[AMDGPU] Pre-allocate WWM registers to reduce VGPR pressure.

[sheredom]

Summary

This change incorporates an effort by Connor Abbot to change how we deal with WWM operations potentially trashing valid values in inactive lanes.

Previously, the SIFixWWMLiveness pass would work out which registers were being trashed within WWM regions, and ensure that the register allocator did not have any values it was depending on resident in those registers if the WWM section would trash them. This worked perfectly well, but would cause sometimes severe register pressure when the WWM section resided before divergent control flow (or at least that is where I mostly observed it).

This fix instead runs through the WWM sections and pre allocates some registers for WWM. It then reserves these registers so that the register allocator cannot use them. This results in a significant register saving on some WWM shaders I'm working with (130 -> 104 VGPRs, with just this change!).

This change was entirely thought up by @cwabbott (I claim no credit for the original idea!), but I had some time to look at it and so we agreed that I could give it a final polish for submission.

[arsenm]

I really don't like introducing new, dynamically reserved registers for this. It's going to introduce hell for dealing with any kind of ABI, and reserved registers are generally a bad idea. There's also nothing guaranteeing there are any free registers available to reserve, since you are just grabbing totally unused ones. This is going to just hit some variant of the problem I've been working on solving for handling SGPR->VGPR spills. Can WWM code be moved into a bundle or something?

include/llvm/IR/IntrinsicsAMDGPU.td

1363–1365 ↗

(On Diff #190402)

This is a separate fix that can be split into its own patch

lib/Target/AMDGPU/SIPreAllocateWWMRegs.cpp

52 ↗

(On Diff #190402)

You can remove this

222–227 ↗

(On Diff #190402)

I don't like this hardcoded opcode check. Why is S_OR_SAVEEXEC_B64 special?

[cwabbott]

In D59295#1427734, @arsenm wrote:

I really don't like introducing new, dynamically reserved registers for this. It's going to introduce hell for dealing with any kind of ABI, and reserved registers are generally a bad idea. There's also nothing guaranteeing there are any free registers available to reserve, since you are just grabbing totally unused ones. This is going to just hit some variant of the problem I've been working on solving for handling SGPR->VGPR spills. Can WWM code be moved into a bundle or something?

No, since the problem the current pass and this new pass are trying to solve affects register allocation for code that is arbitrarily far away from the original WWM sequence. For a more detailed explanation, I can't do any better than the comment at the start of SIFixWWMLiveness.cpp. (There will also be problems if RA decides to split a live interval inside a WWM sequence, which can be fixed by bundling it, but that's a completely different problem).

While it might seem dangerous, in practice this works out, since WWM sequences that the frontends currently emit only require a few registers. Hence this pass is guaranteed to succeed, even if there's very high register pressure. If allocating the registers needed for the WWM sequence fails and RA decides to spill something inside there, then you're pretty much toast anyways since the same invalid-lane-clobbering concerns would reappear.

One idea would be to add a way to tell RA that a certain live range absolutely cannot be split (and probably boost its priority as well, lest we fail to allocate it), pre-allocate one or more of these unsplittable registers for WWM, make every definition in the WWM sequence a partial definition, and add fake definitions of the WWM registers in the closest block with uniform control flow that dominates the WWM sequence in order to prevent definitions whose invalid lanes could be clobbered from using the WWM registers. This gives RA a little more flexibility and means that potentially some other operations could use the WWM registers, but you still basically wind up preallocating them.

[arsenm]

In D59295#1427826, @cwabbott wrote:

One idea would be to add a way to tell RA that a certain live range absolutely cannot be split (and probably boost its priority as well, lest we fail to allocate it), pre-allocate one or more of these unsplittable registers for WWM, make every definition in the WWM sequence a partial definition, and add fake definitions of the WWM registers in the closest block with uniform control flow that dominates the WWM sequence in order to prevent definitions whose invalid lanes could be clobbered from using the WWM registers. This gives RA a little more flexibility and means that potentially some other operations could use the WWM registers, but you still basically wind up preallocating them.

Declaring a certain live range unsplittable is impossible. LiveIntervals sort of supports it, but not all the allocators use it. Particularly, FastRegAlloc doesn't really track liveness and spills all values live out of a block. It's important that anything works correctly without LIveIntervals. If you can express the constraints with some series of uses and defs, that would be preferable.

My main concerns are making sure this works with:

1. -O0/fastregalloc
2. Presence of calls
3. Inline asm or any other physical register constraints

test/CodeGen/AMDGPU/wwm-reserved.ll

2 ↗

(On Diff #190402)

Needs a -O0 run line

57 ↗

(On Diff #190402)

Needs some cases with control flow

[cwabbott]

In D59295#1427864, @arsenm wrote:

If you can express the constraints with some series of uses and defs, that would be preferable.

Unfortunately that's just not possible. The best thing we could come up with that does this was the current SIFixWWMLiveness pass, and even with subsequent modifications, it's still pretty terrible in practice. We just can't get decent performance without doing something like this hack.

[cwabbott]

Actually, now that I think about it, I believe we realized that SIFixWWMLiveness has a giant hole in that if any of the extra live ranges we insert are split, it'll fall over. I don't think anyone has come up with a way to express the constraints only with extra defs and uses in a way that always works, and I'm not sure it's possible. The issue is that we're lying to LLVM RA by pretending that vector instructions always fully clobber their destinations, and while before we were careful to never write to any inactive channels in order to keep up the charade, but WWM instructions force us to deal with it somehow. Fully informing LLVM of what's going on would involve marking every vector instruction as partially clobbering its destination, even the move instructions and load/store instructions LLVM emits during RA, which of course would tank performance unless LLVM is taught about predicated liveness -- but of course that's a whole lot of work that opens another can of worms (register pressure is suddenly not that meaningful anymore...).

In D59295#1427874, @cwabbott wrote:

The best thing we could come up with that does this was the current SIFixWWMLiveness pass, and even with subsequent modifications, it's still pretty terrible in practice.

[cwabbott]

lib/Target/AMDGPU/SIPreAllocateWWMRegs.cpp

222–227 ↗

(On Diff #190402)

IIRC I wrote this since WholeQuadMode currently just emits a S_OR_SAVEEXEC_B64 foo, -1 in order to start WWM, so you have to check it manually in order to know when WWM starts. Maybe it's better to got back and add a ENTER_WWM pseudoinstruction like the pre-existing EXIT_WWM.

[sheredom]

Addressed all review comments.

[sheredom]

lib/Target/AMDGPU/SIPreAllocateWWMRegs.cpp

222–227 ↗

(On Diff #190402)

I've added an ENTER_WWM like was suggested!

test/CodeGen/AMDGPU/wwm-reserved.ll

57 ↗

(On Diff #190402)

Added a CFG and a called function case.

[sheredom]

Remove the explicit pass name.

[arsenm]

test/CodeGen/AMDGPU/wwm-reserved.ll

13 ↗

(On Diff #191077)

Do these tests really need all of this? Can you just have a few sample instructions?

