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 Convert32To16Float(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 ConvertSmallFloatTo32(UINT 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(uint32_t i
)
201 return ConstantVector::getSplat(mVWidth
, cast
<ConstantInt
>(C(i
)));
204 Value
*Builder::VIMMED1(float i
)
206 return ConstantVector::getSplat(mVWidth
, cast
<ConstantFP
>(C(i
)));
209 Value
*Builder::VIMMED1(bool i
)
211 return ConstantVector::getSplat(mVWidth
, cast
<ConstantInt
>(C(i
)));
214 Value
*Builder::VUNDEF_IPTR()
216 return UndefValue::get(VectorType::get(mInt32PtrTy
,mVWidth
));
219 Value
*Builder::VUNDEF_I()
221 return UndefValue::get(VectorType::get(mInt32Ty
, mVWidth
));
224 Value
*Builder::VUNDEF(Type
*ty
, uint32_t size
)
226 return UndefValue::get(VectorType::get(ty
, size
));
229 Value
*Builder::VUNDEF_F()
231 return UndefValue::get(VectorType::get(mFP32Ty
, mVWidth
));
234 Value
*Builder::VUNDEF(Type
* t
)
236 return UndefValue::get(VectorType::get(t
, mVWidth
));
239 #if HAVE_LLVM == 0x306
240 Value
*Builder::VINSERT(Value
*vec
, Value
*val
, uint64_t index
)
242 return VINSERT(vec
, val
, C((int64_t)index
));
246 Value
*Builder::VBROADCAST(Value
*src
)
248 // check if src is already a vector
249 if (src
->getType()->isVectorTy())
254 return VECTOR_SPLAT(mVWidth
, src
);
257 uint32_t Builder::IMMED(Value
* v
)
259 SWR_ASSERT(isa
<ConstantInt
>(v
));
260 ConstantInt
*pValConst
= cast
<ConstantInt
>(v
);
261 return pValConst
->getZExtValue();
264 int32_t Builder::S_IMMED(Value
* v
)
266 SWR_ASSERT(isa
<ConstantInt
>(v
));
267 ConstantInt
*pValConst
= cast
<ConstantInt
>(v
);
268 return pValConst
->getSExtValue();
271 Value
*Builder::GEP(Value
* ptr
, const std::initializer_list
<Value
*> &indexList
)
273 std::vector
<Value
*> indices
;
274 for (auto i
: indexList
)
275 indices
.push_back(i
);
276 return GEPA(ptr
, indices
);
279 Value
*Builder::GEP(Value
* ptr
, const std::initializer_list
<uint32_t> &indexList
)
281 std::vector
<Value
*> indices
;
282 for (auto i
: indexList
)
283 indices
.push_back(C(i
));
284 return GEPA(ptr
, indices
);
287 LoadInst
*Builder::LOAD(Value
*basePtr
, const std::initializer_list
<uint32_t> &indices
, const llvm::Twine
& name
)
289 std::vector
<Value
*> valIndices
;
290 for (auto i
: indices
)
291 valIndices
.push_back(C(i
));
292 return LOAD(GEPA(basePtr
, valIndices
), name
);
295 LoadInst
*Builder::LOADV(Value
*basePtr
, const std::initializer_list
<Value
*> &indices
, const llvm::Twine
& name
)
297 std::vector
<Value
*> valIndices
;
298 for (auto i
: indices
)
299 valIndices
.push_back(i
);
300 return LOAD(GEPA(basePtr
, valIndices
), name
);
303 StoreInst
*Builder::STORE(Value
*val
, Value
*basePtr
, const std::initializer_list
<uint32_t> &indices
)
305 std::vector
<Value
*> valIndices
;
306 for (auto i
: indices
)
307 valIndices
.push_back(C(i
));
308 return STORE(val
, GEPA(basePtr
, valIndices
));
311 StoreInst
*Builder::STOREV(Value
*val
, Value
*basePtr
, const std::initializer_list
<Value
*> &indices
)
313 std::vector
<Value
*> valIndices
;
314 for (auto i
: indices
)
315 valIndices
.push_back(i
);
316 return STORE(val
, GEPA(basePtr
, valIndices
));
319 CallInst
*Builder::CALL(Value
*Callee
, const std::initializer_list
<Value
*> &argsList
)
321 std::vector
<Value
*> args
;
322 for (auto arg
: argsList
)
324 return CALLA(Callee
, args
);
327 #if HAVE_LLVM > 0x306
328 CallInst
*Builder::CALL(Value
*Callee
, Value
* arg
)
330 std::vector
<Value
*> args
;
332 return CALLA(Callee
, args
);
335 CallInst
*Builder::CALL2(Value
*Callee
, Value
* arg1
, Value
* arg2
)
337 std::vector
<Value
*> args
;
338 args
.push_back(arg1
);
339 args
.push_back(arg2
);
340 return CALLA(Callee
, args
);
343 CallInst
*Builder::CALL3(Value
*Callee
, Value
* arg1
, Value
* arg2
, Value
* arg3
)
345 std::vector
<Value
*> args
;
346 args
.push_back(arg1
);
347 args
.push_back(arg2
);
348 args
.push_back(arg3
);
349 return CALLA(Callee
, args
);
353 //////////////////////////////////////////////////////////////////////////
354 Value
*Builder::DEBUGTRAP()
356 Function
*func
= Intrinsic::getDeclaration(JM()->mpCurrentModule
, Intrinsic::debugtrap
);
360 Value
*Builder::VRCP(Value
*va
)
362 return FDIV(VIMMED1(1.0f
), va
); // 1 / a
365 Value
*Builder::VPLANEPS(Value
* vA
, Value
* vB
, Value
* vC
, Value
* &vX
, Value
* &vY
)
367 Value
* vOut
= FMADDPS(vA
, vX
, vC
);
368 vOut
= FMADDPS(vB
, vY
, vOut
);
372 //////////////////////////////////////////////////////////////////////////
373 /// @brief Generate an i32 masked load operation in LLVM IR. If not
374 /// supported on the underlying platform, emulate it with float masked load
375 /// @param src - base address pointer for the load
376 /// @param vMask - SIMD wide mask that controls whether to access memory load 0
377 Value
*Builder::MASKLOADD(Value
* src
,Value
* mask
)
380 // use avx2 gather instruction is available
381 if(JM()->mArch
.AVX2())
383 Function
*func
= Intrinsic::getDeclaration(JM()->mpCurrentModule
, Intrinsic::x86_avx2_maskload_d_256
);
384 vResult
= CALL(func
,{src
,mask
});
388 // maskload intrinsic expects integer mask operand in llvm >= 3.8
389 #if (LLVM_VERSION_MAJOR > 3) || (LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 8)
390 mask
= BITCAST(mask
,VectorType::get(mInt32Ty
,mVWidth
));
392 mask
= BITCAST(mask
,VectorType::get(mFP32Ty
,mVWidth
));
394 Function
*func
= Intrinsic::getDeclaration(JM()->mpCurrentModule
,Intrinsic::x86_avx_maskload_ps_256
);
395 vResult
= BITCAST(CALL(func
,{src
,mask
}), VectorType::get(mInt32Ty
,mVWidth
));
400 //////////////////////////////////////////////////////////////////////////
401 /// @brief insert a JIT call to CallPrint
402 /// - outputs formatted string to both stdout and VS output window
403 /// - DEBUG builds only
405 /// PRINT("index %d = 0x%p\n",{C(lane), pIndex});
406 /// where C(lane) creates a constant value to print, and pIndex is the Value*
407 /// result from a GEP, printing out the pointer to memory
408 /// @param printStr - constant string to print, which includes format specifiers
409 /// @param printArgs - initializer list of Value*'s to print to std out
410 CallInst
*Builder::PRINT(const std::string
&printStr
,const std::initializer_list
