1 /****************************************************************************
2 * Copyright (C) 2014-2015 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
23 * @file builder_misc.cpp
25 * @brief Implementation for miscellaneous builder functions
29 ******************************************************************************/
31 #include "common/rdtsc_buckets.h"
37 void __cdecl
CallPrint(const char* fmt
, ...);
39 //////////////////////////////////////////////////////////////////////////
40 /// @brief Convert an IEEE 754 32-bit single precision float to an
41 /// 16 bit float with 5 exponent bits and a variable
42 /// number of mantissa bits.
43 /// @param val - 32-bit float
44 /// @todo Maybe move this outside of this file into a header?
45 static uint16_t ConvertFloat32ToFloat16(float val
)
47 uint32_t sign
, exp
, mant
;
50 // Extract the sign, exponent, and mantissa
51 uint32_t uf
= *(uint32_t*)&val
;
52 sign
= (uf
& 0x80000000) >> 31;
53 exp
= (uf
& 0x7F800000) >> 23;
54 mant
= uf
& 0x007FFFFF;
56 // Check for out of range
61 sign
= 1; // set the sign bit for NANs
63 else if (std::isinf(val
))
68 else if (exp
> (0x70 + 0x1E)) // Too big to represent -> max representable value
73 else if ((exp
<= 0x70) && (exp
>= 0x66)) // It's a denorm
76 for (; exp
<= 0x70; mant
>>= 1, exp
++)
81 else if (exp
< 0x66) // Too small to represent -> Zero
88 // Saves bits that will be shifted off for rounding
89 roundBits
= mant
& 0x1FFFu
;
90 // convert exponent and mantissa to 16 bit format
94 // Essentially RTZ, but round up if off by only 1 lsb
95 if (roundBits
== 0x1FFFu
)
99 if ((mant
& 0xC00u
) != 0)
101 // make sure only the needed bits are used
106 uint32_t tmpVal
= (sign
<< 15) | (exp
<< 10) | mant
;
107 return (uint16_t)tmpVal
;
110 //////////////////////////////////////////////////////////////////////////
111 /// @brief Convert an IEEE 754 16-bit float to an 32-bit single precision
113 /// @param val - 16-bit float
114 /// @todo Maybe move this outside of this file into a header?
115 static float ConvertFloat16ToFloat32(uint32_t val
)
118 if ((val
& 0x7fff) == 0)
120 result
= ((uint32_t)(val
& 0x8000)) << 16;
122 else if ((val
& 0x7c00) == 0x7c00)
124 result
= ((val
& 0x3ff) == 0) ? 0x7f800000 : 0x7fc00000;
125 result
|= ((uint32_t)val
& 0x8000) << 16;
129 uint32_t sign
= (val
& 0x8000) << 16;
130 uint32_t mant
= (val
& 0x3ff) << 13;
131 uint32_t exp
= (val
>> 10) & 0x1f;
132 if ((exp
== 0) && (mant
!= 0)) // Adjust exponent and mantissa for denormals
135 while (mant
< (0x400 << 13))
140 mant
&= (0x3ff << 13);
142 exp
= ((exp
- 15 + 127) & 0xff) << 23;
143 result
= sign
| exp
| mant
;
146 return *(float*)&result
;
149 Constant
*Builder::C(bool i
)
151 return ConstantInt::get(IRB()->getInt1Ty(), (i
? 1 : 0));
154 Constant
*Builder::C(char i
)
156 return ConstantInt::get(IRB()->getInt8Ty(), i
);
159 Constant
*Builder::C(uint8_t i
)
161 return ConstantInt::get(IRB()->getInt8Ty(), i
);
164 Constant
*Builder::C(int i
)
166 return ConstantInt::get(IRB()->getInt32Ty(), i
);
169 Constant
*Builder::C(int64_t i
)
171 return ConstantInt::get(IRB()->getInt64Ty(), i
);
174 Constant
*Builder::C(uint16_t i
)
176 return ConstantInt::get(mInt16Ty
,i
);
179 Constant
*Builder::C(uint32_t i
)
181 return ConstantInt::get(IRB()->getInt32Ty(), i
);
184 Constant
*Builder::C(float i
)
186 return ConstantFP::get(IRB()->getFloatTy(), i
);
189 Constant
*Builder::PRED(bool pred
)
191 return ConstantInt::get(IRB()->getInt1Ty(), (pred
? 1 : 0));
194 Value
*Builder::VIMMED1(int i
)
196 return ConstantVector::getSplat(mVWidth
, cast
<ConstantInt
>(C(i
)));
199 Value
*Builder::VIMMED1_16(int i
)
201 return ConstantVector::getSplat(mVWidth16
, cast
<ConstantInt
>(C(i
)));
204 Value
*Builder::VIMMED1(uint32_t i
)
206 return ConstantVector::getSplat(mVWidth
, cast
<ConstantInt
>(C(i
)));
209 Value
*Builder::VIMMED1_16(uint32_t i
)
211 return ConstantVector::getSplat(mVWidth16
, cast
<ConstantInt
>(C(i
)));
214 Value
*Builder::VIMMED1(float i
)
216 return ConstantVector::getSplat(mVWidth
, cast
<ConstantFP
>(C(i
)));
219 Value
*Builder::VIMMED1_16(float i
)
221 return ConstantVector::getSplat(mVWidth16
, cast
<ConstantFP
>(C(i
)));
224 Value
*Builder::VIMMED1(bool i
)
226 return ConstantVector::getSplat(mVWidth
, cast
<ConstantInt
>(C(i
)));
229 Value
*Builder::VIMMED1_16(bool i
)
231 return ConstantVector::getSplat(mVWidth16
, cast
<ConstantInt
>(C(i
)));
234 Value
*Builder::VUNDEF_IPTR()
236 return UndefValue::get(VectorType::get(mInt32PtrTy
,mVWidth
));
239 Value
*Builder::VUNDEF(Type
* t
)
241 return UndefValue::get(VectorType::get(t
, mVWidth
));
244 Value
*Builder::VUNDEF_I()
246 return UndefValue::get(VectorType::get(mInt32Ty
, mVWidth
));
249 Value
*Builder::VUNDEF_I_16()
251 return UndefValue::get(VectorType::get(mInt32Ty
, mVWidth16
));
254 Value
*Builder::VUNDEF_F()
256 return UndefValue::get(VectorType::get(mFP32Ty
, mVWidth
));
259 Value
*Builder::VUNDEF_F_16()
261 return UndefValue::get(VectorType::get(mFP32Ty
, mVWidth16
));
264 Value
*Builder::VUNDEF(Type
*ty
, uint32_t size
)
266 return UndefValue::get(VectorType::get(ty
, size
));
269 Value
*Builder::VBROADCAST(Value
*src
)
271 // check if src is already a vector
272 if (src
->getType()->isVectorTy())
277 return VECTOR_SPLAT(mVWidth
, src
);
280 Value
*Builder::VBROADCAST_16(Value
*src
)
282 // check if src is already a vector
283 if (src
->getType()->isVectorTy())
288 return VECTOR_SPLAT(mVWidth16
, src
);
291 uint32_t Builder::IMMED(Value
* v
)
293 SWR_ASSERT(isa
<ConstantInt
>(v
));
294 ConstantInt
*pValConst
= cast
<ConstantInt
>(v
);
295 return pValConst
->getZExtValue();
298 int32_t Builder::S_IMMED(Value
* v
)
300 SWR_ASSERT(isa
<ConstantInt
>(v
));
301 ConstantInt
*pValConst
= cast
<ConstantInt
>(v
);
302 return pValConst
->getSExtValue();
305 Value
*Builder::GEP(Value
* ptr
, const std::initializer_list
<Value
*> &indexList
)
307 std::vector
<Value
*> indices
;
308 for (auto i
: indexList
)
309 indices
.push_back(i
);
310 return GEPA(ptr
, indices
);
313 Value
*Builder::GEP(Value
* ptr
, const std::initializer_list
<uint32_t> &indexList
)
315 std::vector
<Value
*> indices
;
316 for (auto i
: indexList
)
317 indices
.push_back(C(i
));
318 return GEPA(ptr
, indices
);
321 Value
*Builder::IN_BOUNDS_GEP(Value
* ptr
, const std::initializer_list
<Value
*> &indexList
)
323 std::vector
<Value
*> indices
;
324 for (auto i
: indexList
)
325 indices
.push_back(i
);
326 return IN_BOUNDS_GEP(ptr
, indices
);
329 Value
*Builder::IN_BOUNDS_GEP(Value
* ptr
, const std::initializer_list
<uint32_t> &indexList
)
331 std::vector
<Value
*> indices
;
332 for (auto i
: indexList
)
333 indices
.push_back(C(i
));
334 return IN_BOUNDS_GEP(ptr
, indices
);
337 LoadInst
*Builder::LOAD(Value
*basePtr
, const std::initializer_list
<uint32_t> &indices
, const llvm::Twine
& name
)
339 std::vector
<Value
*> valIndices
;
340 for (auto i
: indices
)
341 valIndices
.push_back(C(i
));
342 return LOAD(GEPA(basePtr
, valIndices
), name
);
345 LoadInst
*Builder::LOADV(Value
*basePtr
, const std::initializer_list
<Value
*> &indices
, const llvm::Twine
& name
)
347 std::vector
<Value
*> valIndices
;
348 for (auto i
: indices
)
349 valIndices
.push_back(i
);
350 return LOAD(GEPA(basePtr