[sheredom]

test/CodeGen/AMDGPU/wwm-reserved.ll

13 ↗

(On Diff #191077)

I can probably cut it down to just a single DPP inst in each WWM section - good idea!

[sheredom]

Reduce the number of DPP calls in the test for cleanliness, and reintroduce the convergent on WWM. The CFG test contains the bug that was exposed by the lack of convergent on WWM, LLVM will sink the WWM statement out of the branch which totally messes up all calculations.

[sheredom]

include/llvm/IR/IntrinsicsAMDGPU.td

1363–1365 ↗

(On Diff #190402)

So I tried to remove this (forgetting why I needed it) and LLVM will sink the WWM out of the branch which totally messes up the WWM calculation. So this is actually a requirement for the patch, not a separate thing.

[arsenm]

include/llvm/IR/IntrinsicsAMDGPU.td

1363–1365 ↗

(On Diff #190402)

You can commit that first then

[sheredom]

include/llvm/IR/IntrinsicsAMDGPU.td

1363–1365 ↗

(On Diff #190402)

https://reviews.llvm.org/D59536

[sheredom]

include/llvm/IR/IntrinsicsAMDGPU.td

1363–1365 ↗

(On Diff #190402)

I've submitted that other change, so this should be good to go once I get a sign off!

[arsenm]

lib/Target/AMDGPU/SIPreAllocateWWMRegs.cpp

154 ↗

(On Diff #191099)

Can't you just hardcoded this to v_mov_b32?

156 ↗

(On Diff #191099)

MRI->reg_instructions()? It would also save the check that the operand matches later

160 ↗

(On Diff #191099)

.isCopy()

183–184 ↗

(On Diff #191099)

I know we have a helper to do this somewhere

[arsenm]

lib/Target/AMDGPU/SIPreAllocateWWMRegs.cpp

154 ↗

(On Diff #191099)

I'm confused on the constraints for the WWM intrinsic, or lack thereof. The WWM instruction just uses "unknown" and the intrinsic allows any type. Can this be a 64-bit register or greater?

[sheredom]

lib/Target/AMDGPU/SIPreAllocateWWMRegs.cpp

154 ↗

(On Diff #191099)

We only use WWM with 32-bit & 64-bit types in our stack - nothing else.

183–184 ↗

(On Diff #191099)

Couldn't find it myself in the code.

[sheredom]

Fix review comments.

[sheredom]

Update for two reasons:

• was missing V_SET_INACTIVE_B64 in the set inactive check, causing miscompile with double/int64
• changed where the mov's are injected to tie them explicitly to WWM statements rather than all copy's within a WWM section (which produces better codegen and actually matches more the intent of the WWM intrinsic in the first place).

[sheredom]

Found a fun little bug whereby the phys vgprs were being coalesced onto previous instructions, and then shouldClusterMemOps was assuming only virt regs. Added a workaround for that.

[tpr]

LGTM modulo the wrong license on the new file.

lib/Target/AMDGPU/SIPreAllocateWWMRegs.cpp

5 ↗

(On Diff #192842)

Wrong license.

[arsenm]

Can you add a test where allocation is impossible? e.g. use

call void asm sideeffect "",
  "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7},~{v8},~{v9}
  ,~{v10},~{v11},~{v12},~{v13},~{v14},~{v15},~{v16},~{v17},~{v18},~{v19}
  ,~{v20},~{v21},~{v22},~{v23},~{v24},~{v25},~{v26},~{v27},~{v28},~{v29}
  ,~{v30},~{v31},~{v32},~{v33},~{v34},~{v35},~{v36},~{v37},~{v38},~{v39}
  ,~{v40},~{v41},~{v42},~{v43},~{v44},~{v45},~{v46},~{v47},~{v48},~{v49}
  ,~{v50},~{v51},~{v52},~{v53},~{v54},~{v55},~{v56},~{v57},~{v58},~{v59}
  ,~{v60},~{v61},~{v62},~{v63},~{v64},~{v65},~{v66},~{v67},~{v68},~{v69}
  ,~{v70},~{v71},~{v72},~{v73},~{v74},~{v75},~{v76},~{v77},~{v78},~{v79}
  ,~{v80},~{v81},~{v82},~{v83},~{v84},~{v85},~{v86},~{v87},~{v88},~{v89}
  ,~{v90},~{v91},~{v92},~{v93},~{v94},~{v95},~{v96},~{v97},~{v98},~{v99}
  ,~{v100},~{v101},~{v102},~{v103},~{v104},~{v105},~{v106},~{v107},~{v108},~{v109}
  ,~{v110},~{v111},~{v112},~{v113},~{v114},~{v115},~{v116},~{v117},~{v118},~{v119}
  ,~{v120},~{v121},~{v122},~{v123},~{v124},~{v125},~{v126},~{v127},~{v128},~{v129}
  ,~{v130},~{v131},~{v132},~{v133},~{v134},~{v135},~{v136},~{v137},~{v138},~{v139}
  ,~{v140},~{v141},~{v142},~{v143},~{v144},~{v145},~{v146},~{v147},~{v148},~{v149}
  ,~{v150},~{v151},~{v152},~{v153},~{v154},~{v155},~{v156},~{v157},~{v158},~{v159}
  ,~{v160},~{v161},~{v162},~{v163},~{v164},~{v165},~{v166},~{v167},~{v168},~{v169}
  ,~{v170},~{v171},~{v172},~{v173},~{v174},~{v175},~{v176},~{v177},~{v178},~{v179}
  ,~{v180},~{v181},~{v182},~{v183},~{v184},~{v185},~{v186},~{v187},~{v188},~{v189}
  ,~{v190},~{v191},~{v192},~{v193},~{v194},~{v195},~{v196},~{v197},~{v198},~{v199}
  ,~{v200},~{v201},~{v202},~{v203},~{v204},~{v205},~{v206},~{v207},~{v208},~{v209}
  ,~{v210},~{v211},~{v212},~{v213},~{v214},~{v215},~{v216},~{v217},~{v218},~{v219}
  ,~{v220},~{v221},~{v222},~{v223},~{v224},~{v225},~{v226},~{v227},~{v228},~{v229}
  ,~{v230},~{v231},~{v232},~{v233},~{v234},~{v235},~{v236},~{v237},~{v238},~{v239}
  ,~{v240},~{v241},~{v242},~{v243},~{v244},~{v245},~{v246},~{v247},~{v248},~{v249}
  ,~{v250},~{v251},~{v252},~{v253},~{v254},~{v255}"() #1

[arsenm]

lib/Target/AMDGPU/SIInstrInfo.cpp

454–458 ↗

(On Diff #192842)

This is a separate change

lib/Target/AMDGPU/SIMachineFunctionInfo.h

260 ↗

(On Diff #192842)

This should be serialized into YAML. Maybe this should also be in terms of RegUnits?

lib/Target/AMDGPU/SIPreAllocateWWMRegs.cpp

103 ↗

(On Diff #192842)

