diff options
author | Akira Hatanaka <ahatanaka@apple.com> | 2017-09-23 05:02:02 +0000 |
---|---|---|
committer | Akira Hatanaka <ahatanaka@apple.com> | 2017-09-23 05:02:02 +0000 |
commit | 34b5dbca0a449d96daa782a884418669b6d7736d (patch) | |
tree | 2528cf8192a189b1f6e84062eb0ca394101fd777 /clang/lib/CodeGen/CGExprScalar.cpp | |
parent | 41c4a109d893604fe2570598ee62fceff6e1aee8 (diff) | |
download | bcm5719-llvm-34b5dbca0a449d96daa782a884418669b6d7736d.tar.gz bcm5719-llvm-34b5dbca0a449d96daa782a884418669b6d7736d.zip |
Promote storage-only __fp16 vector operands to float vectors.
This commit fixes a bug in the handling of storage-only __fp16 vectors
where clang didn't promote __fp16 vector operands to float vectors.
Conceptually, it performs the following transformation on the AST in
CreateBuiltinBinOp and CreateBuiltinUnaryOp:
(Before)
typedef __fp16 half4 __attribute__ ((vector_size (8)));
typedef float float4 __attribute__ ((vector_size (16)));
half4 hv0, hv1, hv2, hv3;
hv0 = hv1 + hv2 + hv3;
(After)
float4 t0 = (float4)hv1 + (float4)hv2;
float4 t1 = t0 + (float4)hv3;
hv0 = (half4)t1;
Note that this commit fixes the bug for targets that set
HalfArgsAndReturns to true (ARM and ARM64). Targets using intrinsics
such as llvm.convert.to.fp16 to handle __fp16 are still broken.
rdar://problem/20625184
Differential Revision: https://reviews.llvm.org/D32520
llvm-svn: 314056
Diffstat (limited to 'clang/lib/CodeGen/CGExprScalar.cpp')
-rw-r--r-- | clang/lib/CodeGen/CGExprScalar.cpp | 39 |
1 files changed, 35 insertions, 4 deletions
diff --git a/clang/lib/CodeGen/CGExprScalar.cpp b/clang/lib/CodeGen/CGExprScalar.cpp index a488c80979f..d9d619e802b 100644 --- a/clang/lib/CodeGen/CGExprScalar.cpp +++ b/clang/lib/CodeGen/CGExprScalar.cpp @@ -1014,10 +1014,41 @@ Value *ScalarExprEmitter::EmitScalarConversion(Value *Src, QualType SrcType, return Builder.CreateVectorSplat(NumElements, Src, "splat"); } - // Allow bitcast from vector to integer/fp of the same size. - if (isa<llvm::VectorType>(SrcTy) || - isa<llvm::VectorType>(DstTy)) - return Builder.CreateBitCast(Src, DstTy, "conv"); + if (isa<llvm::VectorType>(SrcTy) || isa<llvm::VectorType>(DstTy)) { + // Allow bitcast from vector to integer/fp of the same size. + unsigned SrcSize = SrcTy->getPrimitiveSizeInBits(); + unsigned DstSize = DstTy->getPrimitiveSizeInBits(); + if (SrcSize == DstSize) + return Builder.CreateBitCast(Src, DstTy, "conv"); + + // Conversions between vectors of different sizes are not allowed except + // when vectors of half are involved. Operations on storage-only half + // vectors require promoting half vector operands to float vectors and + // truncating the result, which is either an int or float vector, to a + // short or half vector. + + // Source and destination are both expected to be vectors. + llvm::Type *SrcElementTy = SrcTy->getVectorElementType(); + llvm::Type *DstElementTy = DstTy->getVectorElementType(); + + assert(((SrcElementTy->isIntegerTy() && + DstElementTy->isIntegerTy()) || + (SrcElementTy->isFloatingPointTy() && + DstElementTy->isFloatingPointTy())) && + "unexpected conversion between a floating-point vector and an " + "integer vector"); + + // Truncate an i32 vector to an i16 vector. + if (SrcElementTy->isIntegerTy()) + return Builder.CreateIntCast(Src, DstTy, false, "conv"); + + // Truncate a float vector to a half vector. + if (SrcSize > DstSize) + return Builder.CreateFPTrunc(Src, DstTy, "conv"); + + // Promote a half vector to a float vector. + return Builder.CreateFPExt(Src, DstTy, "conv"); + } // Finally, we have the arithmetic types: real int/float. Value *Res = nullptr; |