[llvm] a1c05fe - [InstCombine] exclude bitcast of ppc_fp128 in icmp signbit fold

[Eric Christopher via llvm-commits]

Looks like that wasn't sufficient, I've uploaded
https://reviews.llvm.org/D77911 as a harder hammer.

-eric

On Wed, Apr 8, 2020 at 5:57 AM Sanjay Patel via llvm-commits <
llvm-commits at lists.llvm.org> wrote:

>
> Author: Sanjay Patel
> Date: 2020-04-08T08:56:19-04:00
> New Revision: a1c05fe20f3def1f1be9f50d2adefc6b6f1578ad
>
> URL:
> https://github.com/llvm/llvm-project/commit/a1c05fe20f3def1f1be9f50d2adefc6b6f1578ad
> DIFF:
> https://github.com/llvm/llvm-project/commit/a1c05fe20f3def1f1be9f50d2adefc6b6f1578ad.diff
>
> LOG: [InstCombine] exclude bitcast of ppc_fp128 in icmp signbit fold
>
> Based on the post-commit comments for rG0f56bbc, there might
> be a problem with this transform:
>
> (bitcast (fpext/fptrunc X)) to iX) < 0 --> (bitcast X to iY) < 0
>
> ...and the ppc_fp128 data type, so conservatively bypass if we
> are bitcasting a ppc_fp128.
>
> We might be able to account for endian or other differences to
> enable this for PowerPC again if that is useful.
>
> Differential Revision: https://reviews.llvm.org/D77642
>
> Added:
>
>
> Modified:
>     llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
>     llvm/test/Transforms/InstCombine/icmp.ll
>
> Removed:
>
>
>
>
> ################################################################################
> diff  --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
> b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
> index bb7d0c930755..e1e262821ef8 100644
> --- a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
> +++ b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
> @@ -2763,11 +2763,13 @@ static Instruction *foldICmpBitCast(ICmpInst &Cmp,
>          return new ICmpInst(Pred, X,
> ConstantInt::getNullValue(X->getType()));
>
>      // If this is a sign-bit test of a bitcast of a casted FP value,
> eliminate
> -    // the FP cast because the FP cast does not change the sign-bit.
> +    // the FP extend/truncate because that cast does not change the
> sign-bit.
> +    // This is true for all standard IEEE-754 types and the X86 80-bit
> type.
> +    // The sign-bit is always the most significant bit in those types.
>      const APInt *C;
>      bool TrueIfSigned;
> -    if (match(Op1, m_APInt(C)) && Bitcast->hasOneUse() &&
> -        isSignBitCheck(Pred, *C, TrueIfSigned)) {
> +    if (!BCSrcOp->getType()->isPPC_FP128Ty() && match(Op1, m_APInt(C)) &&
> +        Bitcast->hasOneUse() && isSignBitCheck(Pred, *C, TrueIfSigned)) {
>        if (match(BCSrcOp, m_FPExt(m_Value(X))) ||
>            match(BCSrcOp, m_FPTrunc(m_Value(X)))) {
>          // (bitcast (fpext/fptrunc X)) to iX) < 0 --> (bitcast X to iY) <
> 0
>
> diff  --git a/llvm/test/Transforms/InstCombine/icmp.ll
> b/llvm/test/Transforms/InstCombine/icmp.ll
> index 2da5c696d046..96673eb9e451 100644
> --- a/llvm/test/Transforms/InstCombine/icmp.ll
> +++ b/llvm/test/Transforms/InstCombine/icmp.ll
> @@ -3700,8 +3700,9 @@ define i1 @signbit_bitcast_fpext_extra_use(float %x,
> i64* %p) {
>
>  define i1 @signbit_bitcast_fpext_ppc_fp128(float %x) {
>  ; CHECK-LABEL: @signbit_bitcast_fpext_ppc_fp128(
> -; CHECK-NEXT:    [[TMP1:%.*]] = bitcast float [[X:%.*]] to i32
> -; CHECK-NEXT:    [[S4:%.*]] = icmp slt i32 [[TMP1]], 0
> +; CHECK-NEXT:    [[S2:%.*]] = fpext float [[X:%.*]] to ppc_fp128
> +; CHECK-NEXT:    [[S3:%.*]] = bitcast ppc_fp128 [[S2]] to i128
> +; CHECK-NEXT:    [[S4:%.*]] = icmp slt i128 [[S3]], 0
>  ; CHECK-NEXT:    ret i1 [[S4]]
>  ;
>    %s2 = fpext float %x to ppc_fp128
>
>
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200410/41d2ff57/attachment.html>