1 /****************************************************************************
2 * Copyright (C) 2014-2018 Intel Corporation. All Rights Reserved.
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:
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
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
25 * @brief llvm pass to lower meta code to x86
29 ******************************************************************************/
31 #include "jit_pch.hpp"
33 #include "JitManager.h"
35 #include "common/simdlib.hpp"
37 #include <unordered_map>
39 extern "C" void ScatterPS_256(uint8_t*, SIMD256::Integer
, SIMD256::Float
, uint8_t, uint32_t);
43 // foward declare the initializer
44 void initializeLowerX86Pass(PassRegistry
&);
67 typedef std::function
<Instruction
*(LowerX86
*, TargetArch
, TargetWidth
, CallInst
*)> EmuFunc
;
71 IntrinsicID intrin
[NUM_WIDTHS
];
75 // Map of intrinsics that haven't been moved to the new mechanism yet. If used, these get the
76 // previous behavior of mapping directly to avx/avx2 intrinsics.
77 static std::map
<std::string
, IntrinsicID
> intrinsicMap
= {
78 {"meta.intrinsic.BEXTR_32", Intrinsic::x86_bmi_bextr_32
},
79 {"meta.intrinsic.VPSHUFB", Intrinsic::x86_avx2_pshuf_b
},
80 {"meta.intrinsic.VCVTPS2PH", Intrinsic::x86_vcvtps2ph_256
},
81 {"meta.intrinsic.VPTESTC", Intrinsic::x86_avx_ptestc_256
},
82 {"meta.intrinsic.VPTESTZ", Intrinsic::x86_avx_ptestz_256
},
83 {"meta.intrinsic.VPHADDD", Intrinsic::x86_avx2_phadd_d
},
84 {"meta.intrinsic.PDEP32", Intrinsic::x86_bmi_pdep_32
},
85 {"meta.intrinsic.RDTSC", Intrinsic::x86_rdtsc
},
89 Instruction
* NO_EMU(LowerX86
* pThis
, TargetArch arch
, TargetWidth width
, CallInst
* pCallInst
);
91 VPERM_EMU(LowerX86
* pThis
, TargetArch arch
, TargetWidth width
, CallInst
* pCallInst
);
93 VGATHER_EMU(LowerX86
* pThis
, TargetArch arch
, TargetWidth width
, CallInst
* pCallInst
);
95 VSCATTER_EMU(LowerX86
* pThis
, TargetArch arch
, TargetWidth width
, CallInst
* pCallInst
);
97 VROUND_EMU(LowerX86
* pThis
, TargetArch arch
, TargetWidth width
, CallInst
* pCallInst
);
99 VHSUB_EMU(LowerX86
* pThis
, TargetArch arch
, TargetWidth width
, CallInst
* pCallInst
);
101 VCONVERT_EMU(LowerX86
* pThis
, TargetArch arch
, TargetWidth width
, CallInst
* pCallInst
);
103 Instruction
* DOUBLE_EMU(LowerX86
* pThis
,
107 Intrinsic::ID intrin
);
109 static Intrinsic::ID DOUBLE
= (Intrinsic::ID
)-1;
112 static std::map
<std::string
, X86Intrinsic
> intrinsicMap2
[] = {
116 {"meta.intrinsic.VRCPPS", {{Intrinsic::x86_avx_rcp_ps_256
, DOUBLE
}, NO_EMU
}},
117 {"meta.intrinsic.VPERMPS", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VPERM_EMU
}},
118 {"meta.intrinsic.VPERMD", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VPERM_EMU
}},
119 {"meta.intrinsic.VGATHERPD", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VGATHER_EMU
}},
120 {"meta.intrinsic.VGATHERPS", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VGATHER_EMU
}},
121 {"meta.intrinsic.VGATHERDD", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VGATHER_EMU
}},
122 {"meta.intrinsic.VSCATTERPS", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VSCATTER_EMU
}},
123 {"meta.intrinsic.VCVTPD2PS", {{Intrinsic::x86_avx_cvt_pd2_ps_256
, Intrinsic::not_intrinsic
}, NO_EMU
}},
124 {"meta.intrinsic.VROUND", {{Intrinsic::x86_avx_round_ps_256
, DOUBLE
}, NO_EMU
}},
125 {"meta.intrinsic.VHSUBPS", {{Intrinsic::x86_avx_hsub_ps_256
, DOUBLE
}, NO_EMU
}},
129 {"meta.intrinsic.VRCPPS", {{Intrinsic::x86_avx_rcp_ps_256
, DOUBLE
}, NO_EMU
}},
130 {"meta.intrinsic.VPERMPS", {{Intrinsic::x86_avx2_permps
, Intrinsic::not_intrinsic
}, VPERM_EMU
}},
131 {"meta.intrinsic.VPERMD", {{Intrinsic::x86_avx2_permd
, Intrinsic::not_intrinsic
}, VPERM_EMU
}},
132 {"meta.intrinsic.VGATHERPD", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VGATHER_EMU
}},
133 {"meta.intrinsic.VGATHERPS", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VGATHER_EMU