Is the isPhysRegUsed check really necessary? It's going to break if you have multiple WWM sections

llvm/trunk/lib/Target/AMDGPU/AMDGPU.h

 FunctionPass *createSIShrinkInstructionsPass();
 FunctionPass *createSILoadStoreOptimizerPass();
 FunctionPass *createSIWholeQuadModePass();
 FunctionPass *createSIFixControlFlowLiveIntervalsPass();
 FunctionPass *createSIOptimizeExecMaskingPreRAPass();
 FunctionPass *createSIFixSGPRCopiesPass();
 FunctionPass *createSIMemoryLegalizerPass();
 FunctionPass *createSIInsertWaitcntsPass();
-FunctionPass *createSIFixWWMLivenessPass();
+FunctionPass *createSIPreAllocateWWMRegsPass();
 FunctionPass *createSIFormMemoryClausesPass();
 FunctionPass *createAMDGPUSimplifyLibCallsPass(const TargetOptions &);
 FunctionPass *createAMDGPUUseNativeCallsPass();
 FunctionPass *createAMDGPUCodeGenPreparePass();
 FunctionPass *createAMDGPUMachineCFGStructurizerPass();
 FunctionPass *createAMDGPURewriteOutArgumentsPass();
 FunctionPass *createSIModeRegisterPass();
 

 extern char &SILowerControlFlowID;
 
 void initializeSIInsertSkipsPass(PassRegistry &);
 extern char &SIInsertSkipsPassID;
 
 void initializeSIOptimizeExecMaskingPass(PassRegistry &);
 extern char &SIOptimizeExecMaskingID;
 
-void initializeSIFixWWMLivenessPass(PassRegistry &);
-extern char &SIFixWWMLivenessID;
+void initializeSIPreAllocateWWMRegsPass(PassRegistry &);
+extern char &SIPreAllocateWWMRegsID;
 
 void initializeAMDGPUSimplifyLibCallsPass(PassRegistry &);
 extern char &AMDGPUSimplifyLibCallsID;
 
 void initializeAMDGPUUseNativeCallsPass(PassRegistry &);
 extern char &AMDGPUUseNativeCallsID;
 
 void initializeSIAddIMGInitPass(PassRegistry &);

llvm/trunk/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

extern "C" void LLVMInitializeAMDGPUTarget() {
   initializeSIAnnotateControlFlowPass(*PR);
   initializeSIInsertWaitcntsPass(*PR);
   initializeSIModeRegisterPass(*PR);
   initializeSIWholeQuadModePass(*PR);
   initializeSILowerControlFlowPass(*PR);
   initializeSIInsertSkipsPass(*PR);
   initializeSIMemoryLegalizerPass(*PR);
   initializeSIOptimizeExecMaskingPass(*PR);
-  initializeSIFixWWMLivenessPass(*PR);
+  initializeSIPreAllocateWWMRegsPass(*PR);
   initializeSIFormMemoryClausesPass(*PR);
   initializeAMDGPUUnifyDivergentExitNodesPass(*PR);
   initializeAMDGPUAAWrapperPassPass(*PR);
   initializeAMDGPUExternalAAWrapperPass(*PR);
   initializeAMDGPUUseNativeCallsPass(*PR);
   initializeAMDGPUSimplifyLibCallsPass(*PR);
   initializeAMDGPUInlinerPass(*PR);
 }
void GCNPassConfig::addFastRegAlloc() {
   // FIXME: We have to disable the verifier here because of PHIElimination +
   // TwoAddressInstructions disabling it.
 
   // This must be run immediately after phi elimination and before
   // TwoAddressInstructions, otherwise the processing of the tied operand of
   // SI_ELSE will introduce a copy of the tied operand source after the else.
   insertPass(&PHIEliminationID, &SILowerControlFlowID, false);
 
-  // This must be run after SILowerControlFlow, since it needs to use the
-  // machine-level CFG, but before register allocation.
-  insertPass(&SILowerControlFlowID, &SIFixWWMLivenessID, false);
+  // This must be run just after RegisterCoalescing.
+  insertPass(&RegisterCoalescerID, &SIPreAllocateWWMRegsID, false);
 
   TargetPassConfig::addFastRegAlloc();
 }
 
 void GCNPassConfig::addOptimizedRegAlloc() {
   if (OptExecMaskPreRA) {
     insertPass(&MachineSchedulerID, &SIOptimizeExecMaskingPreRAID);
     insertPass(&SIOptimizeExecMaskingPreRAID, &SIFormMemoryClausesID);
   } else {
     insertPass(&MachineSchedulerID, &SIFormMemoryClausesID);
   }
 
   // This must be run immediately after phi elimination and before
   // TwoAddressInstructions, otherwise the processing of the tied operand of
   // SI_ELSE will introduce a copy of the tied operand source after the else.
   insertPass(&PHIEliminationID, &SILowerControlFlowID, false);
 
-  // This must be run after SILowerControlFlow, since it needs to use the
-  // machine-level CFG, but before register allocation.
-  insertPass(&SILowerControlFlowID, &SIFixWWMLivenessID, false);
+  // This must be run just after RegisterCoalescing.
+  insertPass(&RegisterCoalescerID, &SIPreAllocateWWMRegsID, false);
 
   TargetPassConfig::addOptimizedRegAlloc();
 }
 
 void GCNPassConfig::addPostRegAlloc() {
   addPass(&SIFixVGPRCopiesID);
   if (getOptLevel() > CodeGenOpt::None)
     addPass(&SIOptimizeExecMaskingID);

llvm/trunk/lib/Target/AMDGPU/CMakeLists.txt

add_llvm_target(AMDGPUCodeGen
   R600OptimizeVectorRegisters.cpp
   R600Packetizer.cpp
   R600RegisterInfo.cpp
   SIAddIMGInit.cpp
   SIAnnotateControlFlow.cpp
   SIFixSGPRCopies.cpp
   SIFixupVectorISel.cpp
   SIFixVGPRCopies.cpp
-  SIFixWWMLiveness.cpp
+  SIPreAllocateWWMRegs.cpp
   SIFoldOperands.cpp
   SIFormMemoryClauses.cpp
   SIFrameLowering.cpp
   SIInsertSkips.cpp
   SIInsertWaitcnts.cpp
   SIInstrInfo.cpp
   SIISelLowering.cpp
   SILoadStoreOptimizer.cpp

