From dab10b37356ae5eda7f837243f69422821ef887b Mon Sep 17 00:00:00 2001
From: Quentin Colombet <qcolombet@apple.com>
Date: Wed, 13 Jan 2016 00:02:40 +0000
Subject: [PATCH] [ARM] Mark VMOV with immediate: isAsCheapAsMove. VMOVs are
 not strictly speaking cheap, but they are as expensive as a vector copy
 (VORR), so we should prefer rematerialization over splitting when it applies.

rdar://problem/23754176


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@257545 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/ARM/ARMInstrNEON.td      |  7 +++-
 test/CodeGen/ARM/zero-cycle-zero.ll | 58 +++++++++++------------------
 2 files changed, 26 insertions(+), 39 deletions(-)

diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
index 7020ffb41b6..defef4ea907 100644
--- a/lib/Target/ARM/ARMInstrNEON.td
+++ b/lib/Target/ARM/ARMInstrNEON.td
@@ -5689,7 +5689,10 @@ def : NEONInstAlias<"vmov${p} $Vd, $Vm",
 
 //   VMOV     : Vector Move (Immediate)
 
-let isReMaterializable = 1 in {
+// Although VMOVs are not strictly speaking cheap, they are as expensive
+// as their copies counterpart (VORR), so we should prefer rematerialization
+// over splitting when it applies.
+let isReMaterializable = 1, isAsCheapAsAMove=1 in {
 def VMOVv8i8  : N1ModImm<1, 0b000, 0b1110, 0, 0, 0, 1, (outs DPR:$Vd),
                          (ins nImmSplatI8:$SIMM), IIC_VMOVImm,
                          "vmov", "i8", "$Vd, $SIMM", "",
@@ -5744,7 +5747,7 @@ def VMOVv4f32 : N1ModImm<1, 0b000, 0b1111, 0, 1, 0, 1, (outs QPR:$Vd),
                          (ins nImmVMOVF32:$SIMM), IIC_VMOVImm,
                          "vmov", "f32", "$Vd, $SIMM", "",
                          [(set QPR:$Vd, (v4f32 (NEONvmovFPImm timm:$SIMM)))]>;
-} // isReMaterializable
+} // isReMaterializable, isAsCheapAsAMove
 
 // Add support for bytes replication feature, so it could be GAS compatible.
 // E.g. instructions below:
diff --git a/test/CodeGen/ARM/zero-cycle-zero.ll b/test/CodeGen/ARM/zero-cycle-zero.ll
index 121a87f5b84..4e8696f4418 100644
--- a/test/CodeGen/ARM/zero-cycle-zero.ll
+++ b/test/CodeGen/ARM/zero-cycle-zero.ll
@@ -1,26 +1,19 @@
-; RUN: llc -mtriple=armv8 -mcpu=cyclone < %s | FileCheck %s --check-prefix=CHECK-CYCLONE
-; RUN: llc -mtriple=armv8 -mcpu=swift < %s | FileCheck %s --check-prefix=CHECK-SWIFT
+; RUN: llc -mtriple=armv8 -mcpu=cyclone < %s | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-NOTSWIFT
+; RUN: llc -mtriple=armv8 -mcpu=swift < %s | FileCheck %s --check-prefix=CHECK
+; RUN: llc -mtriple=armv8 -mcpu=cortex-a57 < %s | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-NOTSWIFT
 
 declare arm_aapcs_vfpcc void @take_vec64(<2 x i32>)
 
 define void @test_vec64() {
-; CHECK-CYCLONE-LABEL: test_vec64:
-; CHECK-SWIFT-LABEL: test_vec64:
+; CHECK-LABEL: test_vec64:
 
   call arm_aapcs_vfpcc void @take_vec64(<2 x i32> <i32 0, i32 0>)
   call arm_aapcs_vfpcc void @take_vec64(<2 x i32> <i32 0, i32 0>)
-; CHECK-CYCLONE-NOT: vmov.f64 d0,
-; CHECK-CYCLONE: vmov.i32 d0, #0
-; CHECK-CYCLONE: bl
-; CHECK-CYCLONE: vmov.i32 d0, #0
-; CHECK-CYCLONE: bl
-
-; CHECK-SWIFT: vmov.f64 [[ZEROREG:d[0-9]+]],
-; CHECK-SWIFT: vmov.i32 [[ZEROREG]], #0
-; CHECK-SWIFT: vorr d0, [[ZEROREG]], [[ZEROREG]]
-; CHECK-SWIFT: bl
-; CHECK-SWIFT: vorr d0, [[ZEROREG]], [[ZEROREG]]
-; CHECK-SWIFT: bl
+; CHECK-NOTSWIFT-NOT: vmov.f64 d0,
+; CHECK: vmov.i32 d0, #0
+; CHECK: bl
+; CHECK: vmov.i32 d0, #0
+; CHECK: bl
 
   ret void
 }
@@ -28,23 +21,15 @@ define void @test_vec64() {
 declare arm_aapcs_vfpcc void @take_vec128(<8 x i16>)
 
 define void @test_vec128() {
-; CHECK-CYCLONE-LABEL: test_vec128:
-; CHECK-SWIFT-LABEL: test_vec128:
+; CHECK-LABEL: test_vec128:
 
   call arm_aapcs_vfpcc void @take_vec128(<8 x i16> <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>)
   call arm_aapcs_vfpcc void @take_vec128(<8 x i16> <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>)
-; CHECK-CYCLONE-NOT: vmov.f64 [[ZEROREG:d[0-9]+]],
-; CHECK-CYCLONE: vmov.i32 q0, #0
-; CHECK-CYCLONE: bl
-; CHECK-CYCLONE: vmov.i32 q0, #0
-; CHECK-CYCLONE: bl
-
-; CHECK-SWIFT-NOT: vmov.f64 [[ZEROREG:d[0-9]+]],
-; CHECK-SWIFT: vmov.i32 [[ZEROREG:q[0-9]+]], #0
-; CHECK-SWIFT: vorr q0, [[ZEROREG]], [[ZEROREG]]
-; CHECK-SWIFT: bl
-; CHECK-SWIFT: vorr q0, [[ZEROREG]], [[ZEROREG]]
-; CHECK-SWIFT: bl
+; CHECK-NOT: vmov.f64 [[ZEROREG:d[0-9]+]],
+; CHECK: vmov.i32 q0, #0
+; CHECK: bl
+; CHECK: vmov.i32 q0, #0
+; CHECK: bl
 
   ret void
 }
@@ -52,16 +37,15 @@ define void @test_vec128() {
 declare void @take_i32(i32)
 
 define void @test_i32() {
-; CHECK-CYCLONE-LABEL: test_i32:
-; CHECK-SWIFT-LABEL: test_i32:
+; CHECK-LABEL: test_i32:
 
   call arm_aapcs_vfpcc void @take_i32(i32 0)
   call arm_aapcs_vfpcc void @take_i32(i32 0)
-; CHECK-CYCLONE-NOT: vmov.f64 [[ZEROREG:d[0-9]+]],
-; CHECK-CYCLONE: mov r0, #0
-; CHECK-CYCLONE: bl
-; CHECK-CYCLONE: mov r0, #0
-; CHECK-CYCLONE: bl
+; CHECK-NOTSWIFT-NOT: vmov.f64 [[ZEROREG:d[0-9]+]],
+; CHECK: mov r0, #0
+; CHECK: bl
+; CHECK: mov r0, #0
+; CHECK: bl
 
 ; It doesn't particularly matter what Swift does here, there isn't carefully
 ; crafted behaviour that we might break in Cyclone.
-- 
2.34.1