Skip to content

Conversation

dtcxzyw
Copy link
Member

@dtcxzyw dtcxzyw commented Sep 9, 2023

This patch folds the pattern a ne/eq (zext/sext (a ne/eq c)) into a boolean constant or a compare.
Clang vs GCC: https://godbolt.org/z/4ro817WE8
Proof for zext: https://alive2.llvm.org/ce/z/6z9NRF
Proof for sext: https://alive2.llvm.org/ce/z/tv5wuE
Fixes #65073.

@dtcxzyw dtcxzyw requested a review from a team as a code owner September 9, 2023 16:03
@Endilll Endilll added llvm:instcombine Covers the InstCombine, InstSimplify and AggressiveInstCombine passes and removed llvm:transforms labels Sep 9, 2023
@llvmbot
Copy link
Member

llvmbot commented Sep 18, 2023

@llvm/pr-subscribers-llvm-transforms

Changes

This patch folds the pattern a ne/eq (zext (a ne/eq c)) into a boolean constant or a compare.
Clang vs GCC: https://godbolt.org/z/4ro817WE8
Alive2: https://alive2.llvm.org/ce/z/6z9NRF
Fixes #65073.


Full diff: https://github.com/llvm/llvm-project/pull/65852.diff

2 Files Affected:

  • (modified) llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp (+62)
  • (modified) llvm/test/Transforms/InstCombine/icmp-range.ll (+18-58)
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
index a219dac7acfbe16..d0b62c17ec94358 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -6380,7 +6380,69 @@ Instruction *InstCombinerImpl::foldICmpUsingBoolRange(ICmpInst &I) {
       Y->getType()->isIntOrIntVectorTy(1) && Pred == ICmpInst::ICMP_ULE)
     return BinaryOperator::CreateOr(Builder.CreateIsNull(X), Y);
 
