From a1c34a9d6a5cbb25826455f67d98c2099d27391c Mon Sep 17 00:00:00 2001 From: David Green Date: Fri, 19 Feb 2021 14:43:51 +0000 Subject: [PATCH] [ARM] Correct vector predicate type in MVE getCmpSelInstrCost --- .../lib/Target/ARM/ARMTargetTransformInfo.cpp | 2 +- .../LoopVectorize/ARM/mve-icmpcost.ll | 29 ++++++++++++++++++- 2 files changed, 29 insertions(+), 2 deletions(-) diff --git a/llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp b/llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp index 9553b8edc9ae..806aff9f2310 100644 --- a/llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp +++ b/llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp @@ -943,7 +943,7 @@ int ARMTTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy, BaseT::getScalarizationOverhead(VecCondTy, true, false) + VecValTy->getNumElements() * getCmpSelInstrCost(Opcode, ValTy->getScalarType(), - CondTy->getScalarType(), VecPred, CostKind, + VecCondTy->getScalarType(), VecPred, CostKind, I); } diff --git a/llvm/test/Transforms/LoopVectorize/ARM/mve-icmpcost.ll b/llvm/test/Transforms/LoopVectorize/ARM/mve-icmpcost.ll index fd1476743ae2..d2961326ba71 100644 --- a/llvm/test/Transforms/LoopVectorize/ARM/mve-icmpcost.ll +++ b/llvm/test/Transforms/LoopVectorize/ARM/mve-icmpcost.ll @@ -237,4 +237,31 @@ while.end: ; preds = %while.end.loopexit, ret void } -attributes #0 = { "target-features"="+mve.fp" } +; CHECK: LV: Found an estimated cost of 1 for VF 1 For instruction: %cmp1 = fcmp +; CHECK: LV: Found an estimated cost of 10 for VF 2 For instruction: %cmp1 = fcmp +; CHECK: LV: Found an estimated cost of 36 for VF 4 For instruction: %cmp1 = fcmp +define void @floatcmp(float* nocapture readonly %pSrc, i32* nocapture %pDst, i32 %blockSize) #0 { +entry: + %cmp.not7 = icmp eq i32 %blockSize, 0 + br i1 %cmp.not7, label %while.end, label %while.body + +while.body: ; preds = %entry, %while.body + %pSrc.addr.010 = phi float* [ %incdec.ptr2, %while.body ], [ %pSrc, %entry ] + %blockSize.addr.09 = phi i32 [ %dec, %while.body ], [ %blockSize, %entry ] + %pDst.addr.08 = phi i32* [ %incdec.ptr, %while.body ], [ %pDst, %entry ] + %0 = load float, float* %pSrc.addr.010, align 4 + %cmp1 = fcmp nnan ninf nsz olt float %0, 0.000000e+00 + %cond = select nnan ninf nsz i1 %cmp1, float 1.000000e+01, float %0 + %conv = fptosi float %cond to i32 + %incdec.ptr = getelementptr inbounds i32, i32* %pDst.addr.08, i32 1 + store i32 %conv, i32* %pDst.addr.08, align 4 + %incdec.ptr2 = getelementptr inbounds float, float* %pSrc.addr.010, i32 1 + %dec = add i32 %blockSize.addr.09, -1 + %cmp.not = icmp eq i32 %dec, 0 + br i1 %cmp.not, label %while.end, label %while.body + +while.end: ; preds = %while.body, %entry + ret void +} + +attributes #0 = { "target-features"="+mve" } -- 2.34.1