}},
134 {"meta.intrinsic.VGATHERDD", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VGATHER_EMU
}},
135 {"meta.intrinsic.VSCATTERPS", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VSCATTER_EMU
}},
136 {"meta.intrinsic.VCVTPD2PS", {{Intrinsic::x86_avx_cvt_pd2_ps_256
, DOUBLE
}, NO_EMU
}},
137 {"meta.intrinsic.VROUND", {{Intrinsic::x86_avx_round_ps_256
, DOUBLE
}, NO_EMU
}},
138 {"meta.intrinsic.VHSUBPS", {{Intrinsic::x86_avx_hsub_ps_256
, DOUBLE
}, NO_EMU
}},
142 {"meta.intrinsic.VRCPPS", {{Intrinsic::x86_avx512_rcp14_ps_256
, Intrinsic::x86_avx512_rcp14_ps_512
}, NO_EMU
}},
143 #if LLVM_VERSION_MAJOR < 7
144 {"meta.intrinsic.VPERMPS", {{Intrinsic::x86_avx512_mask_permvar_sf_256
, Intrinsic::x86_avx512_mask_permvar_sf_512
}, NO_EMU
}},
145 {"meta.intrinsic.VPERMD", {{Intrinsic::x86_avx512_mask_permvar_si_256
, Intrinsic::x86_avx512_mask_permvar_si_512
}, NO_EMU
}},
147 {"meta.intrinsic.VPERMPS", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VPERM_EMU
}},
148 {"meta.intrinsic.VPERMD", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VPERM_EMU
}},
150 {"meta.intrinsic.VGATHERPD", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VGATHER_EMU
}},
151 {"meta.intrinsic.VGATHERPS", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VGATHER_EMU
}},
152 {"meta.intrinsic.VGATHERDD", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VGATHER_EMU
}},
153 {"meta.intrinsic.VSCATTERPS", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VSCATTER_EMU
}},
154 #if LLVM_VERSION_MAJOR < 7
155 {"meta.intrinsic.VCVTPD2PS", {{Intrinsic::x86_avx512_mask_cvtpd2ps_256
, Intrinsic::x86_avx512_mask_cvtpd2ps_512
}, NO_EMU
}},
157 {"meta.intrinsic.VCVTPD2PS", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VCONVERT_EMU
}},
159 {"meta.intrinsic.VROUND", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VROUND_EMU
}},
160 {"meta.intrinsic.VHSUBPS", {{Intrinsic::not_intrinsic
, Intrinsic::not_intrinsic
}, VHSUB_EMU
}},
164 static uint32_t getBitWidth(VectorType
*pVTy
)
166 #if LLVM_VERSION_MAJOR >= 11
167 return pVTy
->getNumElements() * pVTy
->getElementType()->getPrimitiveSizeInBits();
169 return pVTy
->getBitWidth();
173 struct LowerX86
: public FunctionPass
175 LowerX86(Builder
* b
= nullptr) : FunctionPass(ID
), B(b
)
177 initializeLowerX86Pass(*PassRegistry::getPassRegistry());
179 // Determine target arch
180 if (JM()->mArch
.AVX512F())
184 else if (JM()->mArch
.AVX2())
188 else if (JM()->mArch
.AVX())
194 SWR_ASSERT(false, "Unsupported AVX architecture.");
198 // Setup scatter function for 256 wide
199 uint32_t curWidth
= B
->mVWidth
;
200 B
->SetTargetWidth(8);
201 std::vector
<Type
*> args
= {
202 B
->mInt8PtrTy
, // pBase
203 B
->mSimdInt32Ty
, // vIndices
204 B
->mSimdFP32Ty
, // vSrc
209 FunctionType
* pfnScatterTy
= FunctionType::get(B
->mVoidTy
, args
, false);
210 mPfnScatter256
= cast
<Function
>(
211 #if LLVM_VERSION_MAJOR >= 9
212 B
->JM()->mpCurrentModule
->getOrInsertFunction("ScatterPS_256", pfnScatterTy
).getCallee());
214 B
->JM()->mpCurrentModule
->getOrInsertFunction("ScatterPS_256", pfnScatterTy
));
216 if (sys::DynamicLibrary::SearchForAddressOfSymbol("ScatterPS_256") == nullptr)
218 sys::DynamicLibrary::AddSymbol("ScatterPS_256", (void*)&ScatterPS_256
);
221 B
->SetTargetWidth(curWidth
);
224 // Try to decipher the vector type of the instruction. This does not work properly
225 // across all intrinsics, and will have to be rethought. Probably need something
226 // similar to llvm's getDeclaration() utility to map a set of inputs to a specific typed
228 void GetRequestedWidthAndType(CallInst
* pCallInst
,
229 const StringRef intrinName
,
234 Type
* pVecTy
= pCallInst
->getType();
236 // Check for intrinsic specific types
237 // VCVTPD2PS type comes from src, not dst
238 if (intrinName
.equals("meta.intrinsic.VCVTPD2PS"))
240 Value
* pOp
= pCallInst
->getOperand(0);
242 pVecTy
= pOp
->getType();