+  ICmpInst::Predicate Pred1, Pred2;
   const APInt *C;
+  // icmp eq/ne X, (zext (icmp eq/ne X, C))
+  if (match(&I, m_c_ICmp(Pred1, m_Value(X),
+                         m_ZExt(m_ICmp(Pred2, m_Deferred(X), m_APInt(C))))) &&
+      ICmpInst::isEquality(Pred1) && ICmpInst::isEquality(Pred2)) {
+    if (C->isZero()) {
+      if (Pred2 == ICmpInst::ICMP_EQ) {
+        // icmp eq X, (zext (icmp eq X, 0)) --> false
+        // icmp ne X, (zext (icmp eq X, 0)) --> true
+        return replaceInstUsesWith(
+            I,
+            Constant::getIntegerValue(
+                I.getType(),
+                APInt(1U, static_cast<uint64_t>(Pred1 == ICmpInst::ICMP_NE))));
+      } else {
+        // icmp eq X, (zext (icmp ne X, 0)) --> icmp ult X, 2
+        // icmp ne X, (zext (icmp ne X, 0)) --> icmp ugt X, 1
+        return ICmpInst::Create(
+            Instruction::ICmp,
+            Pred1 == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGT
+                                       : ICmpInst::ICMP_ULT,
+            X,
+            Constant::getIntegerValue(
+                X->getType(), APInt(X->getType()->getScalarSizeInBits(),
+                                    Pred1 == ICmpInst::ICMP_NE ? 1 : 2)));
+      }
+    } else if (C->isOne()) {
+      if (Pred2 == ICmpInst::ICMP_NE) {
+        // icmp eq X, (zext (icmp ne X, 1)) --> false
+        // icmp ne X, (zext (icmp ne X, 1)) --> true
+        return replaceInstUsesWith(
+            I,
+            Constant::getIntegerValue(
+                I.getType(),
+                APInt(1U, static_cast<uint64_t>(Pred1 == ICmpInst::ICMP_NE))));
+      } else {
+        // icmp eq X, (zext (icmp eq X, 1)) --> icmp ult X, 2
+        // icmp ne X, (zext (icmp eq X, 1)) --> icmp ugt X, 1
+        return ICmpInst::Create(
+            Instruction::ICmp,
+            Pred1 == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGT
+                                       : ICmpInst::ICMP_ULT,
+            X,
+            Constant::getIntegerValue(
+                X->getType(), APInt(X->getType()->getScalarSizeInBits(),
+                                    Pred1 == ICmpInst::ICMP_NE ? 1 : 2)));
+      }
+    } else {
+      // C != 0 && C != 1
+      // icmp eq X, (zext (icmp eq X, C)) --> icmp eq X, 0
+      // icmp eq X, (zext (icmp ne X, C)) --> icmp eq X, 1
+      // icmp ne X, (zext (icmp eq X, C)) --> icmp ne X, 0
+      // icmp ne X, (zext (icmp ne X, C)) --> icmp ne X, 1
+      return ICmpInst::Create(
+          Instruction::ICmp, Pred1, X,
+          Constant::getIntegerValue(
+              X->getType(),
+              APInt(X->getType()->getScalarSizeInBits(),
+                    static_cast<uint64_t>(Pred2 == ICmpInst::ICMP_NE))));
+    }
+  }
+
   if (match(I.getOperand(0), m_c_Add(m_ZExt(m_Value(X)), m_SExt(m_Value(Y)))) &&
       match(I.getOperand(1), m_APInt(C)) &&
       X->getType()->isIntOrIntVectorTy(1) &&
diff --git a/llvm/test/Transforms/InstCombine/icmp-range.ll b/llvm/test/Transforms/InstCombine/icmp-range.ll
index a26e760059b43fe..88d5a723747ad5f 100644
--- a/llvm/test/Transforms/InstCombine/icmp-range.ll
+++ b/llvm/test/Transforms/InstCombine/icmp-range.ll
@@ -1037,10 +1037,7 @@ define i1 @icmp_ne_bool_1(ptr %ptr) {
 ; Tests from PR65073
 define i1 @icmp_ne_zext_eq_zero(i32 %a) {
 ; CHECK-LABEL: @icmp_ne_zext_eq_zero(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i32 [[A:%.*]], 0
-; CHECK-NEXT:    [[CONV:%.*]] = zext i1 [[CMP]] to i32
-; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne i32 [[CONV]], [[A]]
-; CHECK-NEXT:    ret i1 [[CMP1]]
+; CHECK-NEXT:    ret i1 true
 ;
   %cmp = icmp eq i32 %a, 0
   %conv = zext i1 %cmp to i32
@@ -1050,9 +1047,7 @@ define i1 @icmp_ne_zext_eq_zero(i32 %a) {
 
 define i1 @icmp_ne_zext_ne_zero(i32 %a) {
 ; CHECK-LABEL: @icmp_ne_zext_ne_zero(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[A:%.*]], 0
-; CHECK-NEXT:    [[CONV:%.*]] = zext i1 [[CMP]] to i32
-; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne i32 [[CONV]], [[A]]
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp ugt i32 [[A:%.*]], 1
 ; CHECK-NEXT:    ret i1 [[CMP1]]
 ;
   %cmp = icmp ne i32 %a, 0
@@ -1063,10 +1058,7 @@ define i1 @icmp_ne_zext_ne_zero(i32 %a) {
 
 define i1 @icmp_eq_zext_eq_zero(i32 %a) {
 ; CHECK-LABEL: @icmp_eq_zext_eq_zero(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i32 [[A:%.*]], 0
-; CHECK-NEXT:    [[CONV:%.*]] = zext i1 [[CMP]] to i32
-; CHECK-NEXT:    [[CMP1:%.*]] = icmp eq i32 [[CONV]], [[A]]
-; CHECK-NEXT:    ret i1 [[CMP1]]
+; CHECK-NEXT:    ret i1 false
 ;
   %cmp = icmp eq i32 %a, 0
   %conv = zext i1 %cmp to i32
@@ -1076,9 +1068,7 @@ define i1 @icmp_eq_zext_eq_zero(i32 %a) {
 
