[InstCombine] fold icmp of zext bool based on limited range

X <u (zext i1 Y) --> (X == 0) && Y

https://alive2.llvm.org/ce/z/avQDRY

This is a generalization of 4069cccf3b4ff4a based on the post-commit suggestion.
This also adds the i1 type check and tests that were missing from the earlier
attempt; that commit caused several bot fails and was reverted.

Differential Revision: https://reviews.llvm.org/D126171

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
index 4f20a06..f885859 100644 (file)
--- a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -5631,6 +5631,24 @@ Instruction *InstCombinerImpl::foldICmpUsingKnownBits(ICmpInst &I) {
   return nullptr;
 }
 
+/// If one operand of an icmp is effectively a bool (value range of {0,1}),
+/// then try to reduce patterns based on that limit.
+static Instruction *foldICmpUsingBoolRange(ICmpInst &I,
+                                           InstCombiner::BuilderTy &Builder) {
+  Value *X, *Y;
+  ICmpInst::Predicate Pred;
+
+  // X must be 0 and bool must be true for "ULT":
+  // X <u (zext i1 Y) --> (X == 0) && Y
+  if (match(&I, m_c_ICmp(Pred, m_Value(X), m_OneUse(m_ZExt(m_Value(Y))))) &&
+      Y->getType()->isIntOrIntVectorTy(1) && Pred == ICmpInst::ICMP_ULT)
+    return BinaryOperator::CreateAnd(Builder.CreateIsNull(X), Y);
+
+  // TODO: Handle the related pattern with UGE/sext.
+
+  return nullptr;
+}
+
 llvm::Optional<std::pair<CmpInst::Predicate, Constant *>>
 InstCombiner::getFlippedStrictnessPredicateAndConstant(CmpInst::Predicate Pred,
                                                        Constant *C) {
@@ -6058,6 +6076,9 @@ Instruction *InstCombinerImpl::visitICmpInst(ICmpInst &I) {
   if (Instruction *Res = foldICmpWithDominatingICmp(I))
     return Res;
 
+  if (Instruction *Res = foldICmpUsingBoolRange(I, Builder))
+    return Res;
+
   if (Instruction *Res = foldICmpUsingKnownBits(I))
     return Res;
 

diff --git a/llvm/test/Transforms/InstCombine/icmp-range.ll b/llvm/test/Transforms/InstCombine/icmp-range.ll
index 53253b4..d55787e 100644 (file)
--- a/llvm/test/Transforms/InstCombine/icmp-range.ll
+++ b/llvm/test/Transforms/InstCombine/icmp-range.ll
@@ -173,8 +173,8 @@ define i1 @test_two_ranges3(i32* nocapture readonly %arg1, i32* nocapture readon
 
 define i1 @ugt_zext(i1 %b, i8 %x) {
 ; CHECK-LABEL: @ugt_zext(
-; CHECK-NEXT:    [[Z:%.*]] = zext i1 [[B:%.*]] to i8
-; CHECK-NEXT:    [[R:%.*]] = icmp ugt i8 [[Z]], [[X:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp eq i8 [[X:%.*]], 0
+; CHECK-NEXT:    [[R:%.*]] = and i1 [[TMP1]], [[B:%.*]]
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
   %z = zext i1 %b to i8
@@ -185,8 +185,8 @@ define i1 @ugt_zext(i1 %b, i8 %x) {
 define <2 x i1> @ult_zext(<2 x i1> %b, <2 x i8> %p) {
 ; CHECK-LABEL: @ult_zext(
 ; CHECK-NEXT:    [[X:%.*]] = mul <2 x i8> [[P:%.*]], [[P]]
-; CHECK-NEXT:    [[Z:%.*]] = zext <2 x i1> [[B:%.*]] to <2 x i8>
-; CHECK-NEXT:    [[R:%.*]] = icmp ult <2 x i8> [[X]], [[Z]]
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp eq <2 x i8> [[X]], zeroinitializer
+; CHECK-NEXT:    [[R:%.*]] = and <2 x i1> [[TMP1]], [[B:%.*]]
 ; CHECK-NEXT:    ret <2 x i1> [[R]]
 ;
   %x = mul <2 x i8> %p, %p ; thwart complexity-based canonicalization
@@ -195,6 +195,8 @@ define <2 x i1> @ult_zext(<2 x i1> %b, <2 x i8> %p) {
   ret <2 x i1> %r
 }
 
+; negative test - need ult/ugt
+
 define i1 @uge_zext(i1 %b, i8 %x) {
 ; CHECK-LABEL: @uge_zext(
 ; CHECK-NEXT:    [[Z:%.*]] = zext i1 [[B:%.*]] to i8
@@ -206,6 +208,8 @@ define i1 @uge_zext(i1 %b, i8 %x) {
   ret i1 %r
 }
 
+; negative test - need ult/ugt
+
 define i1 @ule_zext(i1 %b, i8 %p) {
 ; CHECK-LABEL: @ule_zext(
 ; CHECK-NEXT:    [[X:%.*]] = mul i8 [[P:%.*]], [[P]]
@@ -219,6 +223,8 @@ define i1 @ule_zext(i1 %b, i8 %p) {
   ret i1 %r
 }
 