llvm/trunk/lib/Target/AMDGPU/SIFixWWMLiveness.cpp

-//===-- SIFixWWMLiveness.cpp - Fix WWM live intervals ---------===//
-//
-// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
-// See https://llvm.org/LICENSE.txt for license information.
-// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
-//
-//===----------------------------------------------------------------------===//
-//
-/// \file
-/// Computations in WWM can overwrite values in inactive channels for
-/// variables that the register allocator thinks are dead. This pass adds fake
-/// uses of those variables to their def(s) to make sure that they aren't
-/// overwritten.
-///
-/// As an example, consider this snippet:
-/// %vgpr0 = V_MOV_B32_e32 0.0
-/// if (...) {
-///   %vgpr1 = ...
-///   %vgpr2 = WWM killed %vgpr1
-///   ... = killed %vgpr2
-///   %vgpr0 = V_MOV_B32_e32 1.0
-/// }
-/// ... = %vgpr0
-///
-/// The live intervals of %vgpr0 don't overlap with those of %vgpr1. Normally,
-/// we can safely allocate %vgpr0 and %vgpr1 in the same register, since
-/// writing %vgpr1 would only write to channels that would be clobbered by the
-/// second write to %vgpr0 anyways. But if %vgpr1 is written with WWM enabled,
-/// it would clobber even the inactive channels for which the if-condition is
-/// false, for which %vgpr0 is supposed to be 0. This pass adds an implicit use
-/// of %vgpr0 to its def to make sure they aren't allocated to the
-/// same register.
-///
-/// In general, we need to figure out what registers might have their inactive
-/// channels which are eventually used accidentally clobbered by a WWM
-/// instruction. We do that by spotting three separate cases of registers:
-///
-/// 1. A "then phi": the value resulting from phi elimination of a phi node at
-///    the end of an if..endif. If there is WWM code in the "then", then we
-///    make the def at the end of the "then" branch a partial def by adding an
-///    implicit use of the register.
-///
-/// 2. A "loop exit register": a value written inside a loop but used outside the
-///    loop, where there is WWM code inside the loop (the case in the example
-///    above). We add an implicit_def of the register in the loop pre-header,
-///    and make the original def a partial def by adding an implicit use of the
-///    register.
-///
-/// 3. A "loop exit phi": the value resulting from phi elimination of a phi node
-///    in a loop header. If there is WWM code inside the loop, then we make all
-///    defs inside the loop partial defs by adding an implicit use of the
-///    register on each one.
-///
-/// Note that we do not need to consider an if..else..endif phi. We only need to
-/// consider non-uniform control flow, and control flow structurization would
-/// have transformed a non-uniform if..else..endif into two if..endifs.
-///
-/// The analysis to detect these cases relies on a property of the MIR
-/// arising from this pass running straight after PHIElimination and before any
-/// coalescing: that any virtual register with more than one definition must be
-/// the new register added to lower a phi node by PHIElimination.
-///
-/// FIXME: We should detect whether a register in one of the above categories is
-/// already live at the WWM code before deciding to add the implicit uses to
-/// synthesize its liveness.
-///
-/// FIXME: I believe this whole scheme may be flawed due to the possibility of
-/// the register allocator doing live interval splitting.
-///
-//===----------------------------------------------------------------------===//
-
-#include "AMDGPU.h"
-#include "AMDGPUSubtarget.h"
-#include "SIInstrInfo.h"
-#include "SIRegisterInfo.h"
-#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
-#include "llvm/ADT/DepthFirstIterator.h"
-#include "llvm/ADT/SparseBitVector.h"
-#include "llvm/CodeGen/LiveIntervals.h"
-#include "llvm/CodeGen/MachineDominators.h"
-#include "llvm/CodeGen/MachineFunctionPass.h"
-#include "llvm/CodeGen/MachineLoopInfo.h"
-#include "llvm/CodeGen/Passes.h"
-#include "llvm/CodeGen/TargetRegisterInfo.h"
-
-using namespace llvm;
-
-#define DEBUG_TYPE "si-fix-wwm-liveness"
-
-namespace {
-
-class SIFixWWMLiveness : public MachineFunctionPass {
-private:
-  MachineDominatorTree *DomTree;
-  MachineLoopInfo *LoopInfo;
-  LiveIntervals *LIS = nullptr;
-  const SIInstrInfo *TII;
-  const SIRegisterInfo *TRI;
-  MachineRegisterInfo *MRI;
-
-  std::vector<MachineInstr *> WWMs;
-  std::vector<MachineOperand *> ThenDefs;
-  std::vector<std::pair<MachineOperand *, MachineLoop *>> LoopExitDefs;
-  std::vector<std::pair<MachineOperand *, MachineLoop *>> LoopPhiDefs;
-
-public:
-  static char ID;
-
-  SIFixWWMLiveness() : MachineFunctionPass(ID) {
-    initializeSIFixWWMLivenessPass(*PassRegistry::getPassRegistry());
-  }
-
-  bool runOnMachineFunction(MachineFunction &MF) override;
-
-  StringRef getPassName() const override { return "SI Fix WWM Liveness"; }
-
-  void getAnalysisUsage(AnalysisUsage &AU) const override {
-    AU.addRequiredID(MachineDominatorsID);
-    AU.addRequiredID(MachineLoopInfoID);
-    // Should preserve the same set that TwoAddressInstructions does.
-    AU.addPreserved<SlotIndexes>();
-    AU.addPreserved<LiveIntervals>();
-    AU.addPreservedID(LiveVariablesID);
-    AU.addPreservedID(MachineLoopInfoID);
-    AU.addPreservedID(MachineDominatorsID);
-    AU.setPreservesCFG();
-    MachineFunctionPass::getAnalysisUsage(AU);
-  }
-
-private:
-  void processDef(MachineOperand &DefOpnd);
-  bool processThenDef(MachineOperand *DefOpnd);
-  bool processLoopExitDef(MachineOperand *DefOpnd, MachineLoop *Loop);
-  bool processLoopPhiDef(MachineOperand *DefOpnd, MachineLoop *Loop);
-};
-
-} // End anonymous namespace.
-
-INITIALIZE_PASS_BEGIN(SIFixWWMLiveness, DEBUG_TYPE,
-                "SI fix WWM liveness", false, false)
-INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
-INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
-INITIALIZE_PASS_END(SIFixWWMLiveness, DEBUG_TYPE,
-                "SI fix WWM liveness", false, false)
-
-char SIFixWWMLiveness::ID = 0;
-
-char &llvm::SIFixWWMLivenessID = SIFixWWMLiveness::ID;
-
-FunctionPass *llvm::createSIFixWWMLivenessPass() {
-  return new SIFixWWMLiveness();
-}
-
-bool SIFixWWMLiveness::runOnMachineFunction(MachineFunction &MF) {
-  LLVM_DEBUG(dbgs() << "SIFixWWMLiveness: function " << MF.getName() << "\n");
-  bool Modified = false;
-
-  // This doesn't actually need LiveIntervals, but we can preserve them.
-  LIS = getAnalysisIfAvailable<LiveIntervals>();
-
-  const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
-
-  TII = ST.getInstrInfo();
-  TRI = &TII->getRegisterInfo();
-  MRI = &MF.getRegInfo();
-
-  DomTree = &getAnalysis<MachineDominatorTree>();
-  LoopInfo = &getAnalysis<MachineLoopInfo>();
-
-  // Scan the function to find the WWM sections and the candidate registers for
-  // having liveness modified.
-  for (MachineBasicBlock &MBB : MF) {
-    for (MachineInstr &MI : MBB) {
-      if (MI.getOpcode() == AMDGPU::EXIT_WWM)
-        WWMs.push_back(&MI);
-      else {
-        for (MachineOperand &DefOpnd : MI.defs()) {
-          if (DefOpnd.isReg()) {
-            unsigned Reg = DefOpnd.getReg();
-            if (TRI->isVGPR(*MRI, Reg))
-              processDef(DefOpnd);
-          }
-        }
-      }
-    }
-  }
-  if (!WWMs.empty()) {
-    // Synthesize liveness over WWM sections as required.
-    for (auto ThenDef : ThenDefs)
-      Modified |= processThenDef(ThenDef);
-    for (auto LoopExitDef : LoopExitDefs)
-      Modified |= processLoopExitDef(LoopExitDef.first, LoopExitDef.second);
-    for (auto LoopPhiDef : LoopPhiDefs)
-      Modified |= processLoopPhiDef(LoopPhiDef.first, LoopPhiDef.second);
-  }
-
-  WWMs.clear();
-  ThenDefs.clear();
-  LoopExitDefs.clear();
-  LoopPhiDefs.clear();
-
-  return Modified;
-}
-
-// During the function scan, process an operand that defines a VGPR.
-// This categorizes the register and puts it in the appropriate list for later
-// use when processing a WWM section.
-void SIFixWWMLiveness::processDef(MachineOperand &DefOpnd) {
-  unsigned Reg = DefOpnd.getReg();
-  // Get all the defining instructions. For convenience, make Defs[0] the def
-  // we are on now.
-  SmallVector<const MachineInstr *, 4> Defs;
-  Defs.push_back(DefOpnd.getParent());
-  for (auto &MI : MRI->def_instructions(Reg)) {