<Value
*> &printArgs
)
412 // push the arguments to CallPrint into a vector
413 std::vector
<Value
*> printCallArgs
;
414 // save room for the format string. we still need to modify it for vectors
415 printCallArgs
.resize(1);
417 // search through the format string for special processing
419 std::string
tempStr(printStr
);
420 pos
= tempStr
.find('%', pos
);
421 auto v
= printArgs
.begin();
423 while ((pos
!= std::string::npos
) && (v
!= printArgs
.end()))
426 Type
* pType
= pArg
->getType();
428 if (pType
->isVectorTy())
430 Type
* pContainedType
= pType
->getContainedType(0);
432 if (toupper(tempStr
[pos
+ 1]) == 'X')
435 tempStr
[pos
+ 1] = 'x';
436 tempStr
.insert(pos
+ 2, "%08X ");
439 printCallArgs
.push_back(VEXTRACT(pArg
, C(0)));
441 std::string vectorFormatStr
;
442 for (uint32_t i
= 1; i
< pType
->getVectorNumElements(); ++i
)
444 vectorFormatStr
+= "0x%08X ";
445 printCallArgs
.push_back(VEXTRACT(pArg
, C(i
)));
448 tempStr
.insert(pos
, vectorFormatStr
);
449 pos
+= vectorFormatStr
.size();
451 else if ((tempStr
[pos
+ 1] == 'f') && (pContainedType
->isFloatTy()))
454 for (; i
< (pArg
->getType()->getVectorNumElements()) - 1; i
++)
456 tempStr
.insert(pos
, std::string("%f "));
458 printCallArgs
.push_back(FP_EXT(VEXTRACT(pArg
, C(i
)), Type::getDoubleTy(JM()->mContext
)));
460 printCallArgs
.push_back(FP_EXT(VEXTRACT(pArg
, C(i
)), Type::getDoubleTy(JM()->mContext
)));
462 else if ((tempStr
[pos
+ 1] == 'd') && (pContainedType
->isIntegerTy()))
465 for (; i
< (pArg
->getType()->getVectorNumElements()) - 1; i
++)
467 tempStr
.insert(pos
, std::string("%d "));
469 printCallArgs
.push_back(VEXTRACT(pArg
, C(i
)));
471 printCallArgs
.push_back(VEXTRACT(pArg
, C(i
)));
476 if (toupper(tempStr
[pos
+ 1]) == 'X')
479 tempStr
.insert(pos
+ 1, "x%08");
480 printCallArgs
.push_back(pArg
);
483 // for %f we need to cast float Values to doubles so that they print out correctly
484 else if ((tempStr
[pos
+ 1] == 'f') && (pType
->isFloatTy()))
486 printCallArgs
.push_back(FP_EXT(pArg
, Type::getDoubleTy(JM()->mContext
)));
491 printCallArgs
.push_back(pArg
);
495 // advance to the next arguement
497 pos
= tempStr
.find('%', ++pos
);
500 // create global variable constant string
501 Constant
*constString
= ConstantDataArray::getString(JM()->mContext
,tempStr
,true);
502 GlobalVariable
*gvPtr
= new GlobalVariable(constString
->getType(),true,GlobalValue::InternalLinkage
,constString
,"printStr");
503 JM()->mpCurrentModule
->getGlobalList().push_back(gvPtr
);
505 // get a pointer to the first character in the constant string array
506 std::vector
<Constant
*> geplist
{C(0),C(0)};
507 #if HAVE_LLVM == 0x306
508 Constant
*strGEP
= ConstantExpr::getGetElementPtr(gvPtr
,geplist
,false);
510 Constant
*strGEP
= ConstantExpr::getGetElementPtr(nullptr, gvPtr
,geplist
,false);
513 // insert the pointer to the format string in the argument vector
514 printCallArgs
[0] = strGEP
;
516 // get pointer to CallPrint function and insert decl into the module if needed
517 std::vector
<Type
*> args
;
518 args
.push_back(PointerType::get(mInt8Ty
,0));
519 FunctionType
* callPrintTy
= FunctionType::get(Type::getVoidTy(JM()->mContext
),args
,true);
520 Function
*callPrintFn
= cast
<Function
>(JM()->mpCurrentModule
->getOrInsertFunction("CallPrint", callPrintTy
));
522 // if we haven't yet added the symbol to the symbol table
523 if((sys::DynamicLibrary::SearchForAddressOfSymbol("CallPrint")) == nullptr)
525 sys::DynamicLibrary::AddSymbol("CallPrint", (void *)&CallPrint
);
528 // insert a call to CallPrint
529 return CALLA(callPrintFn
,printCallArgs
);
532 //////////////////////////////////////////////////////////////////////////
533 /// @brief Wrapper around PRINT with initializer list.
534 CallInst
* Builder::PRINT(const std::string
&printStr
)
536 return PRINT(printStr
, {});
539 //////////////////////////////////////////////////////////////////////////
540 /// @brief Generate a masked gather operation in LLVM IR. If not
541 /// supported on the underlying platform, emulate it with loads
542 /// @param vSrc - SIMD wide value that will be loaded if mask is invalid
543 /// @param pBase - Int8* base VB address pointer value
544 /// @param vIndices - SIMD wide value of VB byte offsets
545 /// @param vMask - SIMD wide mask that controls whether to access memory or the src values
546 /// @param scale - value to scale indices by
547 Value
*Builder::GATHERPS(Value
* vSrc
, Value
* pBase
, Value
* vIndices
, Value
* vMask
, Value
* scale
)
551 // use avx2 gather instruction if available
552 if(JM()->mArch
.AVX2())
554 // force mask to <N x float>, required by vgather
555 vMask
= BITCAST(vMask
, mSimdFP32Ty
);
556 vGather
= VGATHERPS(vSrc
,pBase
,vIndices
,vMask
,scale
);
560 Value
* pStack
= STACKSAVE();
562 // store vSrc on the stack. this way we can select between a valid load address and the vSrc address
563 Value
* vSrcPtr
= ALLOCA(vSrc
->getType());
564 STORE(vSrc
, vSrcPtr
);
566 vGather
= VUNDEF_F();
567 Value
*vScaleVec
= VBROADCAST(Z_EXT(scale
,mInt32Ty
));
568 Value
*vOffsets
= MUL(vIndices
,vScaleVec
);
569 Value
*mask
= MASK(vMask
);
570 for(uint32_t i
= 0; i
< mVWidth
; ++i
)
572 // single component byte index
573 Value
*offset
= VEXTRACT(vOffsets
,C(i
));
574 // byte pointer to component
575 Value
*loadAddress
= GEP(pBase
,offset
);
576 loadAddress
= BITCAST(loadAddress
,PointerType::get(mFP32Ty
,0));
577 // pointer to the value to load if we're masking off a component
578 Value
*maskLoadAddress
= GEP(vSrcPtr
,{C(0), C(i
)});
579 Value
*selMask
= VEXTRACT(mask
,C(i
));
580 // switch in a safe address to load if we're trying to access a vertex
581 Value
*validAddress
= SELECT(selMask
, loadAddress
, maskLoadAddress
);
582 Value
*val
= LOAD(validAddress
);
583 vGather
= VINSERT(vGather
,val
,C(i
));
585 STACKRESTORE(pStack
);
591 //////////////////////////////////////////////////////////////////////////
592 /// @brief Generate a masked gather operation in LLVM IR. If not
593 /// supported on the underlying platform, emulate it with loads
594 /// @param vSrc - SIMD wide value that will be loaded if mask is invalid
595 /// @param pBase - Int8* base VB address pointer value
596 /// @param vIndices - SIMD wide value of VB byte offsets