, valIndices
), name
);
353 StoreInst
*Builder::STORE(Value
*val
, Value
*basePtr
, const std::initializer_list
<uint32_t> &indices
)
355 std::vector
<Value
*> valIndices
;
356 for (auto i
: indices
)
357 valIndices
.push_back(C(i
));
358 return STORE(val
, GEPA(basePtr
, valIndices
));
361 StoreInst
*Builder::STOREV(Value
*val
, Value
*basePtr
, const std::initializer_list
<Value
*> &indices
)
363 std::vector
<Value
*> valIndices
;
364 for (auto i
: indices
)
365 valIndices
.push_back(i
);
366 return STORE(val
, GEPA(basePtr
, valIndices
));
369 CallInst
*Builder::CALL(Value
*Callee
, const std::initializer_list
<Value
*> &argsList
)
371 std::vector
<Value
*> args
;
372 for (auto arg
: argsList
)
374 return CALLA(Callee
, args
);
377 CallInst
*Builder::CALL(Value
*Callee
, Value
* arg
)
379 std::vector
<Value
*> args
;
381 return CALLA(Callee
, args
);
384 CallInst
*Builder::CALL2(Value
*Callee
, Value
* arg1
, Value
* arg2
)
386 std::vector
<Value
*> args
;
387 args
.push_back(arg1
);
388 args
.push_back(arg2
);
389 return CALLA(Callee
, args
);
392 CallInst
*Builder::CALL3(Value
*Callee
, Value
* arg1
, Value
* arg2
, Value
* arg3
)
394 std::vector
<Value
*> args
;
395 args
.push_back(arg1
);
396 args
.push_back(arg2
);
397 args
.push_back(arg3
);
398 return CALLA(Callee
, args
);
401 //////////////////////////////////////////////////////////////////////////
402 Value
*Builder::DEBUGTRAP()
404 Function
*func
= Intrinsic::getDeclaration(JM()->mpCurrentModule
, Intrinsic::debugtrap
);
408 Value
*Builder::VRCP(Value
*va
)
410 return FDIV(VIMMED1(1.0f
), va
); // 1 / a
413 Value
*Builder::VPLANEPS(Value
* vA
, Value
* vB
, Value
* vC
, Value
* &vX
, Value
* &vY
)
415 Value
* vOut
= FMADDPS(vA
, vX
, vC
);
416 vOut
= FMADDPS(vB
, vY
, vOut
);
420 //////////////////////////////////////////////////////////////////////////
421 /// @brief Generate an i32 masked load operation in LLVM IR. If not
422 /// supported on the underlying platform, emulate it with float masked load
423 /// @param src - base address pointer for the load
424 /// @param vMask - SIMD wide mask that controls whether to access memory load 0
425 Value
*Builder::MASKLOADD(Value
* src
,Value
* mask
)
428 // use avx2 gather instruction is available
429 if(JM()->mArch
.AVX2())
431 Function
*func
= Intrinsic::getDeclaration(JM()->mpCurrentModule
, Intrinsic::x86_avx2_maskload_d_256
);
432 vResult
= CALL(func
,{src
,mask
});
436 // maskload intrinsic expects integer mask operand in llvm >= 3.8
437 #if (LLVM_VERSION_MAJOR > 3) || (LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 8)
438 mask
= BITCAST(mask
,VectorType::get(mInt32Ty
,mVWidth
));
440 mask
= BITCAST(mask
,VectorType::get(mFP32Ty
,mVWidth
));
442 Function
*func
= Intrinsic::getDeclaration(JM()->mpCurrentModule
,Intrinsic::x86_avx_maskload_ps_256
);
443 vResult
= BITCAST(CALL(func
,{src
,mask
}), VectorType::get(mInt32Ty
,mVWidth
));
448 //////////////////////////////////////////////////////////////////////////
449 /// @brief insert a JIT call to CallPrint
450 /// - outputs formatted string to both stdout and VS output window
451 /// - DEBUG builds only
453 /// PRINT("index %d = 0x%p\n",{C(lane), pIndex});
454 /// where C(lane) creates a constant value to print, and pIndex is the Value*
455 /// result from a GEP, printing out the pointer to memory
456 /// @param printStr - constant string to print, which includes format specifiers
457 /// @param printArgs - initializer list of Value*'s to print to std out
458 CallInst
*Builder::PRINT(const std::string
&printStr
,const std::initializer_list
<Value
*> &printArgs
)
460 // push the arguments to CallPrint into a vector
461 std::vector
<Value
*> printCallArgs
;
462 // save room for the format string. we still need to modify it for vectors
463 printCallArgs
.resize(1);
465 // search through the format string for special processing
467 std::string
tempStr(printStr
);
468 pos
= tempStr
.find('%', pos
);
469 auto v
= printArgs
.begin();
471 while ((pos
!= std::string::npos
) && (v
!= printArgs
.end()))
474 Type
* pType
= pArg
->getType();
476 if (pType
->isVectorTy())
478 Type
* pContainedType
= pType
->getContainedType(0);
480 if (toupper(tempStr
[pos
+ 1]) == 'X')
483 tempStr
[pos
+ 1] = 'x';
484 tempStr
.insert(pos
+ 2, "%08X ");
487 printCallArgs
.push_back(VEXTRACT(pArg
, C(0)));
489 std::string vectorFormatStr
;
490 for (uint32_t i
= 1; i
< pType
->getVectorNumElements(); ++i
)
492 vectorFormatStr
+= "0x%08X ";
493 printCallArgs
.push_back(VEXTRACT(pArg
, C(i
)));
496 tempStr
.insert(pos
, vectorFormatStr
);
497 pos
+= vectorFormatStr
.size();
499 else if ((tempStr
[pos
+ 1] == 'f') && (pContainedType
->isFloatTy()))
502 for (; i
< (pArg
->getType()->getVectorNumElements()) - 1; i
++)
504 tempStr
.insert(pos
, std::string("%f "));
506 printCallArgs
.push_back(FP_EXT(VEXTRACT(pArg
, C(i
)), Type::getDoubleTy(JM()->mContext
)));
508 printCallArgs
.push_back(FP_EXT(VEXTRACT(pArg
, C(i
)), Type::getDoubleTy(JM()->mContext
)));
510 else if ((tempStr
[pos
+ 1] == 'd') && (pContainedType
->isIntegerTy()))
513 for (; i
< (pArg
->getType()->getVectorNumElements()) - 1; i
++)
515 tempStr
.insert(pos
, std::string("%d "));
517 printCallArgs
.push_back(VEXTRACT(pArg
, C(i
)));
519 printCallArgs
.push_back(VEXTRACT(pArg
, C(i
)));
524 if (toupper(tempStr
[pos
+ 1]) == 'X')
527 tempStr
.insert(pos
+ 1, "x%08");
528 printCallArgs
.push_back(pArg
);
531 // for %f we need to cast float Values to doubles so that they print out correctly
532 else if ((tempStr
[pos
+ 1] == 'f') && (pType
->isFloatTy()))
534 printCallArgs
.push_back(FP_EXT(pArg
, Type::getDoubleTy(JM()->mContext
)));
539 printCallArgs
.push_back(pArg
);
543 // advance to the next arguement
545 pos
= tempStr
.find('%', ++pos
);
548 // create global variable constant string
549 Constant
*constString
= ConstantDataArray::getString(JM()->mContext
,tempStr
,true);
550 GlobalVariable
*gvPtr
= new GlobalVariable(constString
->getType(),true,GlobalValue::InternalLinkage
,constString
,"printStr");
551 JM()->mpCurrentModule
->getGlobalList().push_back(gvPtr
);
553 // get a pointer to the first character in the constant string array
554 std::vector
<Constant
*> geplist
{C(0),C(0)};
555 Constant
*strGEP
= ConstantExpr::getGetElementPtr(nullptr, gvPtr
,geplist
,false);
557 // insert the pointer to the format string in the argument vector
558 printCallArgs
[0] = strGEP
;
560 // get pointer to CallPrint function and insert decl into the module if needed
561 std::vector
<Type
*> args
;
562 args
.push_back(PointerType::get(mInt8Ty
,0));
563 FunctionType
* callPrintTy
= FunctionType::get(Type::getVoidTy(JM()->mContext
),args
,true);
564 Function
*callPrintFn
= cast
<Function
>(JM()->mpCurrentModule
->getOrInsertFunction("CallPrint", callPrintTy
));
566 // if we haven't yet added the symbol to the symbol table
567 if((sys::DynamicLibrary::SearchForAddressOfSymbol("CallPrint")) == nullptr)
569 sys::DynamicLibrary::AddSymbol("CallPrint", (void *)&CallPrint
);
572 // insert a call to CallPrint
573 return CALLA(callPrintFn
,printCallArgs
);
576 //////////////////////////////////////////////////////////////////////////
577 /// @brief Wrapper around PRINT with initializer list.
