Skip to content

[ValueTracking] Improve isImpliedCondCommonOperandWithConstants to handle truncated LHS #69829

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 3 commits into from

Conversation

dtcxzyw
Copy link
Member

@dtcxzyw dtcxzyw commented Oct 21, 2023

This patch improves isImpliedCondCommonOperandWithConstants to handle patterns like L0 == trunc(R0) or R0 == trunc(L0). I am sure that we don't need to handle sext/zext cases since icmp pred sext/zext(X), C will be simplified into icmp pred X, C1.

Alive2: https://alive2.llvm.org/ce/z/FLcm2h

Fixes #68514.

@dtcxzyw dtcxzyw requested a review from goldsteinn October 21, 2023 09:51
@dtcxzyw dtcxzyw requested a review from nikic as a code owner October 21, 2023 09:51
@llvmbot
Copy link
Member

llvmbot commented Oct 21, 2023

@llvm/pr-subscribers-backend-amdgpu

@llvm/pr-subscribers-llvm-analysis

Author: Yingwei Zheng (dtcxzyw)

Changes

This patch improves isImpliedCondCommonOperandWithConstants to handle patterns like L0 == trunc(R0) or R0 == trunc(L0). I am sure that we don't need to handle sext/zext cases since they have been handled by InstCombinerImpl::foldAndOrOfICmpsUsingRanges.

Alive2: https://alive2.llvm.org/ce/z/FLcm2h

Fixes #68514.


Full diff: https://github.com/llvm/llvm-project/pull/69829.diff

2 Files Affected:

  • (modified) llvm/lib/Analysis/ValueTracking.cpp (+34-7)
  • (added) llvm/test/Analysis/ValueTracking/pr68514.ll (+70)
diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index 8824a05e3aa6ccd..188faf00329ece2 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -8231,14 +8231,30 @@ isImpliedCondMatchingOperands(CmpInst::Predicate LPred,
   return std::nullopt;
 }
 
-/// Return true if "icmp LPred X, LC" implies "icmp RPred X, RC" is true.
-/// Return false if "icmp LPred X, LC" implies "icmp RPred X, RC" is false.
-/// Otherwise, return std::nullopt if we can't infer anything.
+/// Return true if "icmp LPred X, LC" implies "icmp RPred cast(X), RC" is true.
+/// Return false if "icmp LPred X, LC" implies "icmp RPred cast(X), RC" is
+/// false. Otherwise, return std::nullopt if we can't infer anything.
 static std::optional<bool> isImpliedCondCommonOperandWithConstants(
-    CmpInst::Predicate LPred, const APInt &LC, CmpInst::Predicate RPred,
-    const APInt &RC) {
+    const Value *L0, CmpInst::Predicate LPred, const APInt &LC, const Value *R0,
+    CmpInst::Predicate RPred, const APInt &RC) {
   ConstantRange DomCR = ConstantRange::makeExactICmpRegion(LPred, LC);
   ConstantRange CR = ConstantRange::makeExactICmpRegion(RPred, RC);
+
+  if (L0 == R0)
+    ; // noop
+  // Example: icmp eq X, 3 --> icmp sgt trunc(X), 2
+  else if (match(R0, m_Trunc(m_Specific(L0))))
+    DomCR = DomCR.truncate(RC.getBitWidth());
+  // Example: icmp slt trunc(X), 3 --> icmp ne X, 3
+  else if (match(L0, m_Trunc(m_Specific(R0)))) {
+    // Try to prove by negation
+    DomCR = DomCR.inverse();
+    CR = CR.inverse();
+    std::swap(DomCR, CR);
+    DomCR = DomCR.truncate(LC.getBitWidth());
+  } else
+    return std::nullopt;
+
   ConstantRange Intersection = DomCR.intersectWith(CR);
   ConstantRange Difference = DomCR.difference(CR);
   if (Intersection.isEmptySet())
@@ -8272,8 +8288,19 @@ static std::optional<bool> isImpliedCondICmps(const ICmpInst *LHS,
   // Can we infer anything when the 0-operands match and the 1-operands are
   // constants (not necessarily matching)?
   const APInt *LC, *RC;
-  if (L0 == R0 && match(L1, m_APInt(LC)) && match(R1, m_APInt(RC)))
-    return isImpliedCondCommonOperandWithConstants(LPred, *LC, RPred, *RC);
+  if (match(L1, m_APInt(LC)) && match(R1, m_APInt(RC))) {
+    if (auto Res = isImpliedCondCommonOperandWithConstants(L0, LPred, *LC, R0,
+                                                           RPred, *RC))
+      return Res;
+
+    if (match(L0, m_Trunc(m_Specific(R0)))) {
+      // When L0 == trunc(R0), we use the law of excluded middle to cover some
+      // missing cases.
+      if (auto Res = isImpliedCondCommonOperandWithConstants(
+              L0, LPred, *LC, R0, ICmpInst::getInversePredicate(RPred), *RC))
+        return !*Res;
+    }
+  }
 
