[X86] Implement canceling out of XOR with sub nodes

AZero13 · AZero13 · commit 8301e5db1468 · 2025-08-23T12:48:48.000-04:00
This can happen when comparing the two bit positions.
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -49009,6 +49009,28 @@ static SDValue combineSetCCMOVMSK(SDValue EFLAGS, X86::CondCode &CC,
 static SDValue combineSetCCEFLAGS(SDValue EFLAGS, X86::CondCode &CC,
                                   SelectionDAG &DAG,
                                   const X86Subtarget &Subtarget) {
+  // For EQ/NE, rewrite SUB(xor A,B, xor C,B) -> SUB(A,C) and
+  //            SUB(add A,K, add C,K)       -> SUB(A,C)
+  // Only safe if the arithmetic result is unused, since we only
+  // care about flags for EQ/NE.
+
+  // TODO: Can we do this for other comparions like 31 ^ A > 31 ^ B ?
+  if ((CC == X86::COND_E || CC == X86::COND_NE) &&
+      EFLAGS.getOpcode() == X86ISD::SUB &&
+      !EFLAGS.getNode()->hasAnyUseOfValue(0)) {
+    using namespace llvm::SDPatternMatch;
+    SDValue LHS = EFLAGS.getOperand(0);
+    SDValue RHS = EFLAGS.getOperand(1);
+    SDValue A, B, C;
+
+    // (xor A, B) - (xor C, B) -> A - C
+    if (sd_match(LHS, m_Xor(m_Value(A), m_Value(B))) &&
+        sd_match(RHS, m_Xor(m_Value(C), m_Specific(B)))) {
+      DAG.UpdateNodeOperands(EFLAGS.getNode(), A, C);
+      return SDValue(EFLAGS.getNode(), 1);
+    }
+  }
+
   if (CC == X86::COND_B)
     if (SDValue Flags = combineCarryThroughADD(EFLAGS, DAG))
       return Flags;
diff --git a/llvm/test/CodeGen/X86/ctlz.ll b/llvm/test/CodeGen/X86/ctlz.ll
@@ -1339,9 +1339,7 @@ define i32 @bsr_eq(i32 %a, i32 %b) {
 ; X86-LABEL: bsr_eq:
 ; X86:       # %bb.0:
 ; X86-NEXT:    bsrl {{[0-9]+}}(%esp), %ecx
-; X86-NEXT:    xorl $31, %ecx
 ; X86-NEXT:    bsrl {{[0-9]+}}(%esp), %edx
-; X86-NEXT:    xorl $31, %edx
 ; X86-NEXT:    xorl %eax, %eax
 ; X86-NEXT:    cmpl %edx, %ecx
 ; X86-NEXT:    sete %al
@@ -1350,9 +1348,7 @@ define i32 @bsr_eq(i32 %a, i32 %b) {
 ; X64-LABEL: bsr_eq:
 ; X64:       # %bb.0:
 ; X64-NEXT:    bsrl %edi, %ecx
-; X64-NEXT:    xorl $31, %ecx
 ; X64-NEXT:    bsrl %esi, %edx
-; X64-NEXT:    xorl $31, %edx
 ; X64-NEXT:    xorl %eax, %eax
 ; X64-NEXT:    cmpl %edx, %ecx
 ; X64-NEXT:    sete %al
