From: Sanjay Patel Date: Wed, 21 Mar 2018 17:17:13 +0000 (+0000) Subject: [InstCombine] add folds for xor-of-icmp signbit tests (PR36682) X-Git-Tag: llvmorg-7.0.0-rc1~10028 X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=778032f39d8aa5105f92a6ed70040e029d63397f;p=platform%2Fupstream%2Fllvm.git [InstCombine] add folds for xor-of-icmp signbit tests (PR36682) This is part of solving: https://bugs.llvm.org/show_bug.cgi?id=36682 There's also a leftover improvement from the long-ago-closed: https://bugs.llvm.org/show_bug.cgi?id=5438 https://rise4fun.com/Alive/dC1 llvm-svn: 328119 --- diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp index 4a5a6d1..7c720d8 100644 --- a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp +++ b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp @@ -2363,6 +2363,36 @@ Value *InstCombiner::foldXorOfICmps(ICmpInst *LHS, ICmpInst *RHS) { } } + // TODO: This can be generalized to compares of non-signbits using + // decomposeBitTestICmp(). It could be enhanced more by using (something like) + // foldLogOpOfMaskedICmps(). + ICmpInst::Predicate PredL = LHS->getPredicate(); + ICmpInst::Predicate PredR = RHS->getPredicate(); + Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1); + Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1); + if ((LHS->hasOneUse() || RHS->hasOneUse()) && PredL == PredR) { + // (X > -1) ^ (Y > -1) --> (X ^ Y) < 0 + // (X < 0) ^ (Y < 0) --> (X ^ Y) < 0 + if ((match(LHS1, m_AllOnes()) && match(RHS1, m_AllOnes()) && + PredL == CmpInst::ICMP_SGT) || + (match(LHS1, m_Zero()) && match(RHS1, m_Zero()) && + PredL == CmpInst::ICMP_SLT)) { + Value *Zero = ConstantInt::getNullValue(LHS0->getType()); + return Builder.CreateICmpSLT(Builder.CreateXor(LHS0, RHS0), Zero); + } + } + if ((LHS->hasOneUse() || RHS->hasOneUse())) { + // (X > -1) ^ (Y < 0) --> (X ^ Y) > -1 + // (X < 0) ^ (Y > -1) --> (X ^ Y) > -1 + if ((match(LHS1, m_AllOnes()) && match(RHS1, m_Zero()) && + PredL == CmpInst::ICMP_SGT && PredR == CmpInst::ICMP_SLT) || + (match(LHS1, m_Zero()) && match(RHS1, m_AllOnes()) && + PredL == CmpInst::ICMP_SLT && PredR == CmpInst::ICMP_SGT)) { + Value *MinusOne = ConstantInt::getAllOnesValue(LHS0->getType()); + return Builder.CreateICmpSGT(Builder.CreateXor(LHS0, RHS0), MinusOne); + } + } + // Instead of trying to imitate the folds for and/or, decompose this 'xor' // into those logic ops. That is, try to turn this into an and-of-icmps // because we have many folds for that pattern. diff --git a/llvm/test/Transforms/InstCombine/compare-signs.ll b/llvm/test/Transforms/InstCombine/compare-signs.ll index 5d69d35..c6c56f2 100644 --- a/llvm/test/Transforms/InstCombine/compare-signs.ll +++ b/llvm/test/Transforms/InstCombine/compare-signs.ll @@ -1,15 +1,14 @@ ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py ; RUN: opt -instcombine -S < %s | FileCheck %s + ; PR5438 -; TODO: This should also optimize down. define i32 @test1(i32 %a, i32 %b) nounwind readnone { ; CHECK-LABEL: @test1( -; CHECK-NEXT: [[T0:%.