   // L0 = R0 = L1 + R1, L0 >=u L1 implies R0 >=u R1, L0 <u L1 implies R0 <u R1
   if (ICmpInst::isUnsigned(LPred) && ICmpInst::isUnsigned(RPred)) {
diff --git a/llvm/test/Analysis/ValueTracking/pr68514.ll b/llvm/test/Analysis/ValueTracking/pr68514.ll
new file mode 100644
index 000000000000000..482cad942624aaa
--- /dev/null
+++ b/llvm/test/Analysis/ValueTracking/pr68514.ll
@@ -0,0 +1,70 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 3
+; RUN: opt < %s -passes=instcombine -S | FileCheck %s
+
+; Tests from PR68514
+define i1 @f(i32 %a) {
+; CHECK-LABEL: define i1 @f(
+; CHECK-SAME: i32 [[A:%.*]]) {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    ret i1 false
+;
+entry:
+  %cmp = icmp eq i32 %a, 0
+  %b = trunc i32 %a to i16
+  %cmp2 = icmp sgt i16 %b, 2
+  %and9 = and i1 %cmp, %cmp2
+  ret i1 %and9
+}
+
+define i1 @g(i32 %a) {
+; CHECK-LABEL: define i1 @g(
+; CHECK-SAME: i32 [[A:%.*]]) {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[B:%.*]] = trunc i32 [[A]] to i16
+; CHECK-NEXT:    [[CMP2:%.*]] = icmp sgt i16 [[B]], 2
+; CHECK-NEXT:    ret i1 [[CMP2]]
+;
+entry:
+  %cmp = icmp eq i32 %a, 3
+  %b = trunc i32 %a to i16
+  %cmp2 = icmp sgt i16 %b, 2
+  %or9 = or i1 %cmp, %cmp2
+  ret i1 %or9
+}
+
+; b = trunc(a)
+
+define i1 @fold_trunc(i32 %a) {
+; CHECK-LABEL: define i1 @fold_trunc(
+; CHECK-SAME: i32 [[A:%.*]]) {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ugt i32 [[A]], 2
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+entry:
+  %cmp = icmp ugt i32 %a, 2
+  %b = trunc i32 %a to i16
+  %cmp2 = icmp ugt i16 %b, 2
+  %and9 = or i1 %cmp, %cmp2
+  ret i1 %and9
+}
+
+; Negative tests
+
+define i1 @nofold_trunc(i32 %a) {
+; CHECK-LABEL: define i1 @nofold_trunc(
+; CHECK-SAME: i32 [[A:%.*]]) {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ugt i32 [[A]], 3
+; CHECK-NEXT:    [[B:%.*]] = trunc i32 [[A]] to i16
+; CHECK-NEXT:    [[CMP2:%.*]] = icmp ugt i16 [[B]], 2
+; CHECK-NEXT:    [[AND9:%.*]] = or i1 [[CMP]], [[CMP2]]
+; CHECK-NEXT:    ret i1 [[AND9]]
+;
+entry:
+  %cmp = icmp ugt i32 %a, 3
+  %b = trunc i32 %a to i16
+  %cmp2 = icmp ugt i16 %b, 2
+  %and9 = or i1 %cmp, %cmp2
+  ret i1 %and9
+}

…ts` to handle truncated LHS

Fix AMDGPU tests. NFC.
@dtcxzyw
Copy link
Member Author

dtcxzyw commented Nov 4, 2023

Ping.

%b = trunc i32 %a to i16
%cmp2 = icmp ugt i16 %b, 2
%and9 = or i1 %cmp, %cmp2
ret i1 %and9
Copy link
Contributor

Choose a reason for hiding this comment

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

Missing test for lhs = trunc(rhs)

@dtcxzyw
Copy link
Member Author

dtcxzyw commented Nov 12, 2023

Closed as the original issue has been fixed by #70335.

@dtcxzyw dtcxzyw closed this Nov 12, 2023
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

Simplification for (A ==/!= B) &/| (((cast)A) CMP C)
3 participants