578 CallInst
* Builder::PRINT(const std::string
&printStr
)
580 return PRINT(printStr
, {});
583 //////////////////////////////////////////////////////////////////////////
584 /// @brief Generate a masked gather operation in LLVM IR. If not
585 /// supported on the underlying platform, emulate it with loads
586 /// @param vSrc - SIMD wide value that will be loaded if mask is invalid
587 /// @param pBase - Int8* base VB address pointer value
588 /// @param vIndices - SIMD wide value of VB byte offsets
589 /// @param vMask - SIMD wide mask that controls whether to access memory or the src values
590 /// @param scale - value to scale indices by
591 Value
*Builder::GATHERPS(Value
*vSrc
, Value
*pBase
, Value
*vIndices
, Value
*vMask
, uint8_t scale
)
595 // use avx2 gather instruction if available
596 if(JM()->mArch
.AVX2())
598 // force mask to <N x float>, required by vgather
599 Value
*mask
= BITCAST(VMASK(vMask
), mSimdFP32Ty
);
601 vGather
= VGATHERPS(vSrc
, pBase
, vIndices
, mask
, C(scale
));
605 Value
* pStack
= STACKSAVE();
607 // store vSrc on the stack. this way we can select between a valid load address and the vSrc address
608 Value
* vSrcPtr
= ALLOCA(vSrc
->getType());
609 STORE(vSrc
, vSrcPtr
);
611 vGather
= VUNDEF_F();
612 Value
*vScaleVec
= VIMMED1((uint32_t)scale
);
613 Value
*vOffsets
= MUL(vIndices
,vScaleVec
);
614 for(uint32_t i
= 0; i
< mVWidth
; ++i
)
616 // single component byte index
617 Value
*offset
= VEXTRACT(vOffsets
,C(i
));
618 // byte pointer to component
619 Value
*loadAddress
= GEP(pBase
,offset
);
620 loadAddress
= BITCAST(loadAddress
,PointerType::get(mFP32Ty
,0));
621 // pointer to the value to load if we're masking off a component
622 Value
*maskLoadAddress
= GEP(vSrcPtr
,{C(0), C(i
)});
623 Value
*selMask
= VEXTRACT(vMask
,C(i
));
624 // switch in a safe address to load if we're trying to access a vertex
625 Value
*validAddress
= SELECT(selMask
, loadAddress
, maskLoadAddress
);
626 Value
*val
= LOAD(validAddress
);
627 vGather
= VINSERT(vGather
,val
,C(i
));
630 STACKRESTORE(pStack
);
636 Value
*Builder::GATHERPS_16(Value
*vSrc
, Value
*pBase
, Value
*vIndices
, Value
*vMask
, uint8_t scale
)
638 Value
*vGather
= VUNDEF_F_16();
640 // use AVX512F gather instruction if available
641 if (JM()->mArch
.AVX512F())
643 // force mask to <N-bit Integer>, required by vgather2
644 Value
*mask
= BITCAST(vMask
, mInt16Ty
);
646 vGather
= VGATHERPS_16(vSrc
, pBase
, vIndices
, mask
, C((uint32_t)scale
));
650 Value
*src0
= EXTRACT_16(vSrc
, 0);
651 Value
*src1
= EXTRACT_16(vSrc
, 1);
653 Value
*indices0
= EXTRACT_16(vIndices
, 0);
654 Value
*indices1
= EXTRACT_16(vIndices
, 1);
656 Value
*mask0
= EXTRACT_16(vMask
, 0);
657 Value
*mask1
= EXTRACT_16(vMask
, 1);
659 Value
*gather0
= GATHERPS(src0
, pBase
, indices0
, mask0
, scale
);
660 Value
*gather1
= GATHERPS(src1
, pBase
, indices1
, mask1
, scale
);
662 vGather
= JOIN_16(gather0
, gather1
);
668 //////////////////////////////////////////////////////////////////////////
669 /// @brief Generate a masked gather operation in LLVM IR. If not
670 /// supported on the underlying platform, emulate it with loads
671 /// @param vSrc - SIMD wide value that will be loaded if mask is invalid
672 /// @param pBase - Int8* base VB address pointer value
673 /// @param vIndices - SIMD wide value of VB byte offsets
674 /// @param vMask - SIMD wide mask that controls whether to access memory or the src values
675 /// @param scale - value to scale indices by
676 Value
*Builder::GATHERDD(Value
* vSrc
, Value
* pBase
, Value
* vIndices
, Value
* vMask
, uint8_t scale
)
680 // use avx2 gather instruction if available
681 if(JM()->mArch
.AVX2())
683 vGather
= VGATHERDD(vSrc
, pBase
, vIndices
, VMASK(vMask
), C(scale
));
687 Value
* pStack
= STACKSAVE();
689 // store vSrc on the stack. this way we can select between a valid load address and the vSrc address
690 Value
* vSrcPtr
= ALLOCA(vSrc
->getType());
691 STORE(vSrc
, vSrcPtr
);
693 vGather
= VUNDEF_I();
694 Value
*vScaleVec
= VIMMED1((uint32_t)scale
);
695 Value
*vOffsets
= MUL(vIndices
, vScaleVec
);
696 for(uint32_t i
= 0; i
< mVWidth
; ++i
)
698 // single component byte index
699 Value
*offset
= VEXTRACT(vOffsets
, C(i
));
700 // byte pointer to component
701 Value
*loadAddress
= GEP(pBase
, offset
);
702 loadAddress
= BITCAST(loadAddress
, PointerType::get(mInt32Ty
, 0));
703 // pointer to the value to load if we're masking off a component
704 Value
*maskLoadAddress
= GEP(vSrcPtr
, {C(0), C(i
)});
705 Value
*selMask
= VEXTRACT(vMask
, C(i
));
706 // switch in a safe address to load if we're trying to access a vertex
707 Value
*validAddress
= SELECT(selMask
, loadAddress
, maskLoadAddress
);
708 Value
*val
= LOAD(validAddress
, C(0));
709 vGather
= VINSERT(vGather
, val
, C(i
));
712 STACKRESTORE(pStack
);
718 Value
*Builder::GATHERDD_16(Value
*vSrc
, Value
*pBase
, Value
*vIndices
, Value
*vMask
, uint8_t scale
)
720 Value
*vGather
= VUNDEF_I_16();
722 // use AVX512F gather instruction if available
723 if (JM()->mArch
.AVX512F())
725 // force mask to <N-bit Integer>, required by vgather2
726 Value
*mask
= BITCAST(vMask
, mInt16Ty
);
728 vGather
= VGATHERDD_16(vSrc
, pBase
, vIndices
, mask
, C((uint32_t)scale
));
732 Value
*src0
= EXTRACT_16(vSrc
, 0);
733 Value
*src1
= EXTRACT_16(vSrc
, 1);
735 Value
*indices0
= EXTRACT_16(vIndices
, 0);
736 Value
*indices1
= EXTRACT_16(vIndices
, 1);
738 Value
*mask0
= EXTRACT_16(vMask
, 0);
739 Value
*mask1
= EXTRACT_16(vMask
, 1);
741 Value
*gather0
= GATHERDD(src0
, pBase
, indices0
, mask0
, scale
);
742 Value
*gather1
= GATHERDD(src1
, pBase
, indices1
, mask1
, scale
);
744 vGather
= JOIN_16(gather0
, gather1
);
750 //////////////////////////////////////////////////////////////////////////
751 /// @brief Generate a masked gather operation in LLVM IR. If not
752 /// supported on the underlying platform, emulate it with loads
753 /// @param vSrc - SIMD wide value that will be loaded if mask is invalid
754 /// @param pBase - Int8* base VB address pointer value
755 /// @param vIndices - SIMD wide value of VB byte offsets