*]] = icmp sgt i32 [[A:%.*]], -1 -; CHECK-NEXT: [[T1:%.*]] = icmp slt i32 [[B:%.*]], 0 -; CHECK-NEXT: [[T2:%.*]] = xor i1 [[T1]], [[T0]] -; CHECK-NEXT: [[T3:%.*]] = zext i1 [[T2]] to i32 -; CHECK-NEXT: ret i32 [[T3]] +; CHECK-NEXT: [[TMP1:%.*]] = xor i32 [[B:%.*]], [[A:%.*]] +; CHECK-NEXT: [[DOTLOBIT:%.*]] = lshr i32 [[TMP1]], 31 +; CHECK-NEXT: [[DOTLOBIT_NOT:%.*]] = xor i32 [[DOTLOBIT]], 1 +; CHECK-NEXT: ret i32 [[DOTLOBIT_NOT]] ; %t0 = icmp sgt i32 %a, -1 %t1 = icmp slt i32 %b, 0 @@ -36,7 +35,7 @@ define i32 @test2(i32 %a, i32 %b) nounwind readnone { define i32 @test3(i32 %a, i32 %b) nounwind readnone { ; CHECK-LABEL: @test3( -; CHECK-NEXT: [[T2_UNSHIFTED:%.*]] = xor i32 %a, %b +; CHECK-NEXT: [[T2_UNSHIFTED:%.*]] = xor i32 [[A:%.*]], [[B:%.*]] ; CHECK-NEXT: [[T2_UNSHIFTED_LOBIT:%.*]] = lshr i32 [[T2_UNSHIFTED]], 31 ; CHECK-NEXT: [[T2_UNSHIFTED_LOBIT_NOT:%.*]] = xor i32 [[T2_UNSHIFTED_LOBIT]], 1 ; CHECK-NEXT: ret i32 [[T2_UNSHIFTED_LOBIT_NOT]] @@ -68,7 +67,7 @@ define <2 x i32> @test3vec(<2 x i32> %a, <2 x i32> %b) nounwind readnone { ; is one, not zero. define i32 @test3i(i32 %a, i32 %b) nounwind readnone { ; CHECK-LABEL: @test3i( -; CHECK-NEXT: [[T01:%.*]] = xor i32 %a, %b +; CHECK-NEXT: [[T01:%.*]] = xor i32 [[A:%.*]], [[B:%.*]] ; CHECK-NEXT: [[TMP1:%.*]] = lshr i32 [[T01]], 31 ; CHECK-NEXT: [[T4:%.*]] = xor i32 [[TMP1]], 1 ; CHECK-NEXT: ret i32 [[T4]] @@ -84,7 +83,7 @@ define i32 @test3i(i32 %a, i32 %b) nounwind readnone { define i1 @test4a(i32 %a) { ; CHECK-LABEL: @test4a( -; CHECK-NEXT: [[C:%.*]] = icmp slt i32 %a, 1 +; CHECK-NEXT: [[C:%.*]] = icmp slt i32 [[A:%.*]], 1 ; CHECK-NEXT: ret i1 [[C]] ; %l = ashr i32 %a, 31 @@ -97,7 +96,7 @@ define i1 @test4a(i32 %a) { define <2 x i1> @test4a_vec(<2 x i32> %a) { ; CHECK-LABEL: @test4a_vec( -; CHECK-NEXT: [[C:%.*]] = icmp slt <2 x i32> %a, +; CHECK-NEXT: [[C:%.*]] = icmp slt <2 x i32> [[A:%.*]], ; CHECK-NEXT: ret <2 x i1> [[C]] ; %l = ashr <2 x i32> %a, @@ -110,7 +109,7 @@ define <2 x i1> @test4a_vec(<2 x i32> %a) { define i1 @test4b(i64 %a) { ; CHECK-LABEL: @test4b( -; CHECK-NEXT: [[C:%.*]] = icmp slt i64 %a, 1 +; CHECK-NEXT: [[C:%.*]] = icmp slt i64 [[A:%.*]], 1 ; CHECK-NEXT: ret i1 [[C]] ; %l = ashr i64 %a, 63 @@ -123,7 +122,7 @@ define i1 @test4b(i64 %a) { define i1 @test4c(i64 %a) { ; CHECK-LABEL: @test4c( -; CHECK-NEXT: [[C:%.*]] = icmp slt i64 %a, 1 +; CHECK-NEXT: [[C:%.*]] = icmp slt i64 [[A:%.