Skip to content

Conversation

dtcxzyw
Copy link
Member

@dtcxzyw dtcxzyw commented Sep 25, 2023

This patch handles the pattern (icmp ugt X, mask) & (icmp eq (X & ~mask), C2) in foldAndOrOfICmpsUsingRanges.
It interprets the icmp ugt X, mask idiom into the canonical form icmp ne (X & ~mask), 0 using getMaskedTypeForICmpPair.
Fixes #59555.

@llvmbot
Copy link
Member

llvmbot commented Sep 25, 2023

@llvm/pr-subscribers-llvm-transforms

Changes

This patch handles the pattern (icmp ugt X, mask) & (icmp eq (X & ~mask), C2) in foldAndOrOfICmpsUsingRanges.
It interprets the icmp ugt X, mask idiom into the canonical form icmp ne (X & ~mask), 0 using getMaskedTypeForICmpPair.
Fixes #59555.


Full diff: https://github.com/llvm/llvm-project/pull/67327.diff

2 Files Affected:

  • (modified) llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp (+40-14)
  • (modified) llvm/test/Transforms/InstCombine/icmp-range.ll (+2-5)
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
index c04fe827207fd0d..a092258579aa1b9 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
@@ -1273,23 +1273,46 @@ Value *InstCombinerImpl::foldAndOrOfICmpsUsingRanges(ICmpInst *ICmp1,
   ICmpInst::Predicate Pred1, Pred2;
   Value *V1, *V2;
   const APInt *C1, *C2;
-  if (!match(ICmp1, m_ICmp(Pred1, m_Value(V1), m_APInt(C1))) ||
-      !match(ICmp2, m_ICmp(Pred2, m_Value(V2), m_APInt(C2))))
-    return nullptr;
-
-  // Look through add of a constant offset on V1, V2, or both operands. This
-  // allows us to interpret the V + C' < C'' range idiom into a proper range.
   const APInt *Offset1 = nullptr, *Offset2 = nullptr;
-  if (V1 != V2) {
-    Value *X;
-    if (match(V1, m_Add(m_Value(X), m_APInt(Offset1))))
-      V1 = X;
-    if (match(V2, m_Add(m_Value(X), m_APInt(Offset2))))
-      V2 = X;
+  bool Matched = false;
+
+  if (match(ICmp1, m_ICmp(Pred1, m_Value(V1), m_APInt(C1))) &&
+      match(ICmp2, m_ICmp(Pred2, m_Value(V2), m_APInt(C2)))) {
+    // Look through add of a constant offset on V1, V2, or both operands. This
+    // allows us to interpret the V + C' < C'' range idiom into a proper range.
+    if (V1 != V2) {
+      Value *X;
+      if (match(V1, m_Add(m_Value(X), m_APInt(Offset1))))
+        V1 = X;
+      if (match(V2, m_Add(m_Value(X), m_APInt(Offset2))))
+        V2 = X;
+    }
+
+    Matched = V1 == V2;
   }
 
-  if (V1 != V2)
+  if (!Matched) {
+    Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr;
+    // Match (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
+    auto MaskPair =
+        getMaskedTypeForICmpPair(A, B, C, D, E, ICmp1, ICmp2, Pred1, Pred2);
+    // Match (icmp(A & B) ==/!= C1) &/| (icmp(A & B) ==/!= C2)
+    if (MaskPair && B == D && match(C, m_APIntAllowUndef(C1)) &&
+        match(E, m_APIntAllowUndef(C2)) &&
+        (match(ICmp1->getOperand(0),
+               m_CombineAnd(m_Value(V1),
+                            m_c_And(m_Specific(A), m_Specific(B)))) ||
+         match(ICmp2->getOperand(0),
+               m_CombineAnd(m_Value(V1),
+                            m_c_And(m_Specific(A), m_Specific(D)))))) {
+      V2 = V1;
+      Matched = true;
+    }
+  }
+
+  if (!Matched)
     return nullptr;
+  assert(V1 == V2);
 
   ConstantRange CR1 = ConstantRange::makeExactICmpRegion(
       IsAnd ? ICmpInst::getInversePredicate(Pred1) : Pred1, *C1);
@@ -3183,6 +3206,9 @@ Value *InstCombinerImpl::foldAndOrOfICmps(ICmpInst *LHS, ICmpInst *RHS,
                               Constant::getAllOnesValue(LHS0->getType()));
   }
 