 define i1 @icmp_eq_zext_ne_zero(i32 %a) {
 ; CHECK-LABEL: @icmp_eq_zext_ne_zero(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[A:%.*]], 0
-; CHECK-NEXT:    [[CONV:%.*]] = zext i1 [[CMP]] to i32
-; CHECK-NEXT:    [[CMP1:%.*]] = icmp eq i32 [[CONV]], [[A]]
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp ult i32 [[A:%.*]], 2
 ; CHECK-NEXT:    ret i1 [[CMP1]]
 ;
   %cmp = icmp ne i32 %a, 0
@@ -1089,9 +1079,7 @@ define i1 @icmp_eq_zext_ne_zero(i32 %a) {
 
 define i1 @icmp_ne_zext_eq_one(i32 %a) {
 ; CHECK-LABEL: @icmp_ne_zext_eq_one(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i32 [[A:%.*]], 1
-; CHECK-NEXT:    [[CONV:%.*]] = zext i1 [[CMP]] to i32
-; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne i32 [[CONV]], [[A]]
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp ugt i32 [[A:%.*]], 1
 ; CHECK-NEXT:    ret i1 [[CMP1]]
 ;
   %cmp = icmp eq i32 %a, 1
@@ -1102,10 +1090,7 @@ define i1 @icmp_ne_zext_eq_one(i32 %a) {
 
 define i1 @icmp_ne_zext_ne_one(i32 %a) {
 ; CHECK-LABEL: @icmp_ne_zext_ne_one(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[A:%.*]], 1
-; CHECK-NEXT:    [[CONV:%.*]] = zext i1 [[CMP]] to i32
-; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne i32 [[CONV]], [[A]]
-; CHECK-NEXT:    ret i1 [[CMP1]]
+; CHECK-NEXT:    ret i1 true
 ;
   %cmp = icmp ne i32 %a, 1
   %conv = zext i1 %cmp to i32
@@ -1115,9 +1100,7 @@ define i1 @icmp_ne_zext_ne_one(i32 %a) {
 
 define i1 @icmp_eq_zext_eq_one(i32 %a) {
 ; CHECK-LABEL: @icmp_eq_zext_eq_one(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i32 [[A:%.*]], 1
-; CHECK-NEXT:    [[CONV:%.*]] = zext i1 [[CMP]] to i32
-; CHECK-NEXT:    [[CMP1:%.*]] = icmp eq i32 [[CONV]], [[A]]
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp ult i32 [[A:%.*]], 2
 ; CHECK-NEXT:    ret i1 [[CMP1]]
 ;
   %cmp = icmp eq i32 %a, 1
@@ -1128,10 +1111,7 @@ define i1 @icmp_eq_zext_eq_one(i32 %a) {
 
 define i1 @icmp_eq_zext_ne_one(i32 %a) {
 ; CHECK-LABEL: @icmp_eq_zext_ne_one(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[A:%.*]], 1
-; CHECK-NEXT:    [[CONV:%.*]] = zext i1 [[CMP]] to i32
-; CHECK-NEXT:    [[CMP1:%.*]] = icmp eq i32 [[CONV]], [[A]]
-; CHECK-NEXT:    ret i1 [[CMP1]]
+; CHECK-NEXT:    ret i1 false
 ;
   %cmp = icmp ne i32 %a, 1
   %conv = zext i1 %cmp to i32
@@ -1141,9 +1121,7 @@ define i1 @icmp_eq_zext_ne_one(i32 %a) {
 
 define i1 @icmp_ne_zext_eq_non_boolean(i32 %a) {
 ; CHECK-LABEL: @icmp_ne_zext_eq_non_boolean(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i32 [[A:%.*]], 2
-; CHECK-NEXT:    [[CONV:%.*]] = zext i1 [[CMP]] to i32
-; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne i32 [[CONV]], [[A]]
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne i32 [[A:%.*]], 0
 ; CHECK-NEXT:    ret i1 [[CMP1]]
 ;
   %cmp = icmp eq i32 %a, 2
@@ -1154,9 +1132,7 @@ define i1 @icmp_ne_zext_eq_non_boolean(i32 %a) {
 