756 /// @param vMask - SIMD wide mask that controls whether to access memory or the src values
757 /// @param scale - value to scale indices by
758 Value
*Builder::GATHERPD(Value
* vSrc
, Value
* pBase
, Value
* vIndices
, Value
* vMask
, uint8_t scale
)
762 // use avx2 gather instruction if available
763 if(JM()->mArch
.AVX2())
765 vMask
= BITCAST(S_EXT(vMask
, VectorType::get(mInt64Ty
, mVWidth
/2)), VectorType::get(mDoubleTy
, mVWidth
/2));
766 vGather
= VGATHERPD(vSrc
, pBase
, vIndices
, vMask
, C(scale
));
770 Value
* pStack
= STACKSAVE();
772 // store vSrc on the stack. this way we can select between a valid load address and the vSrc address
773 Value
* vSrcPtr
= ALLOCA(vSrc
->getType());
774 STORE(vSrc
, vSrcPtr
);
776 vGather
= UndefValue::get(VectorType::get(mDoubleTy
, 4));
777 Value
*vScaleVec
= VECTOR_SPLAT(4, C((uint32_t)scale
));
778 Value
*vOffsets
= MUL(vIndices
,vScaleVec
);
779 for(uint32_t i
= 0; i
< mVWidth
/2; ++i
)
781 // single component byte index
782 Value
*offset
= VEXTRACT(vOffsets
,C(i
));
783 // byte pointer to component
784 Value
*loadAddress
= GEP(pBase
,offset
);
785 loadAddress
= BITCAST(loadAddress
,PointerType::get(mDoubleTy
,0));
786 // pointer to the value to load if we're masking off a component
787 Value
*maskLoadAddress
= GEP(vSrcPtr
,{C(0), C(i
)});
788 Value
*selMask
= VEXTRACT(vMask
,C(i
));
789 // switch in a safe address to load if we're trying to access a vertex
790 Value
*validAddress
= SELECT(selMask
, loadAddress
, maskLoadAddress
);
791 Value
*val
= LOAD(validAddress
);
792 vGather
= VINSERT(vGather
,val
,C(i
));
794 STACKRESTORE(pStack
);
799 Value
*Builder::EXTRACT_16(Value
*x
, uint32_t imm
)
803 return VSHUFFLE(x
, UndefValue::get(x
->getType()), { 0, 1, 2, 3, 4, 5, 6, 7 });
807 return VSHUFFLE(x
, UndefValue::get(x
->getType()), { 8, 9, 10, 11, 12, 13, 14, 15 });
811 Value
*Builder::JOIN_16(Value
*a
, Value
*b
)
813 return VSHUFFLE(a
, b
, { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 });
816 //////////////////////////////////////////////////////////////////////////
817 /// @brief convert x86 <N x float> mask to llvm <N x i1> mask
818 Value
*Builder::MASK(Value
*vmask
)
820 Value
*src
= BITCAST(vmask
, mSimdInt32Ty
);
821 return ICMP_SLT(src
, VIMMED1(0));
824 Value
*Builder::MASK_16(Value
*vmask
)
826 Value
*src
= BITCAST(vmask
, mSimd16Int32Ty
);
827 return ICMP_SLT(src
, VIMMED1_16(0));
830 //////////////////////////////////////////////////////////////////////////
831 /// @brief convert llvm <N x i1> mask to x86 <N x i32> mask
832 Value
*Builder::VMASK(Value
*mask
)
834 return S_EXT(mask
, mSimdInt32Ty
);
837 Value
*Builder::VMASK_16(Value
*mask
)
839 return S_EXT(mask
, mSimd16Int32Ty
);
842 //////////////////////////////////////////////////////////////////////////
843 /// @brief Generate a VPSHUFB operation in LLVM IR. If not
844 /// supported on the underlying platform, emulate it
845 /// @param a - 256bit SIMD(32x8bit) of 8bit integer values
846 /// @param b - 256bit SIMD(32x8bit) of 8bit integer mask values
847 /// Byte masks in lower 128 lane of b selects 8 bit values from lower
848 /// 128bits of a, and vice versa for the upper lanes. If the mask
849 /// value is negative, '0' is inserted.
850 Value
*Builder::PSHUFB(Value
* a
, Value
* b
)
853 // use avx2 pshufb instruction if available
854 if(JM()->mArch
.AVX2())
860 Constant
* cB
= dyn_cast
<Constant
>(b
);
861 // number of 8 bit elements in b
862 uint32_t numElms
= cast
<VectorType
>(cB
->getType())->getNumElements();
864 Value
* vShuf
= UndefValue::get(VectorType::get(mInt8Ty
, numElms
));
866 // insert an 8 bit value from the high and low lanes of a per loop iteration
868 for(uint32_t i
= 0; i
< numElms
; i
++)
870 ConstantInt
* cLow128b
= cast
<ConstantInt
>(cB
->getAggregateElement(i
));
871 ConstantInt
* cHigh128b
= cast
<ConstantInt
>(cB
->getAggregateElement(i
+ numElms
));
873 // extract values from constant mask
874 char valLow128bLane
= (char)(cLow128b
->getSExtValue());
875 char valHigh128bLane
= (char)(cHigh128b
->getSExtValue());
877 Value
* insertValLow128b
;
878 Value
* insertValHigh128b
;
880 // if the mask value is negative, insert a '0' in the respective output position
881 // otherwise, lookup the value at mask position (bits 3..0 of the respective mask byte) in a and insert in output vector
882 insertValLow128b
= (valLow128bLane
< 0) ? C((char)0) : VEXTRACT(a
, C((valLow128bLane
& 0xF)));
883 insertValHigh128b
= (valHigh128bLane
< 0) ? C((char)0) : VEXTRACT(a
, C((valHigh128bLane
& 0xF) + numElms
));
885 vShuf
= VINSERT(vShuf
, insertValLow128b
, i
);
886 vShuf
= VINSERT(vShuf
, insertValHigh128b
, (i
+ numElms
));
893 //////////////////////////////////////////////////////////////////////////
894 /// @brief Generate a VPSHUFB operation (sign extend 8 8bit values to 32
895 /// bits)in LLVM IR. If not supported on the underlying platform, emulate it
896 /// @param a - 128bit SIMD lane(16x8bit) of 8bit integer values. Only
897 /// lower 8 values are used.
898 Value
*Builder::PMOVSXBD(Value
* a
)
900 // VPMOVSXBD output type
901 Type
* v8x32Ty
= VectorType::get(mInt32Ty
, 8);
902 // Extract 8 values from 128bit lane and sign extend
903 return S_EXT(VSHUFFLE(a
, a
, C
<int>({0, 1, 2, 3, 4, 5, 6, 7})), v8x32Ty
);
906 //////////////////////////////////////////////////////////////////////////
907 /// @brief Generate a VPSHUFB operation (sign extend 8 16bit values to 32
908 /// bits)in LLVM IR. If not supported on the underlying platform, emulate it
909 /// @param a - 128bit SIMD lane(8x16bit) of 16bit integer values.
910 Value
*Builder::PMOVSXWD(Value
* a
)
912 // VPMOVSXWD output type
913 Type
* v8x32Ty
= VectorType::get(mInt32Ty
, 8);
914 // Extract 8 values from 128bit lane and sign extend
915 return S_EXT(VSHUFFLE(a
, a
, C
<int>({0, 1, 2, 3, 4, 5, 6, 7})), v8x32Ty
);
918 //////////////////////////////////////////////////////////////////////////
919 /// @brief Generate a VPERMD operation (shuffle 32 bit integer values
920 /// across 128 bit lanes) in LLVM IR. If not supported on the underlying
921 /// platform, emulate it
922 /// @param a - 256bit SIMD lane(8x32bit) of integer values.
