[llvm] [AArch64] Use isKnownNonZero to optimize eligible compares to cmn (PR #96349)

via llvm-commits llvm-commits at lists.llvm.org
Tue Jul 16 07:53:23 PDT 2024


https://github.com/AtariDreams updated https://github.com/llvm/llvm-project/pull/96349

>From ba333474c6ce67dd4a0d62fb33ff822b4ab4c899 Mon Sep 17 00:00:00 2001
From: Rose <gfunni234 at gmail.com>
Date: Fri, 21 Jun 2024 15:12:41 -0400
Subject: [PATCH 1/2] Pre-commit tests (NFC)

---
 llvm/test/CodeGen/AArch64/cmp-select-sign.ll | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/llvm/test/CodeGen/AArch64/cmp-select-sign.ll b/llvm/test/CodeGen/AArch64/cmp-select-sign.ll
index 09a6e26fe5a40..10ad2155efc50 100644
--- a/llvm/test/CodeGen/AArch64/cmp-select-sign.ll
+++ b/llvm/test/CodeGen/AArch64/cmp-select-sign.ll
@@ -262,4 +262,19 @@ define <4 x i65> @sign_4xi65(<4 x i65> %a) {
   ret <4 x i65> %res
 }
 
+define i32 @or_neg(i32 %x, i32 %y) {
+; CHECK-LABEL: or_neg:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    orr w8, w0, #0x1
+; CHECK-NEXT:    neg w8, w8
+; CHECK-NEXT:    cmp w8, w1
+; CHECK-NEXT:    cset w0, gt
+; CHECK-NEXT:    ret
+  %3 = or i32 %x, 1
+  %4 = sub i32 0, %3
+  %5 = icmp sgt i32 %4, %y
+  %6 = zext i1 %5 to i32
+  ret i32 %6
+}
+
 declare void @use_4xi1(<4 x i1>)

>From 2968edb305b92af6ef3d920687b0543b52b36ad1 Mon Sep 17 00:00:00 2001
From: Rose <gfunni234 at gmail.com>
Date: Fri, 21 Jun 2024 15:26:02 -0400
Subject: [PATCH 2/2] [AArch64] Use isKnownNonZero to optimize to cmn instead
 of cmp

Turning a cmp into cmn saves an extra mov and negate instruction, so take that into account when choosing when to flip the compare operands.

Also do not consider right-hand operands whose absolute value can be encoded into a cmn.
---
 .../Target/AArch64/AArch64ISelLowering.cpp    | 55 ++++++++++++-------
 llvm/test/CodeGen/AArch64/cmp-select-sign.ll  |  5 +-
 2 files changed, 36 insertions(+), 24 deletions(-)

diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index df9b0ae1a632f..e52765ee9859b 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -3385,6 +3385,11 @@ static bool isLegalArithImmed(uint64_t C) {
   return IsLegal;
 }
 
+static bool cannotBeIntMin(SDValue CheckedVal, SelectionDAG &DAG) {
+  KnownBits KnownSrc = DAG.computeKnownBits(CheckedVal);
+  return !KnownSrc.getSignedMinValue().isMinSignedValue();
+}
+
 // Can a (CMP op1, (sub 0, op2) be turned into a CMN instruction on
 // the grounds that "op1 - (-op2) == op1 + op2" ? Not always, the C and V flags
 // can be set differently by this operation. It comes down to whether
@@ -3392,12 +3397,15 @@ static bool isLegalArithImmed(uint64_t C) {
 // everything is fine. If not then the optimization is wrong. Thus general
 // comparisons are only valid if op2 != 0.
 //
-// So, finally, the only LLVM-native comparisons that don't mention C and V
-// are SETEQ and SETNE. They're the only ones we can safely use CMN for in
-// the absence of information about op2.
-static bool isCMN(SDValue Op, ISD::CondCode CC) {
-  return Op.getOpcode() == ISD::SUB && isNullConstant(Op.getOperand(0)) &&
-         (CC == ISD::SETEQ || CC == ISD::SETNE);
+// 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
+// safely use CMN for in the absence of information about op2.
+static bool isCMN(SDValue Op, SDValue CheckedVal, ISD::CondCode CC,
+                  SelectionDAG &DAG) {
+  return isNullConstant(Op.getOperand(0)) &&
+         (isIntEqualitySetCC(CC) ||
+          (isUnsignedIntSetCC(CC) ? DAG.isKnownNeverZero(CheckedVal)
+                                  : cannotBeIntMin(CheckedVal, DAG)));
 }
 
 static SDValue emitStrictFPComparison(SDValue LHS, SDValue RHS, const SDLoc &dl,
@@ -3442,11 +3450,12 @@ static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
   // register to WZR/XZR if it ends up being unused.
   unsigned Opcode = AArch64ISD::SUBS;
 
