AMDGPU: Add macro fusion schedule DAG mutation

[arsenm]

Summary

Try to increase opportunities to shrink vcc uses.

[javed.absar]

lib/Target/AMDGPU/AMDGPUMacroFusion.cpp

34

This might be shorter and easier to read. Just a suggestion.

if (opcode == AMDGPU::V_ADDC_U32_e64 || ... || ) {

 ... 
return FirstMI->...

}
return false;

[arsenm]

shaderdb results: https://ghostbin.com/paste/sv5b3

[rampitec]

LGTM

[arsenm]

r307313

lib/Target/AMDGPU/AMDGPUMacroFusion.h

+//===- AMDGPUMacroFusion.h - AMDGPU Macro Fusion ----------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/CodeGen/MachineScheduler.h"
+
+namespace llvm {
+
+/// Note that you have to add:
+///   DAG.addMutation(createAMDGPUMacroFusionDAGMutation());
+/// to AMDGPUPassConfig::createMachineScheduler() to have an effect.
+std::unique_ptr<ScheduleDAGMutation> createAMDGPUMacroFusionDAGMutation();
+
+} // llvm

lib/Target/AMDGPU/AMDGPUMacroFusion.cpp

+//===--- AMDGPUMacroFusion.cpp - AMDGPU Macro Fusion ----------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file This file contains the AMDGPU implementation of the DAG scheduling
+///  mutation to pair instructions back to back.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPUMacroFusion.h"
+#include "AMDGPUSubtarget.h"
+#include "SIInstrInfo.h"
+
+#include "llvm/CodeGen/MacroFusion.h"
+
+using namespace llvm;
+
+namespace {
+
+/// \brief Check if the instr pair, FirstMI and SecondMI, should be fused
+/// together. Given SecondMI, when FirstMI is unspecified, then check if
+/// SecondMI may be part of a fused pair at all.
+static bool shouldScheduleAdjacent(const TargetInstrInfo &TII_,
+                                   const TargetSubtargetInfo &TSI,
+                                   const MachineInstr *FirstMI,
+                                   const MachineInstr &SecondMI) {
+  const SIInstrInfo &TII = static_cast<const SIInstrInfo&>(TII_);
+
+  switch (SecondMI.getOpcode()) {
+  case AMDGPU::V_ADDC_U32_e64:
+  case AMDGPU::V_SUBB_U32_e64:
+  case AMDGPU::V_CNDMASK_B32_e64: {
+    // Try to cluster defs of condition registers to their uses. This improves
+    // the chance VCC will be available which will allow shrinking to VOP2
+    // encodings.
+    if (!FirstMI)
+      return true;
+
+    const MachineOperand *Src2 = TII.getNamedOperand(SecondMI,
+                                                     AMDGPU::OpName::src2);
+    return FirstMI->definesRegister(Src2->getReg());
+  }
+  default:
+    return false;
+  }
+
+  return false;
+}
+
+} // end namespace
+
+
+namespace llvm {
+
+std::unique_ptr<ScheduleDAGMutation> createAMDGPUMacroFusionDAGMutation () {
+  return createMacroFusionDAGMutation(shouldScheduleAdjacent);
+}
+
+} // end namespace llvm

lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

 //===----------------------------------------------------------------------===//
 
 #include "AMDGPUTargetMachine.h"
 #include "AMDGPU.h"
 #include "AMDGPUAliasAnalysis.h"
 #include "AMDGPUCallLowering.h"
 #include "AMDGPUInstructionSelector.h"
 #include "AMDGPULegalizerInfo.h"
+#include "AMDGPUMacroFusion.h"
+
 #ifdef LLVM_BUILD_GLOBAL_ISEL
 #include "AMDGPURegisterBankInfo.h"
 #endif
 #include "AMDGPUTargetObjectFile.h"
 #include "AMDGPUTargetTransformInfo.h"
 #include "GCNIterativeScheduler.h"
 #include "GCNSchedStrategy.h"
 #include "R600MachineScheduler.h"

 }
 
 static ScheduleDAGInstrs *
 createGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
   ScheduleDAGMILive *DAG =
     new GCNScheduleDAGMILive(C, make_unique<GCNMaxOccupancySchedStrategy>(C));
   DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
   DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
+  DAG->addMutation(createAMDGPUMacroFusionDAGMutation());
   return DAG;
 }
 
 static ScheduleDAGInstrs *
 createIterativeGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
   auto DAG = new GCNIterativeScheduler(C,
     GCNIterativeScheduler::SCHEDULE_LEGACYMAXOCCUPANCY);
   DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));

lib/Target/AMDGPU/CMakeLists.txt

add_llvm_target(AMDGPUCodeGen
   AMDGPUAnnotateUniformValues.cpp
   AMDGPUAsmPrinter.cpp
   AMDGPUCodeGenPrepare.cpp
   AMDGPUFrameLowering.cpp
   AMDGPUTargetObjectFile.cpp
   AMDGPUIntrinsicInfo.cpp
   AMDGPUISelDAGToDAG.cpp
   AMDGPULowerIntrinsics.cpp
+  AMDGPUMacroFusion.cpp
   AMDGPUMCInstLower.cpp
   AMDGPUMachineCFGStructurizer.cpp
   AMDGPUMachineFunction.cpp
   AMDGPUUnifyMetadata.cpp
   AMDGPUOpenCLImageTypeLoweringPass.cpp
   AMDGPUSubtarget.cpp
   AMDGPUTargetMachine.cpp
   AMDGPUTargetTransformInfo.cpp

test/CodeGen/AMDGPU/addrspacecast.ll

 ; RUN: llc -march=amdgcn -mtriple=amdgcn-amd-amdhsa -mattr=-promote-alloca -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=HSA -check-prefix=CI %s
 ; RUN: llc -march=amdgcn -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -mattr=-promote-alloca -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=HSA -check-prefix=GFX9 %s
 
 ; HSA-LABEL: {{^}}use_group_to_flat_addrspacecast:
 ; HSA: enable_sgpr_private_segment_buffer = 1
 ; HSA: enable_sgpr_dispatch_ptr = 0
 ; CI: enable_sgpr_queue_ptr = 1
 ; GFX9: enable_sgpr_queue_ptr = 0
 
 ; CI-DAG: s_load_dword [[PTR:s[0-9]+]], s[6:7], 0x0{{$}}
 ; CI-DAG: s_load_dword [[APERTURE:s[0-9]+]], s[4:5], 0x10{{$}}
 ; CI-DAG: v_mov_b32_e32 [[VAPERTURE:v[0-9]+]], [[APERTURE]]
+; CI-DAG: v_cmp_ne_u32_e64 vcc, [[PTR]], -1
+; CI-DAG: v_cndmask_b32_e32 v[[HI:[0-9]+]], 0, [[VAPERTURE]], vcc
+; CI-DAG: v_mov_b32_e32 [[VPTR:v[0-9]+]], [[PTR]]
+; CI-DAG: v_cndmask_b32_e32 v[[LO:[0-9]+]], 0, [[VPTR]]
 
+; HSA-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 7
 ; GFX9-DAG: s_load_dword [[PTR:s[0-9]+]], s[4:5], 0x0{{$}}
 ; GFX9-DAG: s_getreg_b32 [[SSRC_SHARED:s[0-9]+]], hwreg(15, 16, 16)
 ; GFX9-DAG: s_lshl_b32 [[SSRC_SHARED_BASE:s[0-9]+]], [[SSRC_SHARED]], 16
 ; GFX9-DAG: v_mov_b32_e32 [[VAPERTURE:v[0-9]+]], [[SSRC_SHARED_BASE]]
 
 ; GFX9-XXX: v_mov_b32_e32 [[VAPERTURE:v[0-9]+]], src_shared_base
-
-; HSA-DAG: v_mov_b32_e32 [[VPTR:v[0-9]+]], [[PTR]]
-
-; HSA-DAG: v_cmp_ne_u32_e64 vcc, [[PTR]], -1
-; HSA-DAG: v_cndmask_b32_e32 v[[HI:[0-9]+]], 0, [[VAPERTURE]]
-; HSA-DAG: v_cndmask_b32_e32 v[[LO:[0-9]+]], 0, [[VPTR]]
-; HSA-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 7
+; GFX9: v_cmp_ne_u32_e64 vcc, [[PTR]], -1
+; GFX9: v_cndmask_b32_e32 v[[HI:[0-9]+]], 0, [[VAPERTURE]], vcc
+; GFX9-DAG: v_mov_b32_e32 [[VPTR:v[0-9]+]], [[PTR]]
+; GFX9-DAG: v_cndmask_b32_e32 v[[LO:[0-9]+]], 0, [[VPTR]]
 
 ; HSA: flat_store_dword v{{\[}}[[LO]]:[[HI]]{{\]}}, [[K]]
 
 ; At most 2 digits. Make sure src_shared_base is not counted as a high
 ; number SGPR.
 
 ; CI: NumSgprs: {{[0-9][0-9]+}}
 ; GFX9: NumSgprs: {{[0-9]+}}
 define amdgpu_kernel void @use_group_to_flat_addrspacecast(i32 addrspace(3)* %ptr) #0 {
   %stof = addrspacecast i32 addrspace(3)* %ptr to i32 addrspace(4)*
   store volatile i32 7, i32 addrspace(4)* %stof
   ret void
 }
 
 ; HSA-LABEL: {{^}}use_private_to_flat_addrspacecast:
 ; HSA: enable_sgpr_private_segment_buffer = 1
 ; HSA: enable_sgpr_dispatch_ptr = 0
 ; CI: enable_sgpr_queue_ptr = 1
 ; GFX9: enable_sgpr_queue_ptr = 0
 
 ; CI-DAG: s_load_dword [[PTR:s[0-9]+]], s[6:7], 0x0{{$}}
 ; CI-DAG: s_load_dword [[APERTURE:s[0-9]+]], s[4:5], 0x11{{$}}
 ; CI-DAG: v_mov_b32_e32 [[VAPERTURE:v[0-9]+]], [[APERTURE]]
 
+; CI-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 7
+; CI-DAG: v_cmp_ne_u32_e64 vcc, [[PTR]], 0
+; CI-DAG: v_cndmask_b32_e32 v[[HI:[0-9]+]], 0, [[VAPERTURE]], vcc
+; CI-DAG: v_mov_b32_e32 [[VPTR:v[0-9]+]], [[PTR]]
+; CI-DAG: v_cndmask_b32_e32 v[[LO:[0-9]+]], 0, [[VPTR]]
+
 ; GFX9-DAG: s_load_dword [[PTR:s[0-9]+]], s[4:5], 0x0{{$}}
 ; GFX9-DAG: s_getreg_b32 [[SSRC_PRIVATE:s[0-9]+]], hwreg(15, 0, 16)
 ; GFX9-DAG: s_lshl_b32 [[SSRC_PRIVATE_BASE:s[0-9]+]], [[SSRC_PRIVATE]], 16
 ; GFX9-DAG: v_mov_b32_e32 [[VAPERTURE:v[0-9]+]], [[SSRC_PRIVATE_BASE]]
 