923 /// @param idx - 256bit SIMD lane(8x32bit) of 3 bit lane index values
924 Value
*Builder::PERMD(Value
* a
, Value
* idx
)
927 // use avx2 permute instruction if available
928 if(JM()->mArch
.AVX2())
930 res
= VPERMD(a
, idx
);
934 if (isa
<Constant
>(idx
))
936 res
= VSHUFFLE(a
, a
, idx
);
941 for (uint32_t l
= 0; l
< JM()->mVWidth
; ++l
)
943 Value
* pIndex
= VEXTRACT(idx
, C(l
));
944 Value
* pVal
= VEXTRACT(a
, pIndex
);
945 res
= VINSERT(res
, pVal
, C(l
));
952 //////////////////////////////////////////////////////////////////////////
953 /// @brief Generate a VPERMPS operation (shuffle 32 bit float values
954 /// across 128 bit lanes) in LLVM IR. If not supported on the underlying
955 /// platform, emulate it
956 /// @param a - 256bit SIMD lane(8x32bit) of float values.
957 /// @param idx - 256bit SIMD lane(8x32bit) of 3 bit lane index values
958 Value
*Builder::PERMPS(Value
* a
, Value
* idx
)
961 // use avx2 permute instruction if available
962 if (JM()->mArch
.AVX2())
964 // llvm 3.6.0 swapped the order of the args to vpermd
965 res
= VPERMPS(idx
, a
);
969 if (isa
<Constant
>(idx
))
971 res
= VSHUFFLE(a
, a
, idx
);
976 for (uint32_t l
= 0; l
< JM()->mVWidth
; ++l
)
978 Value
* pIndex
= VEXTRACT(idx
, C(l
));
979 Value
* pVal
= VEXTRACT(a
, pIndex
);
980 res
= VINSERT(res
, pVal
, C(l
));
988 //////////////////////////////////////////////////////////////////////////
989 /// @brief Generate a VCVTPH2PS operation (float16->float32 conversion)
990 /// in LLVM IR. If not supported on the underlying platform, emulate it
991 /// @param a - 128bit SIMD lane(8x16bit) of float16 in int16 format.
992 Value
*Builder::CVTPH2PS(Value
* a
)
994 if (JM()->mArch
.F16C())
1000 FunctionType
* pFuncTy
= FunctionType::get(mFP32Ty
, mInt16Ty
);
1001 Function
* pCvtPh2Ps
= cast
<Function
>(JM()->mpCurrentModule
->getOrInsertFunction("ConvertFloat16ToFloat32", pFuncTy
));
1003 if (sys::DynamicLibrary::SearchForAddressOfSymbol("ConvertFloat16ToFloat32") == nullptr)
1005 sys::DynamicLibrary::AddSymbol("ConvertFloat16ToFloat32", (void *)&ConvertFloat16ToFloat32
);
1008 Value
* pResult
= UndefValue::get(mSimdFP32Ty
);
1009 for (uint32_t i
= 0; i
< mVWidth
; ++i
)
1011 Value
* pSrc
= VEXTRACT(a
, C(i
));
1012 Value
* pConv
= CALL(pCvtPh2Ps
, std::initializer_list
<Value
*>{pSrc
});
1013 pResult
= VINSERT(pResult
, pConv
, C(i
));
1020 //////////////////////////////////////////////////////////////////////////
1021 /// @brief Generate a VCVTPS2PH operation (float32->float16 conversion)
1022 /// in LLVM IR. If not supported on the underlying platform, emulate it
1023 /// @param a - 128bit SIMD lane(8x16bit) of float16 in int16 format.
1024 Value
*Builder::CVTPS2PH(Value
* a
, Value
* rounding
)
1026 if (JM()->mArch
.F16C())
1028 return VCVTPS2PH(a
, rounding
);
1032 // call scalar C function for now
1033 FunctionType
* pFuncTy
= FunctionType::get(mInt16Ty
, mFP32Ty
);
1034 Function
* pCvtPs2Ph
= cast
<Function
>(JM()->mpCurrentModule
->getOrInsertFunction("ConvertFloat32ToFloat16", pFuncTy
));
1036 if (sys::DynamicLibrary::SearchForAddressOfSymbol("ConvertFloat32ToFloat16") == nullptr)
1038 sys::DynamicLibrary::AddSymbol("ConvertFloat32ToFloat16", (void *)&ConvertFloat32ToFloat16
);
1041 Value
* pResult
= UndefValue::get(mSimdInt16Ty
);
1042 for (uint32_t i
= 0; i
< mVWidth
; ++i
)
1044 Value
* pSrc
= VEXTRACT(a
, C(i
));
1045 Value
* pConv
= CALL(pCvtPs2Ph
, std::initializer_list
<Value
*>{pSrc
});
1046 pResult
= VINSERT(pResult
, pConv
, C(i
));
1053 Value
*Builder::PMAXSD(Value
* a
, Value
* b
)
1055 Value
* cmp
= ICMP_SGT(a
, b
);
1056 return SELECT(cmp
, a
, b
);
1059 Value
*Builder::PMINSD(Value
* a
, Value
* b
)
1061 Value
* cmp
= ICMP_SLT(a
, b
);
1062 return SELECT(cmp
, a
, b
);
1065 void Builder::Gather4(const SWR_FORMAT format
, Value
* pSrcBase
, Value
* byteOffsets
,
1066 Value
* mask
, Value
* vGatherComponents
[], bool bPackedOutput
)
1068 const SWR_FORMAT_INFO
&info
= GetFormatInfo(format
);
1069 if(info
.type
[0] == SWR_TYPE_FLOAT
&& info
.bpc
[0] == 32)
1071 GATHER4PS(info
, pSrcBase
, byteOffsets
, mask
, vGatherComponents
, bPackedOutput
);
1075 GATHER4DD(info
, pSrcBase
, byteOffsets
, mask
, vGatherComponents
, bPackedOutput
);
1079 void Builder::GATHER4PS(const SWR_FORMAT_INFO
&info
, Value
* pSrcBase
, Value
* byteOffsets
,
1080 Value
* vMask
, Value
* vGatherComponents
[], bool bPackedOutput
)
1082 switch(info
.bpp
/ info
.numComps
)
1086 Value
* vGatherResult
[2];
1088 // TODO: vGatherMaskedVal
1089 Value
* vGatherMaskedVal
= VIMMED1((float)0);
1091 // always have at least one component out of x or y to fetch
1093 vGatherResult
[0] = GATHERPS(vGatherMaskedVal
, pSrcBase
, byteOffsets
, vMask
);
1094 // e.g. result of first 8x32bit integer gather for 16bit components
1095 // 256i - 0 1 2 3 4 5 6 7
1096 // xyxy xyxy xyxy xyxy xyxy xyxy xyxy xyxy
1099 // if we have at least one component out of x or y to fetch
1100 if(info
.numComps
> 2)
1102 // offset base to the next components(zw) in the vertex to gather
1103 pSrcBase
= GEP(pSrcBase
, C((char)4));
1105 vGatherResult
[1] = GATHERPS(vGatherMaskedVal
, pSrcBase
, byteOffsets
, vMask
);
1106 // e.g. result of second 8x32bit integer gather for 16bit components
1107 // 256i - 0 1 2 3 4 5 6 7
1108 // zwzw zwzw zwzw zwzw zwzw zwzw zwzw zwzw
1113 vGatherResult
[1] = vGatherMaskedVal
;
1116 // Shuffle gathered components into place, each row is a component
1117 Shuffle16bpcGather4(info
, vGatherResult
, vGatherComponents
, bPackedOutput
);
1123 for (uint32_t i
= 0; i
< 4; ++i
)
1125 vGatherComponents
[i
] = VIMMED1(*(float*)&info
.defaults
[i
]);
1128 for(uint32_t i
= 0; i
< info
.numComps
; i
++)
1130 uint32_t swizzleIndex
= info
.swizzle
[i
];
1132 // Gather a SIMD of components
1133 vGatherComponents
[swizzleIndex
] = GATHERPS(vGatherComponents
[swizzleIndex
], pSrcBase
, byteOffsets
, vMask
);
1135 // offset base to the next component to gather
1136 pSrcBase
= GEP(pSrcBase
, C((char)4));
1141 SWR_INVALID("Invalid float format");
1146 void Builder::GATHER4DD(const SWR_FORMAT_INFO
&info
, Value
* pSrcBase
, Value
* byteOffsets
,
1147 Value
* vMask
, Value
* vGatherComponents
[], bool bPackedOutput
)
1149 switch (info
.bpp
/ info
.numComps
)
1153 Value