245 if (!pVecTy
->isVectorTy())
247 for (auto& op
: pCallInst
->arg_operands())
249 if (op
.get()->getType()->isVectorTy())
251 pVecTy
= op
.get()->getType();
256 SWR_ASSERT(pVecTy
->isVectorTy(), "Couldn't determine vector size");
258 uint32_t width
= getBitWidth(cast
<VectorType
>(pVecTy
));
268 SWR_ASSERT(false, "Unhandled vector width %d", width
);
272 *pTy
= pVecTy
->getScalarType();
275 Value
* GetZeroVec(TargetWidth width
, Type
* pTy
)
277 uint32_t numElem
= 0;
287 SWR_ASSERT(false, "Unhandled vector width type %d\n", width
);
290 return ConstantVector::getNullValue(getVectorType(pTy
, numElem
));
293 Value
* GetMask(TargetWidth width
)
299 mask
= B
->C((uint8_t)-1);
302 mask
= B
->C((uint16_t)-1);
305 SWR_ASSERT(false, "Unhandled vector width type %d\n", width
);
310 // Convert <N x i1> mask to <N x i32> x86 mask
311 Value
* VectorMask(Value
* vi1Mask
)
313 #if LLVM_VERSION_MAJOR >= 11
314 uint32_t numElem
= cast
<VectorType
>(vi1Mask
->getType())->getNumElements();
316 uint32_t numElem
= vi1Mask
->getType()->getVectorNumElements();
318 return B
->S_EXT(vi1Mask
, getVectorType(B
->mInt32Ty
, numElem
));
321 Instruction
* ProcessIntrinsicAdvanced(CallInst
* pCallInst
)
323 Function
* pFunc
= pCallInst
->getCalledFunction();
326 auto& intrinsic
= intrinsicMap2
[mTarget
][pFunc
->getName().str()];
327 TargetWidth vecWidth
;
329 GetRequestedWidthAndType(pCallInst
, pFunc
->getName(), &vecWidth
, &pElemTy
);
331 // Check if there is a native intrinsic for this instruction
332 IntrinsicID id
= intrinsic
.intrin
[vecWidth
];
335 // Double pump the next smaller SIMD intrinsic
336 SWR_ASSERT(vecWidth
!= 0, "Cannot double pump smallest SIMD width.");
337 Intrinsic::ID id2
= intrinsic
.intrin
[vecWidth
- 1];
338 SWR_ASSERT(id2
!= Intrinsic::not_intrinsic
,
339 "Cannot find intrinsic to double pump.");
340 return DOUBLE_EMU(this, mTarget
, vecWidth
, pCallInst
, id2
);
342 else if (id
!= Intrinsic::not_intrinsic
)
344 Function
* pIntrin
= Intrinsic::getDeclaration(B
->JM()->mpCurrentModule
, id
);
345 SmallVector
<Value
*, 8> args
;
346 for (auto& arg
: pCallInst
->arg_operands())
348 args
.push_back(arg
.get());
351 // If AVX512, all instructions add a src operand and mask. We'll pass in 0 src and
352 // full mask for now Assuming the intrinsics are consistent and place the src
353 // operand and mask last in the argument list.
354 if (mTarget
== AVX512
)
356 if (pFunc
->getName().equals("meta.intrinsic.VCVTPD2PS"))
358 args
.push_back(GetZeroVec(W256
, pCallInst
->getType()->getScalarType()));
359 args
.push_back(GetMask(W256
));
360 // for AVX512 VCVTPD2PS, we also have to add rounding mode
361 args
.push_back(B
->C(_MM_FROUND_TO_NEAREST_INT
| _MM_FROUND_NO_EXC
));
365 args
.push_back(GetZeroVec(vecWidth
, pElemTy
));
366 args
.push_back(GetMask(vecWidth
));
370 return B
->CALLA(pIntrin
, args
);
374 // No native intrinsic, call emulation function
375 return intrinsic
.emuFunc(this, mTarget
, vecWidth
, pCallInst
);
382 Instruction
* ProcessIntrinsic(CallInst
* pCallInst
)
384 Function
* pFunc
= pCallInst
->getCalledFunction();
387 // Forward to the advanced support if found
388 if (intrinsicMap2
[mTarget
].find(pFunc
->getName().str()) != intrinsicMap2
[mTarget
].end())
390 return ProcessIntrinsicAdvanced(pCallInst
);
393 SWR_ASSERT(intrinsicMap
.find(pFunc
->getName().str()) != intrinsicMap
.end(),
394 "Unimplemented intrinsic %s.",
395 pFunc
->getName().str().c_str());
397 Intrinsic::ID x86Intrinsic
= intrinsicMap
[pFunc
->getName().str()];
398 Function
* pX86IntrinFunc
=
399 Intrinsic::getDeclaration(B
->JM()->mpCurrentModule
, x86Intrinsic
);
401 SmallVector
<Value
*, 8> args
;
402 for (auto& arg
: pCallInst
->arg_operands())
404 args
.push_back(arg
.get());
406 return B
->CALLA(pX86IntrinFunc
, args
);
409 //////////////////////////////////////////////////////////////////////////
410 /// @brief LLVM funtion pass run method.