 ; GFX9-XXX: v_mov_b32_e32 [[VAPERTURE:v[0-9]+]], src_private_base
 
-; HSA-DAG: v_mov_b32_e32 [[VPTR:v[0-9]+]], [[PTR]]
-
-; HSA-DAG: v_cmp_ne_u32_e64 vcc, [[PTR]], 0
-; HSA-DAG: v_cndmask_b32_e32 v[[HI:[0-9]+]], 0, [[VAPERTURE]]
-; HSA-DAG: v_cndmask_b32_e32 v[[LO:[0-9]+]], 0, [[VPTR]]
-; HSA-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 7
+; GFX9-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 7
+; GFX9: v_cmp_ne_u32_e64 vcc, [[PTR]], 0
+; GFX9: v_cndmask_b32_e32 v[[HI:[0-9]+]], 0, [[VAPERTURE]], vcc
+; GFX9: v_mov_b32_e32 [[VPTR:v[0-9]+]], [[PTR]]
+; GFX9-DAG: v_cndmask_b32_e32 v[[LO:[0-9]+]], 0, [[VPTR]]
 
 ; HSA: flat_store_dword v{{\[}}[[LO]]:[[HI]]{{\]}}, [[K]]
 
 ; CI: NumSgprs: {{[0-9][0-9]+}}
 ; GFX9: NumSgprs: {{[0-9]+}}
 define amdgpu_kernel void @use_private_to_flat_addrspacecast(i32* %ptr) #0 {
   %stof = addrspacecast i32* %ptr to i32 addrspace(4)*
   store volatile i32 7, i32 addrspace(4)* %stof

test/CodeGen/AMDGPU/ctlz.ll

-; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI -check-prefix=FUNC %s
-; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI -check-prefix=FUNC %s
-; RUN: llc -march=r600 -mcpu=cypress -verify-machineinstrs < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=SI -check-prefix=FUNC %s
+; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=VI -check-prefix=FUNC %s
+; RUN: llc -march=r600 -mcpu=cypress -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=EG -check-prefix=FUNC %s
 
 declare i7 @llvm.ctlz.i7(i7, i1) nounwind readnone
 declare i8 @llvm.ctlz.i8(i8, i1) nounwind readnone
 declare i16 @llvm.ctlz.i16(i16, i1) nounwind readnone
 
 declare i32 @llvm.ctlz.i32(i32, i1) nounwind readnone
 declare <2 x i32> @llvm.ctlz.v2i32(<2 x i32>, i1) nounwind readnone
 declare <4 x i32> @llvm.ctlz.v4i32(<4 x i32>, i1) nounwind readnone

 define amdgpu_kernel void @s_ctlz_i32(i32 addrspace(1)* noalias %out, i32 %val) nounwind {
   %ctlz = call i32 @llvm.ctlz.i32(i32 %val, i1 false) nounwind readnone
   store i32 %ctlz, i32 addrspace(1)* %out, align 4
   ret void
 }
 
 ; FUNC-LABEL: {{^}}v_ctlz_i32:
 ; GCN: {{buffer|flat}}_load_dword [[VAL:v[0-9]+]],
-; GCN-DAG: v_ffbh_u32_e32 [[CTLZ:v[0-9]+]], [[VAL]]
-; GCN-DAG: v_cmp_ne_u32_e32 vcc, 0, [[CTLZ]]
+; GCN: v_ffbh_u32_e32 [[CTLZ:v[0-9]+]], [[VAL]]
+; GCN: v_cmp_ne_u32_e32 vcc, 0, [[VAL]]
 ; GCN: v_cndmask_b32_e32 [[RESULT:v[0-9]+]], 32, [[CTLZ]], vcc
 ; GCN: buffer_store_dword [[RESULT]],
 ; GCN: s_endpgm
 
 ; EG: FFBH_UINT
 ; EG: CNDE_INT
 define amdgpu_kernel void @v_ctlz_i32(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %valptr) nounwind {
   %tid = call i32 @llvm.r600.read.tidig.x()
define amdgpu_kernel void @v_ctlz_v4i32(<4 x i32> addrspace(1)* noalias %out, <4 x i32> addrspace(1)* noalias %valptr) nounwind {
   %val = load <4 x i32>, <4 x i32> addrspace(1)* %in.gep, align 16
   %ctlz = call <4 x i32> @llvm.ctlz.v4i32(<4 x i32> %val, i1 false) nounwind readnone
   store <4 x i32> %ctlz, <4 x i32> addrspace(1)* %out, align 16
   ret void
 }
 
 ; FUNC-LABEL: {{^}}v_ctlz_i8:
 ; GCN: {{buffer|flat}}_load_ubyte [[VAL:v[0-9]+]],
-; SI-DAG: v_ffbh_u32_e32 [[RESULT:v[0-9]+]], [[VAL]]
-; VI-DAG: v_ffbh_u32_sdwa [[RESULT:v[0-9]+]], [[VAL]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
+; SI-DAG: v_ffbh_u32_e32 [[FFBH:v[0-9]+]], [[VAL]]
+; VI-DAG: v_ffbh_u32_sdwa [[FFBH:v[0-9]+]], [[VAL]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
+; SI: v_cmp_ne_u32_e32 vcc, 0, [[VAL]]
+; VI: v_cmp_ne_u16_e32 vcc, 0, [[VAL]]
+
+; GCN: v_cndmask_b32_e32 [[SELECT:v[0-9]+]], 32, [[FFBH]], vcc
+
+; SI: v_subrev_i32_e32 [[RESULT:v[0-9]+]], vcc, 24, [[SELECT]]
+; VI: v_add_i32_e32 [[RESULT:v[0-9]+]], vcc, -16, [[SELECT]]
 ; GCN: buffer_store_byte [[RESULT]],
 ; GCN: s_endpgm
 define amdgpu_kernel void @v_ctlz_i8(i8 addrspace(1)* noalias %out, i8 addrspace(1)* noalias %valptr) nounwind {
   %val = load i8, i8 addrspace(1)* %valptr
   %ctlz = call i8 @llvm.ctlz.i8(i8 %val, i1 false) nounwind readnone
   store i8 %ctlz, i8 addrspace(1)* %out
   ret void
 }
define amdgpu_kernel void @s_ctlz_i64_trunc(i32 addrspace(1)* noalias %out, i64 %val) nounwind {
   %ctlz = call i64 @llvm.ctlz.i64(i64 %val, i1 false)
   %trunc = trunc i64 %ctlz to i32
   store i32 %trunc, i32 addrspace(1)* %out
   ret void
 }
 
 ; FUNC-LABEL: {{^}}v_ctlz_i64:
 ; GCN-DAG: {{buffer|flat}}_load_dwordx2 v{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}}
-; GCN-DAG: v_cmp_eq_u32_e64 [[CMPHI:s\[[0-9]+:[0-9]+\]]], 0, v[[HI]]
+; GCN-DAG: v_cmp_eq_u32_e32 vcc, 0, v[[HI]]
 ; GCN-DAG: v_ffbh_u32_e32 [[FFBH_LO:v[0-9]+]], v[[LO]]
 ; GCN-DAG: v_add_i32_e32 [[ADD:v[0-9]+]], vcc, 32, [[FFBH_LO]]
 ; GCN-DAG: v_ffbh_u32_e32 [[FFBH_HI:v[0-9]+]], v[[HI]]
-; GCN-DAG: v_cndmask_b32_e64 v[[CTLZ:[0-9]+]], [[FFBH_HI]], [[ADD]], [[CMPHI]]
+; GCN-DAG: v_cndmask_b32_e32 v[[CTLZ:[0-9]+]], [[FFBH_HI]], [[ADD]], vcc
 ; GCN-DAG: v_or_b32_e32 [[OR:v[0-9]+]], v[[HI]], v[[LO]]
 ; GCN-DAG: v_cmp_ne_u32_e32 vcc, 0, [[OR]]
 ; GCN-DAG: v_cndmask_b32_e32 v[[CLTZ_LO:[0-9]+]], 64, v[[CTLZ:[0-9]+]], vcc
 ; GCN: {{buffer|flat}}_store_dwordx2 {{.*}}v{{\[}}[[CLTZ_LO]]:[[CTLZ_HI:[0-9]+]]{{\]}}
 define amdgpu_kernel void @v_ctlz_i64(i64 addrspace(1)* noalias %out, i64 addrspace(1)* noalias %in) nounwind {
   %tid = call i32 @llvm.r600.read.tidig.x()
   %in.gep = getelementptr i64, i64 addrspace(1)* %in, i32 %tid
   %out.gep = getelementptr i64, i64 addrspace(1)* %out, i32 %tid

test/CodeGen/AMDGPU/ctlz_zero_undef.ll

define amdgpu_kernel void @s_ctlz_zero_undef_i64_trunc(i32 addrspace(1)* noalias %out, i64 %val) nounwind {
   %ctlz = call i64 @llvm.ctlz.i64(i64 %val, i1 true)
   %trunc = trunc i64 %ctlz to i32
   store i32 %trunc, i32 addrspace(1)* %out
   ret void
 }
 
 ; FUNC-LABEL: {{^}}v_ctlz_zero_undef_i64:
 ; GCN-DAG: {{buffer|flat}}_load_dwordx2 v{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}}
-; GCN-DAG: v_cmp_eq_u32_e64 [[CMPHI:s\[[0-9]+:[0-9]+\]]], 0, v[[HI]]
+; GCN-DAG: v_cmp_eq_u32_e32 vcc, 0, v[[HI]]
 ; GCN-DAG: v_ffbh_u32_e32 [[FFBH_LO:v[0-9]+]], v[[LO]]
 ; GCN-DAG: v_add_i32_e32 [[ADD:v[0-9]+]], vcc, 32, [[FFBH_LO]]
 ; GCN-DAG: v_ffbh_u32_e32 [[FFBH_HI:v[0-9]+]], v[[HI]]
-; GCN-DAG: v_cndmask_b32_e64 v[[CTLZ:[0-9]+]], [[FFBH_HI]], [[FFBH_LO]]
+; GCN-DAG: v_cndmask_b32_e32 v[[CTLZ:[0-9]+]], [[FFBH_HI]], [[FFBH_LO]]
 ; GCN: {{buffer|flat}}_store_dwordx2 {{.*}}v{{\[}}[[CTLZ]]:[[CTLZ_HI:[0-9]+]]{{\]}}
 define amdgpu_kernel void @v_ctlz_zero_undef_i64(i64 addrspace(1)* noalias %out, i64 addrspace(1)* noalias %in) nounwind {
   %tid = call i32 @llvm.r600.read.tidig.x()
   %in.gep = getelementptr i64, i64 addrspace(1)* %in, i32 %tid
   %out.gep = getelementptr i64, i64 addrspace(1)* %out, i32 %tid
   %val = load i64, i64 addrspace(1)* %in.gep
   %ctlz = call i64 @llvm.ctlz.i64(i64 %val, i1 true)
   store i64 %ctlz, i64 addrspace(1)* %out.gep

test/CodeGen/AMDGPU/fcmp.f16.ll

-; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
-; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=fiji -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
+; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=SI %s
+; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=fiji -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=VI %s
 
