[llvm] r257545 - [ARM] Mark VMOV with immediate: isAsCheapAsMove.

[Quentin Colombet via llvm-commits]

Author: qcolombet
Date: Tue Jan 12 18:02:40 2016
New Revision: 257545

URL: http://llvm.org/viewvc/llvm-project?rev=257545&view=rev
Log:
[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

Modified:
    llvm/trunk/lib/Target/ARM/ARMInstrNEON.td
    llvm/trunk/test/CodeGen/ARM/zero-cycle-zero.ll

Modified: llvm/trunk/lib/Target/ARM/ARMInstrNEON.td
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/ARM/ARMInstrNEON.td?rev=257545&r1=257544&r2=257545&view=diff
==============================================================================
--- llvm/trunk/lib/Target/ARM/ARMInstrNEON.td (original)
+++ llvm/trunk/lib/Target/ARM/ARMInstrNEON.td Tue Jan 12 18:02:40 2016
@@ -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, $SIMM", "",
@@ -5744,7 +5747,7 @@ def VMOVv4f32 : N1ModImm<1, 0b000, 0b111
                          (ins nImmVMOVF32:$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:

Modified: llvm/trunk/test/CodeGen/ARM/zero-cycle-zero.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/ARM/zero-cycle-zero.ll?rev=257545&r1=257544&r2=257545&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/ARM/zero-cycle-zero.ll (original)
+++ llvm/trunk/test/CodeGen/ARM/zero-cycle-zero.ll Tue Jan 12 18:02:40 2016
@@ -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.