+; negative test - extra use
+
 define i1 @ugt_zext_use(i1 %b, i8 %x) {
 ; CHECK-LABEL: @ugt_zext_use(
 ; CHECK-NEXT:    [[Z:%.*]] = zext i1 [[B:%.*]] to i8
@@ -232,6 +238,8 @@ define i1 @ugt_zext_use(i1 %b, i8 %x) {
   ret i1 %r
 }
 
+; negative test - must be zext of i1
+
 define i1 @ult_zext_not_i1(i2 %b, i8 %x) {
 ; CHECK-LABEL: @ult_zext_not_i1(
 ; CHECK-NEXT:    [[Z:%.*]] = zext i2 [[B:%.*]] to i8
@@ -243,11 +251,12 @@ define i1 @ult_zext_not_i1(i2 %b, i8 %x) {
   ret i1 %r
 }
 
+; sub is eliminated
+
 define i1 @sub_ult_zext(i1 %b, i8 %x, i8 %y) {
 ; CHECK-LABEL: @sub_ult_zext(
-; CHECK-NEXT:    [[Z:%.*]] = zext i1 [[B:%.*]] to i8
-; CHECK-NEXT:    [[S:%.*]] = sub i8 [[X:%.*]], [[Y:%.*]]
-; CHECK-NEXT:    [[R:%.*]] = icmp ult i8 [[S]], [[Z]]
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp eq i8 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = and i1 [[TMP1]], [[B:%.*]]
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
   %z = zext i1 %b to i8
@@ -259,8 +268,8 @@ define i1 @sub_ult_zext(i1 %b, i8 %x, i8 %y) {
 define i1 @zext_ult_zext(i1 %b, i8 %p) {
 ; CHECK-LABEL: @zext_ult_zext(
 ; CHECK-NEXT:    [[X:%.*]] = mul i8 [[P:%.*]], [[P]]
-; CHECK-NEXT:    [[TMP1:%.*]] = zext i1 [[B:%.*]] to i8
-; CHECK-NEXT:    [[R:%.*]] = icmp ult i8 [[X]], [[TMP1]]
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp eq i8 [[X]], 0
+; CHECK-NEXT:    [[R:%.*]] = and i1 [[TMP1]], [[B:%.*]]
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
   %x = mul i8 %p, %p ; thwart complexity-based canonicalization
@@ -270,12 +279,14 @@ define i1 @zext_ult_zext(i1 %b, i8 %p) {
   ret i1 %r
 }
 
+; match and fold even if both sides are zexts (from different source types)
+
 define i1 @zext_ugt_zext(i1 %b, i4 %x) {
 ; CHECK-LABEL: @zext_ugt_zext(
 ; CHECK-NEXT:    [[ZX:%.*]] = zext i4 [[X:%.*]] to i8
 ; CHECK-NEXT:    call void @use(i8 [[ZX]])
-; CHECK-NEXT:    [[TMP1:%.*]] = zext i1 [[B:%.*]] to i4
-; CHECK-NEXT:    [[R:%.*]] = icmp ugt i4 [[TMP1]], [[X]]
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp eq i4 [[X]], 0
+; CHECK-NEXT:    [[R:%.*]] = and i1 [[TMP1]], [[B:%.*]]
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
   %z = zext i1 %b to i8
@@ -285,6 +296,8 @@ define i1 @zext_ugt_zext(i1 %b, i4 %x) {
   ret i1 %r
 }
 
+; negative test - must be zext of i1
+
 define i1 @sub_ult_zext_not_i1(i2 %b, i8 %x, i8 %y) {
 ; CHECK-LABEL: @sub_ult_zext_not_i1(
 ; CHECK-NEXT:    [[Z:%.*]] = zext i2 [[B:%.*]] to i8
@@ -298,6 +311,8 @@ define i1 @sub_ult_zext_not_i1(i2 %b, i8 %x, i8 %y) {
   ret i1 %r
 }
 
+; negative test - extra use (but we could try harder to fold this)
+
 define i1 @sub_ult_zext_use1(i1 %b, i8 %x, i8 %y) {
 ; CHECK-LABEL: @sub_ult_zext_use1(
 ; CHECK-NEXT:    [[Z:%.*]] = zext i1 [[B:%.*]] to i8
@@ -315,10 +330,10 @@ define i1 @sub_ult_zext_use1(i1 %b, i8 %x, i8 %y) {
 
 define <2 x i1> @zext_ugt_sub_use2(<2 x i1> %b, <2 x i8> %x, <2 x i8> %y) {
 ; CHECK-LABEL: @zext_ugt_sub_use2(
-; CHECK-NEXT:    [[Z:%.*]] = zext <2 x i1> [[B:%.*]] to <2 x i8>
 ; CHECK-NEXT:    [[S:%.*]] = sub <2 x i8> [[X:%.*]], [[Y:%.*]]
 ; CHECK-NEXT:    call void @use_vec(<2 x i8> [[S]])
-; CHECK-NEXT:    [[R:%.*]] = icmp ult <2 x i8> [[S]], [[Z]]
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp eq <2 x i8> [[X]], [[Y]]
+; CHECK-NEXT:    [[R:%.*]] = and <2 x i1> [[TMP1]], [[B:%.*]]
 ; CHECK-NEXT:    ret <2 x i1> [[R]]
 ;
   %z = zext <2 x i1> %b to <2 x i8>