* vGatherMaskedVal
= VIMMED1((int32_t)0);
1154 Value
* vGatherResult
= GATHERDD(vGatherMaskedVal
, pSrcBase
, byteOffsets
, vMask
);
1155 // e.g. result of an 8x32bit integer gather for 8bit components
1156 // 256i - 0 1 2 3 4 5 6 7
1157 // xyzw xyzw xyzw xyzw xyzw xyzw xyzw xyzw
1159 Shuffle8bpcGather4(info
, vGatherResult
, vGatherComponents
, bPackedOutput
);
1164 Value
* vGatherResult
[2];
1166 // TODO: vGatherMaskedVal
1167 Value
* vGatherMaskedVal
= VIMMED1((int32_t)0);
1169 // always have at least one component out of x or y to fetch
1171 vGatherResult
[0] = GATHERDD(vGatherMaskedVal
, pSrcBase
, byteOffsets
, vMask
);
1172 // e.g. result of first 8x32bit integer gather for 16bit components
1173 // 256i - 0 1 2 3 4 5 6 7
1174 // xyxy xyxy xyxy xyxy xyxy xyxy xyxy xyxy
1177 // if we have at least one component out of x or y to fetch
1178 if(info
.numComps
> 2)
1180 // offset base to the next components(zw) in the vertex to gather
1181 pSrcBase
= GEP(pSrcBase
, C((char)4));
1183 vGatherResult
[1] = GATHERDD(vGatherMaskedVal
, pSrcBase
, byteOffsets
, vMask
);
1184 // e.g. result of second 8x32bit integer gather for 16bit components
1185 // 256i - 0 1 2 3 4 5 6 7
1186 // zwzw zwzw zwzw zwzw zwzw zwzw zwzw zwzw
1191 vGatherResult
[1] = vGatherMaskedVal
;
1194 // Shuffle gathered components into place, each row is a component
1195 Shuffle16bpcGather4(info
, vGatherResult
, vGatherComponents
, bPackedOutput
);
1202 for (uint32_t i
= 0; i
< 4; ++i
)
1204 vGatherComponents
[i
] = VIMMED1((int)info
.defaults
[i
]);
1207 for(uint32_t i
= 0; i
< info
.numComps
; i
++)
1209 uint32_t swizzleIndex
= info
.swizzle
[i
];
1211 // Gather a SIMD of components
1212 vGatherComponents
[swizzleIndex
] = GATHERDD(vGatherComponents
[swizzleIndex
], pSrcBase
, byteOffsets
, vMask
);
1214 // offset base to the next component to gather
1215 pSrcBase
= GEP(pSrcBase
, C((char)4));
1220 SWR_INVALID("unsupported format");
1225 void Builder::Shuffle16bpcGather4(const SWR_FORMAT_INFO
&info
, Value
* vGatherInput
[2], Value
* vGatherOutput
[4], bool bPackedOutput
)
1228 Type
* vGatherTy
= VectorType::get(IntegerType::getInt32Ty(JM()->mContext
), mVWidth
);
1229 Type
* v32x8Ty
= VectorType::get(mInt8Ty
, mVWidth
* 4); // vwidth is units of 32 bits
1231 // input could either be float or int vector; do shuffle work in int
1232 vGatherInput
[0] = BITCAST(vGatherInput
[0], mSimdInt32Ty
);
1233 vGatherInput
[1] = BITCAST(vGatherInput
[1], mSimdInt32Ty
);
1237 Type
* v128bitTy
= VectorType::get(IntegerType::getIntNTy(JM()->mContext
, 128), mVWidth
/ 4); // vwidth is units of 32 bits
1240 Value
* vConstMask
= C
<char>({0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15,
1241 0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15});
1242 Value
* vShufResult
= BITCAST(PSHUFB(BITCAST(vGatherInput
[0], v32x8Ty
), vConstMask
), vGatherTy
);
1243 // after pshufb: group components together in each 128bit lane
1244 // 256i - 0 1 2 3 4 5 6 7
1245 // xxxx xxxx yyyy yyyy xxxx xxxx yyyy yyyy
1247 Value
* vi128XY
= BITCAST(PERMD(vShufResult
, C
<int32_t>({0, 1, 4, 5, 2, 3, 6, 7})), v128bitTy
);
1248 // after PERMD: move and pack xy components into each 128bit lane
1249 // 256i - 0 1 2 3 4 5 6 7
1250 // xxxx xxxx xxxx xxxx yyyy yyyy yyyy yyyy
1252 // do the same for zw components
1253 Value
* vi128ZW
= nullptr;
1254 if(info
.numComps
> 2)
1256 Value
* vShufResult
= BITCAST(PSHUFB(BITCAST(vGatherInput
[1], v32x8Ty
), vConstMask
), vGatherTy
);
1257 vi128ZW
= BITCAST(PERMD(vShufResult
, C
<int32_t>({0, 1, 4, 5, 2, 3, 6, 7})), v128bitTy
);
1260 for(uint32_t i
= 0; i
< 4; i
++)
1262 uint32_t swizzleIndex
= info
.swizzle
[i
];
1263 // todo: fixed for packed
1264 Value
* vGatherMaskedVal
= VIMMED1((int32_t)(info
.defaults
[i
]));
1265 if(i
>= info
.numComps
)
1267 // set the default component val
1268 vGatherOutput
[swizzleIndex
] = vGatherMaskedVal
;
1272 // if x or z, extract 128bits from lane 0, else for y or w, extract from lane 1
1273 uint32_t lane
= ((i
== 0) || (i
== 2)) ? 0 : 1;
1274 // if x or y, use vi128XY permute result, else use vi128ZW
1275 Value
* selectedPermute
= (i
< 2) ? vi128XY
: vi128ZW
;
1277 // extract packed component 128 bit lanes
1278 vGatherOutput
[swizzleIndex
] = VEXTRACT(selectedPermute
, C(lane
));
1284 // pshufb masks for each component
1285 Value
* vConstMask
[2];
1287 vConstMask
[0] = C
<char>({0, 1, -1, -1, 4, 5, -1, -1, 8, 9, -1, -1, 12, 13, -1, -1,
1288 0, 1, -1, -1, 4, 5, -1, -1, 8, 9, -1, -1, 12, 13, -1, -1, });
1291 vConstMask
[1] = C
<char>({2, 3, -1, -1, 6, 7, -1, -1, 10, 11, -1, -1, 14, 15, -1, -1,
1292 2, 3, -1, -1, 6, 7, -1, -1, 10, 11, -1, -1, 14, 15, -1, -1});
1295 // shuffle enabled components into lower word of each 32bit lane, 0 extending to 32 bits
1297 for (uint32_t i
= 0; i
< 4; ++i
)
1299 vGatherOutput
[i
] = VIMMED1((int32_t)info
.defaults
[i
]);
1302 for(uint32_t i
= 0; i
< info
.numComps
; i
++)
1304 uint32_t swizzleIndex
= info
.swizzle
[i
];
1306 // select correct constMask for x/z or y/w pshufb
1307 uint32_t selectedMask
= ((i
== 0) || (i
== 2)) ? 0 : 1;
1308 // if x or y, use vi128XY permute result, else use vi128ZW
1309 uint32_t selectedGather
= (i
< 2) ? 0 : 1;
1311 vGatherOutput
[swizzleIndex
] = BITCAST(PSHUFB(BITCAST(vGatherInput
[selectedGather
], v32x8Ty
), vConstMask
[selectedMask
]), vGatherTy
);
1312 // after pshufb mask for x channel; z uses the same shuffle from the second gather
1313 // 256i - 0 1 2 3 4 5 6 7
1314 // xx00 xx00 xx00 xx00 xx00 xx00 xx00 xx00
1319 void Builder::Shuffle8bpcGather4(const SWR_FORMAT_INFO
&info
, Value
* vGatherInput
, Value
* vGatherOutput
[], bool bPackedOutput
)
1322 Type
* vGatherTy
= VectorType::get(IntegerType::getInt32Ty(JM()->mContext
), mVWidth
);
1323 Type
* v32x8Ty
= VectorType::get(mInt8Ty
, mVWidth
* 4 ); // vwidth is units of 32 bits
1327 Type
* v128Ty
= VectorType::get(IntegerType::getIntNTy(JM()->mContext
, 128), mVWidth
/ 4); // vwidth is units of 32 bits
1329 Value
* vConstMask
= C
<char>({0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15,
1330 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15});
1331 Value
* vShufResult
= BITCAST(PSHUFB(BITCAST(vGatherInput
, v32x8Ty
), vConstMask
), vGatherTy
);
