespace aco
485