597 /// @param vMask - SIMD wide mask that controls whether to access memory or the src values
598 /// @param scale - value to scale indices by
599 Value
*Builder::GATHERDD(Value
* vSrc
, Value
* pBase
, Value
* vIndices
, Value
* vMask
, Value
* scale
)
603 // use avx2 gather instruction if available
604 if(JM()->mArch
.AVX2())
606 vGather
= VGATHERDD(vSrc
, pBase
, vIndices
, vMask
, scale
);
610 Value
* pStack
= STACKSAVE();
612 // store vSrc on the stack. this way we can select between a valid load address and the vSrc address
613 Value
* vSrcPtr
= ALLOCA(vSrc
->getType());
614 STORE(vSrc
, vSrcPtr
);
616 vGather
= VUNDEF_I();
617 Value
*vScaleVec
= VBROADCAST(Z_EXT(scale
, mInt32Ty
));
618 Value
*vOffsets
= MUL(vIndices
, vScaleVec
);
619 Value
*mask
= MASK(vMask
);
620 for(uint32_t i
= 0; i
< mVWidth
; ++i
)
622 // single component byte index
623 Value
*offset
= VEXTRACT(vOffsets
, C(i
));
624 // byte pointer to component
625 Value
*loadAddress
= GEP(pBase
, offset
);
626 loadAddress
= BITCAST(loadAddress
, PointerType::get(mInt32Ty
, 0));
627 // pointer to the value to load if we're masking off a component
628 Value
*maskLoadAddress
= GEP(vSrcPtr
, {C(0), C(i
)});
629 Value
*selMask
= VEXTRACT(mask
, C(i
));
630 // switch in a safe address to load if we're trying to access a vertex
631 Value
*validAddress
= SELECT(selMask
, loadAddress
, maskLoadAddress
);
632 Value
*val
= LOAD(validAddress
, C(0));
633 vGather
= VINSERT(vGather
, val
, C(i
));
636 STACKRESTORE(pStack
);
641 //////////////////////////////////////////////////////////////////////////
642 /// @brief Generate a masked gather operation in LLVM IR. If not
643 /// supported on the underlying platform, emulate it with loads
644 /// @param vSrc - SIMD wide value that will be loaded if mask is invalid
645 /// @param pBase - Int8* base VB address pointer value
646 /// @param vIndices - SIMD wide value of VB byte offsets
647 /// @param vMask - SIMD wide mask that controls whether to access memory or the src values
648 /// @param scale - value to scale indices by
649 Value
*Builder::GATHERPD(Value
* vSrc
, Value
* pBase
, Value
* vIndices
, Value
* vMask
, Value
* scale
)
653 // use avx2 gather instruction if available
654 if(JM()->mArch
.AVX2())
656 vGather
= VGATHERPD(vSrc
, pBase
, vIndices
, vMask
, scale
);
660 Value
* pStack
= STACKSAVE();
662 // store vSrc on the stack. this way we can select between a valid load address and the vSrc address
663 Value
* vSrcPtr
= ALLOCA(vSrc
->getType());
664 STORE(vSrc
, vSrcPtr
);
666 vGather
= UndefValue::get(VectorType::get(mDoubleTy
, 4));
667 Value
*vScaleVec
= VECTOR_SPLAT(4, Z_EXT(scale
,mInt32Ty
));
668 Value
*vOffsets
= MUL(vIndices
,vScaleVec
);
669 Value
*mask
= MASK(vMask
);
670 for(uint32_t i
= 0; i
< mVWidth
/2; ++i
)
672 // single component byte index
673 Value
*offset
= VEXTRACT(vOffsets
,C(i
));
674 // byte pointer to component
675 Value
*loadAddress
= GEP(pBase
,offset
);
676 loadAddress
= BITCAST(loadAddress
,PointerType::get(mDoubleTy
,0));
677 // pointer to the value to load if we're masking off a component
678 Value
*maskLoadAddress
= GEP(vSrcPtr
,{C(0), C(i
)});
679 Value
*selMask
= VEXTRACT(mask
,C(i
));
680 // switch in a safe address to load if we're trying to access a vertex
681 Value
*validAddress
= SELECT(selMask
, loadAddress
, maskLoadAddress
);
682 Value
*val
= LOAD(validAddress
);
683 vGather
= VINSERT(vGather
,val
,C(i
));
685 STACKRESTORE(pStack
);
690 //////////////////////////////////////////////////////////////////////////
691 /// @brief convert x86 <N x float> mask to llvm <N x i1> mask
692 Value
* Builder::MASK(Value
* vmask
)
694 Value
* src
= BITCAST(vmask
, mSimdInt32Ty
);
695 return ICMP_SLT(src
, VIMMED1(0));
698 //////////////////////////////////////////////////////////////////////////
699 /// @brief convert llvm <N x i1> mask to x86 <N x i32> mask
700 Value
* Builder::VMASK(Value
* mask
)
702 return S_EXT(mask
, mSimdInt32Ty
);
705 //////////////////////////////////////////////////////////////////////////
706 /// @brief Generate a VPSHUFB operation in LLVM IR. If not
707 /// supported on the underlying platform, emulate it
708 /// @param a - 256bit SIMD(32x8bit) of 8bit integer values
709 /// @param b - 256bit SIMD(32x8bit) of 8bit integer mask values
710 /// Byte masks in lower 128 lane of b selects 8 bit values from lower
711 /// 128bits of a, and vice versa for the upper lanes. If the mask
712 /// value is negative, '0' is inserted.
713 Value
*Builder::PSHUFB(Value
* a
, Value
* b
)
716 // use avx2 pshufb instruction if available
717 if(JM()->mArch
.AVX2())
723 Constant
* cB
= dyn_cast
<Constant
>(b
);
724 // number of 8 bit elements in b
725 uint32_t numElms
= cast
<VectorType
>(cB
->getType())->getNumElements();
727 Value
* vShuf
= UndefValue::get(VectorType::get(mInt8Ty
, numElms
));
729 // insert an 8 bit value from the high and low lanes of a per loop iteration
731 for(uint32_t i
= 0; i
< numElms
; i
++)
733 ConstantInt
* cLow128b
= cast
<ConstantInt
>(cB
->getAggregateElement(i
));
734 ConstantInt
* cHigh128b
= cast
<ConstantInt
>(cB
->getAggregateElement(i
+ numElms
));
736 // extract values from constant mask
737 char valLow128bLane
= (char)(cLow128b
->getSExtValue());
738 char valHigh128bLane
= (char)(cHigh128b
->getSExtValue());
740 Value
* insertValLow128b
;
741 Value
* insertValHigh128b
;
743 // if the mask value is negative, insert a '0' in the respective output position
744 // otherwise, lookup the value at mask position (bits 3..0 of the respective mask byte) in a and insert in output vector
745 insertValLow128b
= (valLow128bLane
< 0) ? C((char)0) : VEXTRACT(a
, C((valLow128bLane
& 0xF)));
746 insertValHigh128b
= (valHigh128bLane
< 0) ? C((char)0) : VEXTRACT(a
, C((valHigh128bLane
& 0xF) + numElms
));
748 vShuf
= VINSERT(vShuf
, insertValLow128b
, i
);
749 vShuf
= VINSERT(vShuf
, insertValHigh128b
, (i
+ numElms
));
756 //////////////////////////////////////////////////////////////////////////
757 /// @brief Generate a VPSHUFB operation (sign extend 8 8bit values to 32
758 /// bits)in LLVM IR. If not supported on the underlying platform, emulate it
759 /// @param a - 128bit SIMD lane(16x8bit) of 8bit integer values. Only
760 /// lower 8 values are used.
