File: InstructionHoistingOptimization.cpp

package info (click to toggle)
intel-graphics-compiler2 2.16.0-2
  • links: PTS, VCS
  • area: main
  • in suites: sid
  • size: 106,644 kB
  • sloc: cpp: 805,640; lisp: 287,672; ansic: 16,414; python: 3,952; yacc: 2,588; lex: 1,666; pascal: 313; sh: 186; makefile: 35
file content (318 lines) | stat: -rw-r--r-- 13,230 bytes parent folder | download
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
/*========================== begin_copyright_notice ============================

Copyright (C) 2021-2025 Intel Corporation

SPDX-License-Identifier: MIT

============================= end_copyright_notice ===========================*/

#include "InstructionHoistingOptimization.hpp"
#include "Compiler/CISACodeGen/helper.h"
#include "Compiler/IGCPassSupport.h"
#include "GenISAIntrinsics/GenIntrinsicInst.h"
#include "IGC/Compiler/CodeGenPublic.h"
#include "LLVM3DBuilder/BuiltinsFrontend.hpp"
#include "Probe/Assertion.h"
#include "common/LLVMWarningsPush.hpp"
#include "llvm/PassInfo.h"
#include "llvm/PassRegistry.h"
#include <llvm/Transforms/Utils/BasicBlockUtils.h>
#include "common/LLVMWarningsPop.hpp"