-  if (isCMN(RHS, CC)) {
+  if (RHS.getOpcode() == ISD::SUB && isCMN(RHS, RHS.getOperand(1), CC, DAG)) {
     // Can we combine a (CMP op1, (sub 0, op2) into a CMN instruction ?
     Opcode = AArch64ISD::ADDS;
     RHS = RHS.getOperand(1);
-  } else if (isCMN(LHS, CC)) {
+  } else if (LHS.getOpcode() == ISD::SUB && isNullConstant(LHS.getOperand(0)) &&
+             isIntEqualitySetCC(CC)) {
     // As we are looking for EQ/NE compares, the operands can be commuted ; can
     // we combine a (CMP (sub 0, op1), op2) into a CMN instruction ?
     Opcode = AArch64ISD::ADDS;
@@ -3456,10 +3465,9 @@ static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
       // Similarly, (CMP (and X, Y), 0) can be implemented with a TST
       // (a.k.a. ANDS) except that the flags are only guaranteed to work for one
       // of the signed comparisons.
-      const SDValue ANDSNode = DAG.getNode(AArch64ISD::ANDS, dl,
-                                           DAG.getVTList(VT, MVT_CC),
-                                           LHS.getOperand(0),
-                                           LHS.getOperand(1));
+      const SDValue ANDSNode =
+          DAG.getNode(AArch64ISD::ANDS, dl, DAG.getVTList(VT, MVT_CC),
+                      LHS.getOperand(0), LHS.getOperand(1));
       // Replace all users of (and X, Y) with newly generated (ands X, Y)
       DAG.ReplaceAllUsesWith(LHS, ANDSNode);
       return ANDSNode.getValue(1);
@@ -3548,13 +3556,16 @@ static SDValue emitConditionalComparison(SDValue LHS, SDValue RHS,
       Opcode = AArch64ISD::CCMN;
       RHS = DAG.getConstant(Imm.abs(), DL, Const->getValueType(0));
     }
-  } else if (RHS.getOpcode() == ISD::SUB) {
-    SDValue SubOp0 = RHS.getOperand(0);
-    if (isNullConstant(SubOp0) && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
-      // See emitComparison() on why we can only do this for SETEQ and SETNE.
-      Opcode = AArch64ISD::CCMN;
-      RHS = RHS.getOperand(1);
-    }
+  } else if (RHS.getOpcode() == ISD::SUB &&
+             isCMN(RHS, RHS.getOperand(1), CC, DAG)) {
+    Opcode = AArch64ISD::CCMN;
+    RHS = RHS.getOperand(1);
+  } else if (LHS.getOpcode() == ISD::SUB && isNullConstant(LHS.getOperand(0)) &&
+             isIntEqualitySetCC(CC)) {
+    // As we are looking for EQ/NE compares, the operands can be commuted ; can
+    // we combine a (CMP (sub 0, op1), op2) into a CMN instruction ?
+    Opcode = AArch64ISD::CCMN;
+    LHS = LHS.getOperand(1);
   }
   if (Opcode == 0)
     Opcode = AArch64ISD::CCMP;
@@ -3872,8 +3883,10 @@ static SDValue getAArch64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
   //    cmp     w12, w11, lsl #1
   if (!isa<ConstantSDNode>(RHS) ||
       !isLegalArithImmed(RHS->getAsAPIntVal().abs().getZExtValue())) {
-    bool LHSIsCMN = isCMN(LHS, CC);
-    bool RHSIsCMN = isCMN(RHS, CC);
+    bool LHSIsCMN =
+        LHS.getOpcode() == ISD::SUB && isCMN(LHS, LHS.getOperand(1), CC, DAG);
+    bool RHSIsCMN =
+        RHS.getOpcode() == ISD::SUB && isCMN(RHS, RHS.getOperand(1), CC, DAG);
     SDValue TheLHS = LHSIsCMN ? LHS.getOperand(1) : LHS;
     SDValue TheRHS = RHSIsCMN ? RHS.getOperand(1) : RHS;
 
diff --git a/llvm/test/CodeGen/AArch64/cmp-select-sign.ll b/llvm/test/CodeGen/AArch64/cmp-select-sign.ll
index 10ad2155efc50..65c55da7502fa 100644
--- a/llvm/test/CodeGen/AArch64/cmp-select-sign.ll
+++ b/llvm/test/CodeGen/AArch64/cmp-select-sign.ll
@@ -266,9 +266,8 @@ define i32 @or_neg(i32 %x, i32 %y) {
 ; CHECK-LABEL: or_neg:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    orr w8, w0, #0x1
-; CHECK-NEXT:    neg w8, w8
-; CHECK-NEXT:    cmp w8, w1
-; CHECK-NEXT:    cset w0, gt
+; CHECK-NEXT:    cmn w1, w8
+; CHECK-NEXT:    cset w0, lt
 ; CHECK-NEXT:    ret
   %3 = or i32 %x, 1
   %4 = sub i32 0, %3



More information about the llvm-commits mailing list