761 Value
*Builder::PMOVSXBD(Value
* a
)
763 // llvm-3.9 removed the pmovsxbd intrinsic
764 #if HAVE_LLVM < 0x309
765 // use avx2 byte sign extend instruction if available
766 if(JM()->mArch
.AVX2())
768 Function
*pmovsxbd
= Intrinsic::getDeclaration(JM()->mpCurrentModule
, Intrinsic::x86_avx2_pmovsxbd
);
769 return CALL(pmovsxbd
, std::initializer_list
<Value
*>{a
});
774 // VPMOVSXBD output type
775 Type
* v8x32Ty
= VectorType::get(mInt32Ty
, 8);
776 // Extract 8 values from 128bit lane and sign extend
777 return S_EXT(VSHUFFLE(a
, a
, C
<int>({0, 1, 2, 3, 4, 5, 6, 7})), v8x32Ty
);
781 //////////////////////////////////////////////////////////////////////////
782 /// @brief Generate a VPSHUFB operation (sign extend 8 16bit values to 32
783 /// bits)in LLVM IR. If not supported on the underlying platform, emulate it
784 /// @param a - 128bit SIMD lane(8x16bit) of 16bit integer values.
785 Value
*Builder::PMOVSXWD(Value
* a
)
787 // llvm-3.9 removed the pmovsxwd intrinsic
788 #if HAVE_LLVM < 0x309
789 // use avx2 word sign extend if available
790 if(JM()->mArch
.AVX2())
792 Function
*pmovsxwd
= Intrinsic::getDeclaration(JM()->mpCurrentModule
, Intrinsic::x86_avx2_pmovsxwd
);
793 return CALL(pmovsxwd
, std::initializer_list
<Value
*>{a
});
798 // VPMOVSXWD output type
799 Type
* v8x32Ty
= VectorType::get(mInt32Ty
, 8);
800 // Extract 8 values from 128bit lane and sign extend
801 return S_EXT(VSHUFFLE(a
, a
, C
<int>({0, 1, 2, 3, 4, 5, 6, 7})), v8x32Ty
);
805 //////////////////////////////////////////////////////////////////////////
806 /// @brief Generate a VPERMD operation (shuffle 32 bit integer values
807 /// across 128 bit lanes) in LLVM IR. If not supported on the underlying
808 /// platform, emulate it
809 /// @param a - 256bit SIMD lane(8x32bit) of integer values.
810 /// @param idx - 256bit SIMD lane(8x32bit) of 3 bit lane index values
811 Value
*Builder::PERMD(Value
* a
, Value
* idx
)
814 // use avx2 permute instruction if available
815 if(JM()->mArch
.AVX2())
817 res
= VPERMD(a
, idx
);
821 if (isa
<Constant
>(idx
))
823 res
= VSHUFFLE(a
, a
, idx
);
828 for (uint32_t l
= 0; l
< JM()->mVWidth
; ++l
)
830 Value
* pIndex
= VEXTRACT(idx
, C(l
));
831 Value
* pVal
= VEXTRACT(a
, pIndex
);
832 res
= VINSERT(res
, pVal
, C(l
));
839 //////////////////////////////////////////////////////////////////////////
840 /// @brief Generate a VPERMPS operation (shuffle 32 bit float values
841 /// across 128 bit lanes) in LLVM IR. If not supported on the underlying
842 /// platform, emulate it
843 /// @param a - 256bit SIMD lane(8x32bit) of float values.
844 /// @param idx - 256bit SIMD lane(8x32bit) of 3 bit lane index values
845 Value
*Builder::PERMPS(Value
* a
, Value
* idx
)
848 // use avx2 permute instruction if available
849 if (JM()->mArch
.AVX2())
851 // llvm 3.6.0 swapped the order of the args to vpermd
852 res
= VPERMPS(idx
, a
);
856 if (isa
<Constant
>(idx
))
858 res
= VSHUFFLE(a
, a
, idx
);
863 for (uint32_t l
= 0; l
< JM()->mVWidth
; ++l
)
865 Value
* pIndex
= VEXTRACT(idx
, C(l
));
866 Value
* pVal
= VEXTRACT(a
, pIndex
);
867 res
= VINSERT(res
, pVal
, C(l
));
875 //////////////////////////////////////////////////////////////////////////
876 /// @brief Generate a VCVTPH2PS operation (float16->float32 conversion)
877 /// in LLVM IR. If not supported on the underlying platform, emulate it
878 /// @param a - 128bit SIMD lane(8x16bit) of float16 in int16 format.
879 Value
*Builder::CVTPH2PS(Value
* a
)
881 if (JM()->mArch
.F16C())
887 FunctionType
* pFuncTy
= FunctionType::get(mFP32Ty
, mInt16Ty
);
888 Function
* pCvtPh2Ps
= cast
<Function
>(JM()->mpCurrentModule
->getOrInsertFunction("ConvertSmallFloatTo32", pFuncTy
));
890 if (sys::DynamicLibrary::SearchForAddressOfSymbol("ConvertSmallFloatTo32") == nullptr)
892 sys::DynamicLibrary::AddSymbol("ConvertSmallFloatTo32", (void *)&ConvertSmallFloatTo32
);
895 Value
* pResult
= UndefValue::get(mSimdFP32Ty
);
896 for (uint32_t i
= 0; i
< mVWidth
; ++i
)
898 Value
* pSrc
= VEXTRACT(a
, C(i
));
899 Value
* pConv
= CALL(pCvtPh2Ps
, std::initializer_list
<Value
*>{pSrc
});
900 pResult
= VINSERT(pResult
, pConv
, C(i
));
907 //////////////////////////////////////////////////////////////////////////
908 /// @brief Generate a VCVTPS2PH operation (float32->float16 conversion)
909 /// in LLVM IR. If not supported on the underlying platform, emulate it
910 /// @param a - 128bit SIMD lane(8x16bit) of float16 in int16 format.
