[llvm] r319501 - AMDGPU: Use carry-less adds in FI elimination
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Thu Nov 30 15:42:30 PST 2017
Author: arsenm
Date: Thu Nov 30 15:42:30 2017
New Revision: 319501
URL: http://llvm.org/viewvc/llvm-project?rev=319501&view=rev
Log:
AMDGPU: Use carry-less adds in FI elimination
Modified:
llvm/trunk/lib/Target/AMDGPU/SIInstrInfo.cpp
llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.cpp
llvm/trunk/test/CodeGen/AMDGPU/frame-index-elimination.ll
Modified: llvm/trunk/lib/Target/AMDGPU/SIInstrInfo.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/AMDGPU/SIInstrInfo.cpp?rev=319501&r1=319500&r2=319501&view=diff
==============================================================================
--- llvm/trunk/lib/Target/AMDGPU/SIInstrInfo.cpp (original)
+++ llvm/trunk/lib/Target/AMDGPU/SIInstrInfo.cpp Thu Nov 30 15:42:30 2017
@@ -4764,9 +4764,12 @@ SIInstrInfo::getAddNoCarry(MachineBasicB
MachineBasicBlock::iterator I,
const DebugLoc &DL,
unsigned DestReg) const {
- MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
+ if (ST.hasAddNoCarry())
+ return BuildMI(MBB, I, DL, get(AMDGPU::V_ADD_U32_e64), DestReg);
+ MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
unsigned UnusedCarry = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
+ MRI.setRegAllocationHint(UnusedCarry, 0, AMDGPU::VCC);
return BuildMI(MBB, I, DL, get(AMDGPU::V_ADD_I32_e64), DestReg)
.addReg(UnusedCarry, RegState::Define | RegState::Dead);
Modified: llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.cpp?rev=319501&r1=319500&r2=319501&view=diff
==============================================================================
--- llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.cpp (original)
+++ llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.cpp Thu Nov 30 15:42:30 2017
@@ -1071,8 +1071,6 @@ void SIRegisterInfo::eliminateFrameIndex
.addImm(Log2_32(ST.getWavefrontSize()))
.addReg(DiffReg);
} else {
- unsigned CarryOut
- = MRI.createVirtualRegister(&AMDGPU::SReg_64_XEXECRegClass);
unsigned ScaledReg
= MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
@@ -1082,8 +1080,7 @@ void SIRegisterInfo::eliminateFrameIndex
// TODO: Fold if use instruction is another add of a constant.
if (AMDGPU::isInlinableLiteral32(Offset, ST.hasInv2PiInlineImm())) {
- BuildMI(*MBB, MI, DL, TII->get(AMDGPU::V_ADD_I32_e64), ResultReg)
- .addReg(CarryOut, RegState::Define | RegState::Dead)
+ TII->getAddNoCarry(*MBB, MI, DL, ResultReg)
.addImm(Offset)
.addReg(ScaledReg, RegState::Kill);
} else {
@@ -1092,13 +1089,10 @@ void SIRegisterInfo::eliminateFrameIndex
BuildMI(*MBB, MI, DL, TII->get(AMDGPU::S_MOV_B32), ConstOffsetReg)
.addImm(Offset);
- BuildMI(*MBB, MI, DL, TII->get(AMDGPU::V_ADD_I32_e64), ResultReg)
- .addReg(CarryOut, RegState::Define | RegState::Dead)
+ TII->getAddNoCarry(*MBB, MI, DL, ResultReg)
.addReg(ConstOffsetReg, RegState::Kill)
.addReg(ScaledReg, RegState::Kill);
}
-
- MRI.setRegAllocationHint(CarryOut, 0, AMDGPU::VCC);
}
// Don't introduce an extra copy if we're just materializing in a mov.