1332 // after pshufb: group components together in each 128bit lane
1333 // 256i - 0 1 2 3 4 5 6 7
1334 // xxxx yyyy zzzz wwww xxxx yyyy zzzz wwww
1336 Value
* vi128XY
= BITCAST(PERMD(vShufResult
, C
<int32_t>({0, 4, 0, 0, 1, 5, 0, 0})), v128Ty
);
1337 // after PERMD: move and pack xy and zw components in low 64 bits of each 128bit lane
1338 // 256i - 0 1 2 3 4 5 6 7
1339 // xxxx xxxx dcdc dcdc yyyy yyyy dcdc dcdc (dc - don't care)
1341 // do the same for zw components
1342 Value
* vi128ZW
= nullptr;
1343 if(info
.numComps
> 2)
1345 vi128ZW
= BITCAST(PERMD(vShufResult
, C
<int32_t>({2, 6, 0, 0, 3, 7, 0, 0})), v128Ty
);
1348 // sign extend all enabled components. If we have a fill vVertexElements, output to current simdvertex
1349 for(uint32_t i
= 0; i
< 4; i
++)
1351 uint32_t swizzleIndex
= info
.swizzle
[i
];
1352 // todo: fix for packed
1353 Value
* vGatherMaskedVal
= VIMMED1((int32_t)(info
.defaults
[i
]));
1354 if(i
>= info
.numComps
)
1356 // set the default component val
1357 vGatherOutput
[swizzleIndex
] = vGatherMaskedVal
;
1361 // if x or z, extract 128bits from lane 0, else for y or w, extract from lane 1
1362 uint32_t lane
= ((i
== 0) || (i
== 2)) ? 0 : 1;
1363 // if x or y, use vi128XY permute result, else use vi128ZW
1364 Value
* selectedPermute
= (i
< 2) ? vi128XY
: vi128ZW
;
1367 vGatherOutput
[swizzleIndex
] = VEXTRACT(selectedPermute
, C(lane
));
1372 // shuffle enabled components into lower byte of each 32bit lane, 0 extending to 32 bits
1374 for (uint32_t i
= 0; i
< 4; ++i
)
1376 vGatherOutput
[i
] = VIMMED1((int32_t)info
.defaults
[i
]);
1379 for(uint32_t i
= 0; i
< info
.numComps
; i
++){
1380 uint32_t swizzleIndex
= info
.swizzle
[i
];
1382 // pshufb masks for each component
1388 vConstMask
= C
<char>({0, -1, -1, -1, 4, -1, -1, -1, 8, -1, -1, -1, 12, -1, -1, -1,
1389 0, -1, -1, -1, 4, -1, -1, -1, 8, -1, -1, -1, 12, -1, -1, -1});
1393 vConstMask
= C
<char>({1, -1, -1, -1, 5, -1, -1, -1, 9, -1, -1, -1, 13, -1, -1, -1,
1394 1, -1, -1, -1, 5, -1, -1, -1, 9, -1, -1, -1, 13, -1, -1, -1});
1398 vConstMask
= C
<char>({2, -1, -1, -1, 6, -1, -1, -1, 10, -1, -1, -1, 14, -1, -1, -1,
1399 2, -1, -1, -1, 6, -1, -1, -1, 10, -1, -1, -1, 14, -1, -1, -1});
1403 vConstMask
= C
<char>({3, -1, -1, -1, 7, -1, -1, -1, 11, -1, -1, -1, 15, -1, -1, -1,
1404 3, -1, -1, -1, 7, -1, -1, -1, 11, -1, -1, -1, 15, -1, -1, -1});
1407 vConstMask
= nullptr;
1411 vGatherOutput
[swizzleIndex
] = BITCAST(PSHUFB(BITCAST(vGatherInput
, v32x8Ty
), vConstMask
), vGatherTy
);
1412 // after pshufb for x channel
1413 // 256i - 0 1 2 3 4 5 6 7
1414 // x000 x000 x000 x000 x000 x000 x000 x000
1419 // Helper function to create alloca in entry block of function
1420 Value
* Builder::CreateEntryAlloca(Function
* pFunc
, Type
* pType
)
1422 auto saveIP
= IRB()->saveIP();
1423 IRB()->SetInsertPoint(&pFunc
->getEntryBlock(),
1424 pFunc
->getEntryBlock().begin());
1425 Value
* pAlloca
= ALLOCA(pType
);
1426 if (saveIP
.isSet()) IRB()->restoreIP(saveIP
);
1430 Value
* Builder::CreateEntryAlloca(Function
* pFunc
, Type
* pType
, Value
* pArraySize
)
1432 auto saveIP
= IRB()->saveIP();
1433 IRB()->SetInsertPoint(&pFunc
->getEntryBlock(),
1434 pFunc
->getEntryBlock().begin());
1435 Value
* pAlloca
= ALLOCA(pType
, pArraySize
);
1436 if (saveIP
.isSet()) IRB()->restoreIP(saveIP
);
1440 //////////////////////////////////////////////////////////////////////////
1441 /// @brief emulates a scatter operation.
1442 /// @param pDst - pointer to destination
1443 /// @param vSrc - vector of src data to scatter
1444 /// @param vOffsets - vector of byte offsets from pDst
1445 /// @param vMask - mask of valid lanes
1446 void Builder::SCATTERPS(Value
* pDst
, Value
* vSrc
, Value
* vOffsets
, Value
* vMask
)
1448 /* Scatter algorithm
1450 while(Index = BitScanForward(mask))
1451 srcElem = srcVector[Index]
1452 offsetElem = offsetVector[Index]
1453 *(pDst + offsetElem) = srcElem
1454 Update mask (&= ~(1<<Index)
1458 BasicBlock
* pCurBB
= IRB()->GetInsertBlock();
1459 Function
* pFunc
= pCurBB
->getParent();
1460 Type
* pSrcTy
= vSrc
->getType()->getVectorElementType();
1462 // Store vectors on stack
1463 if (pScatterStackSrc
== nullptr)
1465 // Save off stack allocations and reuse per scatter. Significantly reduces stack
1466 // requirements for shaders with a lot of scatters.
1467 pScatterStackSrc
= CreateEntryAlloca(pFunc
, mSimdInt64Ty
);
1468 pScatterStackOffsets
= CreateEntryAlloca(pFunc
, mSimdInt32Ty
);
1471 Value
* pSrcArrayPtr
= BITCAST(pScatterStackSrc
, PointerType::get(vSrc
->getType(), 0));
1472 Value
* pOffsetsArrayPtr
= pScatterStackOffsets
;
1473 STORE(vSrc
, pSrcArrayPtr
);
1474 STORE(vOffsets
, pOffsetsArrayPtr
);
1476 // Cast to pointers for random access
1477 pSrcArrayPtr
= POINTER_CAST(pSrcArrayPtr
, PointerType::get(pSrcTy
, 0));
1478 pOffsetsArrayPtr
= POINTER_CAST(pOffsetsArrayPtr
, PointerType::get(mInt32Ty
, 0));
1480 Value
* pMask
= VMOVMSKPS(BITCAST(vMask
, mSimdFP32Ty
));
1482 // Get cttz function
1483 Function
* pfnCttz
= Intrinsic::getDeclaration(mpJitMgr
->mpCurrentModule
, Intrinsic::cttz
, { mInt32Ty
});
1485 // Setup loop basic block
1486 BasicBlock
* pLoop
= BasicBlock::Create(mpJitMgr
->mContext
, "Scatter Loop", pFunc
);
1488 // compute first set bit
1489 Value
* pIndex
= CALL(pfnCttz
, { pMask
, C(false) });
1491 Value
* pIsUndef
= ICMP_EQ(pIndex
, C(32));
1493 // Split current block
1494 BasicBlock
* pPostLoop
= pCurBB
->splitBasicBlock(cast
<Instruction
>(pIsUndef
)->getNextNode());
1496 // Remove unconditional jump created by splitBasicBlock
1497 pCurBB
->getTerminator()->eraseFromParent();
1499 // Add terminator to end of original block
1500 IRB()->SetInsertPoint(pCurBB
);
1502 // Add conditional branch
1503 COND_BR(pIsUndef
, pPostLoop
, pLoop
);
1505 // Add loop basic block contents
1506 IRB()->SetInsertPoint(pLoop
);
1507 PHINode
* pIndexPhi
= PHI(mInt32Ty
, 2);
1508 PHINode
* pMaskPhi
= PHI(mInt32Ty
, 2);
1510 pIndexPhi
->addIncoming(pIndex
, pCurBB
);
1511 pMaskPhi
->addIncoming(pMask
, pCurBB
);
1513 // Extract elements for this index
1514 Value
* pSrcElem
= LOADV(pSrcArrayPtr
, { pIndexPhi
});
1515 Value
* pOffsetElem
= LOADV(pOffsetsArrayPtr
, { pIndexPhi
});
1517 // GEP to this offset in dst
1518 Value
* pCurDst