911 Value
*Builder::CVTPS2PH(Value
* a
, Value
* rounding
)
913 if (JM()->mArch
.F16C())
915 return VCVTPS2PH(a
, rounding
);
919 // call scalar C function for now
920 FunctionType
* pFuncTy
= FunctionType::get(mInt16Ty
, mFP32Ty
);
921 Function
* pCvtPs2Ph
= cast
<Function
>(JM()->mpCurrentModule
->getOrInsertFunction("Convert32To16Float", pFuncTy
));
923 if (sys::DynamicLibrary::SearchForAddressOfSymbol("Convert32To16Float") == nullptr)
925 sys::DynamicLibrary::AddSymbol("Convert32To16Float", (void *)&Convert32To16Float
);
928 Value
* pResult
= UndefValue::get(mSimdInt16Ty
);
929 for (uint32_t i
= 0; i
< mVWidth
; ++i
)
931 Value
* pSrc
= VEXTRACT(a
, C(i
));
932 Value
* pConv
= CALL(pCvtPs2Ph
, std::initializer_list
<Value
*>{pSrc
});
933 pResult
= VINSERT(pResult
, pConv
, C(i
));
940 Value
*Builder::PMAXSD(Value
* a
, Value
* b
)
942 // llvm-3.9 removed the pmax intrinsics
943 #if HAVE_LLVM >= 0x309
944 Value
* cmp
= ICMP_SGT(a
, b
);
945 return SELECT(cmp
, a
, b
);
947 if (JM()->mArch
.AVX2())
949 Function
* pmaxsd
= Intrinsic::getDeclaration(JM()->mpCurrentModule
, Intrinsic::x86_avx2_pmaxs_d
);
950 return CALL(pmaxsd
, {a
, b
});
954 // use 4-wide sse max intrinsic on lower/upper halves of 8-wide sources
955 Function
* pmaxsd
= Intrinsic::getDeclaration(JM()->mpCurrentModule
, Intrinsic::x86_sse41_pmaxsd
);
958 Value
* aLo
= VEXTRACTI128(a
, C((uint8_t)0));
959 Value
* bLo
= VEXTRACTI128(b
, C((uint8_t)0));
960 Value
* resLo
= CALL(pmaxsd
, {aLo
, bLo
});
963 Value
* aHi
= VEXTRACTI128(a
, C((uint8_t)1));
964 Value
* bHi
= VEXTRACTI128(b
, C((uint8_t)1));
965 Value
* resHi
= CALL(pmaxsd
, {aHi
, bHi
});
968 Value
* result
= VINSERTI128(VUNDEF_I(), resLo
, C((uint8_t)0));
969 result
= VINSERTI128(result
, resHi
, C((uint8_t)1));
976 Value
*Builder::PMINSD(Value
* a
, Value
* b
)
978 // llvm-3.9 removed the pmin intrinsics
979 #if HAVE_LLVM >= 0x309
980 Value
* cmp
= ICMP_SLT(a
, b
);
981 return SELECT(cmp
, a
, b
);
983 if (JM()->mArch
.AVX2())
985 Function
* pminsd
= Intrinsic::getDeclaration(JM()->mpCurrentModule
, Intrinsic::x86_avx2_pmins_d
);
986 return CALL(pminsd
, {a
, b
});
990 // use 4-wide sse max intrinsic on lower/upper halves of 8-wide sources
991 Function
* pminsd
= Intrinsic::getDeclaration(JM()->mpCurrentModule
, Intrinsic::x86_sse41_pminsd
);
994 Value
* aLo
= VEXTRACTI128(a
, C((uint8_t)0));
995 Value
* bLo
= VEXTRACTI128(b
, C((uint8_t)0));
996 Value
* resLo
= CALL(pminsd
, {aLo
, bLo
});
999 Value
* aHi
= VEXTRACTI128(a
, C((uint8_t)1));
1000 Value
* bHi
= VEXTRACTI128(b
, C((uint8_t)1));
1001 Value
* resHi
= CALL(pminsd
, {aHi
, bHi
});
1004 Value
* result
= VINSERTI128(VUNDEF_I(), resLo
, C((uint8_t)0));
1005 result
= VINSERTI128(result
, resHi
, C((uint8_t)1));
1012 void Builder::Gather4(const SWR_FORMAT format
, Value
* pSrcBase
, Value
* byteOffsets
,
1013 Value
* mask
, Value
* vGatherComponents
[], bool bPackedOutput
)
1015 const SWR_FORMAT_INFO
&info
= GetFormatInfo(format
);
1016 if(info
.type
[0] == SWR_TYPE_FLOAT
&& info
.bpc
[0] == 32)
1018 // ensure our mask is the correct type
1019 mask
= BITCAST(mask
, mSimdFP32Ty
);
1020 GATHER4PS(info
, pSrcBase
, byteOffsets
, mask
, vGatherComponents
, bPackedOutput
);
1024 // ensure our mask is the correct type
1025 mask
= BITCAST(mask
, mSimdInt32Ty
);
1026 GATHER4DD(info
, pSrcBase
, byteOffsets
, mask
, vGatherComponents
, bPackedOutput
);
1030 void Builder::GATHER4PS(const SWR_FORMAT_INFO
&info
, Value
* pSrcBase
, Value
* byteOffsets
,
1031 Value
* mask
, Value
* vGatherComponents
[], bool bPackedOutput
)
1033 switch(info
.bpp
/ info
.numComps
)
1037 Value
* vGatherResult
[2];
1040 // TODO: vGatherMaskedVal
1041 Value
* vGatherMaskedVal
= VIMMED1((float)0);
1043 // always have at least one component out of x or y to fetch
1045 // save mask as it is zero'd out after each gather
1048 vGatherResult
[0] = GATHERPS(vGatherMaskedVal
, pSrcBase
, byteOffsets
, vMask
, C((char)1));
1049 // e.g. result of first 8x32bit integer gather for 16bit components
1050 // 256i - 0 1 2 3 4 5 6 7
1051 // xyxy xyxy xyxy xyxy xyxy xyxy xyxy xyxy
1054 // if we have at least one component out of x or y to fetch
1055 if(info
.numComps
> 2)
1057 // offset base to the next components(zw) in the vertex to gather
1058 pSrcBase
= GEP(pSrcBase
, C((char)4));
1061 vGatherResult
[1] = GATHERPS(vGatherMaskedVal
, pSrcBase
, byteOffsets
, vMask
, C((char)1));
1062 // e.g. result of second 8x32bit integer gather for 16bit components
1063 // 256i - 0 1 2 3 4 5 6 7
1064 // zwzw zwzw zwzw zwzw zwzw zwzw zwzw zwzw
1069 vGatherResult
[1] = vGatherMaskedVal
;
1072 // Shuffle gathered components into place, each row is a component
1073 Shuffle16bpcGather4(info
, vGatherResult
, vGatherComponents
, bPackedOutput
);
1079 for (uint32_t i
= 0; i
< 4; ++i
)
1081 vGatherComponents
[i
] = VIMMED1(*(float*)&info
.defaults
[i
]);
1084 for(uint32_t i
= 0; i
< info
.numComps
; i
++)
1086 uint32_t swizzleIndex
= info
.swizzle
[i
];
1088 // save mask as it is zero'd out after each gather
1089 Value
*vMask
= mask
;
1091 // Gather a SIMD of components
1092 vGatherComponents
[swizzleIndex
] = GATHERPS(vGatherComponents
[swizzleIndex
], pSrcBase
, byteOffsets
, vMask
, C((char)1));
1094 // offset base to the next component to gather
1095 pSrcBase
= GEP(pSrcBase
, C((char)4));
1100 SWR_INVALID("Invalid float format");
1105 void Builder::GATHER4DD(const SWR_FORMAT_INFO
&info
, Value
* pSrcBase
, Value
* byteOffsets
,
1106 Value
* mask
, Value
* vGatherComponents
[], bool bPackedOutput
)
1108 switch (info
.bpp
/ info
.numComps
)
1112 Value
* vGatherMaskedVal
= VIMMED1((int32_t)0);
1113 Value
* vGatherResult
= GATHERDD(vGatherMaskedVal
, pSrcBase
, byteOffsets
, mask
, C((char)1));
1114 // e.g. result of an 8x32bit integer gather for 8bit components
1115 // 256i - 0 1 2 3 4 5 6 7
1116 // xyzw xyzw xyzw xyzw xyzw xyzw xyzw xyzw
1118 Shuffle8bpcGather4(info
, vGatherResult
, vGatherComponents
, bPackedOutput
);
1123 Value