using namespace llvm;
namespace IGC {
// clang-format off
/////////////////////////////////////////////////////////////////////////////////////////////
/// @brief Hoisting instructions (sampler) to earlier place in code after loop
/// unroll The reason on doing this hoisting is because for sampler after LICM,
/// only instructions with loop-invariants can be hoisted, and sampler in the
/// loop is like:
/// 19:                                               ; preds = %281, %0
///    %20 = phi i32[0, %0], [%265, %281]
///    %21 = phi i32[0, %0], [%202, %281]
///    %22 = phi i32[0, %0], [%139, %281]
///    %23 = phi i32[0, %0], [%76,  %281]
///    %24 = phi i32[0, %0], [%283, %281]
///    %25 = phi i32[0, %0], [%284, %281]
///    %26 = phi i32[0, %0], [%285, %281]
///    %27 = add nsw i32 %26, %17
///    %28 = uitofp i32 %27 to float
///    %29 = uitofp i32 %18 to float
///    %30 = fadd fast float %28, 5.000000e-01
///    %31 = fadd fast float %29, 5.000000e-01
///    %32 = call fast <4 x float> @llvm.genx.GenISA.ldrawvector.indexed.v4f32.p2555904v4f32(<4 x float> addrspace(2555904)* %"b0,1", i32 832, i32 4, i1 false)
///    %33 = extractelement <4 x float> %32, i64 0
///    %34 = extractelement <4 x float> %32, i64 1
///    %35 = fmul fast float %30, %33
///    %36 = fmul fast float %34, %31
///    %37 = call fast <4 x float> @llvm.genx.GenISA.sampleLptr.v4f32.f32.p2621446__2D_DIM_Resource.p2621446__2D_DIM_Resource.p655360v4f32(float 0.000000e+00, float %35, float %36, float 0.000000e+00, float 0.000000e+00, %__2D_DIM_Resource addrspace(2621446)* undef, %__2D_DIM_Resource addrspace(2621446)* %"t0,12", <4 x float> addrspace(655360)* null, i32 0, i32 0, i32 0)
/// In this loop, only ldrawvector can be hoisted, and sampleLptr has operands
/// %36, %36 have input dependency that LICM cannot do hoisting. So, after LICM
/// and loop unroll, only one sampleLptr is hoisted (offset 0 in phi), we can do
/// a custom pass to handle the rest sampleLptr hoist as below:
/// ===> Hoistable instruction case (%71 sampler below):
/// ......
///    %36 = call fast <4 x float> @llvm.genx.GenISA.sampleLptr.v4f32.f32.p2621446__2D_DIM_Resource.p2621446__2D_DIM_Resource.p655360v4f32(float 0.000000e+00, float %35, float %20, float 0.000000e+00, float 0.000000e+00, %__2D_DIM_Resource addrspace(2621446)* undef, %__2D_DIM_Resource addrspace(2621446)* %"t0,12", <4 x float> addrspace(655360)* null, i32 0, i32 0, i32 0)
///    %37 = extractelement <4 x float> %36, i64 3
/// ......
/// NodeBlock27:                                      ; preds = %NewDefault, %64, %66, %68
///    %70 = phi i32[%69, %68], [%67, %66], [%65, %64], [%63, %NewDefault]
///    %71 = call fast <4 x float> @llvm.genx.GenISA.sampleLptr.v4f32.f32.p2621446__2D_DIM_Resource.p2621446__2D_DIM_Resource.p655360v4f32(float 0.000000e+00, float %35, float %24, float 0.000000e+00, float 0.000000e+00, %__2D_DIM_Resource addrspace(2621446)* undef, %__2D_DIM_Resource addrspace(2621446)* %"t0,12", <4 x float> addrspace(655360)* null, i32 0, i32 0, i32 0)
///    %72 = extractelement <4 x float> %71, i64 3
/// ......
/// ===> Proposed optimization (the above %71 sampler is hoisted up to be %37 below):
/// ......
///    %36 = call fast <4 x float> @llvm.genx.GenISA.sampleLptr.v4f32.f32.p2621446__2D_DIM_Resource.p2621446__2D_DIM_Resource.p655360v4f32(float 0.000000e+00, float %35, float %20, float 0.000000e+00, float 0.000000e+00, %__2D_DIM_Resource addrspace(2621446)* undef, %__2D_DIM_Resource addrspace(2621446)* %"t0,12", <4 x float> addrspace(655360)* null, i32 0, i32 0, i32 0)
///    %37 = call fast <4 x float> @llvm.genx.GenISA.sampleLptr.v4f32.f32.p2621446__2D_DIM_Resource.p2621446__2D_DIM_Resource.p655360v4f32(float 0.000000e+00, float %35, float %24, float 0.000000e+00, float 0.000000e+00, %__2D_DIM_Resource addrspace(2621446)* undef, %__2D_DIM_Resource addrspace(2621446)* %"t0,12", <4 x float> addrspace(655360)* null, i32 0, i32 0, i32 0)
///......
///    %64 = extractelement <4 x float> %36, i64 3
///......
/// NodeBlock27:                                      ; preds = %NewDefault, %91, %93, %95
///    %97 = phi i32[%96, %95], [%94, %93], [%92, %91], [%90, %NewDefault]
///    %98 = extractelement <4 x float> %37, i64 3
///......
// clang-format on
class InstructionHoistingOptimization : public llvm::FunctionPass,
                                        public llvm::InstVisitor<InstructionHoistingOptimization> {
public:
  static char ID; ///< ID used by the llvm PassManager (the value is not important)

  InstructionHoistingOptimization();

  ////////////////////////////////////////////////////////////////////////
  virtual bool runOnFunction(llvm::Function &F);

  ////////////////////////////////////////////////////////////////////////
  virtual void getAnalysisUsage(llvm::AnalysisUsage &AU) const;

  ////////////////////////////////////////////////////////////////////////
  void visitCallInst(llvm::CallInst &CI);

private:
  ////////////////////////////////////////////////////////////////////////
  bool ProcessFunction(llvm::Function &F);

  ////////////////////////////////////////////////////////////////////////
  void InvalidateMembers();

  ////////////////////////////////////////////////////////////////////////
  void CollectHoistableInstructions();

  ////////////////////////////////////////////////////////////////////////
  bool ProcessHoistableInstructions();