-    if (&MI != DefOpnd.getParent())
-      Defs.push_back(&MI);
-  }
-  // Check whether this def dominates all the others. If not, ignore this def.
-  // Either it is going to be processed when the scan encounters its other def
-  // that dominates all defs, or there is no def that dominates all others.
-  // The latter case is an eliminated phi from an if..else..endif or similar,
-  // which must be for uniform control flow so can be ignored.
-  // Because this pass runs shortly after PHIElimination, we assume that any
-  // multi-def register is a lowered phi, and thus has each def in a separate
-  // basic block.
-  for (unsigned I = 1; I != Defs.size(); ++I) {
-    if (!DomTree->dominates(Defs[0]->getParent(), Defs[I]->getParent()))
-      return;
-  }
-  // Check for the case of an if..endif lowered phi: It has two defs, one
-  // dominates the other, and there is a single use in a successor of the
-  // dominant def.
-  // Later we will spot any WWM code inside
-  // the "then" clause and turn the second def into a partial def so its
-  // liveness goes through the WWM code in the "then" clause.
-  if (Defs.size() == 2) {
-    auto DomDefBlock = Defs[0]->getParent();
-    if (DomDefBlock->succ_size() == 2 && MRI->hasOneUse(Reg)) {
-      auto UseBlock = MRI->use_begin(Reg)->getParent()->getParent();
-      for (auto Succ : DomDefBlock->successors()) {
-        if (Succ == UseBlock) {
-          LLVM_DEBUG(dbgs() << printReg(Reg, TRI) << " is a then phi reg\n");
-          ThenDefs.push_back(&DefOpnd);
-          return;
-        }
-      }
-    }
-  }
-  // Check for the case of a non-lowered-phi register (single def) that exits
-  // a loop, that is, it has a use that is outside a loop that the def is
-  // inside. We find the outermost loop that the def is inside but a use is
-  // outside. Later we will spot any WWM code inside that loop and then make
-  // the def a partial def so its liveness goes round the loop and through the
-  // WWM code.
-  if (Defs.size() == 1) {
-    auto Loop = LoopInfo->getLoopFor(Defs[0]->getParent());
-    if (!Loop)
-      return;
-    bool IsLoopExit = false;
-    for (auto &Use : MRI->use_instructions(Reg)) {
-      auto UseBlock = Use.getParent();
-      if (Loop->contains(UseBlock))
-        continue;
-      IsLoopExit = true;
-      while (auto Parent = Loop->getParentLoop()) {
-        if (Parent->contains(UseBlock))
-          break;
-        Loop = Parent;
-      }
-    }
-    if (!IsLoopExit)
-      return;
-    LLVM_DEBUG(dbgs() << printReg(Reg, TRI)
-        << " is a loop exit reg with loop header at "
-        << "bb." << Loop->getHeader()->getNumber() << "\n");
-    LoopExitDefs.push_back(std::pair<MachineOperand *, MachineLoop *>(
-            &DefOpnd, Loop));
-    return;
-  }
-  // Check for the case of a lowered single-preheader-loop phi, that is, a
-  // multi-def register where the dominating def is in the loop pre-header and
-  // all other defs are in backedges. Later we will spot any WWM code inside
-  // that loop and then make the backedge defs partial defs so the liveness
-  // goes through the WWM code.
-  // Note that we are ignoring multi-preheader loops on the basis that the
-  // structurizer does not allow that for non-uniform loops.
-  // There must be a single use in the loop header.
-  if (!MRI->hasOneUse(Reg))
-    return;
-  auto UseBlock = MRI->use_begin(Reg)->getParent()->getParent();
-  auto Loop = LoopInfo->getLoopFor(UseBlock);
-  if (!Loop || Loop->getHeader() != UseBlock
-      || Loop->contains(Defs[0]->getParent())) {
-    LLVM_DEBUG(dbgs() << printReg(Reg, TRI)
-        << " is multi-def but single use not in loop header\n");
-    return;
-  }
-  for (unsigned I = 1; I != Defs.size(); ++I) {
-    if (!Loop->contains(Defs[I]->getParent()))
-      return;
-  }
-  LLVM_DEBUG(dbgs() << printReg(Reg, TRI)
-      << " is a loop phi reg with loop header at "
-      << "bb." << Loop->getHeader()->getNumber() << "\n");
-  LoopPhiDefs.push_back(
-      std::pair<MachineOperand *, MachineLoop *>(&DefOpnd, Loop));
-}
-
-// Process a then phi def: It has two defs, one dominates the other, and there
-// is a single use in a successor of the dominant def. Here we spot any WWM
-// code inside the "then" clause and turn the second def into a partial def so
-// its liveness goes through the WWM code in the "then" clause.
-bool SIFixWWMLiveness::processThenDef(MachineOperand *DefOpnd) {
-  LLVM_DEBUG(dbgs() << "Processing then def: " << *DefOpnd->getParent());
-  if (DefOpnd->getParent()->getOpcode() == TargetOpcode::IMPLICIT_DEF) {
-    // Ignore if dominating def is undef.
-    LLVM_DEBUG(dbgs() << "  ignoring as dominating def is undef\n");
-    return false;
-  }
-  unsigned Reg = DefOpnd->getReg();
-  // Get the use block, which is the endif block.
-  auto UseBlock = MRI->use_instr_begin(Reg)->getParent();
-  // Check whether there is WWM code inside the then branch. The WWM code must
-  // be dominated by the if but not dominated by the endif.
-  bool ContainsWWM = false;
-  for (auto WWM : WWMs) {
-    if (DomTree->dominates(DefOpnd->getParent()->getParent(), WWM->getParent())
-        && !DomTree->dominates(UseBlock, WWM->getParent())) {
-      LLVM_DEBUG(dbgs() << "  contains WWM: " << *WWM);
-      ContainsWWM = true;
-      break;
-    }
-  }
-  if (!ContainsWWM)
-    return false;
-  // Get the other def.
-  MachineInstr *OtherDef = nullptr;
-  for (auto &MI : MRI->def_instructions(Reg)) {
-    if (&MI != DefOpnd->getParent())
-      OtherDef = &MI;
-  }
-  // Make it a partial def.
-  OtherDef->addOperand(MachineOperand::CreateReg(Reg, false, /*isImp=*/true));
-  LLVM_DEBUG(dbgs() << *OtherDef);
-  return true;
-}
-
-// Process a loop exit def, that is, a register with a single use in a loop
-// that has a use outside the loop.  Here we spot any WWM code inside that loop
-// and then make the def a partial def so its liveness goes round the loop and
-// through the WWM code.
-bool SIFixWWMLiveness::processLoopExitDef(MachineOperand *DefOpnd,
-      MachineLoop *Loop) {
-  LLVM_DEBUG(dbgs() << "Processing loop exit def: " << *DefOpnd->getParent());
-  // Check whether there is WWM code inside the loop.
-  bool ContainsWWM = false;
-  for (auto WWM : WWMs) {
-    if (Loop->contains(WWM->getParent())) {
-      LLVM_DEBUG(dbgs() << "  contains WWM: " << *WWM);
-      ContainsWWM = true;
-      break;
-    }
-  }
-  if (!ContainsWWM)
-    return false;
-  unsigned Reg = DefOpnd->getReg();
-  // Add a new implicit_def in loop preheader(s).
-  for (auto Pred : Loop->getHeader()->predecessors()) {
-    if (!Loop->contains(Pred)) {
-      auto ImplicitDef = BuildMI(*Pred, Pred->getFirstTerminator(), DebugLoc(),
-          TII->get(TargetOpcode::IMPLICIT_DEF), Reg);
-      LLVM_DEBUG(dbgs() << *ImplicitDef);
-      (void)ImplicitDef;
-    }
-  }
-  // Make the original def partial.
-  DefOpnd->getParent()->addOperand(MachineOperand::CreateReg(
-          Reg, false, /*isImp=*/true));
-  LLVM_DEBUG(dbgs() << *DefOpnd->getParent());
-  return true;
-}
-
-// Process a loop phi def, that is, a multi-def register where the dominating
-// def is in the loop pre-header and all other defs are in backedges. Here we
-// spot any WWM code inside that loop and then make the backedge defs partial
-// defs so the liveness goes through the WWM code.
-bool SIFixWWMLiveness::processLoopPhiDef(MachineOperand *DefOpnd,
-      MachineLoop *Loop) {
-  LLVM_DEBUG(dbgs() << "Processing loop phi def: " << *DefOpnd->getParent());
-  // Check whether there is WWM code inside the loop.
-  bool ContainsWWM = false;
-  for (auto WWM : WWMs) {
-    if (Loop->contains(WWM->getParent())) {
-      LLVM_DEBUG(dbgs() << "  contains WWM: " << *WWM);
-      ContainsWWM = true;
-      break;
-    }
-  }
-  if (!ContainsWWM)
-    return false;
-  unsigned Reg = DefOpnd->getReg();
-  // Remove kill mark from uses.
-  for (auto &Use : MRI->use_operands(Reg))
-    Use.setIsKill(false);
-  // Make all defs except the dominating one partial defs.
-  SmallVector<MachineInstr *, 4> Defs;
-  for (auto &Def : MRI->def_instructions(Reg))
-    Defs.push_back(&Def);
-  for (auto Def : Defs) {
-    if (DefOpnd->getParent() == Def)
-      continue;
-    Def->addOperand(MachineOperand::CreateReg(Reg, false, /*isImp=*/true));
-    LLVM_DEBUG(dbgs() << *Def);
-  }
-  return true;
-}
-