411 /// @param f- The function we're working on with this pass.
412 virtual bool runOnFunction(Function
& F
)
414 std::vector
<Instruction
*> toRemove
;
415 std::vector
<BasicBlock
*> bbs
;
417 // Make temp copy of the basic blocks and instructions, as the intrinsic
418 // replacement code might invalidate the iterators
419 for (auto& b
: F
.getBasicBlockList())
426 std::vector
<Instruction
*> insts
;
427 for (auto& i
: BB
->getInstList())
432 for (auto* I
: insts
)
434 if (CallInst
* pCallInst
= dyn_cast
<CallInst
>(I
))
436 Function
* pFunc
= pCallInst
->getCalledFunction();
439 if (pFunc
->getName().startswith("meta.intrinsic"))
441 B
->IRB()->SetInsertPoint(I
);
442 Instruction
* pReplace
= ProcessIntrinsic(pCallInst
);
443 toRemove
.push_back(pCallInst
);
446 pCallInst
->replaceAllUsesWith(pReplace
);
454 for (auto* pInst
: toRemove
)
456 pInst
->eraseFromParent();
459 JitManager::DumpToFile(&F
, "lowerx86");
464 virtual void getAnalysisUsage(AnalysisUsage
& AU
) const {}
466 JitManager
* JM() { return B
->JM(); }
469 Function
* mPfnScatter256
;
471 static char ID
; ///< Needed by LLVM to generate ID for FunctionPass.
474 char LowerX86::ID
= 0; // LLVM uses address of ID as the actual ID.
476 FunctionPass
* createLowerX86Pass(Builder
* b
) { return new LowerX86(b
); }
478 Instruction
* NO_EMU(LowerX86
* pThis
, TargetArch arch
, TargetWidth width
, CallInst
* pCallInst
)
480 SWR_ASSERT(false, "Unimplemented intrinsic emulation.");
484 Instruction
* VPERM_EMU(LowerX86
* pThis
, TargetArch arch
, TargetWidth width
, CallInst
* pCallInst
)
486 // Only need vperm emulation for AVX
487 SWR_ASSERT(arch
== AVX
);
489 Builder
* B
= pThis
->B
;
490 auto v32A
= pCallInst
->getArgOperand(0);
491 auto vi32Index
= pCallInst
->getArgOperand(1);
494 if (isa
<Constant
>(vi32Index
))
496 // Can use llvm shuffle vector directly with constant shuffle indices
497 v32Result
= B
->VSHUFFLE(v32A
, v32A
, vi32Index
);
501 v32Result
= UndefValue::get(v32A
->getType());
502 #if LLVM_VERSION_MAJOR >= 11
503 uint32_t numElem
= cast
<VectorType
>(v32A
->getType())->getNumElements();
505 uint32_t numElem
= v32A
->getType()->getVectorNumElements();
507 for (uint32_t l
= 0; l
< numElem
; ++l
)
509 auto i32Index
= B
->VEXTRACT(vi32Index
, B
->C(l
));
510 auto val
= B
->VEXTRACT(v32A
, i32Index
);
511 v32Result
= B
->VINSERT(v32Result
, val
, B
->C(l
));
514 return cast
<Instruction
>(v32Result
);
518 VGATHER_EMU(LowerX86
* pThis
, TargetArch arch
, TargetWidth width
, CallInst
* pCallInst
)
520 Builder
* B
= pThis
->B
;
521 auto vSrc
= pCallInst
->getArgOperand(0);
522 auto pBase
= pCallInst
->getArgOperand(1);
523 auto vi32Indices
= pCallInst
->getArgOperand(2);
524 auto vi1Mask
= pCallInst
->getArgOperand(3);
525 auto i8Scale
= pCallInst
->getArgOperand(4);
527 pBase
= B
->POINTER_CAST(pBase
, PointerType::get(B
->mInt8Ty
, 0));
528 #if LLVM_VERSION_MAJOR >= 11
529 VectorType
* pVectorType
= cast
<VectorType
>(vSrc
->getType());
530 uint32_t numElem
= pVectorType
->getNumElements();
531 auto srcTy
= pVectorType
->getElementType();
533 uint32_t numElem
= vSrc
->getType()->getVectorNumElements();
534 auto srcTy