  ////////////////////////////////////////////////////////////////////////
  CodeGenContext *m_pCGCtxt = nullptr;
  DominatorTree *m_pDT = nullptr;

  std::vector<llvm::Instruction *> m_SamplerInstructions;
  std::vector<llvm::Instruction *> m_HoistableInstructions;

  llvm::Instruction *m_InsertHoistBack = nullptr;
  llvm::BasicBlock *m_HoistBasicBlock = nullptr;

  inline bool traceOperandInHoistBB(Value *operand, BasicBlock *hoistBB) {
    if (Instruction *inst = dyn_cast<Instruction>(operand)) {
      for (auto &op : inst->operands()) {
        if (isPHINode(op) || !isHoistableInstruction(op, hoistBB)) {
          return false;
        }
      }
    }
    return true;
  }
  inline bool isPHINode(Value *op) { return isa<PHINode>(op); }
  inline bool isHoistableInstruction(Value *op, BasicBlock *hoistBB) {
    if (auto *I = dyn_cast<Instruction>(op)) {
      auto *currBB = I->getParent();
      if (isa<LoadInst>(I) || isa<LdRawIntrinsic>(I) || isa<SampleIntrinsic>(I)) {
        return false;
      }
      if (currBB != hoistBB) {
        if (m_pDT->dominates(hoistBB, currBB)) {
          if (!traceOperandInHoistBB(op, hoistBB)) {
            return false;
          }
          InsertHoistableInstruction(I);
        } else {
          return false;
        }
      } else if (m_InsertHoistBack->comesBefore(I)) {
        return false;
      }
    }
    return true;
  }

  inline void InsertHoistableInstruction(Instruction *hoistInstr) {
    auto isDuplicate = [](std::vector<Instruction *> hoistedInstr, Instruction *newInstr) -> bool {
      auto it = std::find(hoistedInstr.begin(), hoistedInstr.end(), newInstr);
      return it != hoistedInstr.end();
    };

    if (!isDuplicate(m_HoistableInstructions, hoistInstr)) {
      m_HoistableInstructions.push_back(hoistInstr);
    }
  }
};

char InstructionHoistingOptimization::ID = 0;

////////////////////////////////////////////////////////////////////////////
InstructionHoistingOptimization::InstructionHoistingOptimization() : llvm::FunctionPass(ID) {
  initializeInstructionHoistingOptimizationPass(*PassRegistry::getPassRegistry());
}

////////////////////////////////////////////////////////////////////////
bool InstructionHoistingOptimization::runOnFunction(llvm::Function &F) {
  m_pCGCtxt = getAnalysis<CodeGenContextWrapper>().getCodeGenContext();
  m_pDT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();

  // should be no loop, all loop is unrolled
  if (!getAnalysis<LoopInfoWrapperPass>().getLoopInfo().empty()) {
    return false;
  }

  InvalidateMembers();

  return ProcessFunction(F);
}

////////////////////////////////////////////////////////////////////////
void InstructionHoistingOptimization::visitCallInst(llvm::CallInst &CI) {
  if (llvm::GenIntrinsicInst *pIntr = llvm::dyn_cast<llvm::GenIntrinsicInst>(&CI)) {
    if (auto *SI = dyn_cast<SampleIntrinsic>(pIntr)) {
      m_SamplerInstructions.push_back(SI);
    }
  }
}

////////////////////////////////////////////////////////////////////////
void InstructionHoistingOptimization::CollectHoistableInstructions() {
  if (m_SamplerInstructions.empty())
    return;