 ; GCN-LABEL: {{^}}fcmp_f16_lt
 ; GCN: buffer_load_ushort v[[A_F16:[0-9]+]]
 ; GCN: buffer_load_ushort v[[B_F16:[0-9]+]]
 ; SI:  v_cvt_f32_f16_e32 v[[A_F32:[0-9]+]], v[[A_F16]]
 ; SI:  v_cvt_f32_f16_e32 v[[B_F32:[0-9]+]], v[[B_F16]]
 ; SI:  v_cmp_lt_f32_e32 vcc, v[[A_F32]], v[[B_F32]]
 ; VI:  v_cmp_lt_f16_e32 vcc, v[[A_F16]], v[[B_F16]]
entry:
   %a.val = load half, half addrspace(1)* %a
   %b.val = load half, half addrspace(1)* %b
   %r.val = fcmp uge half %a.val, %b.val
   %r.val.sext = sext i1 %r.val to i32
   store i32 %r.val.sext, i32 addrspace(1)* %r
   ret void
 }
 
-; GCN-LABEL: {{^}}fcmp_v2f16_lt
-; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]]
-; GCN: buffer_load_dword v[[B_V2_F16:[0-9]+]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
-; SI:  v_cmp_lt_f32_e32 vcc, v[[A_F32_0]], v[[B_F32_0]]
-; SI:  v_cmp_lt_f32_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F32_1]], v[[B_F32_1]]
-; VI:  v_cmp_lt_f16_e32 vcc, v[[A_V2_F16]], v[[B_V2_F16]]
-; VI:  v_cmp_lt_f16_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F16_1]], v[[B_F16_1]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_0:[0-9]+]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_1:[0-9]+]]
-; GCN: buffer_store_dwordx2 v{{\[}}[[R_I32_0]]:[[R_I32_1]]{{\]}}
-; GCN: s_endpgm
+; GCN-LABEL: {{^}}fcmp_v2f16_lt:
+; SI: v_cmp_lt_f32_e32 vcc,
+; SI: v_cmp_lt_f32_e32 vcc,
+
+; VI: v_cmp_lt_f16_e32 vcc,
+; VI: v_cmp_lt_f16_e32 vcc,
 define amdgpu_kernel void @fcmp_v2f16_lt(
     <2 x i32> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a,
     <2 x half> addrspace(1)* %b) {
 entry:
   %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
   %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
   %r.val = fcmp olt <2 x half> %a.val, %b.val
   %r.val.sext = sext <2 x i1> %r.val to <2 x i32>
   store <2 x i32> %r.val.sext, <2 x i32> addrspace(1)* %r
   ret void
 }
 
 ; GCN-LABEL: {{^}}fcmp_v2f16_eq
-; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]]
-; GCN: buffer_load_dword v[[B_V2_F16:[0-9]+]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
-; SI:  v_cmp_eq_f32_e32 vcc, v[[A_F32_0]], v[[B_F32_0]]
-; SI:  v_cmp_eq_f32_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F32_1]], v[[B_F32_1]]
-; VI:  v_cmp_eq_f16_e32 vcc, v[[A_V2_F16]], v[[B_V2_F16]]
-; VI:  v_cmp_eq_f16_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F16_1]], v[[B_F16_1]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_0:[0-9]+]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_1:[0-9]+]]
-; GCN: buffer_store_dwordx2 v{{\[}}[[R_I32_0]]:[[R_I32_1]]{{\]}}
-; GCN: s_endpgm
+; SI:  v_cmp_eq_f32_e32 vcc,
+; SI:  v_cmp_eq_f32_e32 vcc,
+
+; VI:  v_cmp_eq_f16_e32 vcc,
+; VI:  v_cmp_eq_f16_e32 vcc,
 define amdgpu_kernel void @fcmp_v2f16_eq(
     <2 x i32> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a,
     <2 x half> addrspace(1)* %b) {
 entry:
   %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
   %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
   %r.val = fcmp oeq <2 x half> %a.val, %b.val
   %r.val.sext = sext <2 x i1> %r.val to <2 x i32>
   store <2 x i32> %r.val.sext, <2 x i32> addrspace(1)* %r
   ret void
 }
 
-; GCN-LABEL: {{^}}fcmp_v2f16_le
-; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]]
-; GCN: buffer_load_dword v[[B_V2_F16:[0-9]+]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
-; SI:  v_cmp_le_f32_e32 vcc, v[[A_F32_0]], v[[B_F32_0]]
-; SI:  v_cmp_le_f32_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F32_1]], v[[B_F32_1]]
-; VI:  v_cmp_le_f16_e32 vcc, v[[A_V2_F16]], v[[B_V2_F16]]
-; VI:  v_cmp_le_f16_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F16_1]], v[[B_F16_1]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_0:[0-9]+]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_1:[0-9]+]]
-; GCN: buffer_store_dwordx2 v{{\[}}[[R_I32_0]]:[[R_I32_1]]{{\]}}
-; GCN: s_endpgm
+; GCN-LABEL: {{^}}fcmp_v2f16_le:
+; SI:  v_cmp_le_f32_e32 vcc
+; SI:  v_cmp_le_f32_e32 vcc
+; VI:  v_cmp_le_f16_e32 vcc
+; VI:  v_cmp_le_f16_e32 vcc
 define amdgpu_kernel void @fcmp_v2f16_le(
     <2 x i32> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a,
     <2 x half> addrspace(1)* %b) {
 entry:
   %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
   %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
   %r.val = fcmp ole <2 x half> %a.val, %b.val
   %r.val.sext = sext <2 x i1> %r.val to <2 x i32>
   store <2 x i32> %r.val.sext, <2 x i32> addrspace(1)* %r
   ret void
 }
 
-; GCN-LABEL: {{^}}fcmp_v2f16_gt
-; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]]
-; GCN: buffer_load_dword v[[B_V2_F16:[0-9]+]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
-; SI:  v_cmp_gt_f32_e32 vcc, v[[A_F32_0]], v[[B_F32_0]]
-; SI:  v_cmp_gt_f32_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F32_1]], v[[B_F32_1]]
-; VI:  v_cmp_gt_f16_e32 vcc, v[[A_V2_F16]], v[[B_V2_F16]]
-; VI:  v_cmp_gt_f16_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F16_1]], v[[B_F16_1]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_0:[0-9]+]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_1:[0-9]+]]
-; GCN: buffer_store_dwordx2 v{{\[}}[[R_I32_0]]:[[R_I32_1]]{{\]}}
-; GCN: s_endpgm
+; GCN-LABEL: {{^}}fcmp_v2f16_gt:
+; SI: v_cmp_gt_f32_e32 vcc,
+; SI: v_cmp_gt_f32_e32 vcc,
+
+; VI: v_cmp_gt_f16_e32 vcc,
+; VI: v_cmp_gt_f16_e32 vcc,
 define amdgpu_kernel void @fcmp_v2f16_gt(
     <2 x i32> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a,
     <2 x half> addrspace(1)* %b) {
 entry:
   %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
   %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
   %r.val = fcmp ogt <2 x half> %a.val, %b.val
   %r.val.sext = sext <2 x i1> %r.val to <2 x i32>
   store <2 x i32> %r.val.sext, <2 x i32> addrspace(1)* %r
   ret void
 }
 
-; GCN-LABEL: {{^}}fcmp_v2f16_lg
-; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]]
-; GCN: buffer_load_dword v[[B_V2_F16:[0-9]+]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
-; SI:  v_cmp_lg_f32_e32 vcc, v[[A_F32_0]], v[[B_F32_0]]
-; SI:  v_cmp_lg_f32_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F32_1]], v[[B_F32_1]]
-; VI:  v_cmp_lg_f16_e32 vcc, v[[A_V2_F16]], v[[B_V2_F16]]
-; VI:  v_cmp_lg_f16_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F16_1]], v[[B_F16_1]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_0:[0-9]+]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_1:[0-9]+]]
-; GCN: buffer_store_dwordx2 v{{\[}}[[R_I32_0]]:[[R_I32_1]]{{\]}}
-; GCN: s_endpgm
+; GCN-LABEL: {{^}}fcmp_v2f16_lg:
+; SI: v_cmp_lg_f32_e32 vcc,
+; SI: v_cmp_lg_f32_e32 vcc,
+
+; VI: v_cmp_lg_f16_e32 vcc,
+; VI: v_cmp_lg_f16_e32 vcc,
 define amdgpu_kernel void @fcmp_v2f16_lg(
     <2 x i32> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a,
     <2 x half> addrspace(1)* %b) {
 entry:
   %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
   %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
   %r.val = fcmp one <2 x half> %a.val, %b.val
   %r.val.sext = sext <2 x i1> %r.val to <2 x i32>
   store <2 x i32> %r.val.sext, <2 x i32> addrspace(1)* %r
   ret void
 }
 
-; GCN-LABEL: {{^}}fcmp_v2f16_ge
-; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]]
-; GCN: buffer_load_dword v[[B_V2_F16:[0-9]+]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
-; SI:  v_cmp_ge_f32_e32 vcc, v[[A_F32_0]], v[[B_F32_0]]
-; SI:  v_cmp_ge_f32_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F32_1]], v[[B_F32_1]]
-; VI:  v_cmp_ge_f16_e32 vcc, v[[A_V2_F16]], v[[B_V2_F16]]
-; VI:  v_cmp_ge_f16_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F16_1]], v[[B_F16_1]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_0:[0-9]+]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_1:[0-9]+]]
-; GCN: buffer_store_dwordx2 v{{\[}}[[R_I32_0]]:[[R_I32_1]]{{\]}}
-; GCN: s_endpgm
+; GCN-LABEL: {{^}}fcmp_v2f16_ge:
+; SI:  v_cmp_ge_f32_e32 vcc,
+; SI:  v_cmp_ge_f32_e32 vcc,
+
+; VI:  v_cmp_ge_f16_e32 vcc,
+; VI:  v_cmp_ge_f16_e32 vcc,
 define amdgpu_kernel void @fcmp_v2f16_ge(
     <2 x i32> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a,
     <2 x half> addrspace(1)* %b) {
 entry:
   %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
   %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
   %r.val = fcmp oge <2 x half> %a.val, %b.val
   %r.val.sext = sext <2 x i1> %r.val to <2 x i32>
   store <2 x i32> %r.val.sext, <2 x i32> addrspace(1)* %r
   ret void
 }
 