 define i1 @icmp_ne_zext_ne_non_boolean(i32 %a) {
 ; CHECK-LABEL: @icmp_ne_zext_ne_non_boolean(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[A:%.*]], 2
-; CHECK-NEXT:    [[CONV:%.*]] = zext i1 [[CMP]] to i32
-; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne i32 [[CONV]], [[A]]
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne i32 [[A:%.*]], 1
 ; CHECK-NEXT:    ret i1 [[CMP1]]
 ;
   %cmp = icmp ne i32 %a, 2
@@ -1167,9 +1143,7 @@ define i1 @icmp_ne_zext_ne_non_boolean(i32 %a) {
 
 define i1 @icmp_eq_zext_eq_non_boolean(i32 %a) {
 ; CHECK-LABEL: @icmp_eq_zext_eq_non_boolean(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i32 [[A:%.*]], 2
-; CHECK-NEXT:    [[CONV:%.*]] = zext i1 [[CMP]] to i32
-; CHECK-NEXT:    [[CMP1:%.*]] = icmp eq i32 [[CONV]], [[A]]
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp eq i32 [[A:%.*]], 0
 ; CHECK-NEXT:    ret i1 [[CMP1]]
 ;
   %cmp = icmp eq i32 %a, 2
@@ -1180,9 +1154,7 @@ define i1 @icmp_eq_zext_eq_non_boolean(i32 %a) {
 
 define i1 @icmp_eq_zext_ne_non_boolean(i32 %a) {
 ; CHECK-LABEL: @icmp_eq_zext_ne_non_boolean(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[A:%.*]], 2
-; CHECK-NEXT:    [[CONV:%.*]] = zext i1 [[CMP]] to i32
-; CHECK-NEXT:    [[CMP1:%.*]] = icmp eq i32 [[CONV]], [[A]]
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp eq i32 [[A:%.*]], 1
 ; CHECK-NEXT:    ret i1 [[CMP1]]
 ;
   %cmp = icmp ne i32 %a, 2
@@ -1192,11 +1164,8 @@ define i1 @icmp_eq_zext_ne_non_boolean(i32 %a) {
 }
 
 define <2 x i1> @icmp_ne_zext_eq_zero_vec(<2 x i32> %a) {
-; CHECK-LABEL: @icmp_ne_zext_eq_zero_vec(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq <2 x i32> [[A:%.*]], zeroinitializer
-; CHECK-NEXT:    [[CONV:%.*]] = zext <2 x i1> [[CMP]] to <2 x i32>
-; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne <2 x i32> [[CONV]], [[A]]
-; CHECK-NEXT:    ret <2 x i1> [[CMP1]]
+; CHECK-LABEL: @icmp_ne_zext_eq_zero_vec
+; CHECK-NEXT:    ret <2 x i1> <i1 true, i1 true>
 ;
   %cmp = icmp eq <2 x i32> %a, <i32 0, i32 0>
   %conv = zext <2 x i1> %cmp to <2 x i32>
@@ -1206,9 +1175,7 @@ define <2 x i1> @icmp_ne_zext_eq_zero_vec(<2 x i32> %a) {
 
 define <2 x i1> @icmp_ne_zext_ne_zero_vec(<2 x i32> %a) {
 ; CHECK-LABEL: @icmp_ne_zext_ne_zero_vec(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ne <2 x i32> [[A:%.*]], zeroinitializer
-; CHECK-NEXT:    [[CONV:%.*]] = zext <2 x i1> [[CMP]] to <2 x i32>
-; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne <2 x i32> [[CONV]], [[A]]
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp ugt <2 x i32> [[A:%.*]], <i32 1, i32 1>
 ; CHECK-NEXT:    ret <2 x i1> [[CMP1]]
 ;
   %cmp = icmp ne <2 x i32> %a, <i32 0, i32 0>
@@ -1219,9 +1186,7 @@ define <2 x i1> @icmp_ne_zext_ne_zero_vec(<2 x i32> %a) {
 