+  if (Value *V = foldAndOrOfICmpsUsingRanges(LHS, RHS, IsAnd))
+    return V;
+
   // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
   if (!LHSC || !RHSC)
     return nullptr;
@@ -3256,7 +3282,7 @@ Value *InstCombinerImpl::foldAndOrOfICmps(ICmpInst *LHS, ICmpInst *RHS,
     }
   }
 
-  return foldAndOrOfICmpsUsingRanges(LHS, RHS, IsAnd);
+  return nullptr;
 }
 
 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
diff --git a/llvm/test/Transforms/InstCombine/icmp-range.ll b/llvm/test/Transforms/InstCombine/icmp-range.ll
index a6b629373946e30..24da5c61d741196 100644
--- a/llvm/test/Transforms/InstCombine/icmp-range.ll
+++ b/llvm/test/Transforms/InstCombine/icmp-range.ll
@@ -1480,11 +1480,8 @@ define <2 x i1> @icmp_ne_sext_eq_otherwise_vec(<2 x i32> %a) {
 ; tests from PR59555
 define i1 @isFloat(i64 %0) {
 ; CHECK-LABEL: @isFloat(
-; CHECK-NEXT:    [[TMP2:%.*]] = icmp ugt i64 [[TMP0:%.*]], 281474976710655
-; CHECK-NEXT:    [[TMP3:%.*]] = and i64 [[TMP0]], -281474976710656
-; CHECK-NEXT:    [[TMP4:%.*]] = icmp ne i64 [[TMP3]], 281474976710656
-; CHECK-NEXT:    [[TMP5:%.*]] = and i1 [[TMP2]], [[TMP4]]
-; CHECK-NEXT:    ret i1 [[TMP5]]
+; CHECK-NEXT:    [[TMP2:%.*]] = icmp ugt i64 [[TMP0:%.*]], 562949953421311
+; CHECK-NEXT:    ret i1 [[TMP2]]
 ;
   %2 = icmp ugt i64 %0, 281474976710655
   %3 = and i64 %0, -281474976710656

@nikic
Copy link
Contributor

nikic commented Sep 25, 2023

This looks like something that should be handled in foldLogOpOfMaskedICmps()? Why is this in foldAndOrOfICmpsUsingRanges()?

@dtcxzyw
Copy link
Member Author

dtcxzyw commented Sep 26, 2023

This looks like something that should be handled in foldLogOpOfMaskedICmps()? Why is this in foldAndOrOfICmpsUsingRanges()?

Because the following equivalent form is handled in foldAndOrOfICmpsUsingRanges:

define i1 @isFloat2(i64 %0) {
  %a = and i64 %0, -281474976710656
  %cmp1 = icmp ne i64 %a, 0
  %cmp2 = icmp ne i64 %a, 281474976710656
  %res = and i1 %cmp1, %cmp2
  ret i1 %res
}

godbolt: https://godbolt.org/z/j7YYhGPzr

@dtcxzyw
Copy link
Member Author

dtcxzyw commented Oct 11, 2023

Ping.


if (!Matched)
return nullptr;
assert(V1 == V2);
Copy link
Contributor

Choose a reason for hiding this comment

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

nit: assert msg

; CHECK-NEXT: ret i1 [[TMP5]]
; CHECK-NEXT: [[TMP2:%.*]] = icmp ugt i64 [[TMP0:%.*]], 562949953421311
; CHECK-NEXT: ret i1 [[TMP2]]
;
Copy link
Contributor

Choose a reason for hiding this comment

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

  1. please add more robust tests. Think you need some negative ones. Maybe:
  • B != D
  • some of the explicit pattern (with both sides eq/ne)
  • some non eq/ne patterns with a pred that can't be matched in mask pattern (maybe some signed preds, non-mask + ugt, etc...)
  • some vec tests with undef elements.

@dtcxzyw
Copy link
Member Author

dtcxzyw commented Sep 18, 2025

Covered by #158498.

@dtcxzyw dtcxzyw closed this Sep 18, 2025
@dtcxzyw dtcxzyw deleted the fold-nan-boxing branch September 18, 2025 02:48
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.

Nan boxing misoptimization: missing icmp combine?

4 participants