From 2cf6269d1194c58e48f03bba2a65c165f114bc2e Mon Sep 17 00:00:00 2001
From: David Green <david.green@arm.com>
Date: Fri, 21 Jun 2024 11:02:21 +0100
Subject: [PATCH] [AArch64] Remove superfluous sxtw in peephole opt

Across a basic-block we might have in i32 extract from a value that only
operates on upper bits (for example a sxtw). We can replace the COPY with a
new version skipping the sxtw.
---
 .../Target/AArch64/AArch64MIPeepholeOpt.cpp   | 31 +++++++++++++++++++
 .../CodeGen/AArch64/aarch64-mull-masks.ll     | 12 +++----
 2 files changed, 35 insertions(+), 8 deletions(-)

diff --git a/llvm/lib/Target/AArch64/AArch64MIPeepholeOpt.cpp b/llvm/lib/Target/AArch64/AArch64MIPeepholeOpt.cpp
index 22da7ddef98a2a..ccd3eeb31ffbe7 100644
--- a/llvm/lib/Target/AArch64/AArch64MIPeepholeOpt.cpp
+++ b/llvm/lib/Target/AArch64/AArch64MIPeepholeOpt.cpp
@@ -128,6 +128,7 @@ struct AArch64MIPeepholeOpt : public MachineFunctionPass {
   bool visitINSviGPR(MachineInstr &MI, unsigned Opc);
   bool visitINSvi64lane(MachineInstr &MI);
   bool visitFMOVDr(MachineInstr &MI);
+  bool visitCopy(MachineInstr &MI);
   bool runOnMachineFunction(MachineFunction &MF) override;
 
   StringRef getPassName() const override {
@@ -690,6 +691,33 @@ bool AArch64MIPeepholeOpt::visitFMOVDr(MachineInstr &MI) {
   return true;
 }
 
+// Acrocss a basic-block we might have in i32 extract from a value that only
+// operates on upper bits (for example a sxtw). We can replace the COPY with a
+// new version skipping the sxtw.
+bool AArch64MIPeepholeOpt::visitCopy(MachineInstr &MI) {
+  if (MI.getOperand(1).getSubReg() != AArch64::sub_32 ||
+      !MRI->hasOneNonDBGUse(MI.getOperand(1).getReg()))
+    return false;
+
+  MachineInstr *SrcMI = MRI->getUniqueVRegDef(MI.getOperand(1).getReg());
+  MachineInstr *CopyMI = SrcMI;
+  if (SrcMI && SrcMI->isFullCopy() &&
+      MRI->hasOneNonDBGUse(SrcMI->getOperand(1).getReg()))
+    SrcMI = MRI->getUniqueVRegDef(SrcMI->getOperand(1).getReg());
+
+  if (!SrcMI || SrcMI->getOpcode() != AArch64::SBFMXri ||
+      SrcMI->getOperand(2).getImm() != 0 || SrcMI->getOperand(3).getImm() != 31)
+    return false;
+
+  Register SrcReg = SrcMI->getOperand(1).getReg();
+  MRI->constrainRegClass(SrcReg, MRI->getRegClass(MI.getOperand(1).getReg()));
+  MI.getOperand(1).setReg(SrcReg);
+  if (CopyMI != SrcMI)
+    CopyMI->eraseFromParent();
+  SrcMI->eraseFromParent();
+  return true;
+}
+
 bool AArch64MIPeepholeOpt::runOnMachineFunction(MachineFunction &MF) {
   if (skipFunction(MF.getFunction()))
     return false;
@@ -771,6 +799,9 @@ bool AArch64MIPeepholeOpt::runOnMachineFunction(MachineFunction &MF) {
       case AArch64::FMOVDr:
         Changed |= visitFMOVDr(MI);
         break;
+      case AArch64::COPY:
+        Changed |= visitCopy(MI);
+        break;
       }
     }
   }
diff --git a/llvm/test/CodeGen/AArch64/aarch64-mull-masks.ll b/llvm/test/CodeGen/AArch64/aarch64-mull-masks.ll
index e41eb7d38c370d..058cbbe9ff13ca 100644
--- a/llvm/test/CodeGen/AArch64/aarch64-mull-masks.ll
+++ b/llvm/test/CodeGen/AArch64/aarch64-mull-masks.ll
@@ -281,8 +281,7 @@ define i64 @smull_ldrsw_shift(ptr %x0, i64 %x1) {
 ; CHECK-LABEL: smull_ldrsw_shift:
 ; CHECK:       // %bb.0: // %entry
 ; CHECK-NEXT:    ldrsw x8, [x0]
-; CHECK-NEXT:    sxtw x9, w1
-; CHECK-NEXT:    smull x0, w8, w9
+; CHECK-NEXT:    smull x0, w8, w1
 ; CHECK-NEXT:    ret
 entry:
   %ext64 = load i32, ptr %x0
@@ -490,8 +489,7 @@ define i64 @smaddl_ldrsw_shift(ptr %x0, i64 %x1, i64 %x2) {
 ; CHECK-LABEL: smaddl_ldrsw_shift:
 ; CHECK:       // %bb.0: // %entry
 ; CHECK-NEXT:    ldrsw x8, [x0]
-; CHECK-NEXT:    sxtw x9, w1
-; CHECK-NEXT:    smaddl x0, w8, w9, x2
+; CHECK-NEXT:    smaddl x0, w8, w1, x2
 ; CHECK-NEXT:    ret
 entry:
   %ext64 = load i32, ptr %x0
@@ -654,8 +652,7 @@ define i64 @smnegl_ldrsw_shift(ptr %x0, i64 %x1) {
 ; CHECK-LABEL: smnegl_ldrsw_shift:
 ; CHECK:       // %bb.0: // %entry
 ; CHECK-NEXT:    ldrsw x8, [x0]
-; CHECK-NEXT:    sxtw x9, w1
-; CHECK-NEXT:    smnegl x0, w8, w9
+; CHECK-NEXT:    smnegl x0, w8, w1
 ; CHECK-NEXT:    ret
 entry:
   %ext64 = load i32, ptr %x0
@@ -818,8 +815,7 @@ define i64 @smsubl_ldrsw_shift(ptr %x0, i64 %x1, i64 %x2) {
 ; CHECK-LABEL: smsubl_ldrsw_shift:
 ; CHECK:       // %bb.0: // %entry
 ; CHECK-NEXT:    ldrsw x8, [x0]
-; CHECK-NEXT:    sxtw x9, w1
-; CHECK-NEXT:    smsubl x0, w8, w9, x2
+; CHECK-NEXT:    smsubl x0, w8, w1, x2
 ; CHECK-NEXT:    ret
 entry:
   %ext64 = load i32, ptr %x0