-; GCN-LABEL: {{^}}fcmp_v2f16_o
-; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]]
-; GCN: buffer_load_dword v[[B_V2_F16:[0-9]+]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
-; SI:  v_cmp_o_f32_e32 vcc, v[[A_F32_0]], v[[B_F32_0]]
-; SI:  v_cmp_o_f32_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F32_1]], v[[B_F32_1]]
-; VI:  v_cmp_o_f16_e32 vcc, v[[A_V2_F16]], v[[B_V2_F16]]
-; VI:  v_cmp_o_f16_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F16_1]], v[[B_F16_1]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_0:[0-9]+]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_1:[0-9]+]]
-; GCN: buffer_store_dwordx2 v{{\[}}[[R_I32_0]]:[[R_I32_1]]{{\]}}
-; GCN: s_endpgm
+; GCN-LABEL: {{^}}fcmp_v2f16_o:
+; SI:  v_cmp_o_f32_e32 vcc,
+; SI:  v_cmp_o_f32_e32 vcc,
+
+; VI:  v_cmp_o_f16_e32 vcc,
+; VI:  v_cmp_o_f16_e32 vcc,
 define amdgpu_kernel void @fcmp_v2f16_o(
     <2 x i32> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a,
     <2 x half> addrspace(1)* %b) {
 entry:
   %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
   %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
   %r.val = fcmp ord <2 x half> %a.val, %b.val
   %r.val.sext = sext <2 x i1> %r.val to <2 x i32>
   store <2 x i32> %r.val.sext, <2 x i32> addrspace(1)* %r
   ret void
 }
 
-; GCN-LABEL: {{^}}fcmp_v2f16_u
-; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]]
-; GCN: buffer_load_dword v[[B_V2_F16:[0-9]+]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
-; SI:  v_cmp_u_f32_e32 vcc, v[[A_F32_0]], v[[B_F32_0]]
-; SI:  v_cmp_u_f32_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F32_1]], v[[B_F32_1]]
-; VI:  v_cmp_u_f16_e32 vcc, v[[A_V2_F16]], v[[B_V2_F16]]
-; VI:  v_cmp_u_f16_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F16_1]], v[[B_F16_1]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_0:[0-9]+]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_1:[0-9]+]]
-; GCN: buffer_store_dwordx2 v{{\[}}[[R_I32_0]]:[[R_I32_1]]{{\]}}
-; GCN: s_endpgm
+; GCN-LABEL: {{^}}fcmp_v2f16_u:
+; SI:  v_cmp_u_f32_e32 vcc,
+; SI:  v_cmp_u_f32_e32 vcc,
+
+; VI:  v_cmp_u_f16_e32 vcc,
+; VI:  v_cmp_u_f16_e32 vcc,
 define amdgpu_kernel void @fcmp_v2f16_u(
     <2 x i32> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a,
     <2 x half> addrspace(1)* %b) {
 entry:
   %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
   %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
   %r.val = fcmp uno <2 x half> %a.val, %b.val
   %r.val.sext = sext <2 x i1> %r.val to <2 x i32>
   store <2 x i32> %r.val.sext, <2 x i32> addrspace(1)* %r
   ret void
 }
 
 ; GCN-LABEL: {{^}}fcmp_v2f16_nge
-; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]]
-; GCN: buffer_load_dword v[[B_V2_F16:[0-9]+]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
-; SI:  v_cmp_nge_f32_e32 vcc, v[[A_F32_0]], v[[B_F32_0]]
-; SI:  v_cmp_nge_f32_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F32_1]], v[[B_F32_1]]
-; VI:  v_cmp_nge_f16_e32 vcc, v[[A_V2_F16]], v[[B_V2_F16]]
-; VI:  v_cmp_nge_f16_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F16_1]], v[[B_F16_1]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_0:[0-9]+]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_1:[0-9]+]]
-; GCN: buffer_store_dwordx2 v{{\[}}[[R_I32_0]]:[[R_I32_1]]{{\]}}
-; GCN: s_endpgm
+; SI:  v_cmp_nge_f32_e32 vcc,
+; SI:  v_cmp_nge_f32_e32 vcc,
+
+; VI:  v_cmp_nge_f16_e32 vcc,
+; VI:  v_cmp_nge_f16_e32 vcc,
 define amdgpu_kernel void @fcmp_v2f16_nge(
     <2 x i32> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a,
     <2 x half> addrspace(1)* %b) {
 entry:
   %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
   %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
   %r.val = fcmp ult <2 x half> %a.val, %b.val
   %r.val.sext = sext <2 x i1> %r.val to <2 x i32>
   store <2 x i32> %r.val.sext, <2 x i32> addrspace(1)* %r
   ret void
 }
 
 ; GCN-LABEL: {{^}}fcmp_v2f16_nlg
-; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]]
-; GCN: buffer_load_dword v[[B_V2_F16:[0-9]+]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
-; SI:  v_cmp_nlg_f32_e32 vcc, v[[A_F32_0]], v[[B_F32_0]]
-; SI:  v_cmp_nlg_f32_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F32_1]], v[[B_F32_1]]
-; VI:  v_cmp_nlg_f16_e32 vcc, v[[A_V2_F16]], v[[B_V2_F16]]
-; VI:  v_cmp_nlg_f16_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F16_1]], v[[B_F16_1]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_0:[0-9]+]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_1:[0-9]+]]
-; GCN: buffer_store_dwordx2 v{{\[}}[[R_I32_0]]:[[R_I32_1]]{{\]}}
-; GCN: s_endpgm
+; SI:  v_cmp_nlg_f32_e32 vcc
+; SI:  v_cmp_nlg_f32_e32 vcc
+
+; VI:  v_cmp_nlg_f16_e32 vcc
+; VI:  v_cmp_nlg_f16_e32 vcc
 define amdgpu_kernel void @fcmp_v2f16_nlg(
     <2 x i32> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a,
     <2 x half> addrspace(1)* %b) {
 entry:
   %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
   %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
   %r.val = fcmp ueq <2 x half> %a.val, %b.val
   %r.val.sext = sext <2 x i1> %r.val to <2 x i32>
   store <2 x i32> %r.val.sext, <2 x i32> addrspace(1)* %r
   ret void
 }
 
 ; GCN-LABEL: {{^}}fcmp_v2f16_ngt
-; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]]
-; GCN: buffer_load_dword v[[B_V2_F16:[0-9]+]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
-; SI:  v_cmp_ngt_f32_e32 vcc, v[[A_F32_0]], v[[B_F32_0]]
-; SI:  v_cmp_ngt_f32_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F32_1]], v[[B_F32_1]]
-; VI:  v_cmp_ngt_f16_e32 vcc, v[[A_V2_F16]], v[[B_V2_F16]]
-; VI:  v_cmp_ngt_f16_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F16_1]], v[[B_F16_1]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_0:[0-9]+]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_1:[0-9]+]]
-; GCN: buffer_store_dwordx2 v{{\[}}[[R_I32_0]]:[[R_I32_1]]{{\]}}
-; GCN: s_endpgm
+; SI:  v_cmp_ngt_f32_e32 vcc,
+; SI:  v_cmp_ngt_f32_e32 vcc,
+
+; VI:  v_cmp_ngt_f16_e32 vcc,
+; VI:  v_cmp_ngt_f16_e32 vcc,
 define amdgpu_kernel void @fcmp_v2f16_ngt(
     <2 x i32> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a,
     <2 x half> addrspace(1)* %b) {
 entry:
   %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
   %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
   %r.val = fcmp ule <2 x half> %a.val, %b.val
   %r.val.sext = sext <2 x i1> %r.val to <2 x i32>
   store <2 x i32> %r.val.sext, <2 x i32> addrspace(1)* %r
   ret void
 }
 
 ; GCN-LABEL: {{^}}fcmp_v2f16_nle
-; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]]
-; GCN: buffer_load_dword v[[B_V2_F16:[0-9]+]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
-; SI:  v_cmp_nle_f32_e32 vcc, v[[A_F32_0]], v[[B_F32_0]]
-; SI:  v_cmp_nle_f32_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F32_1]], v[[B_F32_1]]
-; VI:  v_cmp_nle_f16_e32 vcc, v[[A_V2_F16]], v[[B_V2_F16]]
-; VI:  v_cmp_nle_f16_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F16_1]], v[[B_F16_1]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_0:[0-9]+]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_1:[0-9]+]]
-; GCN: buffer_store_dwordx2 v{{\[}}[[R_I32_0]]:[[R_I32_1]]{{\]}}
-; GCN: s_endpgm
+; SI: v_cmp_nle_f32_e32 vcc, v{{[0-9]+}}, v{{[0-9]+}}
+; SI: v_cmp_nle_f32_e32 vcc, v{{[0-9]+}}, v{{[0-9]+}}
+
+; VI: v_cmp_nle_f16_e32 vcc, v{{[0-9]+}}, v{{[0-9]+}}
+; VI: v_cmp_nle_f16_e32 vcc, v{{[0-9]+}}, v{{[0-9]+}}
 define amdgpu_kernel void @fcmp_v2f16_nle(
     <2 x i32> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a,
     <2 x half> addrspace(1)* %b) {
 entry:
   %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
   %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
   %r.val = fcmp ugt <2 x half> %a.val, %b.val
   %r.val.sext = sext <2 x i1> %r.val to <2 x i32>
   store <2 x i32> %r.val.sext, <2 x i32> addrspace(1)* %r
   ret void
 }
 
 ; GCN-LABEL: {{^}}fcmp_v2f16_neq
-; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]]
-; GCN: buffer_load_dword v[[B_V2_F16:[0-9]+]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
-; SI:  v_cmp_neq_f32_e32 vcc, v[[A_F32_0]], v[[B_F32_0]]
-; SI:  v_cmp_neq_f32_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F32_1]], v[[B_F32_1]]
-; VI:  v_cmp_neq_f16_e32 vcc, v[[A_V2_F16]], v[[B_V2_F16]]
-; VI:  v_cmp_neq_f16_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F16_1]], v[[B_F16_1]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_0:[0-9]+]]
-; GCN: v_cndmask_b32_e64 v[[R_I32_1:[0-9]+]]
-; GCN: buffer_store_dwordx2 v{{\[}}[[R_I32_0]]:[[R_I32_1]]{{\]}}
-; GCN: s_endpgm
+; SI:  v_cmp_neq_f32_e32 vcc, v{{[0-9]+}}, v{{[0-9]+}}
+; SI:  v_cmp_neq_f32_e32 vcc, v{{[0-9]+}}, v{{[0-9]+}}
+
+; VI:  v_cmp_neq_f16_e32 vcc, v{{[0-9]+}}, v{{[0-9]+}}
+; VI:  v_cmp_neq_f16_e32 vcc, v{{[0-9]+}}, v{{[0-9]+}}
 define amdgpu_kernel void @fcmp_v2f16_neq(
     <2 x i32> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a,
     <2 x half> addrspace(1)* %b) {
 entry:
   %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
   %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
   %r.val = fcmp une <2 x half> %a.val, %b.val
   %r.val.sext = sext <2 x i1> %r.val to <2 x i32>
   store <2 x i32> %r.val.sext, <2 x i32> addrspace(1)* %r
   ret void
 }
 
