diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp index 688601a8ffa54..964616a4eb35e 100644 --- a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp +++ b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp @@ -955,9 +955,11 @@ static Value *foldIsPowerOf2OrZero(ICmpInst *Cmp0, ICmpInst *Cmp1, bool IsAnd, } /// Reduce a pair of compares that check if a value has exactly 1 bit set. -/// Also used for logical and/or, must be poison safe. +/// Also used for logical and/or, must be poison safe if range attributes are +/// dropped. static Value *foldIsPowerOf2(ICmpInst *Cmp0, ICmpInst *Cmp1, bool JoinedByAnd, - InstCombiner::BuilderTy &Builder) { + InstCombiner::BuilderTy &Builder, + InstCombinerImpl &IC) { // Handle 'and' / 'or' commutation: make the equality check the first operand. if (JoinedByAnd && Cmp1->getPredicate() == ICmpInst::ICMP_NE) std::swap(Cmp0, Cmp1); @@ -971,7 +973,10 @@ static Value *foldIsPowerOf2(ICmpInst *Cmp0, ICmpInst *Cmp1, bool JoinedByAnd, match(Cmp1, m_SpecificICmp(ICmpInst::ICMP_ULT, m_Intrinsic(m_Specific(X)), m_SpecificInt(2)))) { - Value *CtPop = Cmp1->getOperand(0); + auto *CtPop = cast(Cmp1->getOperand(0)); + // Drop range attributes and re-infer them in the next iteration. + CtPop->dropPoisonGeneratingAnnotations(); + IC.addToWorklist(CtPop); return Builder.CreateICmpEQ(CtPop, ConstantInt::get(CtPop->getType(), 1)); } // (X == 0) || (ctpop(X) u> 1) --> ctpop(X) != 1 @@ -980,7 +985,10 @@ static Value *foldIsPowerOf2(ICmpInst *Cmp0, ICmpInst *Cmp1, bool JoinedByAnd, match(Cmp1, m_SpecificICmp(ICmpInst::ICMP_UGT, m_Intrinsic(m_Specific(X)), m_SpecificInt(1)))) { - Value *CtPop = Cmp1->getOperand(0); + auto *CtPop = cast(Cmp1->getOperand(0)); + // Drop range attributes and re-infer them in the next iteration. + CtPop->dropPoisonGeneratingAnnotations(); + IC.addToWorklist(CtPop); return Builder.CreateICmpNE(CtPop, ConstantInt::get(CtPop->getType(), 1)); } return nullptr; @@ -3375,7 +3383,7 @@ Value *InstCombinerImpl::foldAndOrOfICmps(ICmpInst *LHS, ICmpInst *RHS, if (Value *V = foldSignedTruncationCheck(LHS, RHS, I, Builder)) return V; - if (Value *V = foldIsPowerOf2(LHS, RHS, IsAnd, Builder)) + if (Value *V = foldIsPowerOf2(LHS, RHS, IsAnd, Builder, *this)) return V; if (Value *V = foldPowerOf2AndShiftedMask(LHS, RHS, IsAnd, Builder)) diff --git a/llvm/test/Transforms/InstCombine/ispow2.ll b/llvm/test/Transforms/InstCombine/ispow2.ll index c21ad95f83a1c..832c066370b0f 100644 --- a/llvm/test/Transforms/InstCombine/ispow2.ll +++ b/llvm/test/Transforms/InstCombine/ispow2.ll @@ -1522,3 +1522,35 @@ define <2 x i1> @not_pow2_or_z_known_bits_fail_wrong_cmp(<2 x i32> %xin) { %r = icmp ugt <2 x i32> %cnt, ret <2 x i1> %r } + +; Make sure that range attributes on return values are dropped after merging these two icmps + +define i1 @has_single_bit(i32 %x) { +; CHECK-LABEL: @has_single_bit( +; CHECK-NEXT: entry: +; CHECK-NEXT: [[POPCNT:%.*]] = call range(i32 0, 33) i32 @llvm.ctpop.i32(i32 [[X:%.*]]) +; CHECK-NEXT: [[SEL:%.*]] = icmp eq i32 [[POPCNT]], 1 +; CHECK-NEXT: ret i1 [[SEL]] +; +entry: + %cmp1 = icmp ne i32 %x, 0 + %popcnt = call range(i32 1, 33) i32 @llvm.ctpop.i32(i32 %x) + %cmp2 = icmp ult i32 %popcnt, 2 + %sel = select i1 %cmp1, i1 %cmp2, i1 false + ret i1 %sel +} + +define i1 @has_single_bit_inv(i32 %x) { +; CHECK-LABEL: @has_single_bit_inv( +; CHECK-NEXT: entry: +; CHECK-NEXT: [[POPCNT:%.*]] = call range(i32 0, 33) i32 @llvm.ctpop.i32(i32 [[X:%.*]]) +; CHECK-NEXT: [[SEL:%.*]] = icmp ne i32 [[POPCNT]], 1 +; CHECK-NEXT: ret i1 [[SEL]] +; +entry: + %cmp1 = icmp eq i32 %x, 0 + %popcnt = call range(i32 1, 33) i32 @llvm.ctpop.i32(i32 %x) + %cmp2 = icmp ugt i32 %popcnt, 1 + %sel = select i1 %cmp1, i1 true, i1 %cmp2 + ret i1 %sel +}