* vGatherResult
[2];
1126 // TODO: vGatherMaskedVal
1127 Value
* vGatherMaskedVal
= VIMMED1((int32_t)0);
1129 // always have at least one component out of x or y to fetch
1131 // save mask as it is zero'd out after each gather
1134 vGatherResult
[0] = GATHERDD(vGatherMaskedVal
, pSrcBase
, byteOffsets
, vMask
, C((char)1));
1135 // e.g. result of first 8x32bit integer gather for 16bit components
1136 // 256i - 0 1 2 3 4 5 6 7
1137 // xyxy xyxy xyxy xyxy xyxy xyxy xyxy xyxy
1140 // if we have at least one component out of x or y to fetch
1141 if(info
.numComps
> 2)
1143 // offset base to the next components(zw) in the vertex to gather
1144 pSrcBase
= GEP(pSrcBase
, C((char)4));
1147 vGatherResult
[1] = GATHERDD(vGatherMaskedVal
, pSrcBase
, byteOffsets
, vMask
, C((char)1));
1148 // e.g. result of second 8x32bit integer gather for 16bit components
1149 // 256i - 0 1 2 3 4 5 6 7
1150 // zwzw zwzw zwzw zwzw zwzw zwzw zwzw zwzw
1155 vGatherResult
[1] = vGatherMaskedVal
;
1158 // Shuffle gathered components into place, each row is a component
1159 Shuffle16bpcGather4(info
, vGatherResult
, vGatherComponents
, bPackedOutput
);
1166 for (uint32_t i
= 0; i
< 4; ++i
)
1168 vGatherComponents
[i
] = VIMMED1((int)info
.defaults
[i
]);
1171 for(uint32_t i
= 0; i
< info
.numComps
; i
++)
1173 uint32_t swizzleIndex
= info
.swizzle
[i
];
1175 // save mask as it is zero'd out after each gather
1176 Value
*vMask
= mask
;
1178 // Gather a SIMD of components
1179 vGatherComponents
[swizzleIndex
] = GATHERDD(vGatherComponents
[swizzleIndex
], pSrcBase
, byteOffsets
, vMask
, C((char)1));
1181 // offset base to the next component to gather
1182 pSrcBase
= GEP(pSrcBase
, C((char)4));
1187 SWR_INVALID("unsupported format");
1192 void Builder::Shuffle16bpcGather4(const SWR_FORMAT_INFO
&info
, Value
* vGatherInput
[2], Value
* vGatherOutput
[4], bool bPackedOutput
)
1195 Type
* vGatherTy
= VectorType::get(IntegerType::getInt32Ty(JM()->mContext
), mVWidth
);
1196 Type
* v32x8Ty
= VectorType::get(mInt8Ty
, mVWidth
* 4); // vwidth is units of 32 bits
1198 // input could either be float or int vector; do shuffle work in int
1199 vGatherInput
[0] = BITCAST(vGatherInput
[0], mSimdInt32Ty
);
1200 vGatherInput
[1] = BITCAST(vGatherInput
[1], mSimdInt32Ty
);
1204 Type
* v128bitTy
= VectorType::get(IntegerType::getIntNTy(JM()->mContext
, 128), mVWidth
/ 4); // vwidth is units of 32 bits
1207 Value
* vConstMask
= C
<char>({0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15,
1208 0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15});
1209 Value
* vShufResult
= BITCAST(PSHUFB(BITCAST(vGatherInput
[0], v32x8Ty
), vConstMask
), vGatherTy
);
1210 // after pshufb: group components together in each 128bit lane
1211 // 256i - 0 1 2 3 4 5 6 7
1212 // xxxx xxxx yyyy yyyy xxxx xxxx yyyy yyyy
1214 Value
* vi128XY
= BITCAST(PERMD(vShufResult
, C
<int32_t>({0, 1, 4, 5, 2, 3, 6, 7})), v128bitTy
);
1215 // after PERMD: move and pack xy components into each 128bit lane
1216 // 256i - 0 1 2 3 4 5 6 7
1217 // xxxx xxxx xxxx xxxx yyyy yyyy yyyy yyyy
1219 // do the same for zw components
1220 Value
* vi128ZW
= nullptr;
1221 if(info
.numComps
> 2)
1223 Value
* vShufResult
= BITCAST(PSHUFB(BITCAST(vGatherInput
[1], v32x8Ty
), vConstMask
), vGatherTy
);
1224 vi128ZW
= BITCAST(PERMD(vShufResult
, C
<int32_t>({0, 1, 4, 5, 2, 3, 6, 7})), v128bitTy
);
1227 for(uint32_t i
= 0; i
< 4; i
++)
1229 uint32_t swizzleIndex
= info
.swizzle
[i
];
1230 // todo: fixed for packed
1231 Value
* vGatherMaskedVal
= VIMMED1((int32_t)(info
.defaults
[i
]));
1232 if(i
>= info
.numComps
)
1234 // set the default component val
1235 vGatherOutput
[swizzleIndex
] = vGatherMaskedVal
;
1239 // if x or z, extract 128bits from lane 0, else for y or w, extract from lane 1
1240 uint32_t lane
= ((i
== 0) || (i
== 2)) ? 0 : 1;
1241 // if x or y, use vi128XY permute result, else use vi128ZW
1242 Value
* selectedPermute
= (i
< 2) ? vi128XY
: vi128ZW
;
1244 // extract packed component 128 bit lanes
1245 vGatherOutput
[swizzleIndex
] = VEXTRACT(selectedPermute
, C(lane
));
1251 // pshufb masks for each component
1252 Value
* vConstMask
[2];
1254 vConstMask
[0] = C
<char>({0, 1, -1, -1, 4, 5, -1, -1, 8, 9, -1, -1, 12, 13, -1, -1,
1255 0, 1, -1, -1, 4, 5, -1, -1, 8, 9, -1, -1, 12, 13, -1, -1, });
1258 vConstMask
[1] = C
<char>({2, 3, -1, -1, 6, 7, -1, -1, 10, 11, -1, -1, 14, 15, -1, -1,
1259 2, 3, -1, -1, 6, 7, -1, -1, 10, 11, -1, -1, 14, 15, -1, -1});
1262 // shuffle enabled components into lower word of each 32bit lane, 0 extending to 32 bits
1264 for (uint32_t i
= 0; i
< 4; ++i
)
1266 vGatherOutput
[i
] = VIMMED1((int32_t)info
.defaults
[i
]);
1269 for(uint32_t i
= 0; i
< info
.numComps
; i
++)
1271 uint32_t swizzleIndex
= info
.swizzle
[i
];
1273 // select correct constMask for x/z or y/w pshufb
1274 uint32_t selectedMask
= ((i
== 0) || (i
== 2)) ? 0 : 1;
1275 // if x or y, use vi128XY permute result, else use vi128ZW
1276 uint32_t selectedGather
= (i
< 2) ? 0 : 1;
1278 vGatherOutput
[swizzleIndex
] = BITCAST(PSHUFB(BITCAST(vGatherInput
[selectedGather
], v32x8Ty
), vConstMask
[selectedMask
]), vGatherTy
);
1279 // after pshufb mask for x channel; z uses the same shuffle from the second gather
1280 // 256i - 0 1 2 3 4 5 6 7
1281 // xx00 xx00 xx00 xx00 xx00 xx00 xx00 xx00
1286 void Builder::Shuffle8bpcGather4(const SWR_FORMAT_INFO
&info
, Value
* vGatherInput
, Value
* vGatherOutput
[], bool bPackedOutput
)
1289 Type
* vGatherTy
= VectorType::get(IntegerType::getInt32Ty(JM()->mContext
), mVWidth
);
1290 Type
* v32x8Ty
= VectorType::get(mInt8Ty
, mVWidth
* 4 ); // vwidth is units of 32 bits
1294 Type
* v128Ty
= VectorType::get(IntegerType::getIntNTy(JM()->mContext
, 128), mVWidth
/ 4); // vwidth is units of 32 bits
1296 Value
* vConstMask
= C
<char>({0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15,
1297 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15});
1298 Value
* vShufResult
= BITCAST(PSHUFB(BITCAST(vGatherInput
