[llvm] [AArch64] Use ccmn to compare negative imm constants (PR #95825)

via llvm-commits llvm-commits at lists.llvm.org
Wed Jun 19 06:08:54 PDT 2024


https://github.com/AtariDreams updated https://github.com/llvm/llvm-project/pull/95825

>From 29f12c7c688f9c138ac55fec0bc4692188be32ec Mon Sep 17 00:00:00 2001
From: Rose <gfunni234 at gmail.com>
Date: Mon, 17 Jun 2024 14:25:39 -0400
Subject: [PATCH 1/2] Pre-commit test (NFC)

---
 llvm/test/CodeGen/AArch64/cmp-chains.ll | 25 +++++++++++++++++++++++++
 1 file changed, 25 insertions(+)

diff --git a/llvm/test/CodeGen/AArch64/cmp-chains.ll b/llvm/test/CodeGen/AArch64/cmp-chains.ll
index 1d9f39e518593..51b3fab43fe79 100644
--- a/llvm/test/CodeGen/AArch64/cmp-chains.ll
+++ b/llvm/test/CodeGen/AArch64/cmp-chains.ll
@@ -242,5 +242,30 @@ define i32 @true_or3(i32 %0, i32 %1, i32 %2) {
   %9 = zext i1 %8 to i32
   ret i32 %9
 }
+
+; (b > -3 && a < c)
+define i32 @neg_range_int(i32 %a, i32 %b, i32 %c) {
+; SDISEL-LABEL: neg_range_int:
+; SDISEL:       // %bb.0:
+; SDISEL-NEXT:    cmp w0, w2
+; SDISEL-NEXT:    mov w8, #-3 // =0xfffffffd
+; SDISEL-NEXT:    ccmp w1, w8, #4, lt
+; SDISEL-NEXT:    csel w0, w1, w0, gt
+; SDISEL-NEXT:    ret
+;
+; GISEL-LABEL: neg_range_int:
+; GISEL:       // %bb.0:
+; GISEL-NEXT:    mov w8, #-3 // =0xfffffffd
+; GISEL-NEXT:    cmp w0, w2
+; GISEL-NEXT:    ccmp w1, w8, #4, lt
+; GISEL-NEXT:    csel w0, w1, w0, gt
+; GISEL-NEXT:    ret
+  %cmp = icmp sgt i32 %b, -3
+  %cmp1 = icmp slt i32 %a, %c
+  %or.cond = and i1 %cmp, %cmp1
+  %retval.0 = select i1 %or.cond, i32 %b, i32 %a
+  ret i32 %retval.0
+}
+
 ;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
 ; CHECK: {{.*}}

>From e6256e2e41369604b3639751ab0f5b4a425fef09 Mon Sep 17 00:00:00 2001
From: Rose <gfunni234 at gmail.com>
Date: Mon, 17 Jun 2024 14:28:44 -0400
Subject: [PATCH 2/2] [AArch64] Use ccmn to compare negative imm constants

Because ccmn is like a + b, it works when using cmp but with values that when negated are in the range of 1 through 31.
---
 .../Target/AArch64/AArch64ISelLowering.cpp    |  6 +++++
 .../GISel/AArch64InstructionSelector.cpp      |  9 +++++---
 llvm/test/CodeGen/AArch64/arm64-ccmp.ll       |  3 +--
 llvm/test/CodeGen/AArch64/cmp-chains.ll       | 23 +++++--------------
 4 files changed, 19 insertions(+), 22 deletions(-)

diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index 9f6f66e9e0c70..d82a5453bf9f3 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -3515,6 +3515,12 @@ static SDValue emitConditionalComparison(SDValue LHS, SDValue RHS,
       RHS = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, RHS);
     }
     Opcode = AArch64ISD::FCCMP;
+  } else if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(RHS)) {
+    APInt Imm = Const->getAPIntValue();
+    if (imm.isNegative() && imm.sgt(-32)) {
+      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)) {
diff --git a/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp b/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp
index d32007ec45fb6..0357a7206c478 100644
--- a/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp
+++ b/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp
@@ -4667,7 +4667,6 @@ MachineInstr *AArch64InstructionSelector::emitConditionalComparison(
     Register LHS, Register RHS, CmpInst::Predicate CC,
     AArch64CC::CondCode Predicate, AArch64CC::CondCode OutCC,
     MachineIRBuilder &MIB) const {
-  // TODO: emit CMN as an optimization.
   auto &MRI = *MIB.getMRI();
   LLT OpTy = MRI.getType(LHS);
   unsigned CCmpOpc;
@@ -4675,10 +4674,12 @@ MachineInstr *AArch64InstructionSelector::emitConditionalComparison(
   if (CmpInst::isIntPredicate(CC)) {
     assert(OpTy.getSizeInBits() == 32 || OpTy.getSizeInBits() == 64);
     C = getIConstantVRegValWithLookThrough(RHS, MRI);
-    if (C && C->Value.ult(32))
+    if (!C || C->Value.sgt(31) || C->Value.slt(-31))
+      CCmpOpc = OpTy.getSizeInBits() == 32 ? AArch64::CCMPWr : AArch64::CCMPXr;
+    else if (C->Value.ule(31))
       CCmpOpc = OpTy.getSizeInBits() == 32 ? AArch64::CCMPWi : AArch64::CCMPXi;
     else
-      CCmpOpc = OpTy.getSizeInBits() == 32 ? AArch64::CCMPWr : AArch64::CCMPXr;
+      CCmpOpc = OpTy.getSizeInBits() == 32 ? AArch64::CCMNWi : AArch64::CCMNXi;
   } else {
     assert(OpTy.getSizeInBits() == 16 || OpTy.getSizeInBits() == 32 ||
            OpTy.getSizeInBits() == 64);
@@ -4703,6 +4704,8 @@ MachineInstr *AArch64InstructionSelector::emitConditionalComparison(
       MIB.buildInstr(CCmpOpc, {}, {LHS});
   if (CCmpOpc == AArch64::CCMPWi || CCmpOpc == AArch64::CCMPXi)
     CCmp.addImm(C->Value.getZExtValue());
+  else if (CCmpOpc == AArch64::CCMNWi || CCmpOpc == AArch64::CCMNXi)
+    CCmp.addImm(C->Value.abs().getZExtValue());
   else
     CCmp.addReg(RHS);
   CCmp.addImm(NZCV).addImm(Predicate);
diff --git a/llvm/test/CodeGen/AArch64/arm64-ccmp.ll b/llvm/test/CodeGen/AArch64/arm64-ccmp.ll
index 5d3b2d3649e1b..b6702bba1598c 100644
--- a/llvm/test/CodeGen/AArch64/arm64-ccmp.ll
+++ b/llvm/test/CodeGen/AArch64/arm64-ccmp.ll
@@ -706,10 +706,9 @@ define i32 @select_noccmp3(i32 %v0, i32 %v1, i32 %v2) {
 ; GISEL:       ; %bb.0:
 ; GISEL-NEXT:    mov w8, #99 ; =0x63
 ; GISEL-NEXT:    sub w9, w0, #45
-; GISEL-NEXT:    mov w10, #-23 ; =0xffffffe9
 ; GISEL-NEXT:    cmp w0, #77
 ; GISEL-NEXT:    ccmp w0, w8, #4, ne
-; GISEL-NEXT:    ccmp w9, w10, #2, eq
+; GISEL-NEXT:    ccmn w9, #23, #2, eq
 ; GISEL-NEXT:    ccmp w0, #14, #0, lo
 ; GISEL-NEXT:    csel w0, w1, w2, hs
 ; GISEL-NEXT:    ret
diff --git a/llvm/test/CodeGen/AArch64/cmp-chains.ll b/llvm/test/CodeGen/AArch64/cmp-chains.ll
index 51b3fab43fe79..14cb0c82b1c03 100644
--- a/llvm/test/CodeGen/AArch64/cmp-chains.ll
+++ b/llvm/test/CodeGen/AArch64/cmp-chains.ll
@@ -245,21 +245,12 @@ define i32 @true_or3(i32 %0, i32 %1, i32 %2) {
 
 ; (b > -3 && a < c)
 define i32 @neg_range_int(i32 %a, i32 %b, i32 %c) {
-; SDISEL-LABEL: neg_range_int:
-; SDISEL:       // %bb.0:
-; SDISEL-NEXT:    cmp w0, w2
-; SDISEL-NEXT:    mov w8, #-3 // =0xfffffffd
-; SDISEL-NEXT:    ccmp w1, w8, #4, lt
-; SDISEL-NEXT:    csel w0, w1, w0, gt
-; SDISEL-NEXT:    ret
-;
-; GISEL-LABEL: neg_range_int:
-; GISEL:       // %bb.0:
-; GISEL-NEXT:    mov w8, #-3 // =0xfffffffd
-; GISEL-NEXT:    cmp w0, w2
-; GISEL-NEXT:    ccmp w1, w8, #4, lt
-; GISEL-NEXT:    csel w0, w1, w0, gt
-; GISEL-NEXT:    ret
+; CHECK-LABEL: neg_range_int:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    cmp w0, w2
+; CHECK-NEXT:    ccmn w1, #3, #4, lt
+; CHECK-NEXT:    csel w0, w1, w0, gt
+; CHECK-NEXT:    ret
   %cmp = icmp sgt i32 %b, -3
   %cmp1 = icmp slt i32 %a, %c
   %or.cond = and i1 %cmp, %cmp1
@@ -267,5 +258,3 @@ define i32 @neg_range_int(i32 %a, i32 %b, i32 %c) {
   ret i32 %retval.0
 }
 
-;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
-; CHECK: {{.*}}



More information about the llvm-commits mailing list