 ; GCN-LABEL: {{^}}fcmp_v2f16_nlt
 ; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]]
 ; GCN: buffer_load_dword v[[B_V2_F16:[0-9]+]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
-; GCN: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
-; SI:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
-; SI:  v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
-; SI:  v_cmp_nlt_f32_e32 vcc, v[[A_F32_0]], v[[B_F32_0]]
-; SI:  v_cmp_nlt_f32_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F32_1]], v[[B_F32_1]]
-; VI:  v_cmp_nlt_f16_e32 vcc, v[[A_V2_F16]], v[[B_V2_F16]]
-; VI:  v_cmp_nlt_f16_e64 s[{{[0-9]+}}:{{[0-9]+}}], v[[A_F16_1]], v[[B_F16_1]]
+; SI-DAG:  v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
+; GCN-DAG: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
+; SI-DAG:  v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
+; SI-DAG:  v_cmp_nlt_f32_e32 vcc, v[[A_F32_0]], v[[B_F32_0]]
+
+; GCN-DAG: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
+; SI-DAG:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
+; SI-DAG:  v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
+; SI-DAG:  v_cmp_nlt_f32_e32 vcc, v[[A_F32_1]], v[[B_F32_1]]
+; VI-DAG:  v_cmp_nlt_f16_e32 vcc, v[[A_V2_F16]], v[[B_V2_F16]]
 ; GCN: v_cndmask_b32_e64 v[[R_I32_0:[0-9]+]]
+
+; VI:  v_cmp_nlt_f16_e32 vcc, v[[A_F16_1]], v[[B_F16_1]]
 ; GCN: v_cndmask_b32_e64 v[[R_I32_1:[0-9]+]]
 ; GCN: buffer_store_dwordx2 v{{\[}}[[R_I32_0]]:[[R_I32_1]]{{\]}}
 ; GCN: s_endpgm
 define amdgpu_kernel void @fcmp_v2f16_nlt(
     <2 x i32> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a,
     <2 x half> addrspace(1)* %b) {
 entry:

test/CodeGen/AMDGPU/inline-asm.ll

entry:
   %val = load i1, i1 addrspace(1)* undef
   call void asm sideeffect "; use $0 ", "{v0}"(i1 %val)
   ret void
 }
 
 ; FIXME: Should be scheduled to shrink vcc
 ; CHECK-LABEL: {{^}}i1_input_phys_vgpr_x2:
 ; CHECK: v_cmp_eq_u32_e32 vcc, 1, v0
-; CHECK: v_cmp_eq_u32_e64 s[0:1], 1, v1
 ; CHECK: v_cndmask_b32_e64 v0, 0, -1, vcc
-; CHECK: v_cndmask_b32_e64 v1, 0, -1, s[0:1]
+; CHECK: v_cmp_eq_u32_e32 vcc, 1, v1
+; CHECK: v_cndmask_b32_e64 v1, 0, -1, vcc
 define amdgpu_kernel void @i1_input_phys_vgpr_x2() {
 entry:
   %val0 = load volatile i1, i1 addrspace(1)* undef
   %val1 = load volatile i1, i1 addrspace(1)* undef
   call void asm sideeffect "; use $0 $1 ", "{v0}, {v1}"(i1 %val0, i1 %val1)
   ret void
 }
 

test/CodeGen/AMDGPU/macro-fusion-cluster-vcc-uses.mir

+# RUN: llc -march=amdgcn -verify-machineinstrs -run-pass machine-scheduler -o - %s | FileCheck -check-prefix=GCN %s
+
+# GCN-LABEL: name: cluster_add_addc
+# GCN: S_NOP 0, implicit-def %vcc
+# GCN: dead %2, %3 = V_ADD_I32_e64 %0, %1, implicit %exec
+# GCN: dead %4, dead %5 = V_ADDC_U32_e64 %6, %7, %3, implicit %exec
+name: cluster_add_addc
+registers:
+  - { id: 0, class: vgpr_32 }
+  - { id: 1, class: vgpr_32 }
+  - { id: 2, class: vgpr_32 }
+  - { id: 3, class: sreg_64 }
+  - { id: 4, class: vgpr_32 }
+  - { id: 5, class: sreg_64 }
+  - { id: 6, class: vgpr_32 }
+  - { id: 7, class: vgpr_32 }
+
+body: |
+  bb.0:
+    %0 = V_MOV_B32_e32 0, implicit %exec
+    %1 = V_MOV_B32_e32 0, implicit %exec
+    %2, %3 = V_ADD_I32_e64 %0, %1, implicit %exec
+    %6 = V_MOV_B32_e32 0, implicit %exec
+    %7 = V_MOV_B32_e32 0, implicit %exec
+    S_NOP 0, implicit def %vcc
+    %4, %5 = V_ADDC_U32_e64 %6, %7, %3, implicit %exec
+...
+
+# GCN-LABEL: name: interleave_add64s
+# GCN: dead %8, %9 = V_ADD_I32_e64 %0, %1, implicit %exec
+# GCN-NEXT: dead %12, dead %13 = V_ADDC_U32_e64 %4, %5, %9, implicit %exec
+# GCN-NEXT: dead %10, %11 = V_ADD_I32_e64 %2, %3, implicit %exec
+# GCN-NEXT: dead %14, dead %15 = V_ADDC_U32_e64 %6, %7, %11, implicit %exec
+name: interleave_add64s
+registers:
+  - { id: 0, class: vgpr_32 }
+  - { id: 1, class: vgpr_32 }
+  - { id: 2, class: vgpr_32 }
+  - { id: 3, class: vgpr_32 }
+  - { id: 4, class: vgpr_32 }
+  - { id: 5, class: vgpr_32 }
+  - { id: 6, class: vgpr_32 }
+  - { id: 7, class: vgpr_32 }
+  - { id: 8, class: vgpr_32 }
+  - { id: 9, class: sreg_64 }
+  - { id: 10, class: vgpr_32 }
+  - { id: 11, class: sreg_64 }
+  - { id: 12, class: vgpr_32 }
+  - { id: 13, class: sreg_64 }
+  - { id: 14, class: vgpr_32 }
+  - { id: 15, class: sreg_64 }
+
+body: |
+  bb.0:
+    %0 = V_MOV_B32_e32 0, implicit %exec
+    %1 = V_MOV_B32_e32 0, implicit %exec
+    %2 = V_MOV_B32_e32 0, implicit %exec
+    %3 = V_MOV_B32_e32 0, implicit %exec
+    %4 = V_MOV_B32_e32 0, implicit %exec
+    %5 = V_MOV_B32_e32 0, implicit %exec
+    %6 = V_MOV_B32_e32 0, implicit %exec
+    %7 = V_MOV_B32_e32 0, implicit %exec
+
+    %8, %9 = V_ADD_I32_e64 %0, %1, implicit %exec
+    %10, %11 = V_ADD_I32_e64 %2, %3, implicit %exec
+
+
+    %12, %13 = V_ADDC_U32_e64 %4, %5, %9, implicit %exec
+    %14, %15 = V_ADDC_U32_e64 %6, %7, %11, implicit %exec
+...
+
+# GCN-LABEL: name: cluster_mov_addc
+# GCN: S_NOP 0, implicit-def %vcc
+# GCN-NEXT: %2 = S_MOV_B64 0
+# GCN-NEXT: dead %3, dead %4 = V_ADDC_U32_e64 %0, %1, %2, implicit %exec
+name: cluster_mov_addc
+registers:
+  - { id: 0, class: vgpr_32 }
+  - { id: 1, class: vgpr_32 }
+  - { id: 2, class: sreg_64 }
+  - { id: 3, class: vgpr_32 }
+  - { id: 4, class: sreg_64 }
+  - { id: 6, class: vgpr_32 }
+  - { id: 7, class: vgpr_32 }
+
+body: |
+  bb.0:
+    %0 = V_MOV_B32_e32 0, implicit %exec
+    %1 = V_MOV_B32_e32 0, implicit %exec
+    %2 = S_MOV_B64 0
+    S_NOP 0, implicit def %vcc
+    %3, %4 = V_ADDC_U32_e64 %0, %1, %2, implicit %exec
+...
+
+# GCN-LABEL: name: no_cluster_add_addc_diff_sgpr
+# GCN: dead %2, dead %3 = V_ADD_I32_e64 %0, %1, implicit %exec
+# GCN-NEXT: %6 = V_MOV_B32_e32 0, implicit %exec
+# GCN-NEXT: %7 = V_MOV_B32_e32 0, implicit %exec
+# GCN-NEXT: S_NOP 0, implicit-def %vcc
+# GCN-NEXT: %8 = S_MOV_B64 0
+# GCN-NEXT: dead %4, dead %5 = V_ADDC_U32_e64 %6, %7, %8, implicit %exec
+name: no_cluster_add_addc_diff_sgpr
+registers:
+  - { id: 0, class: vgpr_32 }
+  - { id: 1, class: vgpr_32 }
+  - { id: 2, class: vgpr_32 }
+  - { id: 3, class: sreg_64 }
+  - { id: 4, class: vgpr_32 }
+  - { id: 5, class: sreg_64 }
+  - { id: 6, class: vgpr_32 }
+  - { id: 7, class: vgpr_32 }
+  - { id: 8, class: sreg_64 }
+body: |
+  bb.0:
+    %0 = V_MOV_B32_e32 0, implicit %exec
+    %1 = V_MOV_B32_e32 0, implicit %exec
+    %8 = S_MOV_B64 0
+    %2, %3 = V_ADD_I32_e64 %0, %1, implicit %exec
+    %6 = V_MOV_B32_e32 0, implicit %exec
+    %7 = V_MOV_B32_e32 0, implicit %exec
+    S_NOP 0, implicit def %vcc
+    %4, %5 = V_ADDC_U32_e64 %6, %7, %8, implicit %exec
+...
+# GCN-LABEL: name: cluster_sub_subb
+# GCN: S_NOP 0, implicit-def %vcc
+# GCN: dead %2, %3 = V_SUB_I32_e64 %0, %1, implicit %exec
+# GCN: dead %4, dead %5 = V_SUBB_U32_e64 %6, %7, %3, implicit %exec
+name: cluster_sub_subb
+registers:
+  - { id: 0, class: vgpr_32 }
+  - { id: 1, class: vgpr_32 }
+  - { id: 2, class: vgpr_32 }
+  - { id: 3, class: sreg_64 }
+  - { id: 4, class: vgpr_32 }
+  - { id: 5, class: sreg_64 }
+  - { id: 6, class: vgpr_32 }
+  - { id: 7, class: vgpr_32 }
+
+body: |
+  bb.0:
+    %0 = V_MOV_B32_e32 0, implicit %exec
+    %1 = V_MOV_B32_e32 0, implicit %exec
+    %2, %3 = V_SUB_I32_e64 %0, %1, implicit %exec
+    %6 = V_MOV_B32_e32 0, implicit %exec
+    %7 = V_MOV_B32_e32 0, implicit %exec
+    S_NOP 0, implicit def %vcc
+    %4, %5 = V_SUBB_U32_e64 %6, %7, %3, implicit %exec
+...
+
+# GCN-LABEL: name: cluster_cmp_cndmask
+# GCN: S_NOP 0, implicit-def %vcc
+# GCN-NEXT: %3 = V_CMP_EQ_I32_e64 %0, %1, implicit %exec
+# GCN-NEXT: dead %4 = V_CNDMASK_B32_e64 %0, %1, %3, implicit %exec
+name: cluster_cmp_cndmask
+registers:
+  - { id: 0, class: vgpr_32 }
+  - { id: 1, class: vgpr_32 }
+  - { id: 2, class: vgpr_32 }
+  - { id: 3, class: sreg_64 }
+  - { id: 4, class: vgpr_32 }
+  - { id: 5, class: sreg_64 }
+  - { id: 6, class: vgpr_32 }
+  - { id: 7, class: vgpr_32 }
+
+body: |
+  bb.0:
+    %0 = V_MOV_B32_e32 0, implicit %exec
+    %1 = V_MOV_B32_e32 0, implicit %exec
+    %3 = V_CMP_EQ_I32_e64 %0, %1, implicit %exec
+    S_NOP 0, implicit def %vcc
+    %4 = V_CNDMASK_B32_e64 %0, %1, %3, implicit %exec
+...
+
+# GCN-LABEL: name: cluster_multi_use_cmp_cndmask
+# GCN: %4 = V_CMP_EQ_I32_e64 %0, %1, implicit %exec
+# GCN-NEXT: dead %5 = V_CNDMASK_B32_e64 %2, %1, %4, implicit %exec
+# GCN-NEXT: dead %6 = V_CNDMASK_B32_e64 %1, %3, %4, implicit %exec
+name: cluster_multi_use_cmp_cndmask
+registers:
+  - { id: 0, class: vgpr_32 }
+  - { id: 1, class: vgpr_32 }
+  - { id: 2, class: vgpr_32 }
+  - { id: 3, class: vgpr_32 }
+  - { id: 4, class: sreg_64 }
+  - { id: 5, class: vgpr_32 }
+  - { id: 6, class: vgpr_32 }
+  - { id: 7, class: vgpr_32 }
+
+body: |
+  bb.0:
+    %0 = V_MOV_B32_e32 0, implicit %exec
+    %1 = V_MOV_B32_e32 0, implicit %exec
+    %2 = V_MOV_B32_e32 0, implicit %exec
+    %3 = V_MOV_B32_e32 0, implicit %exec
+
+    %4 = V_CMP_EQ_I32_e64 %0, %1, implicit %exec
+    S_NOP 0, implicit def %vcc
+    %5 = V_CNDMASK_B32_e64 %2, %1, %4, implicit %exec
+    %6 = V_CNDMASK_B32_e64 %1, %3, %4, implicit %exec
+...
+
+# GCN-LABEL: name: cluster_multi_use_cmp_cndmask2
+# GCN: %4 = V_CMP_EQ_I32_e64 %0, %1, implicit %exec
+# GCN-NEXT: dead %5 = V_CNDMASK_B32_e64 %2, %1, %4, implicit %exec
+# GCN-NEXT: %3 = V_MOV_B32_e32 0, implicit %exec
+# GCN-NEXT: dead %6 = V_CNDMASK_B32_e64 %1, %3, %4, implicit %exec
+name: cluster_multi_use_cmp_cndmask2
+registers:
+  - { id: 0, class: vgpr_32 }
+  - { id: 1, class: vgpr_32 }
+  - { id: 2, class: vgpr_32 }
+  - { id: 3, class: vgpr_32 }
+  - { id: 4, class: sreg_64 }
+  - { id: 5, class: vgpr_32 }
+  - { id: 6, class: vgpr_32 }
+  - { id: 7, class: vgpr_32 }
+
+body: |
+  bb.0:
+    %0 = V_MOV_B32_e32 0, implicit %exec
+    %1 = V_MOV_B32_e32 0, implicit %exec
+    %4 = V_CMP_EQ_I32_e64 %0, %1, implicit %exec
+    %2 = V_MOV_B32_e32 0, implicit %exec
+    %5 = V_CNDMASK_B32_e64 %2, %1, %4, implicit %exec
+    %3 = V_MOV_B32_e32 0, implicit %exec
+    %6 = V_CNDMASK_B32_e64 %1, %3, %4, implicit %exec
+...

