[LLVM][InstCombine] not (bitcast (cmp A, B) --> bitcast (!cmp A, B) (llvm#167693)

Author: paulwalker-arm

llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

@@ -5096,9 +5096,17 @@ Instruction *InstCombinerImpl::foldNot(BinaryOperator &I) {
     return &I;
   }
 
+  // not (bitcast (cmp A, B) --> bitcast (!cmp A, B)
+  if (match(NotOp, m_OneUse(m_BitCast(m_Value(X)))) &&
+      match(X, m_OneUse(m_Cmp(Pred, m_Value(), m_Value())))) {
+    cast<CmpInst>(X)->setPredicate(CmpInst::getInversePredicate(Pred));
+    return new BitCastInst(X, Ty);
+  }
+
   // Move a 'not' ahead of casts of a bool to enable logic reduction:
   // not (bitcast (sext i1 X)) --> bitcast (sext (not i1 X))
-  if (match(NotOp, m_OneUse(m_BitCast(m_OneUse(m_SExt(m_Value(X)))))) && X->getType()->isIntOrIntVectorTy(1)) {
+  if (match(NotOp, m_OneUse(m_BitCast(m_OneUse(m_SExt(m_Value(X)))))) &&
+      X->getType()->isIntOrIntVectorTy(1)) {
     Type *SextTy = cast<BitCastOperator>(NotOp)->getSrcTy();
     Value *NotX = Builder.CreateNot(X);
     Value *Sext = Builder.CreateSExt(NotX, SextTy);

llvm/test/Transforms/InstCombine/not.ll

@@ -1061,3 +1061,81 @@ if.else:
   call void @f2()
   unreachable
 }
+
+define i8 @invert_bitcasted_icmp(<8 x i32> %a, <8 x i32> %b) {
+; CHECK-LABEL: @invert_bitcasted_icmp(
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt <8 x i32> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[MASK_AS_INT:%.*]] = bitcast <8 x i1> [[CMP]] to i8
+; CHECK-NEXT:    ret i8 [[MASK_AS_INT]]
+;
+  %cmp = icmp sle <8 x i32> %a, %b
+  %mask.as.int = bitcast <8 x i1> %cmp to i8
+  %not = xor i8 %mask.as.int, 255
+  ret i8 %not
+}
+
+define i8 @invert_bitcasted_icmp_samesign(<8 x i32> %a, <8 x i32> %b) {
+; CHECK-LABEL: @invert_bitcasted_icmp_samesign(
+; CHECK-NEXT:    [[CMP:%.*]] = icmp samesign sgt <8 x i32> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[MASK_AS_INT:%.*]] = bitcast <8 x i1> [[CMP]] to i8
+; CHECK-NEXT:    ret i8 [[MASK_AS_INT]]
+;
+  %cmp = icmp samesign sle <8 x i32> %a, %b
+  %mask.as.int = bitcast <8 x i1> %cmp to i8
+  %not = xor i8 %mask.as.int, 255
+  ret i8 %not
+}
+
+define i8 @invert_bitcasted_icmp_multi_use_1(<8 x i32> %a, <8 x i32> %b) {
+; CHECK-LABEL: @invert_bitcasted_icmp_multi_use_1(
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sle <8 x i32> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    call void (...) @llvm.fake.use(<8 x i1> [[CMP]])
+; CHECK-NEXT:    [[MASK_AS_INT:%.*]] = bitcast <8 x i1> [[CMP]] to i8
+; CHECK-NEXT:    [[NOT:%.*]] = xor i8 [[MASK_AS_INT]], -1
+; CHECK-NEXT:    ret i8 [[NOT]]
+;
+  %cmp = icmp sle <8 x i32> %a, %b
+  call void (...) @llvm.fake.use(<8 x i1> %cmp)
+  %mask.as.int = bitcast <8 x i1> %cmp to i8
+  %not = xor i8 %mask.as.int, -1
+  ret i8 %not
+}
+
+define i8 @invert_bitcasted_icmp_multi_use_2(<8 x i32> %a, <8 x i32> %b) {
+; CHECK-LABEL: @invert_bitcasted_icmp_multi_use_2(
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sle <8 x i32> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[MASK_AS_INT:%.*]] = bitcast <8 x i1> [[CMP]] to i8
+; CHECK-NEXT:    call void (...) @llvm.fake.use(i8 [[MASK_AS_INT]])
+; CHECK-NEXT:    [[NOT:%.*]] = xor i8 [[MASK_AS_INT]], -1
+; CHECK-NEXT:    ret i8 [[NOT]]
+;
+  %cmp = icmp sle <8 x i32> %a, %b
+  %mask.as.int = bitcast <8 x i1> %cmp to i8
+  call void (...) @llvm.fake.use(i8 %mask.as.int)
+  %not = xor i8 %mask.as.int, -1
+  ret i8 %not
+}
+
+define i8 @invert_bitcasted_fcmp(<8 x float> %a, <8 x float> %b) {
+; CHECK-LABEL: @invert_bitcasted_fcmp(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp uge <8 x float> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[MASK_AS_INT:%.*]] = bitcast <8 x i1> [[CMP]] to i8
+; CHECK-NEXT:    ret i8 [[MASK_AS_INT]]
+;
+  %cmp = fcmp olt <8 x float> %a, %b
+  %mask.as.int = bitcast <8 x i1> %cmp to i8
+  %not = xor i8 %mask.as.int, 255
+  ret i8 %not
+}
+
+define i8 @invert_bitcasted_fcmp_fast(<8 x float> %a, <8 x float> %b) {
+; CHECK-LABEL: @invert_bitcasted_fcmp_fast(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp fast uge <8 x float> [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[MASK_AS_INT:%.*]] = bitcast <8 x i1> [[CMP]] to i8
+; CHECK-NEXT:    ret i8 [[MASK_AS_INT]]
+;
+  %cmp = fcmp fast olt <8 x float> %a, %b
+  %mask.as.int = bitcast <8 x i1> %cmp to i8
+  %not = xor i8 %mask.as.int, 255
+  ret i8 %not
+}