= vSrc
->getType()->getVectorElementType();
536 auto i32Scale
= B
->Z_EXT(i8Scale
, B
->mInt32Ty
);
538 Value
* v32Gather
= nullptr;
541 // Full emulation for AVX
542 // Store source on stack to provide a valid address to load from inactive lanes
543 auto pStack
= B
->STACKSAVE();
544 auto pTmp
= B
->ALLOCA(vSrc
->getType());
545 B
->STORE(vSrc
, pTmp
);
547 v32Gather
= UndefValue::get(vSrc
->getType());
548 #if LLVM_VERSION_MAJOR > 10
549 auto vi32Scale
= ConstantVector::getSplat(ElementCount::get(numElem
, false), cast
<ConstantInt
>(i32Scale
));
551 auto vi32Scale
= ConstantVector::getSplat(numElem
, cast
<ConstantInt
>(i32Scale
));
553 auto vi32Offsets
= B
->MUL(vi32Indices
, vi32Scale
);
555 for (uint32_t i
= 0; i
< numElem
; ++i
)
557 auto i32Offset
= B
->VEXTRACT(vi32Offsets
, B
->C(i
));
558 auto pLoadAddress
= B
->GEP(pBase
, i32Offset
);
559 pLoadAddress
= B
->BITCAST(pLoadAddress
, PointerType::get(srcTy
, 0));
560 auto pMaskedLoadAddress
= B
->GEP(pTmp
, {0, i
});
561 auto i1Mask
= B
->VEXTRACT(vi1Mask
, B
->C(i
));
562 auto pValidAddress
= B
->SELECT(i1Mask
, pLoadAddress
, pMaskedLoadAddress
);
563 auto val
= B
->LOAD(pValidAddress
);
564 v32Gather
= B
->VINSERT(v32Gather
, val
, B
->C(i
));
567 B
->STACKRESTORE(pStack
);
569 else if (arch
== AVX2
|| (arch
== AVX512
&& width
== W256
))
571 Function
* pX86IntrinFunc
= nullptr;
572 if (srcTy
== B
->mFP32Ty
)
574 pX86IntrinFunc
= Intrinsic::getDeclaration(B
->JM()->mpCurrentModule
,
575 Intrinsic::x86_avx2_gather_d_ps_256
);
577 else if (srcTy
== B
->mInt32Ty
)
579 pX86IntrinFunc
= Intrinsic::getDeclaration(B
->JM()->mpCurrentModule
,
580 Intrinsic::x86_avx2_gather_d_d_256
);
582 else if (srcTy
== B
->mDoubleTy
)
584 pX86IntrinFunc
= Intrinsic::getDeclaration(B
->JM()->mpCurrentModule
,
585 Intrinsic::x86_avx2_gather_d_q_256
);
589 SWR_ASSERT(false, "Unsupported vector element type for gather.");
594 auto v32Mask
= B
->BITCAST(pThis
->VectorMask(vi1Mask
), vSrc
->getType());
595 v32Gather
= B
->CALL(pX86IntrinFunc
, {vSrc
, pBase
, vi32Indices
, v32Mask
, i8Scale
});
597 else if (width
== W512
)
599 // Double pump 4-wide for 64bit elements
600 #if LLVM_VERSION_MAJOR >= 11
601 if (cast
<VectorType
>(vSrc
->getType())->getElementType() == B
->mDoubleTy
)
603 if (vSrc
->getType()->getVectorElementType() == B
->mDoubleTy
)
606 auto v64Mask
= pThis
->VectorMask(vi1Mask
);
607 #if LLVM_VERSION_MAJOR >= 11
608 uint32_t numElem
= cast
<VectorType
>(v64Mask
->getType())->getNumElements();
610 uint32_t numElem
= v64Mask
->getType()->getVectorNumElements();
612 v64Mask
= B
->S_EXT(v64Mask
, getVectorType(B
->mInt64Ty
, numElem
));
613 v64Mask
= B
->BITCAST(v64Mask
, vSrc
->getType());
615 Value
* src0
= B
->VSHUFFLE(vSrc
, vSrc
, B
->C({0, 1, 2, 3}));
616 Value
* src1
= B
->VSHUFFLE(vSrc
, vSrc
, B
->C({4, 5, 6, 7}));
618 Value
* indices0
= B
->VSHUFFLE(vi32Indices
, vi32Indices
, B
->C({0, 1, 2, 3}));
619 Value
* indices1
= B
->VSHUFFLE(vi32Indices
, vi32Indices
, B
->C({4, 5, 6, 7}));
621 Value
* mask0
= B
->VSHUFFLE(v64Mask
, v64Mask
, B
->C({0, 1, 2, 3}));
622 Value
* mask1
= B
->VSHUFFLE(v64Mask
, v64Mask
, B
->C({4, 5, 6, 7}));