llvm/trunk/lib/Target/AMDGPU/SIInstrInfo.cpp

bool SIInstrInfo::shouldClusterMemOps(MachineOperand &BaseOp1,
   // same number of bytes as FirstLdSt.
 
   // The unit of this value is bytes.
   // FIXME: This needs finer tuning.
   unsigned LoadClusterThreshold = 16;
 
   const MachineRegisterInfo &MRI =
       FirstLdSt.getParent()->getParent()->getRegInfo();
-  const TargetRegisterClass *DstRC = MRI.getRegClass(FirstDst->getReg());
+
+  const unsigned Reg = FirstDst->getReg();
+
+  const TargetRegisterClass *DstRC = TargetRegisterInfo::isVirtualRegister(Reg)
+                                         ? MRI.getRegClass(Reg)
+                                         : RI.getPhysRegClass(Reg);
 
   return (NumLoads * (RI.getRegSizeInBits(*DstRC) / 8)) <= LoadClusterThreshold;
 }
 
 // FIXME: This behaves strangely. If, for example, you have 32 load + stores,
 // the first 16 loads will be interleaved with the stores, and the next 16 will
 // be clustered as expected. It should really split into 2 16 store batches.
 //
else
       MIB.add(MI.getOperand(2));
 
     Bundler.append(MIB);
     finalizeBundle(MBB, Bundler.begin());
 
     MI.eraseFromParent();
     break;
   }
+  case AMDGPU::ENTER_WWM: {
+    // This only gets its own opcode so that SIPreAllocateWWMRegs can tell when
+    // WWM is entered.
+    MI.setDesc(get(AMDGPU::S_OR_SAVEEXEC_B64));
+    break;
+  }
   case AMDGPU::EXIT_WWM: {
-    // This only gets its own opcode so that SIFixWWMLiveness can tell when WWM
-    // is exited.
+    // This only gets its own opcode so that SIPreAllocateWWMRegs can tell when
+    // WWM is exited.
     MI.setDesc(get(AMDGPU::S_MOV_B64));
     break;
   }
   case TargetOpcode::BUNDLE: {
     if (!MI.mayLoad())
       return false;
 
     // If it is a load it must be a memory clause

llvm/trunk/lib/Target/AMDGPU/SIInstructions.td

 // the instruction that defines $src0 (which is run in WWM) doesn't
 // accidentally clobber inactive channels of $vdst.
 let Constraints = "@earlyclobber $vdst" in {
 def WWM : PseudoInstSI <(outs unknown:$vdst), (ins unknown:$src0)>;
 }
 