  // use hash.getAsmHash() to debug specific shader

  // To make the hoisted instruction up as early as possible, check the first
  // BB, and check each sample operand and determined the latest location to as
  // hoist point. For the rest BBs, to simplify the process to sample hoist,
  // only pick the hoist point in first BB with sample, and use it to determine
  // the rest sample hoistable choice.
  for (Instruction *I : m_SamplerInstructions) {
    for (auto &Op : I->operands()) {
      if (Instruction *Inst = dyn_cast<Instruction>(Op)) {
        // pick the first op (inst) as the initial hoist point
        if (!m_InsertHoistBack) {
          m_InsertHoistBack = Inst;
          continue;
        }

        if (m_InsertHoistBack->getParent() == Inst->getParent()) {
          if (m_InsertHoistBack->comesBefore(Inst)) {
            m_InsertHoistBack = Inst;
          }
        }
      }
    }
  }

  if (!m_InsertHoistBack)
    return;

  m_HoistBasicBlock = m_InsertHoistBack->getParent();

  for (auto *pSI : m_SamplerInstructions) {
    bool canInstrHoisted = true;

    for (auto &Op : pSI->operands()) {
      if (isa<PHINode>(Op)) {
        canInstrHoisted = false;
        break;
      }

      if (Instruction *I = dyn_cast<Instruction>(Op)) {
        if (isa<LoadInst>(I) || isa<LdRawIntrinsic>(I) || isa<SampleIntrinsic>(I)) {
          canInstrHoisted = false;
          break;
        }

        if (m_HoistBasicBlock == I->getParent()) {
          // If m_HoistBasicBlock comes before any operand of pSI,
          // this pSI won't be hoisted in this BB
          if (m_InsertHoistBack->comesBefore(I)) {
            canInstrHoisted = false;
            break;
          }
        } else {
          // check if op can be traced back to hoistInBB
          if (traceOperandInHoistBB(Op, m_HoistBasicBlock)) {
            InsertHoistableInstruction(I);
          } else {
            canInstrHoisted = false;
            break;
          }
        }
      }
    }

    if (canInstrHoisted) {
      InsertHoistableInstruction(pSI);
    }
  }
}

////////////////////////////////////////////////////////////////////////
bool InstructionHoistingOptimization::ProcessHoistableInstructions() {
  if (m_HoistableInstructions.empty())
    return false;

  auto *hoist_point = m_InsertHoistBack;

  for (auto *I : m_HoistableInstructions) {
    I->moveAfter(hoist_point);
    hoist_point = I;
  }

  return true;
}

////////////////////////////////////////////////////////////////////////
bool InstructionHoistingOptimization::ProcessFunction(llvm::Function &F) {
  visit(F);

  CollectHoistableInstructions();

  return ProcessHoistableInstructions();
}

////////////////////////////////////////////////////////////////////////
void InstructionHoistingOptimization::getAnalysisUsage(llvm::AnalysisUsage &AU) const {
  AU.addRequired<CodeGenContextWrapper>();
  AU.addRequired<DominatorTreeWrapperPass>();
  AU.addRequired<LoopInfoWrapperPass>();
}

////////////////////////////////////////////////////////////////////////
void InstructionHoistingOptimization::InvalidateMembers() {
  m_SamplerInstructions.clear();
  m_HoistableInstructions.clear();
}

////////////////////////////////////////////////////////////////////////
llvm::Pass *createInstructionHoistingOptimization() { return new InstructionHoistingOptimization(); }

} // namespace IGC

using namespace llvm;
using namespace IGC;

#define PASS_FLAG "igc-instruction-hoisting-optimization"
#define PASS_DESCRIPTION "InstructionHoistingOptimization"
#define PASS_CFG_ONLY false
#define PASS_ANALYSIS false
IGC_INITIALIZE_PASS_BEGIN(InstructionHoistingOptimization, PASS_FLAG, PASS_DESCRIPTION, PASS_CFG_ONLY, PASS_ANALYSIS)
IGC_INITIALIZE_PASS_DEPENDENCY(CodeGenContextWrapper)
IGC_INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
IGC_INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
IGC_INITIALIZE_PASS_END(InstructionHoistingOptimization, PASS_FLAG, PASS_DESCRIPTION, PASS_CFG_ONLY, PASS_ANALYSIS)