624 #if LLVM_VERSION_MAJOR >= 11
625 uint32_t numElemSrc0
= cast
<VectorType
>(src0
->getType())->getNumElements();
626 uint32_t numElemMask0
= cast
<VectorType
>(mask0
->getType())->getNumElements();
627 uint32_t numElemSrc1
= cast
<VectorType
>(src1
->getType())->getNumElements();
628 uint32_t numElemMask1
= cast
<VectorType
>(mask1
->getType())->getNumElements();
630 uint32_t numElemSrc0
= src0
->getType()->getVectorNumElements();
631 uint32_t numElemMask0
= mask0
->getType()->getVectorNumElements();
632 uint32_t numElemSrc1
= src1
->getType()->getVectorNumElements();
633 uint32_t numElemMask1
= mask1
->getType()->getVectorNumElements();
635 src0
= B
->BITCAST(src0
, getVectorType(B
->mInt64Ty
, numElemSrc0
));
636 mask0
= B
->BITCAST(mask0
, getVectorType(B
->mInt64Ty
, numElemMask0
));
638 B
->CALL(pX86IntrinFunc
, {src0
, pBase
, indices0
, mask0
, i8Scale
});
639 src1
= B
->BITCAST(src1
, getVectorType(B
->mInt64Ty
, numElemSrc1
));
640 mask1
= B
->BITCAST(mask1
, getVectorType(B
->mInt64Ty
, numElemMask1
));
642 B
->CALL(pX86IntrinFunc
, {src1
, pBase
, indices1
, mask1
, i8Scale
});
643 v32Gather
= B
->VSHUFFLE(gather0
, gather1
, B
->C({0, 1, 2, 3, 4, 5, 6, 7}));
644 v32Gather
= B
->BITCAST(v32Gather
, vSrc
->getType());
648 // Double pump 8-wide for 32bit elements
649 auto v32Mask
= pThis
->VectorMask(vi1Mask
);
650 v32Mask
= B
->BITCAST(v32Mask
, vSrc
->getType());
651 Value
* src0
= B
->EXTRACT_16(vSrc
, 0);
652 Value
* src1
= B
->EXTRACT_16(vSrc
, 1);
654 Value
* indices0
= B
->EXTRACT_16(vi32Indices
, 0);
655 Value
* indices1
= B
->EXTRACT_16(vi32Indices
, 1);
657 Value
* mask0
= B
->EXTRACT_16(v32Mask
, 0);
658 Value
* mask1
= B
->EXTRACT_16(v32Mask
, 1);
661 B
->CALL(pX86IntrinFunc
, {src0
, pBase
, indices0
, mask0
, i8Scale
});
663 B
->CALL(pX86IntrinFunc
, {src1
, pBase
, indices1
, mask1
, i8Scale
});
665 v32Gather
= B
->JOIN_16(gather0
, gather1
);
669 else if (arch
== AVX512
)
671 Value
* iMask
= nullptr;
672 Function
* pX86IntrinFunc
= nullptr;
673 if (srcTy
== B
->mFP32Ty
)
675 pX86IntrinFunc
= Intrinsic::getDeclaration(B
->JM()->mpCurrentModule
,
676 Intrinsic::x86_avx512_gather_dps_512
);
677 iMask
= B
->BITCAST(vi1Mask
, B
->mInt16Ty
);
679 else if (srcTy
== B
->mInt32Ty
)
681 pX86IntrinFunc
= Intrinsic::getDeclaration(B
->JM()->mpCurrentModule
,
682 Intrinsic::x86_avx512_gather_dpi_512
);
683 iMask
= B
->BITCAST(vi1Mask
, B
->mInt16Ty
);
685 else if (srcTy
== B
->mDoubleTy
)
687 pX86IntrinFunc
= Intrinsic::getDeclaration(B
->JM()->mpCurrentModule
,
688 Intrinsic::x86_avx512_gather_dpd_512
);
689 iMask
= B
->BITCAST(vi1Mask
, B
->mInt8Ty
);
693 SWR_ASSERT(false, "Unsupported vector element type for gather.");
696 auto i32Scale
= B
->Z_EXT(i8Scale
, B
->mInt32Ty
);
697 v32Gather
= B
->CALL(pX86IntrinFunc
, {vSrc
, pBase
, vi32Indices
, iMask
, i32Scale
});
700 return cast
<Instruction
>(v32Gather
);
703 VSCATTER_EMU(LowerX86
* pThis
, TargetArch arch
, TargetWidth width
, CallInst
* pCallInst
)
705 Builder
* B
= pThis
->B
;
706 auto pBase
= pCallInst
->getArgOperand(0);
707 auto vi1Mask
= pCallInst
->getArgOperand(1);
708 auto vi32Indices
= pCallInst
->getArgOperand(2);
