[llvm] 0eedc9e - [InstCombine] bitrev (zext i1 X) --> select X, SMinC, 0
Sanjay Patel via llvm-commits
llvm-commits at lists.llvm.org
Mon Jan 9 10:59:11 PST 2023
Thanks! That’s how I had drafted it initially, but we canonicalize that to
the select form already. So I just made this go directly to the end result.
On Mon, Jan 9, 2023 at 1:35 PM Philip Reames <listmail at philipreames.com>
wrote:
> I think you could replace the bitreverse with a left-shift by bitwidth-1
> if you wanted.
>
> Not sure if that's better or worse than the select representation.
>
> Philip
>
> On 1/9/23 09:31, Sanjay Patel via llvm-commits wrote:
> > Author: Sanjay Patel
> > Date: 2023-01-09T12:27:37-05:00
> > New Revision: 0eedc9e56712938ce4c01e14ba755dbdb2a0df18
> >
> > URL:
> https://github.com/llvm/llvm-project/commit/0eedc9e56712938ce4c01e14ba755dbdb2a0df18
> > DIFF:
> https://github.com/llvm/llvm-project/commit/0eedc9e56712938ce4c01e14ba755dbdb2a0df18.diff
> >
> > LOG: [InstCombine] bitrev (zext i1 X) --> select X, SMinC, 0
> >
> > https://alive2.llvm.org/ce/z/ZXCtgi
> >
> > This breaks the infinite combine loop for issue #59897,
> > but we may still need more changes to avoid those loops.
> >
> > Added:
> >
> >
> > Modified:
> > llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
> > llvm/test/Transforms/InstCombine/bitreverse.ll
> >
> > Removed:
> >
> >
> >
> >
> ################################################################################
> > diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
> b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
> > index 59486312448e..65248eae6acb 100644
> > --- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
> > +++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
> > @@ -1441,6 +1441,18 @@ Instruction
> *InstCombinerImpl::visitCallInst(CallInst &CI) {
> >
> > break;
> > }
> > + case Intrinsic::bitreverse: {
> > + // bitrev (zext i1 X to ?) --> X ? SignBitC : 0
> > + Value *X;
> > + if (match(II->getArgOperand(0), m_ZExt(m_Value(X))) &&
> > + X->getType()->isIntOrIntVectorTy(1)) {
> > + Type *Ty = II->getType();
> > + APInt SignBit = APInt::getSignMask(Ty->getScalarSizeInBits());
> > + return SelectInst::Create(X, ConstantInt::get(Ty, SignBit),
> > + ConstantInt::getNullValue(Ty));
> > + }
> > + break;
> > + }
> > case Intrinsic::bswap: {
> > Value *IIOperand = II->getArgOperand(0);
> >
> >
> > diff --git a/llvm/test/Transforms/InstCombine/bitreverse.ll
> b/llvm/test/Transforms/InstCombine/bitreverse.ll
> > index 232d69da9a02..ab41debc65c7 100644
> > --- a/llvm/test/Transforms/InstCombine/bitreverse.ll
> > +++ b/llvm/test/Transforms/InstCombine/bitreverse.ll
> > @@ -323,7 +323,7 @@ define i32 @rev_i1(i1 %x) {
> > ; CHECK-LABEL: @rev_i1(
> > ; CHECK-NEXT: [[Z:%.*]] = zext i1 [[X:%.*]] to i32
> > ; CHECK-NEXT: call void @use_i32(i32 [[Z]])
> > -; CHECK-NEXT: [[R:%.*]] = call i32 @llvm.bitreverse.i32(i32 [[Z]])
> > +; CHECK-NEXT: [[R:%.*]] = select i1 [[X]], i32 -2147483648, i32 0
> > ; CHECK-NEXT: ret i32 [[R]]
> > ;
> > %z = zext i1 %x to i32
> > @@ -334,8 +334,7 @@ define i32 @rev_i1(i1 %x) {
> >
> > define <2 x i8> @rev_v2i1(<2 x i1> %x) {
> > ; CHECK-LABEL: @rev_v2i1(
> > -; CHECK-NEXT: [[Z:%.*]] = zext <2 x i1> [[X:%.*]] to <2 x i8>
> > -; CHECK-NEXT: [[R:%.*]] = call <2 x i8> @llvm.bitreverse.v2i8(<2 x
> i8> [[Z]])
> > +; CHECK-NEXT: [[R:%.*]] = select <2 x i1> [[X:%.*]], <2 x i8> <i8
> -128, i8 -128>, <2 x i8> zeroinitializer
> > ; CHECK-NEXT: ret <2 x i8> [[R]]
> > ;
> > %z = zext <2 x i1> %x to <2 x i8>
> > @@ -353,3 +352,19 @@ define i32 @rev_i2(i2 %x) {
> > %r = call i32 @llvm.bitreverse.i32(i32 %z)
> > ret i32 %r
> > }
> > +
> > +; This used to infinite loop.
> > +
> > +define i64 @PR59897(i1 %X1_2) {
> > +; CHECK-LABEL: @PR59897(
> > +; CHECK-NEXT: [[NOT_X1_2:%.*]] = xor i1 [[X1_2:%.*]], true
> > +; CHECK-NEXT: [[X0_3X2X5X0:%.*]] = zext i1 [[NOT_X1_2]] to i64
> > +; CHECK-NEXT: ret i64 [[X0_3X2X5X0]]
> > +;
> > + %X1_3 = zext i1 %X1_2 to i32
> > + %X8_3x2x2x0 = call i32 @llvm.bitreverse.i32(i32 %X1_3)
> > + %X8_4x2x3x0 = xor i32 %X8_3x2x2x0, -1
> > + %X0_3x2x4x0 = lshr i32 %X8_4x2x3x0, 31
> > + %X0_3x2x5x0 = zext i32 %X0_3x2x4x0 to i64
> > + ret i64 %X0_3x2x5x0
> > +}
> >
> >
> >
> > _______________________________________________
> > 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/20230109/4ab628c0/attachment.html>
More information about the llvm-commits
mailing list