, v32x8Ty
), vConstMask
), vGatherTy
);
1299 // after pshufb: group components together in each 128bit lane
1300 // 256i - 0 1 2 3 4 5 6 7
1301 // xxxx yyyy zzzz wwww xxxx yyyy zzzz wwww
1303 Value
* vi128XY
= BITCAST(PERMD(vShufResult
, C
<int32_t>({0, 4, 0, 0, 1, 5, 0, 0})), v128Ty
);
1304 // after PERMD: move and pack xy and zw components in low 64 bits of each 128bit lane
1305 // 256i - 0 1 2 3 4 5 6 7
1306 // xxxx xxxx dcdc dcdc yyyy yyyy dcdc dcdc (dc - don't care)
1308 // do the same for zw components
1309 Value
* vi128ZW
= nullptr;
1310 if(info
.numComps
> 2)
1312 vi128ZW
= BITCAST(PERMD(vShufResult
, C
<int32_t>({2, 6, 0, 0, 3, 7, 0, 0})), v128Ty
);
1315 // sign extend all enabled components. If we have a fill vVertexElements, output to current simdvertex
1316 for(uint32_t i
= 0; i
< 4; i
++)
1318 uint32_t swizzleIndex
= info
.swizzle
[i
];
1319 // todo: fix for packed
1320 Value
* vGatherMaskedVal
= VIMMED1((int32_t)(info
.defaults
[i
]));
1321 if(i
>= info
.numComps
)
1323 // set the default component val
1324 vGatherOutput
[swizzleIndex
] = vGatherMaskedVal
;
1328 // if x or z, extract 128bits from lane 0, else for y or w, extract from lane 1
1329 uint32_t lane
= ((i
== 0) || (i
== 2)) ? 0 : 1;
1330 // if x or y, use vi128XY permute result, else use vi128ZW
1331 Value
* selectedPermute
= (i
< 2) ? vi128XY
: vi128ZW
;
1334 vGatherOutput
[swizzleIndex
] = VEXTRACT(selectedPermute
, C(lane
));
1339 // shuffle enabled components into lower byte of each 32bit lane, 0 extending to 32 bits
1341 for (uint32_t i
= 0; i
< 4; ++i
)
1343 vGatherOutput
[i
] = VIMMED1((int32_t)info
.defaults
[i
]);
1346 for(uint32_t i
= 0; i
< info
.numComps
; i
++){
1347 uint32_t swizzleIndex
= info
.swizzle
[i
];
1349 // pshufb masks for each component
1355 vConstMask
= C
<char>({0, -1, -1, -1, 4, -1, -1, -1, 8, -1, -1, -1, 12, -1, -1, -1,
1356 0, -1, -1, -1, 4, -1, -1, -1, 8, -1, -1, -1, 12, -1, -1, -1});
1360 vConstMask
= C
<char>({1, -1, -1, -1, 5, -1, -1, -1, 9, -1, -1, -1, 13, -1, -1, -1,
1361 1, -1, -1, -1, 5, -1, -1, -1, 9, -1, -1, -1, 13, -1, -1, -1});
1365 vConstMask
= C
<char>({2, -1, -1, -1, 6, -1, -1, -1, 10, -1, -1, -1, 14, -1, -1, -1,
1366 2, -1, -1, -1, 6, -1, -1, -1, 10, -1, -1, -1, 14, -1, -1, -1});
1370 vConstMask
= C
<char>({3, -1, -1, -1, 7, -1, -1, -1, 11, -1, -1, -1, 15, -1, -1, -1,
1371 3, -1, -1, -1, 7, -1, -1, -1, 11, -1, -1, -1, 15, -1, -1, -1});
1374 vConstMask
= nullptr;
1378 vGatherOutput
[swizzleIndex
] = BITCAST(PSHUFB(BITCAST(vGatherInput
, v32x8Ty
), vConstMask
), vGatherTy
);
1379 // after pshufb for x channel
1380 // 256i - 0 1 2 3 4 5 6 7
1381 // x000 x000 x000 x000 x000 x000 x000 x000
1386 // Helper function to create alloca in entry block of function
1387 Value
* Builder::CreateEntryAlloca(Function
* pFunc
, Type
* pType
)
1389 auto saveIP
= IRB()->saveIP();
1390 IRB()->SetInsertPoint(&pFunc
->getEntryBlock(),
1391 pFunc
->getEntryBlock().begin());
1392 Value
* pAlloca
= ALLOCA(pType
);
1393 IRB()->restoreIP(saveIP
);
1397 //////////////////////////////////////////////////////////////////////////
1398 /// @brief emulates a scatter operation.
1399 /// @param pDst - pointer to destination
1400 /// @param vSrc - vector of src data to scatter
1401 /// @param vOffsets - vector of byte offsets from pDst
1402 /// @param vMask - mask of valid lanes
1403 void Builder::SCATTERPS(Value
* pDst
, Value
* vSrc
, Value
* vOffsets
, Value
* vMask
)
1405 /* Scatter algorithm
1407 while(Index = BitScanForward(mask))
1408 srcElem = srcVector[Index]
1409 offsetElem = offsetVector[Index]
1410 *(pDst + offsetElem) = srcElem
1411 Update mask (&= ~(1<<Index)
1415 BasicBlock
* pCurBB
= IRB()->GetInsertBlock();
1416 Function
* pFunc
= pCurBB
->getParent();
1417 Type
* pSrcTy
= vSrc
->getType()->getVectorElementType();
1419 // Store vectors on stack
1420 if (pScatterStackSrc
== nullptr)
1422 // Save off stack allocations and reuse per scatter. Significantly reduces stack
1423 // requirements for shaders with a lot of scatters.
1424 pScatterStackSrc
= CreateEntryAlloca(pFunc
, mSimdInt64Ty
);
1425 pScatterStackOffsets
= CreateEntryAlloca(pFunc
, mSimdInt32Ty
);
1428 Value
* pSrcArrayPtr
= BITCAST(pScatterStackSrc
, PointerType::get(vSrc
->getType(), 0));
1429 Value
* pOffsetsArrayPtr
= pScatterStackOffsets
;
1430 STORE(vSrc
, pSrcArrayPtr
);
1431 STORE(vOffsets
, pOffsetsArrayPtr
);
1433 // Cast to pointers for random access
1434 pSrcArrayPtr
= POINTER_CAST(pSrcArrayPtr
, PointerType::get(pSrcTy
, 0));
1435 pOffsetsArrayPtr
= POINTER_CAST(pOffsetsArrayPtr
, PointerType::get(mInt32Ty
, 0));
1437 Value
* pMask
= VMOVMSKPS(BITCAST(vMask
, mSimdFP32Ty
));
1439 // Get cttz function
1440 Function
* pfnCttz
= Intrinsic::getDeclaration(mpJitMgr
->mpCurrentModule
, Intrinsic::cttz
, { mInt32Ty
});
1442 // Setup loop basic block
1443 BasicBlock
* pLoop
= BasicBlock::Create(mpJitMgr
->mContext
, "Scatter Loop", pFunc
);
1445 // compute first set bit
1446 Value
* pIndex
= CALL(pfnCttz
, { pMask
, C(false) });
1448 Value
* pIsUndef
= ICMP_EQ(pIndex
, C(32));
1450 // Split current block
1451 BasicBlock
* pPostLoop
= pCurBB
->splitBasicBlock(cast
<Instruction
>(pIsUndef
)->getNextNode());
1453 // Remove unconditional jump created by splitBasicBlock
1454 pCurBB
->getTerminator()->eraseFromParent();
1456 // Add terminator to end of original block
1457 IRB()->SetInsertPoint(pCurBB
);
1459 // Add conditional branch
1460 COND_BR(pIsUndef
, pPostLoop
, pLoop
);
1462 // Add loop basic block contents
1463 IRB()->SetInsertPoint(pLoop
);
1464 PHINode
* pIndexPhi
= PHI(mInt32Ty
, 2);
1465 PHINode
* pMaskPhi
= PHI(mInt32Ty
, 2);
1467 pIndexPhi
->addIncoming(pIndex
, pCurBB
);
1468 pMaskPhi
->addIncoming(pMask
, pCurBB
);
1470 // Extract elements for this index
1471 Value
* pSrcElem
= LOADV(pSrcArrayPtr
, { pIndexPhi
});
1472 Value
* pOffsetElem
= LOADV(pOffsetsArrayPtr
, { pIndexPhi
});
1474 // GEP to this offset in dst
1475 Value
* pCurDst
= GEP(pDst
, pOffsetElem