709 auto v32Src
= pCallInst
->getArgOperand(3);
710 auto i32Scale
= pCallInst
->getArgOperand(4);
714 // Call into C function to do the scatter. This has significantly better compile perf
715 // compared to jitting scatter loops for every scatter
718 auto mask
= B
->BITCAST(vi1Mask
, B
->mInt8Ty
);
719 B
->CALL(pThis
->mPfnScatter256
, {pBase
, vi32Indices
, v32Src
, mask
, i32Scale
});
723 // Need to break up 512 wide scatter to two 256 wide
724 auto maskLo
= B
->VSHUFFLE(vi1Mask
, vi1Mask
, B
->C({0, 1, 2, 3, 4, 5, 6, 7}));
726 B
->VSHUFFLE(vi32Indices
, vi32Indices
, B
->C({0, 1, 2, 3, 4, 5, 6, 7}));
727 auto srcLo
= B
->VSHUFFLE(v32Src
, v32Src
, B
->C({0, 1, 2, 3, 4, 5, 6, 7}));
729 auto mask
= B
->BITCAST(maskLo
, B
->mInt8Ty
);
730 B
->CALL(pThis
->mPfnScatter256
, {pBase
, indicesLo
, srcLo
, mask
, i32Scale
});
732 auto maskHi
= B
->VSHUFFLE(vi1Mask
, vi1Mask
, B
->C({8, 9, 10, 11, 12, 13, 14, 15}));
734 B
->VSHUFFLE(vi32Indices
, vi32Indices
, B
->C({8, 9, 10, 11, 12, 13, 14, 15}));
735 auto srcHi
= B
->VSHUFFLE(v32Src
, v32Src
, B
->C({8, 9, 10, 11, 12, 13, 14, 15}));
737 mask
= B
->BITCAST(maskHi
, B
->mInt8Ty
);
738 B
->CALL(pThis
->mPfnScatter256
, {pBase
, indicesHi
, srcHi
, mask
, i32Scale
});
744 Function
* pX86IntrinFunc
;
747 // No direct intrinsic supported in llvm to scatter 8 elem with 32bit indices, but we
748 // can use the scatter of 8 elements with 64bit indices
749 pX86IntrinFunc
= Intrinsic::getDeclaration(B
->JM()->mpCurrentModule
,
750 Intrinsic::x86_avx512_scatter_qps_512
);
752 auto vi32IndicesExt
= B
->Z_EXT(vi32Indices
, B
->mSimdInt64Ty
);
753 iMask
= B
->BITCAST(vi1Mask
, B
->mInt8Ty
);
754 B
->CALL(pX86IntrinFunc
, {pBase
, iMask
, vi32IndicesExt
, v32Src
, i32Scale
});
756 else if (width
== W512
)
758 pX86IntrinFunc
= Intrinsic::getDeclaration(B
->JM()->mpCurrentModule
,
759 Intrinsic::x86_avx512_scatter_dps_512
);
760 iMask
= B
->BITCAST(vi1Mask
, B
->mInt16Ty
);
761 B
->CALL(pX86IntrinFunc
, {pBase
, iMask
, vi32Indices
, v32Src
, i32Scale
});
766 // No support for vroundps in avx512 (it is available in kncni), so emulate with avx
769 VROUND_EMU(LowerX86
* pThis
, TargetArch arch
, TargetWidth width
, CallInst
* pCallInst
)
771 SWR_ASSERT(arch
== AVX512
);
774 auto vf32Src
= pCallInst
->getOperand(0);
776 auto i8Round
= pCallInst
->getOperand(1);
779 Intrinsic::getDeclaration(B
->JM()->mpCurrentModule
, Intrinsic::x86_avx_round_ps_256
);
783 return cast
<Instruction
>(B
->CALL2(pfnFunc
, vf32Src
, i8Round
));
785 else if (width
== W512
)
787 auto v8f32SrcLo
= B
->EXTRACT_16(vf32Src
, 0);
788 auto v8f32SrcHi
= B
->EXTRACT_16(vf32Src
, 1);
790 auto v8f32ResLo
= B
->CALL2(pfnFunc
, v8f32SrcLo
, i8Round
);
791 auto v8f32ResHi
= B
->CALL2(pfnFunc
, v8f32SrcHi
, i8Round
);
793 return cast
<Instruction
>(B
->JOIN_16(v8f32ResLo
, v8f32ResHi
));
797 SWR_ASSERT(false, "Unimplemented vector width.");
804 VCONVERT_EMU(LowerX86
* pThis
, TargetArch arch
, TargetWidth width
, CallInst
* pCallInst
)
806 SWR_ASSERT(arch
== AVX512
);
809 auto vf32Src
= pCallInst
->getOperand(0);
813 auto vf32SrcRound
= Intrinsic::getDeclaration(B
->JM()->mpCurrentModule
,