= GEP(pDst
, pOffsetElem
);
1519 pCurDst
= POINTER_CAST(pCurDst
, PointerType::get(pSrcTy
, 0));
1520 STORE(pSrcElem
, pCurDst
);
1523 Value
* pNewMask
= AND(pMaskPhi
, NOT(SHL(C(1), pIndexPhi
)));
1526 Value
* pNewIndex
= CALL(pfnCttz
, { pNewMask
, C(false) });
1528 pIsUndef
= ICMP_EQ(pNewIndex
, C(32));
1529 COND_BR(pIsUndef
, pPostLoop
, pLoop
);
1532 pIndexPhi
->addIncoming(pNewIndex
, pLoop
);
1533 pMaskPhi
->addIncoming(pNewMask
, pLoop
);
1535 // Move builder to beginning of post loop
1536 IRB()->SetInsertPoint(pPostLoop
, pPostLoop
->begin());
1539 Value
* Builder::VABSPS(Value
* a
)
1541 Value
* asInt
= BITCAST(a
, mSimdInt32Ty
);
1542 Value
* result
= BITCAST(AND(asInt
, VIMMED1(0x7fffffff)), mSimdFP32Ty
);
1546 Value
*Builder::ICLAMP(Value
* src
, Value
* low
, Value
* high
)
1548 Value
*lowCmp
= ICMP_SLT(src
, low
);
1549 Value
*ret
= SELECT(lowCmp
, low
, src
);
1551 Value
*highCmp
= ICMP_SGT(ret
, high
);
1552 ret
= SELECT(highCmp
, high
, ret
);
1557 Value
*Builder::FCLAMP(Value
* src
, Value
* low
, Value
* high
)
1559 Value
*lowCmp
= FCMP_OLT(src
, low
);
1560 Value
*ret
= SELECT(lowCmp
, low
, src
);
1562 Value
*highCmp
= FCMP_OGT(ret
, high
);
1563 ret
= SELECT(highCmp
, high
, ret
);
1568 Value
*Builder::FCLAMP(Value
* src
, float low
, float high
)
1570 Value
* result
= VMAXPS(src
, VIMMED1(low
));
1571 result
= VMINPS(result
, VIMMED1(high
));
1576 //////////////////////////////////////////////////////////////////////////
1577 /// @brief save/restore stack, providing ability to push/pop the stack and
1578 /// reduce overall stack requirements for temporary stack use
1579 Value
* Builder::STACKSAVE()
1581 Function
* pfnStackSave
= Intrinsic::getDeclaration(JM()->mpCurrentModule
, Intrinsic::stacksave
);
1582 return CALLA(pfnStackSave
);
1585 void Builder::STACKRESTORE(Value
* pSaved
)
1587 Function
* pfnStackRestore
= Intrinsic::getDeclaration(JM()->mpCurrentModule
, Intrinsic::stackrestore
);
1588 CALL(pfnStackRestore
, std::initializer_list
<Value
*>{pSaved
});
1591 Value
*Builder::FMADDPS(Value
* a
, Value
* b
, Value
* c
)
1594 // use FMADs if available
1595 if(JM()->mArch
.AVX2())
1597 vOut
= VFMADDPS(a
, b
, c
);
1601 vOut
= FADD(FMUL(a
, b
), c
);
1606 Value
* Builder::POPCNT(Value
* a
)
1608 Function
* pCtPop
= Intrinsic::getDeclaration(JM()->mpCurrentModule
, Intrinsic::ctpop
, { a
->getType() });
1609 return CALL(pCtPop
, std::initializer_list
<Value
*>{a
});
1612 //////////////////////////////////////////////////////////////////////////
1613 /// @brief C functions called by LLVM IR
1614 //////////////////////////////////////////////////////////////////////////
1616 //////////////////////////////////////////////////////////////////////////
1617 /// @brief called in JIT code, inserted by PRINT
1618 /// output to both stdout and visual studio debug console
1619 void __cdecl
CallPrint(const char* fmt
, ...)
1622 va_start(args
, fmt
);
1625 #if defined( _WIN32 )
1627 vsnprintf_s(strBuf
, _TRUNCATE
, fmt
, args
);
1628 OutputDebugStringA(strBuf
);
1634 Value
*Builder::VEXTRACTI128(Value
* a
, Constant
* imm8
)
1636 bool flag
= !imm8
->isZeroValue();
1637 SmallVector
<Constant
*,8> idx
;
1638 for (unsigned i
= 0; i
< mVWidth
/ 2; i
++) {
1639 idx
.push_back(C(flag
? i
+ mVWidth
/ 2 : i
));
1641 return VSHUFFLE(a
, VUNDEF_I(), ConstantVector::get(idx
));
1644 Value
*Builder::VINSERTI128(Value
* a
, Value
* b
, Constant
* imm8
)
1646 bool flag
= !imm8
->isZeroValue();
1647 SmallVector
<Constant
*,8> idx
;
1648 for (unsigned i
= 0; i
< mVWidth
; i
++) {
1649 idx
.push_back(C(i
));
1651 Value
*inter
= VSHUFFLE(b
, VUNDEF_I(), ConstantVector::get(idx
));
1653 SmallVector
<Constant
*,8> idx2
;
1654 for (unsigned i
= 0; i
< mVWidth
/ 2; i
++) {
1655 idx2
.push_back(C(flag
? i
: i
+ mVWidth
));
1657 for (unsigned i
= mVWidth
/ 2; i
< mVWidth
; i
++) {
1658 idx2
.push_back(C(flag
? i
+ mVWidth
/ 2 : i
));
1660 return VSHUFFLE(a
, inter
, ConstantVector::get(idx2
));
1663 // rdtsc buckets macros
1664 void Builder::RDTSC_START(Value
* pBucketMgr
, Value
* pId
)
1666 // @todo due to an issue with thread local storage propagation in llvm, we can only safely call into
1667 // buckets framework when single threaded
1668 if (KNOB_SINGLE_THREADED
)
1670 std::vector
<Type
*> args
{
1671 PointerType::get(mInt32Ty
, 0), // pBucketMgr
1675 FunctionType
* pFuncTy
= FunctionType::get(Type::getVoidTy(JM()->mContext
), args
, false);
1676 Function
* pFunc
= cast
<Function
>(JM()->mpCurrentModule
->getOrInsertFunction("BucketManager_StartBucket", pFuncTy
));
1677 if (sys::DynamicLibrary::SearchForAddressOfSymbol("BucketManager_StartBucket") == nullptr)
1679 sys::DynamicLibrary::AddSymbol("BucketManager_StartBucket", (void*)&BucketManager_StartBucket
);
1682 CALL(pFunc
, { pBucketMgr
, pId
});
1686 void Builder::RDTSC_STOP(Value
* pBucketMgr
, Value
* pId
)
1688 // @todo due to an issue with thread local storage propagation in llvm, we can only safely call into
1689 // buckets framework when single threaded
1690 if (KNOB_SINGLE_THREADED
)
1692 std::vector
<Type
*> args
{
1693 PointerType::get(mInt32Ty
, 0), // pBucketMgr
1697 FunctionType
* pFuncTy
= FunctionType::get(Type::getVoidTy(JM()->mContext
), args
, false);
1698 Function
* pFunc
= cast
<Function
>(JM()->mpCurrentModule
->getOrInsertFunction("BucketManager_StopBucket", pFuncTy
));
1699 if (sys::DynamicLibrary::SearchForAddressOfSymbol("BucketManager_StopBucket") == nullptr)
1701 sys::DynamicLibrary::AddSymbol("BucketManager_StopBucket", (void*)&BucketManager_StopBucket
);
1704 CALL(pFunc
, { pBucketMgr
, pId
});
1709 uint32_t Builder::GetTypeSize(Type
* pType
)
1711 if (pType
->isStructTy())
1713 uint32_t numElems
= pType
->getStructNumElements();
1714 Type
* pElemTy
= pType
->getStructElementType(0);
1715 return numElems
* GetTypeSize(pElemTy
);
1718 if (pType
->isArrayTy())
1720 uint32_t numElems
= pType
->getArrayNumElements();
1721 Type
* pElemTy
= pType
->getArrayElementType();
1722 return numElems
* GetTypeSize(pElemTy
);
1725 if (pType
->isIntegerTy())
1727 uint32_t bitSize
= pType
->getIntegerBitWidth();
1731 if (pType
->isFloatTy())
1736 if (pType
->isHalfTy())
1741 if (pType
->isDoubleTy())
1746 SWR_ASSERT(false, "Unimplemented type.");