Modified: llvm/trunk/test/CodeGen/AMDGPU/frame-index-elimination.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/AMDGPU/frame-index-elimination.ll?rev=319501&r1=319500&r2=319501&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/AMDGPU/frame-index-elimination.ll (original)
+++ llvm/trunk/test/CodeGen/AMDGPU/frame-index-elimination.ll Thu Nov 30 15:42:30 2017
@@ -1,4 +1,5 @@
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=kaveri -mattr=-promote-alloca -amdgpu-sroa=0 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=kaveri -mattr=-promote-alloca -amdgpu-sroa=0 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,CI %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -mattr=-promote-alloca -amdgpu-sroa=0 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX9 %s
; Test that non-entry function frame indices are expanded properly to
; give an index relative to the scratch wave offset register
@@ -7,8 +8,13 @@
; GCN-LABEL: {{^}}func_mov_fi_i32:
; GCN: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN: s_sub_u32 s6, s5, s4
-; GCN-NEXT: v_lshr_b32_e64 [[SCALED:v[0-9]+]], s6, 6
-; GCN-NEXT: v_add_i32_e64 v0, s[6:7], 4, [[SCALED]]
+
+; CI-NEXT: v_lshr_b32_e64 [[SCALED:v[0-9]+]], s6, 6
+; CI-NEXT: v_add_i32_e64 v0, s[6:7], 4, [[SCALED]]
+
+; GFX9-NEXT: v_lshrrev_b32_e64 [[SCALED:v[0-9]+]], 6, s6
+; GFX9-NEXT: v_add_u32_e32 v0, 4, [[SCALED]]
+
; GCN-NOT: v_mov
; GCN: ds_write_b32 v0, v0
define void @func_mov_fi_i32() #0 {
@@ -23,9 +29,16 @@ define void @func_mov_fi_i32() #0 {
; GCN-LABEL: {{^}}func_add_constant_to_fi_i32:
; GCN: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN: s_sub_u32 s6, s5, s4
-; GCN-NEXT: v_lshr_b32_e64 [[SCALED:v[0-9]+]], s6, 6
-; GCN-NEXT: v_add_i32_e64 v0, s[6:7], 4, [[SCALED]]
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 4, v0
+
+; CI-NEXT: v_lshr_b32_e64 [[SCALED:v[0-9]+]], s6, 6
+; CI-NEXT: v_add_i32_e64 v0, s[6:7], 4, [[SCALED]]
+; CI-NEXT: v_add_i32_e32 v0, vcc, 4, v0
+
+; GFX9-NEXT: v_lshrrev_b32_e64 [[SCALED:v[0-9]+]], 6, s6
+; GFX9-NEXT: v_add_u32_e32 v0, 4, [[SCALED]]
+; GFX9-NEXT: v_add_u32_e32 v0, 4, v0
+
+
; GCN-NOT: v_mov
; GCN: ds_write_b32 v0, v0
define void @func_add_constant_to_fi_i32() #0 {
@@ -40,8 +53,13 @@ define void @func_add_constant_to_fi_i32
; GCN-LABEL: {{^}}func_other_fi_user_i32:
; GCN: s_sub_u32 s6, s5, s4
-; GCN-NEXT: v_lshr_b32_e64 [[SCALED:v[0-9]+]], s6, 6
-; GCN-NEXT: v_add_i32_e64 v0, s[6:7], 4, [[SCALED]]
+
+; CI-NEXT: v_lshr_b32_e64 [[SCALED:v[0-9]+]], s6, 6
+; CI-NEXT: v_add_i32_e64 v0, s[6:7], 4, [[SCALED]]
+
+; GFX9-NEXT: v_lshrrev_b32_e64 [[SCALED:v[0-9]+]], 6, s6
+; GFX9-NEXT: v_add_u32_e32 v0, 4, [[SCALED]]
+
; GCN-NEXT: v_mul_lo_i32 v0, v0, 9
; GCN-NOT: v_mov
; GCN: ds_write_b32 v0, v0
@@ -73,9 +91,15 @@ define void @func_load_private_arg_i32_p
; GCN: s_waitcnt
; GCN-NEXT: s_mov_b32 s5, s32
; GCN-NEXT: s_sub_u32 [[SUB_OFFSET:s[0-9]+]], s5, s4
-; GCN-NEXT: v_lshr_b32_e64 [[SHIFT:v[0-9]+]], [[SUB_OFFSET]], 6
-; GCN-NEXT: v_add_i32_e64 [[ADD:v[0-9]+]], {{s\[[0-9]+:[0-9]+\]}}, 4, [[SHIFT]]