 define <2 x i1> @icmp_ne_zext_eq_one_vec(<2 x i32> %a) {
 ; CHECK-LABEL: @icmp_ne_zext_eq_one_vec(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq <2 x i32> [[A:%.*]], <i32 1, i32 1>
-; CHECK-NEXT:    [[CONV:%.*]] = zext <2 x i1> [[CMP]] to <2 x i32>
-; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne <2 x i32> [[CONV]], [[A]]
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp ugt <2 x i32> [[A:%.*]], <i32 1, i32 1>
 ; CHECK-NEXT:    ret <2 x i1> [[CMP1]]
 ;
   %cmp = icmp eq <2 x i32> %a, <i32 1, i32 1>
@@ -1232,10 +1197,7 @@ define <2 x i1> @icmp_ne_zext_eq_one_vec(<2 x i32> %a) {
 
 define <2 x i1> @icmp_ne_zext_ne_one_vec(<2 x i32> %a) {
 ; CHECK-LABEL: @icmp_ne_zext_ne_one_vec(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ne <2 x i32> [[A:%.*]], <i32 1, i32 1>
-; CHECK-NEXT:    [[CONV:%.*]] = zext <2 x i1> [[CMP]] to <2 x i32>
-; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne <2 x i32> [[CONV]], [[A]]
-; CHECK-NEXT:    ret <2 x i1> [[CMP1]]
+; CHECK-NEXT:    ret <2 x i1> <i1 true, i1 true>
 ;
   %cmp = icmp ne <2 x i32> %a, <i32 1, i32 1>
   %conv = zext <2 x i1> %cmp to <2 x i32>
@@ -1245,9 +1207,7 @@ define <2 x i1> @icmp_ne_zext_ne_one_vec(<2 x i32> %a) {
 
 define <2 x i1> @icmp_ne_zext_eq_non_boolean_vec(<2 x i32> %a) {
 ; CHECK-LABEL: @icmp_ne_zext_eq_non_boolean_vec(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq <2 x i32> [[A:%.*]], <i32 2, i32 2>
-; CHECK-NEXT:    [[CONV:%.*]] = zext <2 x i1> [[CMP]] to <2 x i32>
-; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne <2 x i32> [[CONV]], [[A]]
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne <2 x i32> [[A:%.*]], zeroinitializer
 ; CHECK-NEXT:    ret <2 x i1> [[CMP1]]
 ;
   %cmp = icmp eq <2 x i32> %a, <i32 2, i32 2>

@dtcxzyw dtcxzyw removed the request for review from spatel-gh September 18, 2023 17:57
@dtcxzyw dtcxzyw changed the title [InstCombine] Simplify the pattern a ne/eq (zext (a ne/eq c)) [InstCombine] Simplify the pattern a ne/eq (zext/sext (a ne/eq c)) Sep 18, 2023
@dtcxzyw dtcxzyw requested review from goldsteinn and nikic September 18, 2023 20:18
@goldsteinn
Copy link
Contributor

LGTM.

@dtcxzyw
Copy link
Member Author

dtcxzyw commented Sep 30, 2023

Ping.

@dtcxzyw
Copy link
Member Author

dtcxzyw commented Oct 5, 2023

Ping.

@goldsteinn
Copy link
Contributor

Continue to LGTM...

Copy link
Contributor

@nikic nikic left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Basically LGTM, but I think this is still missing negative tests for non-equality pred1/pred2?

@nikic
Copy link
Contributor

nikic commented Oct 6, 2023

Looks like incorrect conflict resolution in the last merge.

@dtcxzyw dtcxzyw force-pushed the fold-icmp-eq-zext-eq-self branch from c9ce7de to 7e81e63 Compare October 6, 2023 11:28
@dtcxzyw
Copy link
Member Author

dtcxzyw commented Oct 6, 2023

The conflict resolution should be correct now.

Copy link
Contributor

@nikic nikic left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, it looks correct now.

@dtcxzyw dtcxzyw merged commit b3b3336 into llvm:main Oct 6, 2023
@dtcxzyw dtcxzyw deleted the fold-icmp-eq-zext-eq-self branch October 6, 2023 12:57
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
llvm:instcombine Covers the InstCombine, InstSimplify and AggressiveInstCombine passes llvm:transforms
Projects
None yet
Development

Successfully merging this pull request may close these issues.

a ne/eq (a ne/eq 0) pattern is not simplified
5 participants