test/CodeGen/AMDGPU/multi-divergent-exit-region.ll

 ; GCN-LABEL: {{^}}multi_divergent_region_exit_ret_ret:
 ; GCN: v_cmp_lt_i32_e32 vcc, 1
 ; GCN: s_and_saveexec_b64
 ; GCN: s_xor_b64
 
 
 ; FIXME: Why is this compare essentially repeated?
 ; GCN: v_cmp_eq_u32_e32 vcc, 1, [[REG:v[0-9]+]]
-; GCN-NEXT: v_cmp_ne_u32_e64 s{{\[[0-9]+:[0-9]+\]}}, 1, [[REG]]
 ; GCN: v_cndmask_b32_e64 v{{[0-9]+}}, 0, -1, vcc
-; GCN: v_cndmask_b32_e64 v{{[0-9]+}}, 0, -1
+; GCN: v_cmp_ne_u32_e32 vcc, 1, [[REG]]
+; GCN: v_cndmask_b32_e64 v{{[0-9]+}}, 0, -1, vcc
 
 ; GCN: ; %Flow1
 ; GCN-NEXT: s_or_b64 exec, exec
 ; GCN: v_cmp_ne_u32_e32 vcc, 0
 
 ; GCN: ; %exit1
 ; GCN: ds_write_b32
 

test/CodeGen/AMDGPU/sad.ll

define amdgpu_kernel void @v_sad_u32_multi_use_sub_pat2(i32 addrspace(1)* %out, i32 %a, i32 %b, i32 %c) {
   %ret = add i32 %ret0, %c
 
   store i32 %ret, i32 addrspace(1)* %out
   ret void
 }
 
 ; GCN-LABEL: {{^}}v_sad_u32_multi_use_select_pat2:
 ; GCN: s_sub_i32 s{{[0-9]+}}, s{{[0-9]+}}, s{{[0-9]+}}
-; GCN: v_cmp_gt_u32_e32 vcc, s{{[0-9]+}}, v{{[0-9]+}}
-; GCN: s_sub_i32 s{{[0-9]+}}, s{{[0-9]+}}, s{{[0-9]+}}
+; GCN-DAG: v_cmp_gt_u32_e32 vcc, s{{[0-9]+}}, v{{[0-9]+}}
+; GCN-DAG: s_sub_i32 s{{[0-9]+}}, s{{[0-9]+}}, s{{[0-9]+}}
 define amdgpu_kernel void @v_sad_u32_multi_use_select_pat2(i32 addrspace(1)* %out, i32 %a, i32 %b, i32 %c) {
   %icmp0 = icmp ugt i32 %a, %b
   %sub0 = sub i32 %a, %b
   %sub1 = sub i32 %b, %a
   %ret0 = select i1 %icmp0, i32 %sub0, i32 %sub1
   store volatile i32 %ret0, i32 *undef
 
   %ret = add i32 %ret0, %c

test/CodeGen/AMDGPU/select.f16.ll

 ; GCN: buffer_load_ushort v[[D_F16:[0-9]+]]
 ; SI:  v_cvt_f32_f16_e32 v[[A_F32:[0-9]+]], v[[A_F16]]
 ; SI:  v_cvt_f32_f16_e32 v[[B_F32:[0-9]+]], v[[B_F16]]
 ; SI:  v_cvt_f32_f16_e32 v[[D_F32:[0-9]+]], v[[D_F16]]
 ; SI:  v_cmp_nlt_f32_e32 vcc, v[[A_F32]], v[[B_F32]]
 ; SI:  v_cndmask_b32_e32 v[[R_F32:[0-9]+]], 0.5, v[[D_F32]], vcc
 ; SI:  v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[R_F32]]
 
-; VI:  v_cmp_nlt_f16_e32 vcc, v[[A_F16]], v[[B_F16]]
 ; VI:  v_mov_b32_e32 v[[C_F16:[0-9]+]], 0x3800{{$}}
+; VI:  v_cmp_nlt_f16_e32 vcc, v[[A_F16]], v[[B_F16]]
 ; VI:  v_cndmask_b32_e32 v[[R_F16:[0-9]+]], v[[C_F16]], v[[D_F16]], vcc
 ; GCN: buffer_store_short v[[R_F16]]
 ; GCN: s_endpgm
 define amdgpu_kernel void @select_f16_imm_c(
     half addrspace(1)* %r,
     half addrspace(1)* %a,
     half addrspace(1)* %b,
     half addrspace(1)* %d) {

 ; GCN: buffer_load_ushort v[[B_F16:[0-9]+]]
 ; GCN: buffer_load_ushort v[[C_F16:[0-9]+]]
 ; SI:  v_cvt_f32_f16_e32 v[[A_F32:[0-9]+]], v[[A_F16]]
 ; SI:  v_cvt_f32_f16_e32 v[[B_F32:[0-9]+]], v[[B_F16]]
 ; SI:  v_cvt_f32_f16_e32 v[[C_F32:[0-9]+]], v[[C_F16]]
 ; SI:  v_cmp_lt_f32_e32 vcc, v[[A_F32]], v[[B_F32]]
 ; SI:  v_cndmask_b32_e32 v[[R_F32:[0-9]+]], 0.5, v[[C_F32]]
 ; SI:  v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[R_F32]]
-; VI:  v_cmp_lt_f16_e32 vcc, v[[A_F16]], v[[B_F16]]
 ; VI:  v_mov_b32_e32 v[[D_F16:[0-9]+]], 0x3800{{$}}
+; VI:  v_cmp_lt_f16_e32 vcc, v[[A_F16]], v[[B_F16]]
 ; VI:  v_cndmask_b32_e32 v[[R_F16:[0-9]+]], v[[D_F16]], v[[C_F16]], vcc
 ; GCN: buffer_store_short v[[R_F16]]
 ; GCN: s_endpgm
 define amdgpu_kernel void @select_f16_imm_d(
     half addrspace(1)* %r,
     half addrspace(1)* %a,
     half addrspace(1)* %b,
     half addrspace(1)* %c) {
 entry:
   %a.val = load half, half addrspace(1)* %a
   %b.val = load half, half addrspace(1)* %b
   %c.val = load half, half addrspace(1)* %c
   %fcmp = fcmp olt half %a.val, %b.val
   %r.val = select i1 %fcmp, half %c.val, half 0xH3800
   store half %r.val, half addrspace(1)* %r
   ret void
 }
 
