[X86] Create special case for (a-b) - (a<b) -> sbb a, b

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -52388,16 +52388,41 @@ static SDValue combineAddOrSubToADCOrSBB(bool IsSub, const SDLoc &DL, EVT VT,
     // Do not flip "e > c", where "c" is a constant, because Cmp instruction
     // cannot take an immediate as its first operand.
     //
-    if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.getNode()->hasOneUse() &&
-        EFLAGS.getValueType().isInteger() &&
-        !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
-      SDValue NewSub =
-          DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS), EFLAGS.getNode()->getVTList(),
-                      EFLAGS.getOperand(1), EFLAGS.getOperand(0));
-      SDValue NewEFLAGS = NewSub.getValue(EFLAGS.getResNo());
+    // If EFLAGS is from a CMP that compares the same operands as the earlier
+    // SUB producing X (i.e. CMP X, Y), we can directly use the carry flag with
+    // SBB/ADC without creating a flipped SUB.
+    if (EFLAGS.getOpcode() == X86ISD::CMP &&
+        EFLAGS.getValueType().isInteger() && X == EFLAGS.getOperand(0)) {
       return DAG.getNode(IsSub ? X86ISD::SBB : X86ISD::ADC, DL,
                          DAG.getVTList(VT, MVT::i32), X,
-                         DAG.getConstant(0, DL, VT), NewEFLAGS);
+                         DAG.getConstant(0, DL, VT), EFLAGS);
+    }
+
+    if (EFLAGS.getOpcode() == X86ISD::SUB &&
+        EFLAGS.getValueType().isInteger() &&
+        !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
+      // Only create NewSub if we know one of the folds will succeed to avoid
+      // introducing a temporary node that may persist and affect one-use checks
+      // below.
+      if (EFLAGS.getNode()->hasOneUse()) {
+        SDValue NewSub = DAG.getNode(
+            X86ISD::SUB, SDLoc(EFLAGS), EFLAGS.getNode()->getVTList(),
+            EFLAGS.getOperand(1), EFLAGS.getOperand(0));
+        SDValue NewEFLAGS = NewSub.getValue(EFLAGS.getResNo());
+        return DAG.getNode(IsSub ? X86ISD::SBB : X86ISD::ADC, DL,
+                           DAG.getVTList(VT, MVT::i32), X,
+                           DAG.getConstant(0, DL, VT), NewEFLAGS);
+      }
+
+      if (IsSub && X == EFLAGS.getValue(0)) {
+        SDValue NewSub = DAG.getNode(
+            X86ISD::SUB, SDLoc(EFLAGS), EFLAGS.getNode()->getVTList(),
+            EFLAGS.getOperand(1), EFLAGS.getOperand(0));
+        SDValue NewEFLAGS = NewSub.getValue(EFLAGS.getResNo());
+        return DAG.getNode(X86ISD::SBB, DL, DAG.getVTList(VT, MVT::i32),
+                           EFLAGS.getOperand(0), EFLAGS.getOperand(1),
+                           NewEFLAGS);
+      }
     }
   }
 

llvm/test/CodeGen/X86/sbb.ll

@@ -365,3 +365,32 @@ define i32 @uge_sext_add(i32 %0, i32 %1, i32 %2) {
   %6 = add nsw i32 %5, %0
   ret i32 %6
 }
+
+define i32 @sub_sub_ugt(i32 %a, i32 %b) {
+; CHECK-LABEL: sub_sub_ugt:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movl %edi, %eax
+; CHECK-NEXT:    cmpl %edi, %esi
+; CHECK-NEXT:    sbbl %esi, %eax
+; CHECK-NEXT:    retq
+  %cmp = icmp ugt i32 %a, %b
+  %conv = zext i1 %cmp to i32
+  %sub = sub i32 %a, %b
+  %res = sub i32 %sub, %conv
+  ret i32 %res
+}
+
+define i32 @sub_sub_ult(i32 %a, i32 %b) {
+; CHECK-LABEL: sub_sub_ult:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movl %edi, %eax
+; CHECK-NEXT:    cmpl %edi, %esi
+; CHECK-NEXT:    sbbl %esi, %eax
+; CHECK-NEXT:    retq
+  %cmp = icmp ult i32 %b, %a
+  %conv = zext i1 %cmp to i32
+  %sub = sub i32 %a, %b
+  %res = sub i32 %sub, %conv
+  ret i32 %res
+}
+