diff options
| author | Matt Arsenault <Matthew.Arsenault@amd.com> | 2023-08-26 10:47:54 -0400 |
|---|---|---|
| committer | Matt Arsenault <Matthew.Arsenault@amd.com> | 2023-09-01 08:22:16 -0400 |
| commit | deefda7074398b7e8b7dae9dace43a8c694a49c1 (patch) | |
| tree | 02246428cfb4aa9e6955eb7c412f39e7627465a1 /llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp | |
| parent | dac8f974b51b076773ec49603df4e627f4d9089c (diff) | |
AMDGPU: Use exp2 and log2 intrinsics directly for f16/f32
These codegen correctly but f64 doesn't. This prevents losing fast
math flags on the way to the underlying intrinsic.
https://reviews.llvm.org/D158997
Diffstat (limited to 'llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp')
| -rw-r--r-- | llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp | 29 |
1 files changed, 21 insertions, 8 deletions
diff --git a/llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp b/llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp index 0fb1474f8012..285768ff28a1 100644 --- a/llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp +++ b/llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp @@ -969,12 +969,19 @@ bool AMDGPULibCalls::fold_pow(FPMathOperator *FPOp, IRBuilder<> &B, return true; } + // If we should use the generic intrinsic instead of emitting a libcall + const bool ShouldUseIntrinsic = eltType->isFloatTy() || eltType->isHalfTy(); + // powr ---> exp2(y * log2(x)) // pown/pow ---> powr(fabs(x), y) | (x & ((int)y << 31)) - FunctionCallee ExpExpr = - getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_EXP2, FInfo)); - if (!ExpExpr) - return false; + FunctionCallee ExpExpr; + if (ShouldUseIntrinsic) + ExpExpr = Intrinsic::getDeclaration(M, Intrinsic::exp2, {FPOp->getType()}); + else { + ExpExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_EXP2, FInfo)); + if (!ExpExpr) + return false; + } bool needlog = false; bool needabs = false; @@ -1043,10 +1050,16 @@ bool AMDGPULibCalls::fold_pow(FPMathOperator *FPOp, IRBuilder<> &B, nval = cnval ? cnval : opr0; } if (needlog) { - FunctionCallee LogExpr = - getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_LOG2, FInfo)); - if (!LogExpr) - return false; + FunctionCallee LogExpr; + if (ShouldUseIntrinsic) { + LogExpr = + Intrinsic::getDeclaration(M, Intrinsic::log2, {FPOp->getType()}); + } else { + LogExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_LOG2, FInfo)); + if (!LogExpr) + return false; + } + nval = CreateCallEx(B,LogExpr, nval, "__log2"); } |