From ad3c6289b6552723bc80d66275b0ed0d771746a4 Mon Sep 17 00:00:00 2001
From: Quentin Colombet <qcolombet@apple.com>
Date: Thu, 21 Aug 2014 18:10:07 +0000
Subject: [PATCH] [AArch64] Run a peephole pass right after AdvSIMD pass. The
 AdvSIMD pass may produce copies that are not coalescer-friendly. The peephole
 optimizer knows how to fix that as demonstrated in the test case.

<rdar://problem/12702965>


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@216200 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/AArch64/AArch64TargetMachine.cpp  |  6 ++++-
 test/CodeGen/AArch64/arm64-AdvSIMD-Scalar.ll | 25 ++++++++++++++++++--
 2 files changed, 28 insertions(+), 3 deletions(-)

diff --git a/lib/Target/AArch64/AArch64TargetMachine.cpp b/lib/Target/AArch64/AArch64TargetMachine.cpp
index d34be77f0ff..f3172a62f03 100644
--- a/lib/Target/AArch64/AArch64TargetMachine.cpp
+++ b/lib/Target/AArch64/AArch64TargetMachine.cpp
@@ -195,8 +195,12 @@ bool AArch64PassConfig::addILPOpts() {
 
 bool AArch64PassConfig::addPreRegAlloc() {
   // Use AdvSIMD scalar instructions whenever profitable.
-  if (TM->getOptLevel() != CodeGenOpt::None && EnableAdvSIMDScalar)
+  if (TM->getOptLevel() != CodeGenOpt::None && EnableAdvSIMDScalar) {
     addPass(createAArch64AdvSIMDScalar());
+    // The AdvSIMD pass may produce copies that can be rewritten to
+    // be register coaleascer friendly.
+    addPass(&PeepholeOptimizerID);
+  }
   return true;
 }
 
diff --git a/test/CodeGen/AArch64/arm64-AdvSIMD-Scalar.ll b/test/CodeGen/AArch64/arm64-AdvSIMD-Scalar.ll
index a3d7727c8ae..6266d1cc9b3 100644
--- a/test/CodeGen/AArch64/arm64-AdvSIMD-Scalar.ll
+++ b/test/CodeGen/AArch64/arm64-AdvSIMD-Scalar.ll
@@ -1,15 +1,36 @@
-; RUN: llc < %s -march=arm64 -aarch64-neon-syntax=apple -aarch64-simd-scalar=true -asm-verbose=false | FileCheck %s
-; RUN: llc < %s -march=arm64 -aarch64-neon-syntax=generic -aarch64-simd-scalar=true -asm-verbose=false | FileCheck %s -check-prefix=GENERIC
+; RUN: llc < %s -march=arm64 -aarch64-neon-syntax=apple -aarch64-simd-scalar=true -asm-verbose=false -disable-adv-copy-opt=true | FileCheck %s -check-prefix=CHECK -check-prefix=CHECK-NOOPT
+; RUN: llc < %s -march=arm64 -aarch64-neon-syntax=apple -aarch64-simd-scalar=true -asm-verbose=false -disable-adv-copy-opt=false | FileCheck %s -check-prefix=CHECK -check-prefix=CHECK-OPT
+; RUN: llc < %s -march=arm64 -aarch64-neon-syntax=generic -aarch64-simd-scalar=true -asm-verbose=false -disable-adv-copy-opt=true | FileCheck %s -check-prefix=GENERIC -check-prefix=GENERIC-NOOPT
+; RUN: llc < %s -march=arm64 -aarch64-neon-syntax=generic -aarch64-simd-scalar=true -asm-verbose=false -disable-adv-copy-opt=false | FileCheck %s -check-prefix=GENERIC -check-prefix=GENERIC-OPT
 
 define <2 x i64> @bar(<2 x i64> %a, <2 x i64> %b) nounwind readnone {
 ; CHECK-LABEL: bar:
 ; CHECK: add.2d	v[[REG:[0-9]+]], v0, v1
 ; CHECK: add	d[[REG3:[0-9]+]], d[[REG]], d1
+; Without advanced copy optimization, we end up with cross register
+; banks copies that cannot be coalesced.
+; CHECK-NOOPT: fmov [[COPY_REG3:x[0-9]+]], d[[REG3]]
+; With advanced copy optimization, we end up with just one copy
+; to insert the computed high part into the V register. 
+; CHECK-OPT-NOT: fmov
 ; CHECK: sub	d[[REG2:[0-9]+]], d[[REG]], d1
+; CHECK: fmov [[COPY_REG2:x[0-9]+]], d[[REG2]]
+; CHECK-NOOPT: fmov d0, [[COPY_REG3]]
+; CHECK-OPT-NOT: fmov
+; CHECK: ins.d v0[1], [[COPY_REG2]]
+; CHECK-NEXT: ret
+;
 ; GENERIC-LABEL: bar:
 ; GENERIC: add	v[[REG:[0-9]+]].2d, v0.2d, v1.2d
 ; GENERIC: add	d[[REG3:[0-9]+]], d[[REG]], d1
+; GENERIC-NOOPT: fmov [[COPY_REG3:x[0-9]+]], d[[REG3]]
+; GENERIC-OPT-NOT: fmov
 ; GENERIC: sub	d[[REG2:[0-9]+]], d[[REG]], d1
+; GENERIC: fmov [[COPY_REG2:x[0-9]+]], d[[REG2]]
+; GENERIC-NOOPT: fmov d0, [[COPY_REG3]]
+; GENERIC-OPT-NOT: fmov
+; GENERIC: ins v0.d[1], [[COPY_REG2]]
+; GENERIC-NEXT: ret
   %add = add <2 x i64> %a, %b
   %vgetq_lane = extractelement <2 x i64> %add, i32 0
   %vgetq_lane2 = extractelement <2 x i64> %b, i32 0
-- 
2.34.1