);
1476 pCurDst
= POINTER_CAST(pCurDst
, PointerType::get(pSrcTy
, 0));
1477 STORE(pSrcElem
, pCurDst
);
1480 Value
* pNewMask
= AND(pMaskPhi
, NOT(SHL(C(1), pIndexPhi
)));
1483 Value
* pNewIndex
= CALL(pfnCttz
, { pNewMask
, C(false) });
1485 pIsUndef
= ICMP_EQ(pNewIndex
, C(32));
1486 COND_BR(pIsUndef
, pPostLoop
, pLoop
);
1489 pIndexPhi
->addIncoming(pNewIndex
, pLoop
);
1490 pMaskPhi
->addIncoming(pNewMask
, pLoop
);
1492 // Move builder to beginning of post loop
1493 IRB()->SetInsertPoint(pPostLoop
, pPostLoop
->begin());
1496 Value
* Builder::VABSPS(Value
* a
)
1498 Value
* asInt
= BITCAST(a
, mSimdInt32Ty
);
1499 Value
* result
= BITCAST(AND(asInt
, VIMMED1(0x7fffffff)), mSimdFP32Ty
);
1503 Value
*Builder::ICLAMP(Value
* src
, Value
* low
, Value
* high
)
1505 Value
*lowCmp
= ICMP_SLT(src
, low
);
1506 Value
*ret
= SELECT(lowCmp
, low
, src
);
1508 Value
*highCmp
= ICMP_SGT(ret
, high
);
1509 ret
= SELECT(highCmp
, high
, ret
);
1514 Value
*Builder::FCLAMP(Value
* src
, Value
* low
, Value
* high
)
1516 Value
*lowCmp
= FCMP_OLT(src
, low
);
1517 Value
*ret
= SELECT(lowCmp
, low
, src
);
1519 Value
*highCmp
= FCMP_OGT(ret
, high
);
1520 ret
= SELECT(highCmp
, high
, ret
);
1525 Value
*Builder::FCLAMP(Value
* src
, float low
, float high
)
1527 Value
* result
= VMAXPS(src
, VIMMED1(low
));
1528 result
= VMINPS(result
, VIMMED1(high
));
1533 //////////////////////////////////////////////////////////////////////////
1534 /// @brief save/restore stack, providing ability to push/pop the stack and
1535 /// reduce overall stack requirements for temporary stack use
1536 Value
* Builder::STACKSAVE()
1538 Function
* pfnStackSave
= Intrinsic::getDeclaration(JM()->mpCurrentModule
, Intrinsic::stacksave
);
1539 #if HAVE_LLVM == 0x306
1540 return CALL(pfnStackSave
);
1542 return CALLA(pfnStackSave
);
1546 void Builder::STACKRESTORE(Value
* pSaved
)
1548 Function
* pfnStackRestore
= Intrinsic::getDeclaration(JM()->mpCurrentModule
, Intrinsic::stackrestore
);
1549 CALL(pfnStackRestore
, std::initializer_list
<Value
*>{pSaved
});
1552 Value
*Builder::FMADDPS(Value
* a
, Value
* b
, Value
* c
)
1555 // use FMADs if available
1556 if(JM()->mArch
.AVX2())
1558 vOut
= VFMADDPS(a
, b
, c
);
1562 vOut
= FADD(FMUL(a
, b
), c
);
1567 Value
* Builder::POPCNT(Value
* a
)
1569 Function
* pCtPop
= Intrinsic::getDeclaration(JM()->mpCurrentModule
, Intrinsic::ctpop
, { a
->getType() });
1570 return CALL(pCtPop
, std::initializer_list
<Value
*>{a
});
1573 //////////////////////////////////////////////////////////////////////////
1574 /// @brief C functions called by LLVM IR
1575 //////////////////////////////////////////////////////////////////////////
1577 //////////////////////////////////////////////////////////////////////////
1578 /// @brief called in JIT code, inserted by PRINT
1579 /// output to both stdout and visual studio debug console
1580 void __cdecl
CallPrint(const char* fmt
, ...)
1583 va_start(args
, fmt
);
1586 #if defined( _WIN32 )
1588 vsnprintf_s(strBuf
, _TRUNCATE
, fmt
, args
);
1589 OutputDebugString(strBuf
);
1595 Value
*Builder::VEXTRACTI128(Value
* a
, Constant
* imm8
)
1597 #if HAVE_LLVM == 0x306
1599 Intrinsic::getDeclaration(JM()->mpCurrentModule
,
1600 Intrinsic::x86_avx_vextractf128_si_256
);
1601 return CALL(func
, {a
, imm8
});
1603 bool flag
= !imm8
->isZeroValue();
1604 SmallVector
<Constant
*,8> idx
;
1605 for (unsigned i
= 0; i
< mVWidth
/ 2; i
++) {
1606 idx
.push_back(C(flag
? i
+ mVWidth
/ 2 : i
));
1608 return VSHUFFLE(a
, VUNDEF_I(), ConstantVector::get(idx
));
1612 Value
*Builder::VINSERTI128(Value
* a
, Value
* b
, Constant
* imm8
)
1614 #if HAVE_LLVM == 0x306
1616 Intrinsic::getDeclaration(JM()->mpCurrentModule
,
1617 Intrinsic::x86_avx_vinsertf128_si_256
);
1618 return CALL(func
, {a
, b
, imm8
});
1620 bool flag
= !imm8
->isZeroValue();
1621 SmallVector
<Constant
*,8> idx
;
1622 for (unsigned i
= 0; i
< mVWidth
; i
++) {
1623 idx
.push_back(C(i
));
1625 Value
*inter
= VSHUFFLE(b
, VUNDEF_I(), ConstantVector::get(idx
));
1627 SmallVector
<Constant
*,8> idx2
;
1628 for (unsigned i
= 0; i
< mVWidth
/ 2; i
++) {
1629 idx2
.push_back(C(flag
? i
: i
+ mVWidth
));
1631 for (unsigned i
= mVWidth
/ 2; i
< mVWidth
; i
++) {
1632 idx2
.push_back(C(flag
? i
+ mVWidth
/ 2 : i
));
1634 return VSHUFFLE(a
, inter
, ConstantVector::get(idx2
));
1638 // rdtsc buckets macros
1639 void Builder::RDTSC_START(Value
* pBucketMgr
, Value
* pId
)
1641 // @todo due to an issue with thread local storage propagation in llvm, we can only safely call into
1642 // buckets framework when single threaded
1643 if (KNOB_SINGLE_THREADED
)
1645 std::vector
<Type
*> args
{
1646 PointerType::get(mInt32Ty
, 0), // pBucketMgr
1650 FunctionType
* pFuncTy
= FunctionType::get(Type::getVoidTy(JM()->mContext
), args
, false);
1651 Function
* pFunc
= cast
<Function
>(JM()->mpCurrentModule
->getOrInsertFunction("BucketManager_StartBucket", pFuncTy
));
1652 if (sys::DynamicLibrary::SearchForAddressOfSymbol("BucketManager_StartBucket") == nullptr)
1654 sys::DynamicLibrary::AddSymbol("BucketManager_StartBucket", (void*)&BucketManager_StartBucket
);
1657 CALL(pFunc
, { pBucketMgr
, pId
});
1661 void Builder::RDTSC_STOP(Value
* pBucketMgr
, Value
* pId
)
1663 // @todo due to an issue with thread local storage propagation in llvm, we can only safely call into
1664 // buckets framework when single threaded
1665 if (KNOB_SINGLE_THREADED
)
1667 std::vector
<Type
*> args
{
1668 PointerType::get(mInt32Ty
, 0), // pBucketMgr
1672 FunctionType
* pFuncTy
= FunctionType::get(Type::getVoidTy(JM()->mContext
), args
, false);
1673 Function
* pFunc
= cast
<Function
>(JM()->mpCurrentModule
->getOrInsertFunction("BucketManager_StopBucket", pFuncTy
));
1674 if (sys::DynamicLibrary::SearchForAddressOfSymbol("BucketManager_StopBucket") == nullptr)
1676 sys::DynamicLibrary::AddSymbol("BucketManager_StopBucket", (void*)&BucketManager_StopBucket
);
1679 CALL(pFunc
, { pBucketMgr
, pId
});