814 Intrinsic::x86_avx_round_ps_256
);
815 return cast
<Instruction
>(B
->FP_TRUNC(vf32SrcRound
, B
->mFP32Ty
));
817 else if (width
== W512
)
819 // 512 can use intrinsic
820 auto pfnFunc
= Intrinsic::getDeclaration(B
->JM()->mpCurrentModule
,
821 Intrinsic::x86_avx512_mask_cvtpd2ps_512
);
822 return cast
<Instruction
>(B
->CALL(pfnFunc
, vf32Src
));
826 SWR_ASSERT(false, "Unimplemented vector width.");
832 // No support for hsub in AVX512
833 Instruction
* VHSUB_EMU(LowerX86
* pThis
, TargetArch arch
, TargetWidth width
, CallInst
* pCallInst
)
835 SWR_ASSERT(arch
== AVX512
);
838 auto src0
= pCallInst
->getOperand(0);
839 auto src1
= pCallInst
->getOperand(1);
841 // 256b hsub can just use avx intrinsic
844 auto pX86IntrinFunc
=
845 Intrinsic::getDeclaration(B
->JM()->mpCurrentModule
, Intrinsic::x86_avx_hsub_ps_256
);
846 return cast
<Instruction
>(B
->CALL2(pX86IntrinFunc
, src0
, src1
));
848 else if (width
== W512
)
850 // 512b hsub can be accomplished with shuf/sub combo
851 auto minuend
= B
->VSHUFFLE(src0
, src1
, B
->C({0, 2, 8, 10, 4, 6, 12, 14}));
852 auto subtrahend
= B
->VSHUFFLE(src0
, src1
, B
->C({1, 3, 9, 11, 5, 7, 13, 15}));
853 return cast
<Instruction
>(B
->SUB(minuend
, subtrahend
));
857 SWR_ASSERT(false, "Unimplemented vector width.");
862 // Double pump input using Intrin template arg. This blindly extracts lower and upper 256 from
863 // each vector argument and calls the 256 wide intrinsic, then merges the results to 512 wide
864 Instruction
* DOUBLE_EMU(LowerX86
* pThis
,
868 Intrinsic::ID intrin
)
871 SWR_ASSERT(width
== W512
);
873 Function
* pX86IntrinFunc
= Intrinsic::getDeclaration(B
->JM()->mpCurrentModule
, intrin
);
874 for (uint32_t i
= 0; i
< 2; ++i
)
876 SmallVector
<Value
*, 8> args
;
877 for (auto& arg
: pCallInst
->arg_operands())
879 auto argType
= arg
.get()->getType();
880 if (argType
->isVectorTy())
882 #if LLVM_VERSION_MAJOR >= 11
883 uint32_t vecWidth
= cast
<VectorType
>(argType
)->getNumElements();
884 auto elemTy
= cast
<VectorType
>(argType
)->getElementType();
886 uint32_t vecWidth
= argType
->getVectorNumElements();
887 auto elemTy
= argType
->getVectorElementType();
889 Value
* lanes
= B
->CInc
<int>(i
* vecWidth
/ 2, vecWidth
/ 2);
890 Value
* argToPush
= B
->VSHUFFLE(arg
.get(), B
->VUNDEF(elemTy
, vecWidth
), lanes
);
891 args
.push_back(argToPush
);
895 args
.push_back(arg
.get());
898 result
[i
] = B
->CALLA(pX86IntrinFunc
, args
);
901 if (result
[0]->getType()->isVectorTy())
903 assert(result
[1]->getType()->isVectorTy());
904 #if LLVM_VERSION_MAJOR >= 11
905 vecWidth
= cast
<VectorType
>(result
[0]->getType())->getNumElements() +
906 cast
<VectorType
>(result
[1]->getType())->getNumElements();
908 vecWidth
= result
[0]->getType()->getVectorNumElements() +
909 result
[1]->getType()->getVectorNumElements();
916 Value
* lanes
= B
->CInc
<int>(0, vecWidth
);
917 return cast
<Instruction
>(B
->VSHUFFLE(result
[0], result
[1], lanes
));
920 } // namespace SwrJit
922 using namespace SwrJit
;
924 INITIALIZE_PASS_BEGIN(LowerX86
, "LowerX86", "LowerX86", false, false)
925 INITIALIZE_PASS_END(LowerX86
, "LowerX86", "LowerX86", false, false)