*]], 1 ; CHECK-NEXT: ret i1 [[C]] ; %l = ashr i64 %a, 63 @@ -137,7 +136,7 @@ define i1 @test4c(i64 %a) { define <2 x i1> @test4c_vec(<2 x i64> %a) { ; CHECK-LABEL: @test4c_vec( -; CHECK-NEXT: [[C:%.*]] = icmp slt <2 x i64> %a, +; CHECK-NEXT: [[C:%.*]] = icmp slt <2 x i64> [[A:%.*]], ; CHECK-NEXT: ret <2 x i1> [[C]] ; %l = ashr <2 x i64> %a, diff --git a/llvm/test/Transforms/InstCombine/xor-icmps.ll b/llvm/test/Transforms/InstCombine/xor-icmps.ll index 834e913..c456632 100644 --- a/llvm/test/Transforms/InstCombine/xor-icmps.ll +++ b/llvm/test/Transforms/InstCombine/xor-icmps.ll @@ -42,10 +42,9 @@ define i1 @eq_ne_zero(i4 %x, i4 %y) { define i1 @slt_zero(i4 %x, i4 %y) { ; CHECK-LABEL: @slt_zero( -; CHECK-NEXT: [[I0:%.*]] = icmp slt i4 [[X:%.*]], 0 -; CHECK-NEXT: [[I1:%.*]] = icmp slt i4 [[Y:%.*]], 0 -; CHECK-NEXT: [[R:%.*]] = xor i1 [[I0]], [[I1]] -; CHECK-NEXT: ret i1 [[R]] +; CHECK-NEXT: [[TMP1:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]] +; CHECK-NEXT: [[TMP2:%.*]] = icmp slt i4 [[TMP1]], 0 +; CHECK-NEXT: ret i1 [[TMP2]] ; %i0 = icmp slt i4 %x, 0 %i1 = icmp slt i4 %y, 0 @@ -68,10 +67,9 @@ define i1 @sgt_zero(i4 %x, i4 %y) { define i1 @sgt_minus1(i4 %x, i4 %y) { ; CHECK-LABEL: @sgt_minus1( -; CHECK-NEXT: [[I0:%.*]] = icmp sgt i4 [[X:%.*]], -1 -; CHECK-NEXT: [[I1:%.*]] = icmp sgt i4 [[Y:%.*]], -1 -; CHECK-NEXT: [[R:%.*]] = xor i1 [[I0]], [[I1]] -; CHECK-NEXT: ret i1 [[R]] +; CHECK-NEXT: [[TMP1:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]] +; CHECK-NEXT: [[TMP2:%.*]] = icmp slt i4 [[TMP1]], 0 +; CHECK-NEXT: ret i1 [[TMP2]] ; %i0 = icmp sgt i4 %x, -1 %i1 = icmp sgt i4 %y, -1 @@ -81,10 +79,9 @@ define i1 @sgt_minus1(i4 %x, i4 %y) { define i1 @slt_zero_sgt_minus1(i4 %x, i4 %y) { ; CHECK-LABEL: @slt_zero_sgt_minus1( -; CHECK-NEXT: [[I0:%.*]] = icmp slt i4 [[X:%.*]], 0 -; CHECK-NEXT: [[I1:%.*]] = icmp sgt i4 [[Y:%.*]], -1 -; CHECK-NEXT: [[R:%.*]] = xor i1 [[I0]], [[I1]] -; CHECK-NEXT: ret i1 [[R]] +; CHECK-NEXT: [[TMP1:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]] +; CHECK-NEXT: [[TMP2:%.*]] = icmp sgt i4 [[TMP1]], -1 +; CHECK-NEXT: ret i1 [[TMP2]] ; %i0 = icmp slt i4 %x, 0 %i1 = icmp sgt i4 %y, -1 @@ -94,10 +91,9 @@ define i1 @slt_zero_sgt_minus1(i4 %x, i4 %y) { define <2 x i1> @sgt_minus1_slt_zero_sgt(<2 x i4> %x, <2 x i4> %y) { ; CHECK-LABEL: @sgt_minus1_slt_zero_sgt( -; CHECK-NEXT: [[I1:%.*]] = icmp sgt <2 x i4> [[X:%.*]], -; CHECK-NEXT: [[I0:%.*]] = icmp slt <2 x i4> [[Y:%.*]], zeroinitializer -; CHECK-NEXT: [[R:%.*]] = xor <2 x i1> [[I0]], [[I1]] -; CHECK-NEXT: ret <2 x i1> [[R]] +; CHECK-NEXT: [[TMP1:%.*]] = xor <2 x i4> [[Y:%.*]], [[X:%.*]] +; CHECK-NEXT: [[TMP2:%.*]] = icmp sgt <2 x i4> [[TMP1]], +; CHECK-NEXT: ret <2 x i1> [[TMP2]] ; %i1 = icmp sgt <2 x i4> %x, %i0 = icmp slt <2 x i4> %y, zeroinitializer