-; GCN-NEXT: v_add_i32_e32 v0, vcc, 4, [[ADD]]
+
+; CI-NEXT: v_lshr_b32_e64 [[SHIFT:v[0-9]+]], [[SUB_OFFSET]], 6
+; CI-NEXT: v_add_i32_e64 [[ADD:v[0-9]+]], {{s\[[0-9]+:[0-9]+\]}}, 4, [[SHIFT]]
+; CI-NEXT: v_add_i32_e32 v0, vcc, 4, [[ADD]]
+
+; GFX9-NEXT: v_lshrrev_b32_e64 [[SHIFT:v[0-9]+]], 6, [[SUB_OFFSET]]
+; GFX9-NEXT: v_add_u32_e32 [[ADD:v[0-9]+]], 4, [[SHIFT]]
+; GFX9-NEXT: v_add_u32_e32 v0, 4, [[ADD]]
+
; GCN-NOT: v_mov
; GCN: ds_write_b32 v0, v0
define void @void_func_byval_struct_i8_i32_ptr({ i8, i32 }* byval %arg0) #0 {
@@ -106,12 +130,21 @@ define void @void_func_byval_struct_i8_i
; GCN-LABEL: {{^}}void_func_byval_struct_i8_i32_ptr_nonentry_block:
; GCN: s_sub_u32 [[SUB_OFFSET:s[0-9]+]], s5, s4
-; GCN: v_lshr_b32_e64 [[SHIFT:v[0-9]+]], [[SUB_OFFSET]], 6
-; GCN: v_add_i32_e64 [[ADD:v[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 4, [[SHIFT]]
+
+; CI: v_lshr_b32_e64 [[SHIFT:v[0-9]+]], [[SUB_OFFSET]], 6
+; CI: v_add_i32_e64 [[ADD:v[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 4, [[SHIFT]]
+
+; GFX9: v_lshrrev_b32_e64 [[SHIFT:v[0-9]+]], 6, [[SUB_OFFSET]]
+; GFX9: v_add_u32_e32 [[ADD:v[0-9]+]], 4, [[SHIFT]]
+
; GCN: s_and_saveexec_b64
-; GCN: v_add_i32_e32 v0, vcc, 4, [[ADD]]
-; GCN: buffer_load_dword v1, v0, s[0:3], s4 offen{{$}}
+; CI: v_add_i32_e32 v0, vcc, 4, [[ADD]]
+; CI: buffer_load_dword v1, v0, s[0:3], s4 offen{{$}}
+
+; GFX9: v_add_u32_e32 v0, 4, [[ADD]]
+; GFX9: buffer_load_dword v1, v{{[0-9]+}}, s[0:3], s4 offen offset:4{{$}}
+
; GCN: ds_write_b32
define void @void_func_byval_struct_i8_i32_ptr_nonentry_block({ i8, i32 }* byval %arg0, i32 %arg2) #0 {
%cmp = icmp eq i32 %arg2, 0
@@ -131,9 +164,14 @@ ret:
; Added offset can't be used with VOP3 add
; GCN-LABEL: {{^}}func_other_fi_user_non_inline_imm_offset_i32:
; GCN: s_sub_u32 s6, s5, s4
-; GCN-DAG: v_lshr_b32_e64 [[SCALED:v[0-9]+]], s6, 6
; GCN-DAG: s_movk_i32 s6, 0x204
-; GCN: v_add_i32_e64 v0, s[6:7], s6, [[SCALED]]
+
+; CI-DAG: v_lshr_b32_e64 [[SCALED:v[0-9]+]], s6, 6
+; CI: v_add_i32_e64 v0, s[6:7], s6, [[SCALED]]
+
+; GFX9-DAG: v_lshrrev_b32_e64 [[SCALED:v[0-9]+]], 6, s6
+; GFX9: v_add_u32_e32 v0, s6, [[SCALED]]
+
; GCN: v_mul_lo_i32 v0, v0, 9
; GCN: ds_write_b32 v0, v0
define void @func_other_fi_user_non_inline_imm_offset_i32() #0 {
@@ -150,9 +188,14 @@ define void @func_other_fi_user_non_inli
; GCN-LABEL: {{^}}func_other_fi_user_non_inline_imm_offset_i32_vcc_live:
; GCN: s_sub_u32 [[DIFF:s[0-9]+]], s5, s4
-; GCN-DAG: v_lshr_b32_e64 [[SCALED:v[0-9]+]], [[DIFF]], 6
; GCN-DAG: s_movk_i32 [[OFFSET:s[0-9]+]], 0x204
-; GCN: v_add_i32_e64 v0, s{{\[[0-9]+:[0-9]+\]}}, [[OFFSET]], [[SCALED]]
+
+; CI-DAG: v_lshr_b32_e64 [[SCALED:v[0-9]+]], [[DIFF]], 6
+; CI: v_add_i32_e64 v0, s{{\[[0-9]+:[0-9]+\]}}, [[OFFSET]], [[SCALED]]
+
+; GFX9-DAG: v_lshrrev_b32_e64 [[SCALED:v[0-9]+]], 6, [[DIFF]]
+; GFX9: v_add_u32_e32 v0, [[OFFSET]], [[SCALED]]
+
; GCN: v_mul_lo_i32 v0, v0, 9
; GCN: ds_write_b32 v0, v0
define void @func_other_fi_user_non_inline_imm_offset_i32_vcc_live() #0 {
More information about the llvm-commits
mailing list