[AArch64] Signed comparison using CMN is safe when the subtraction is nsw (#141993)

nsw means no signed wrap, and 0 - INT_MIN is a signed wrap.

Now, this is going to be a point I need to get out of the way:

So is it okay to always transform a > -b into cmn if it is a signed
comparison, even if b is INT_MIN because -INT_MIN is undefined, at least
in C, because unless fwrapv is specified, opt puts nsw on signed integer
operations, allowing for more folds anyway.

Author: AZero13

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -3392,8 +3392,19 @@ bool isLegalCmpImmed(APInt C) {
   return isLegalArithImmed(C.abs().getZExtValue());
 }
 
-static bool cannotBeIntMin(SDValue CheckedVal, SelectionDAG &DAG) {
-  KnownBits KnownSrc = DAG.computeKnownBits(CheckedVal);
+static bool isSafeSignedCMN(SDValue Op, SelectionDAG &DAG) {
+  // 0 - INT_MIN sign wraps, so no signed wrap means cmn is safe.
+  if (Op->getFlags().hasNoSignedWrap())
+    return true;
+
+  // We can still figure out if the second operand is safe to use
+  // in a CMN instruction by checking if it is known to be not the minimum
+  // signed value. If it is not, then we can safely use CMN.
+  // Note: We can eventually remove this check and simply rely on
+  // Op->getFlags().hasNoSignedWrap() once SelectionDAG/ISelLowering
+  // consistently sets them appropriately when making said nodes.
+
+  KnownBits KnownSrc = DAG.computeKnownBits(Op.getOperand(1));
   return !KnownSrc.getSignedMinValue().isMinSignedValue();
 }
 
@@ -3402,7 +3413,7 @@ static bool cannotBeIntMin(SDValue CheckedVal, SelectionDAG &DAG) {
 // can be set differently by this operation. It comes down to whether
 // "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are then
 // everything is fine. If not then the optimization is wrong. Thus general
-// comparisons are only valid if op2 != 0.
+// comparisons are only valid if op2 != 0 and op2 != INT_MIN.
 //
 // So, finally, the only LLVM-native comparisons that don't mention C or V
 // are the ones that aren't unsigned comparisons. They're the only ones we can
@@ -3411,7 +3422,7 @@ static bool isCMN(SDValue Op, ISD::CondCode CC, SelectionDAG &DAG) {
   return Op.getOpcode() == ISD::SUB && isNullConstant(Op.getOperand(0)) &&
          (isIntEqualitySetCC(CC) ||
           (isUnsignedIntSetCC(CC) && DAG.isKnownNeverZero(Op.getOperand(1))) ||
-          (isSignedIntSetCC(CC) && cannotBeIntMin(Op.getOperand(1), DAG)));
+          (isSignedIntSetCC(CC) && isSafeSignedCMN(Op, DAG)));
 }
 
 static SDValue emitStrictFPComparison(SDValue LHS, SDValue RHS, const SDLoc &dl,

llvm/test/CodeGen/AArch64/cmp-to-cmn.ll

@@ -602,3 +602,49 @@ define i1 @almost_immediate_neg_ugt_64(i64 %x) {
   %cmp = icmp ugt i64 %x, -16773121
   ret i1 %cmp
 }
+
+define i1 @cmn_nsw(i32 %a, i32 %b) {
+; CHECK-LABEL: cmn_nsw:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    cmn w0, w1
+; CHECK-NEXT:    cset w0, gt
+; CHECK-NEXT:    ret
+  %sub = sub nsw i32 0, %b
+  %cmp = icmp sgt i32 %a, %sub
+  ret i1 %cmp
+}
+
+define i1 @cmn_nsw_64(i64 %a, i64 %b) {
+; CHECK-LABEL: cmn_nsw_64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    cmn x0, x1
+; CHECK-NEXT:    cset w0, gt
+; CHECK-NEXT:    ret
+  %sub = sub nsw i64 0, %b
+  %cmp = icmp sgt i64 %a, %sub
+  ret i1 %cmp
+}
+
+define i1 @cmn_nsw_neg(i32 %a, i32 %b) {
+; CHECK-LABEL: cmn_nsw_neg:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    neg w8, w1
+; CHECK-NEXT:    cmp w0, w8
+; CHECK-NEXT:    cset w0, gt
+; CHECK-NEXT:    ret
+  %sub = sub i32 0, %b
+  %cmp = icmp sgt i32 %a, %sub
+  ret i1 %cmp
+}
+
+define i1 @cmn_nsw_neg_64(i64 %a, i64 %b) {
+; CHECK-LABEL: cmn_nsw_neg_64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    neg x8, x1
+; CHECK-NEXT:    cmp x0, x8
+; CHECK-NEXT:    cset w0, gt
+; CHECK-NEXT:    ret
+  %sub = sub i64 0, %b
+  %cmp = icmp sgt i64 %a, %sub
+  ret i1 %cmp
+}