 ; GCN-LABEL: {{^}}select_v2f16:
 ; SI: v_cvt_f32_f16_e32
 ; SI: v_cvt_f32_f16_e32
 ; SI: v_cvt_f32_f16_e32
 ; SI: v_cvt_f32_f16_e32
-; SI: v_cmp_lt_f32_e64
 ; SI: v_cmp_lt_f32_e32
 ; SI: v_cndmask_b32_e32
-; SI: v_cndmask_b32_e64
+; SI: v_cmp_lt_f32_e32
+; SI: v_cndmask_b32_e32
 ; SI: v_cvt_f16_f32_e32
 ; SI: v_cvt_f16_f32_e32
 
-; VI: v_cmp_lt_f16_e64
 ; VI: v_cmp_lt_f16_e32
-; VI: v_cndmask_b32_e64
+; VI: v_cndmask_b32_e32
+; VI: v_cmp_lt_f16_e32
 ; VI: v_cndmask_b32_e32
 
 ; GCN: s_endpgm
 define amdgpu_kernel void @select_v2f16(
     <2 x half> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a,
     <2 x half> addrspace(1)* %b,
     <2 x half> addrspace(1)* %c,

 
 ; GCN-LABEL: {{^}}select_v2f16_imm_a:
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
-; SI-DAG:  v_cmp_gt_f32_e64
-; SI-DAG:  v_cmp_lt_f32_e32 vcc, 0.5
+
+; SI: v_cmp_lt_f32_e32 vcc, 0.5
+; SI: v_cndmask_b32_e32
+; SI: v_cmp_gt_f32_e32
+; SI: v_cndmask_b32_e32
 
-; VI:  v_cmp_lt_f16_e32
-; VI:  v_cmp_gt_f16_e64
-; GCN: v_cndmask_b32_e32
-; GCN: v_cndmask_b32_e64
+; VI: v_cmp_lt_f16_e32
+; VI: v_cndmask_b32_e32
+; VI: v_cmp_gt_f16_e32
+; VI: v_cndmask_b32_e32
+
 ; SI:  v_cvt_f16_f32_e32
 ; SI:  v_cvt_f16_f32_e32
 ; GCN: s_endpgm
 define amdgpu_kernel void @select_v2f16_imm_a(
     <2 x half> addrspace(1)* %r,
     <2 x half> addrspace(1)* %b,
     <2 x half> addrspace(1)* %c,
     <2 x half> addrspace(1)* %d) {

 
 ; GCN-LABEL: {{^}}select_v2f16_imm_b:
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
-; SI-DAG:  v_cmp_lt_f32_e64
-; SI-DAG:  v_cmp_gt_f32_e32 vcc, 0.5
+
+; SI: v_cmp_gt_f32_e32 vcc, 0.5
+; SI: v_cndmask_b32_e32
+; SI: v_cmp_lt_f32_e32
+; SI: v_cndmask_b32_e32
 
-; VI:  v_cmp_gt_f16_e32
-; VI:  v_cmp_lt_f16_e64
-; GCN: v_cndmask_b32_e32
-; GCN: v_cndmask_b32_e64
+; VI: v_cmp_gt_f16_e32
+; VI: v_cndmask_b32_e32
+; VI: v_cmp_lt_f16_e32
+; VI: v_cndmask_b32_e32
 
 ; SI:  v_cvt_f16_f32_e32
 ; SI:  v_cvt_f16_f32_e32
 ; GCN: s_endpgm
 define amdgpu_kernel void @select_v2f16_imm_b(
     <2 x half> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a,
     <2 x half> addrspace(1)* %c,

 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
 
 ; SI: v_cmp_nlt_f32_e32
-; SI: v_cmp_nlt_f32_e64
-; SI: v_cndmask_b32_e64
+; SI: v_cndmask_b32_e32
+; SI: v_cmp_nlt_f32_e32
 ; SI: v_cndmask_b32_e32
 
 ; VI: v_cmp_nlt_f16_e32
 ; VI: v_cndmask_b32_e32
 
 ; VI: v_cmp_nlt_f16_e32
 ; VI: v_cndmask_b32_e32
 

 
 ; GCN-LABEL: {{^}}select_v2f16_imm_d:
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
 ; SI:  v_cvt_f32_f16_e32
-; SI:  v_cmp_lt_f32_e64
+
 ; SI:  v_cmp_lt_f32_e32
+; SI: v_cndmask_b32
+; SI:  v_cmp_lt_f32_e32
+; SI: v_cndmask_b32
 
 ; VI:  v_cmp_lt_f16_e32
-; VI:  v_cmp_lt_f16_e64
-; GCN: v_cndmask_b32
-; GCN: v_cndmask_b32
+; VI: v_cndmask_b32
+; VI:  v_cmp_lt_f16_e32
+; VI: v_cndmask_b32
+
 ; SI:  v_cvt_f16_f32_e32
 ; SI:  v_cvt_f16_f32_e32
 ; GCN: s_endpgm
 define amdgpu_kernel void @select_v2f16_imm_d(
     <2 x half> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a,
     <2 x half> addrspace(1)* %b,
     <2 x half> addrspace(1)* %c) {

test/CodeGen/AMDGPU/setcc.ll

 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=FUNC %s
 ; RUN: llc -march=r600 -mcpu=redwood -verify-machineinstrs < %s | FileCheck -check-prefix=R600 -check-prefix=FUNC %s
 
 declare i32 @llvm.r600.read.tidig.x() nounwind readnone
 
 ; FUNC-LABEL: {{^}}setcc_v2i32:
 ; R600-DAG: SETE_INT * T{{[0-9]+\.[XYZW]}}, KC0[3].X, KC0[3].Z
 ; R600-DAG: SETE_INT * T{{[0-9]+\.[XYZW]}}, KC0[2].W, KC0[3].Y
 
-; GCN-DAG: v_cmp_eq_u32_e32
-; GCN-DAG: v_cmp_eq_u32_e64
+; GCN: v_cmp_eq_u32_e32
+; GCN: v_cmp_eq_u32_e32
 define amdgpu_kernel void @setcc_v2i32(<2 x i32> addrspace(1)* %out, <2 x i32> %a, <2 x i32> %b) #0 {
   %result = icmp eq <2 x i32> %a, %b
   %sext = sext <2 x i1> %result to <2 x i32>
   store <2 x i32> %sext, <2 x i32> addrspace(1)* %out
   ret void
 }
 
 ; FUNC-LABEL: {{^}}setcc_v4i32:
 ; R600-DAG: SETE_INT * T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 ; R600-DAG: SETE_INT * T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 ; R600-DAG: SETE_INT * T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 ; R600-DAG: SETE_INT * T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 
 ; GCN: v_cmp_eq_u32_e32
-; GCN: v_cmp_eq_u32_e64
-; GCN: v_cmp_eq_u32_e64
-; GCN: v_cmp_eq_u32_e64
+; GCN: v_cmp_eq_u32_e32
+; GCN: v_cmp_eq_u32_e32
+; GCN: v_cmp_eq_u32_e32
 define amdgpu_kernel void @setcc_v4i32(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %in) #0 {
   %b_ptr = getelementptr <4 x i32>, <4 x i32> addrspace(1)* %in, i32 1
   %a = load <4 x i32>, <4 x i32> addrspace(1)* %in
   %b = load <4 x i32>, <4 x i32> addrspace(1)* %b_ptr
   %result = icmp eq <4 x i32> %a, %b
   %sext = sext <4 x i1> %result to <4 x i32>
   store <4 x i32> %sext, <4 x i32> addrspace(1)* %out
   ret void

test/CodeGen/AMDGPU/uaddo.ll

define amdgpu_kernel void @v_uaddo_i32(i32 addrspace(1)* %out, i1 addrspace(1)* %carryout, i32 addrspace(1)* %a.ptr, i32 addrspace(1)* %b.ptr) #0 {
   %val = extractvalue { i32, i1 } %uadd, 0
   %carry = extractvalue { i32, i1 } %uadd, 1
   store i32 %val, i32 addrspace(1)* %out, align 4
   store i1 %carry, i1 addrspace(1)* %carryout
   ret void
 }
 
 ; FUNC-LABEL: {{^}}v_uaddo_i32_novcc:
-; GCN: v_add_i32_e64 v{{[0-9]+}}, [[COND:s\[[0-9]+:[0-9]+\]]], v{{[0-9]+}}, v{{[0-9]+}}
-; GCN: v_cndmask_b32_e64 v{{[0-9]+}}, 0, 1, [[COND]]
+; GCN: v_add_i32_e32 v{{[0-9]+}}, vcc, v{{[0-9]+}}, v{{[0-9]+}}
+; GCN: v_cndmask_b32_e64 v{{[0-9]+}}, 0, 1, vcc
 
 ; EG: ADDC_UINT
 ; EG: ADD_INT
 define amdgpu_kernel void @v_uaddo_i32_novcc(i32 addrspace(1)* %out, i1 addrspace(1)* %carryout, i32 addrspace(1)* %a.ptr, i32 addrspace(1)* %b.ptr) #0 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
   %tid.ext = sext i32 %tid to i64
   %a.gep = getelementptr inbounds i32, i32 addrspace(1)* %a.ptr
   %b.gep = getelementptr inbounds i32, i32 addrspace(1)* %b.ptr

test/CodeGen/AMDGPU/usubo.ll

define amdgpu_kernel void @v_usubo_i32(i32 addrspace(1)* %out, i1 addrspace(1)* %carryout, i32 addrspace(1)* %a.ptr, i32 addrspace(1)* %b.ptr) #0 {
   %val = extractvalue { i32, i1 } %usub, 0
   %carry = extractvalue { i32, i1 } %usub, 1
   store i32 %val, i32 addrspace(1)* %out, align 4
   store i1 %carry, i1 addrspace(1)* %carryout
   ret void
 }
 
 ; FUNC-LABEL: {{^}}v_usubo_i32_novcc:
-; GCN: v_sub_i32_e64 v{{[0-9]+}}, [[COND:s\[[0-9]+:[0-9]+\]]], v{{[0-9]+}}, v{{[0-9]+}}
-; GCN: v_cndmask_b32_e64 v{{[0-9]+}}, 0, 1, [[COND]]
+; GCN: v_sub_i32_e32 v{{[0-9]+}}, vcc, v{{[0-9]+}}, v{{[0-9]+}}
+; GCN: v_cndmask_b32_e64 v{{[0-9]+}}, 0, 1, vcc
 
 ; EG-DAG: SUBB_UINT
 ; EG-DAG: SUB_INT
 define amdgpu_kernel void @v_usubo_i32_novcc(i32 addrspace(1)* %out, i1 addrspace(1)* %carryout, i32 addrspace(1)* %a.ptr, i32 addrspace(1)* %b.ptr) #0 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
   %tid.ext = sext i32 %tid to i64
   %a.gep = getelementptr inbounds i32, i32 addrspace(1)* %a.ptr
   %b.gep = getelementptr inbounds i32, i32 addrspace(1)* %b.ptr

test/CodeGen/AMDGPU/v_cndmask.ll

 ; FIXME: Why does VI make the wrong regalloc choice?
 ; GCN-LABEL: {{^}}icmp_vgprX_k0_select_k1_vgprZ_i64:
 ; GCN: {{buffer|flat}}_load_dwordx2 v{{\[}}[[X_LO:[0-9]+]]:[[X_HI:[0-9]+]]{{\]}}
 ; GCN-DAG: {{buffer|flat}}_load_dwordx2 v{{\[}}[[Z_LO:[0-9]+]]:[[Z_HI:[0-9]+]]{{\]}}
 ; SI-DAG: v_cmp_lt_i64_e32 vcc, -1, v{{\[}}[[X_LO]]:[[X_HI]]{{\]}}
 ; SI-DAG: v_cndmask_b32_e32 v{{[0-9]+}}, 0, v[[Z_HI]], vcc
 ; SI-DAG: v_cndmask_b32_e32 v{{[0-9]+}}, 2, v[[Z_LO]], vcc
 
-; VI-DAG: v_cmp_lt_i64_e64 s{{\[[0-9]+:[0-9]+\]}}, -1, v{{\[}}[[X_LO]]:[[X_HI]]{{\]}}
-; VI-DAG: v_cndmask_b32_e64 v{{[0-9]+}}, 0, v[[Z_HI]], s
-; VI-DAG: v_cndmask_b32_e64 v{{[0-9]+}}, 2, v[[Z_LO]], s
+; VI-DAG: v_cmp_lt_i64_e32 vcc, -1, v{{\[}}[[X_LO]]:[[X_HI]]{{\]}}
+; VI-DAG: v_cndmask_b32_e32 v{{[0-9]+}}, 0, v[[Z_HI]], vcc
+; VI-DAG: v_cndmask_b32_e32 v{{[0-9]+}}, 2, v[[Z_LO]], vcc
 define amdgpu_kernel void @icmp_vgprX_k0_select_k1_vgprZ_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %x.ptr, i64 addrspace(1)* %z.ptr) #0 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x() #1
   %tid.ext = sext i32 %tid to i64
   %x.gep = getelementptr inbounds i64, i64 addrspace(1)* %x.ptr, i64 %tid.ext
   %z.gep = getelementptr inbounds i64, i64 addrspace(1)* %z.ptr, i64 %tid.ext
   %out.gep = getelementptr inbounds i64, i64 addrspace(1)* %out, i64 %tid.ext
   %x = load volatile i64, i64 addrspace(1)* %x.gep
   %z = load volatile i64, i64 addrspace(1)* %z.gep
define amdgpu_kernel void @fcmp_k0_vgprX_select_k1_vgprZ_v4f32(<4 x float> addrspace(1)* %out, float addrspace(1)* %x.ptr, <4 x float> addrspace(1)* %z.ptr) #0 {
   %select = select i1 %setcc, <4 x float> <float 1.0, float 2.0, float -0.5, float 4.0>, <4 x float> %z
   store <4 x float> %select, <4 x float> addrspace(1)* %out.gep
   ret void
 }
 
 ; GCN-LABEL: {{^}}icmp_vgprX_k0_select_k1_vgprZ_i1:
 ; GCN: load_dword
 ; GCN: load_ubyte
-; GCN-DAG: v_cmp_gt_i32_e64 s{{\[[0-9]+:[0-9]+\]}}, 0, v
+; GCN-DAG: v_cmp_gt_i32_e32 vcc, 0, v
 ; DCN-DAG: v_and_b32_e32 v{{[0-9]+}}, 1,
-; GCN-DAG: v_cmp_eq_u32_e32 vcc, 1, v
-; GCN-DAG: s_or_b64 s{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, vcc
+; GCN-DAG: v_cmp_eq_u32_e64 s{{\[[0-9]+:[0-9]+\]}}, 1, v
+; GCN-DAG: s_or_b64 s{{\[[0-9]+:[0-9]+\]}}, vcc, s{{\[[0-9]+:[0-9]+\]}}
 ; GCN: v_cndmask_b32_e64 v{{[0-9]+}}, 0, 1, s
 ; GCN: store_byte
 define amdgpu_kernel void @icmp_vgprX_k0_select_k1_vgprZ_i1(i1 addrspace(1)* %out, i32 addrspace(1)* %x.ptr, i1 addrspace(1)* %z.ptr) #0 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x() #1
   %tid.ext = sext i32 %tid to i64
   %x.gep = getelementptr inbounds i32, i32 addrspace(1)* %x.ptr, i64 %tid.ext
   %z.gep = getelementptr inbounds i1, i1 addrspace(1)* %z.ptr, i64 %tid.ext
   %out.gep = getelementptr inbounds i1, i1 addrspace(1)* %out, i64 %tid.ext

test/CodeGen/AMDGPU/vselect.ll

 ;RUN: llc < %s -march=amdgcn -verify-machineinstrs | FileCheck --check-prefix=SI --check-prefix=FUNC %s
 ;RUN: llc < %s -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs | FileCheck --check-prefix=SI --check-prefix=VI --check-prefix=FUNC %s
 ;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck --check-prefix=EG --check-prefix=FUNC %s
 
 ; FUNC-LABEL: {{^}}test_select_v2i32:
 
 ; EG-DAG: CNDE_INT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW], T[0-9]+\.[XYZW]}}, KC0[3].Z
 ; EG-DAG: CNDE_INT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW], T[0-9]+\.[XYZW]}}, KC0[3].Y
 
-; SI: v_cndmask_b32_e64
+; SI: v_cmp_gt_i32_e32 vcc
+; SI: v_cndmask_b32_e32
+; SI: v_cmp_gt_i32_e32 vcc
 ; SI: v_cndmask_b32_e32
 
 define amdgpu_kernel void @test_select_v2i32(<2 x i32> addrspace(1)* %out, <2 x i32> addrspace(1)* %in0, <2 x i32> addrspace(1)* %in1, <2 x i32> %val) {
 entry:
   %load0 = load <2 x i32>, <2 x i32> addrspace(1)* %in0
   %load1 = load <2 x i32>, <2 x i32> addrspace(1)* %in1
   %cmp = icmp sgt <2 x i32> %load0, %load1
   %result = select <2 x i1> %cmp, <2 x i32> %val, <2 x i32> %load0
   store <2 x i32> %result, <2 x i32> addrspace(1)* %out
   ret void
 }
 
 ; FUNC-LABEL: {{^}}test_select_v2f32:
 
 ; EG: CNDE_INT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 ; EG: CNDE_INT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 
-;SI: v_cndmask_b32_e64
-;SI: v_cndmask_b32_e32
+
+; SI: v_cmp_neq_f32_e32 vcc
+; SI: v_cndmask_b32_e32
+; SI: v_cmp_neq_f32_e32 vcc
+; SI: v_cndmask_b32_e32
 
 define amdgpu_kernel void @test_select_v2f32(<2 x float> addrspace(1)* %out, <2 x float> addrspace(1)* %in0, <2 x float> addrspace(1)* %in1) {
 entry:
   %0 = load <2 x float>, <2 x float> addrspace(1)* %in0
   %1 = load <2 x float>, <2 x float> addrspace(1)* %in1
   %cmp = fcmp une <2 x float> %0, %1
   %result = select <2 x i1> %cmp, <2 x float> %0, <2 x float> %1
   store <2 x float> %result, <2 x float> addrspace(1)* %out
   ret void
 }
 
 ;FUNC-LABEL: {{^}}test_select_v4i32:
 
 ; EG-DAG: CNDE_INT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW], T[0-9]+\.[XYZW]}}, KC0[4].X
 ; EG-DAG: CNDE_INT {{\** *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}, KC0[3].W
 ; EG-DAG: CNDE_INT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW], T[0-9]+\.[XYZW]}}, KC0[3].Z
 ; EG-DAG: CNDE_INT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW], T[0-9]+\.[XYZW]}}, KC0[3].Y
 
-; FIXME: The shrinking does not happen on tonga
-
-; SI: v_cndmask_b32
-; SI: v_cndmask_b32
-; SI: v_cndmask_b32
-; SI: v_cndmask_b32
+; SI: v_cndmask_b32_e32
+; SI: v_cndmask_b32_e32
+; SI: v_cndmask_b32_e32
+; SI: v_cndmask_b32_e32
 
 define amdgpu_kernel void @test_select_v4i32(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %in0, <4 x i32> addrspace(1)* %in1, <4 x i32> %val) {
 entry:
   %load0 = load <4 x i32>, <4 x i32> addrspace(1)* %in0
   %load1 = load <4 x i32>, <4 x i32> addrspace(1)* %in1
   %cmp = icmp sgt <4 x i32> %load0, %load1
   %result = select <4 x i1> %cmp, <4 x i32> %val, <4 x i32> %load0
   store <4 x i32> %result, <4 x i32> addrspace(1)* %out
   ret void
 }
 
 ;FUNC-LABEL: {{^}}test_select_v4f32:
 ;EG: CNDE_INT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 ;EG: CNDE_INT {{\** *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 ;EG: CNDE_INT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 ;EG: CNDE_INT {{\** *}}T{{[0-9]+\.[XYZW], PV\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 
+; SI: v_cndmask_b32_e32
+; SI: v_cndmask_b32_e32
+; SI: v_cndmask_b32_e32
+; SI: v_cndmask_b32_e32
 define amdgpu_kernel void @test_select_v4f32(<4 x float> addrspace(1)* %out, <4 x float> addrspace(1)* %in0, <4 x float> addrspace(1)* %in1) {
 entry:
   %0 = load <4 x float>, <4 x float> addrspace(1)* %in0
   %1 = load <4 x float>, <4 x float> addrspace(1)* %in1
   %cmp = fcmp une <4 x float> %0, %1
   %result = select <4 x i1> %cmp, <4 x float> %0, <4 x float> %1
   store <4 x float> %result, <4 x float> addrspace(1)* %out
   ret void
 }