 } // End let hasSideEffects = 0, mayLoad = 0, mayStore = 0, Uses = [EXEC]
 
+def ENTER_WWM : SPseudoInstSI <(outs SReg_64:$sdst), (ins i64imm:$src0)> {
+  let Defs = [EXEC];
+  let hasSideEffects = 0;
+  let mayLoad = 0;
+  let mayStore = 0;
+}
+
 def EXIT_WWM : SPseudoInstSI <(outs SReg_64:$sdst), (ins SReg_64:$src0)> {
   let hasSideEffects = 0;
   let mayLoad = 0;
   let mayStore = 0;
 }
 
 // Invert the exec mask and overwrite the inactive lanes of dst with inactive,
 // restoring it after we're done.

llvm/trunk/lib/Target/AMDGPU/SIMachineFunctionInfo.h

 #include "AMDGPUMachineFunction.h"
 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
 #include "SIInstrInfo.h"
 #include "SIRegisterInfo.h"
 #include "llvm/ADT/ArrayRef.h"
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/ADT/Optional.h"
 #include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/SparseBitVector.h"
 #include "llvm/CodeGen/MIRYamlMapping.h"
 #include "llvm/CodeGen/PseudoSourceValue.h"
 #include "llvm/CodeGen/TargetInstrInfo.h"
 #include "llvm/MC/MCRegisterInfo.h"
 #include "llvm/Support/ErrorHandling.h"
 #include <array>
 #include <cassert>
 #include <utility>
struct SGPRSpillVGPRCSR {
 
     // If the VGPR is a CSR, the stack slot used to save/restore it in the
     // prolog/epilog.
     Optional<int> FI;
 
     SGPRSpillVGPRCSR(unsigned V, Optional<int> F) : VGPR(V), FI(F) {}
   };
 
+  SparseBitVector<> WWMReservedRegs;
+
+  void ReserveWWMRegister(unsigned reg) { WWMReservedRegs.set(reg); }
+
 private:
   // SGPR->VGPR spilling support.
   using SpillRegMask = std::pair<unsigned, unsigned>;
 
   // Track VGPR + wave index for each subregister of the SGPR spilled to
   // frameindex key.
   DenseMap<int, std::vector<SpilledReg>> SGPRToVGPRSpills;
   unsigned NumVGPRSpillLanes = 0;

llvm/trunk/lib/Target/AMDGPU/SIPreAllocateWWMRegs.cpp

+//===- SIPreAllocateWWMRegs.cpp - WWM Register Pre-allocation -------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// Pass to pre-allocated WWM registers
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "AMDGPUSubtarget.h"
+#include "SIInstrInfo.h"
+#include "SIRegisterInfo.h"
+#include "SIMachineFunctionInfo.h"
+#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
+#include "llvm/ADT/PostOrderIterator.h"
+#include "llvm/CodeGen/VirtRegMap.h"
+#include "llvm/CodeGen/LiveInterval.h"
+#include "llvm/CodeGen/LiveIntervals.h"
+#include "llvm/CodeGen/LiveRegMatrix.h"
+#include "llvm/CodeGen/MachineDominators.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/RegisterClassInfo.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "si-pre-allocate-wwm-regs"
+
+namespace {
+
+class SIPreAllocateWWMRegs : public MachineFunctionPass {
+private:
+  const SIInstrInfo *TII;
+  const SIRegisterInfo *TRI;
+  MachineRegisterInfo *MRI;
+  LiveIntervals *LIS;
+  LiveRegMatrix *Matrix;
+  VirtRegMap *VRM;
+  RegisterClassInfo RegClassInfo;
+
+  std::vector<unsigned> RegsToRewrite;
+
+public:
+  static char ID;
+
+  SIPreAllocateWWMRegs() : MachineFunctionPass(ID) {
+    initializeSIPreAllocateWWMRegsPass(*PassRegistry::getPassRegistry());
+  }
+
+  bool runOnMachineFunction(MachineFunction &MF) override;
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.addRequired<LiveIntervals>();
+    AU.addPreserved<LiveIntervals>();
+    AU.addRequired<VirtRegMap>();
+    AU.addRequired<LiveRegMatrix>();
+    AU.addPreserved<SlotIndexes>();
+    AU.setPreservesCFG();
+    MachineFunctionPass::getAnalysisUsage(AU);
+  }
+
+private:
+  bool processDef(MachineOperand &MO);
+  void rewriteRegs(MachineFunction &MF);
+};
+
+} // End anonymous namespace.
+
+INITIALIZE_PASS_BEGIN(SIPreAllocateWWMRegs, DEBUG_TYPE,
+                "SI Pre-allocate WWM Registers", false, false)
+INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
+INITIALIZE_PASS_DEPENDENCY(VirtRegMap)
+INITIALIZE_PASS_DEPENDENCY(LiveRegMatrix)
+INITIALIZE_PASS_END(SIPreAllocateWWMRegs, DEBUG_TYPE,
+                "SI Pre-allocate WWM Registers", false, false)
+
+char SIPreAllocateWWMRegs::ID = 0;
+
+char &llvm::SIPreAllocateWWMRegsID = SIPreAllocateWWMRegs::ID;
+
+FunctionPass *llvm::createSIPreAllocateWWMRegsPass() {
+  return new SIPreAllocateWWMRegs();
+}
+
+bool SIPreAllocateWWMRegs::processDef(MachineOperand &MO) {
+  if (!MO.isReg())
+    return false;
+
+  unsigned Reg = MO.getReg();
+
+  if (!TRI->isVGPR(*MRI, Reg))
+    return false;
+
+  if (TRI->isPhysicalRegister(Reg))
+    return false;
+
+  if (VRM->hasPhys(Reg))
+    return false;
+
+  LiveInterval &LI = LIS->getInterval(Reg);
+
+  for (unsigned PhysReg : RegClassInfo.getOrder(MRI->getRegClass(Reg))) {
+    if (!MRI->isPhysRegUsed(PhysReg) &&
+        Matrix->checkInterference(LI, PhysReg) == LiveRegMatrix::IK_Free) {
+      Matrix->assign(LI, PhysReg);
+      assert(PhysReg != 0);
+      RegsToRewrite.push_back(Reg);
+      return true;
+    }
+  }
+
+  llvm_unreachable("physreg not found for WWM expression");
+  return false;
+}
+
+void SIPreAllocateWWMRegs::rewriteRegs(MachineFunction &MF) {
+  for (MachineBasicBlock &MBB : MF) {
+    for (MachineInstr &MI : MBB) {
+      for (MachineOperand &MO : MI.operands()) {
+        if (!MO.isReg())
+          continue;
+
+        const unsigned VirtReg = MO.getReg();
+        if (TRI->isPhysicalRegister(VirtReg))
+          continue;
+
+        if (!VRM->hasPhys(VirtReg))
+          continue;
+
+        unsigned PhysReg = VRM->getPhys(VirtReg);
+        const unsigned SubReg = MO.getSubReg();
+        if (SubReg != 0) {
+          PhysReg = TRI->getSubReg(PhysReg, SubReg);
+          MO.setSubReg(0);
+        }
+
+        MO.setReg(PhysReg);
+        MO.setIsRenamable(false);
+      }
+    }
+  }
+
+  SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
+
+  for (unsigned Reg : RegsToRewrite) {
+    LIS->removeInterval(Reg);
+
+    const unsigned PhysReg = VRM->getPhys(Reg);
+    assert(PhysReg != 0);
+    MFI->ReserveWWMRegister(PhysReg);
+  }
+
+  RegsToRewrite.clear();
+
+  // Update the set of reserved registers to include WWM ones.
+  MRI->freezeReservedRegs(MF);
+}
+
+bool SIPreAllocateWWMRegs::runOnMachineFunction(MachineFunction &MF) {
+  LLVM_DEBUG(dbgs() << "SIPreAllocateWWMRegs: function " << MF.getName() << "\n");
+
+  const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
+
+  TII = ST.getInstrInfo();
+  TRI = &TII->getRegisterInfo();
+  MRI = &MF.getRegInfo();
+
+  LIS = &getAnalysis<LiveIntervals>();
+  Matrix = &getAnalysis<LiveRegMatrix>();
+  VRM = &getAnalysis<VirtRegMap>();
+
+  RegClassInfo.runOnMachineFunction(MF);
+
+  bool RegsAssigned = false;
+
+  // We use a reverse post-order traversal of the control-flow graph to
+  // guarantee that we visit definitions in dominance order. Since WWM
+  // expressions are guaranteed to never involve phi nodes, and we can only
+  // escape WWM through the special WWM instruction, this means that this is a
+  // perfect elimination order, so we can never do any better.
+  ReversePostOrderTraversal<MachineFunction*> RPOT(&MF);
+
+  for (MachineBasicBlock *MBB : RPOT) {
+    bool InWWM = false;
+    for (MachineInstr &MI : *MBB) {
+      if (MI.getOpcode() == AMDGPU::V_SET_INACTIVE_B32 ||
+          MI.getOpcode() == AMDGPU::V_SET_INACTIVE_B64)
+        RegsAssigned |= processDef(MI.getOperand(0));
+
+      if (MI.getOpcode() == AMDGPU::ENTER_WWM) {
+        LLVM_DEBUG(dbgs() << "entering WWM region: " << MI << "\n");
+        InWWM = true;
+        continue;
+      }
+
+      if (MI.getOpcode() == AMDGPU::EXIT_WWM) {
+        LLVM_DEBUG(dbgs() << "exiting WWM region: " << MI << "\n");
+        InWWM = false;
+      }
+
+      if (!InWWM)
+        continue;
+
+      LLVM_DEBUG(dbgs() << "processing " << MI << "\n");
+
+      for (MachineOperand &DefOpnd : MI.defs()) {
+        RegsAssigned |= processDef(DefOpnd);
+      }
+    }
+  }
+
+  if (!RegsAssigned)
+    return false;
+
+  rewriteRegs(MF);
+  return true;
+}

llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.cpp

BitVector SIRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
   }
 
   unsigned FrameReg = MFI->getFrameOffsetReg();
   if (FrameReg != AMDGPU::NoRegister) {
     reserveRegisterTuples(Reserved, FrameReg);
     assert(!isSubRegister(ScratchRSrcReg, FrameReg));
   }
 
+  for (unsigned Reg : MFI->WWMReservedRegs) {
+    reserveRegisterTuples(Reserved, Reg);
+  }
+
   return Reserved;
 }
 
 bool SIRegisterInfo::requiresRegisterScavenging(const MachineFunction &Fn) const {
   const SIMachineFunctionInfo *Info = Fn.getInfo<SIMachineFunctionInfo>();
   if (Info->isEntryFunction()) {
     const MachineFrameInfo &MFI = Fn.getFrameInfo();
     return MFI.hasStackObjects() || MFI.hasCalls();

llvm/trunk/lib/Target/AMDGPU/SIWholeQuadMode.cpp

 }
 
 void SIWholeQuadMode::toWWM(MachineBasicBlock &MBB,
                             MachineBasicBlock::iterator Before,
                             unsigned SaveOrig) {
   MachineInstr *MI;
 
   assert(SaveOrig);
-  MI = BuildMI(MBB, Before, DebugLoc(), TII->get(AMDGPU::S_OR_SAVEEXEC_B64),
-               SaveOrig)
+  MI = BuildMI(MBB, Before, DebugLoc(), TII->get(AMDGPU::ENTER_WWM), SaveOrig)
            .addImm(-1);
   LIS->InsertMachineInstrInMaps(*MI);
 }
 
 void SIWholeQuadMode::fromWWM(MachineBasicBlock &MBB,
                               MachineBasicBlock::iterator Before,
                               unsigned SavedOrig) {
   MachineInstr *MI;
for (MachineInstr *MI : LiveMaskQueries) {
     MI->eraseFromParent();
   }
 }
 
 void SIWholeQuadMode::lowerCopyInstrs() {
   for (MachineInstr *MI : LowerToCopyInstrs) {
     for (unsigned i = MI->getNumExplicitOperands() - 1; i > 1; i--)
       MI->RemoveOperand(i);
+
+    const unsigned Reg = MI->getOperand(0).getReg();
+
+    if (TRI->isVGPR(*MRI, Reg)) {
+      const TargetRegisterClass *regClass =
+          TargetRegisterInfo::isVirtualRegister(Reg)
+              ? MRI->getRegClass(Reg)
+              : TRI->getPhysRegClass(Reg);
+
+      const unsigned MovOp = TII->getMovOpcode(regClass);
+      MI->setDesc(TII->get(MovOp));
+
+      // And make it implicitly depend on exec (like all VALU movs should do).
+      MI->addOperand(MachineOperand::CreateReg(AMDGPU::EXEC, false, true));
+    } else {
       MI->setDesc(TII->get(AMDGPU::COPY));
     }
   }
+}
 
 bool SIWholeQuadMode::runOnMachineFunction(MachineFunction &MF) {
   Instructions.clear();
   Blocks.clear();
   LiveMaskQueries.clear();
   LowerToCopyInstrs.clear();
   CallingConv = MF.getFunction().getCallingConv();
 

llvm/trunk/test/CodeGen/AMDGPU/atomic_optimizations_buffer.ll

entry:
   ret void
 }
 
 ; GCN-LABEL: sub_i32_varying_vdata:
 ; GFX7LESS-NOT: v_mbcnt_lo_u32_b32
 ; GFX7LESS-NOT: v_mbcnt_hi_u32_b32
 ; GFX7LESS-NOT: s_bcnt1_i32_b64
 ; GFX7LESS: buffer_atomic_sub v{{[0-9]+}}
-; GFX8MORE: v_mov_b32_dpp v[[wave_shr1:[0-9]+]], v[[sub_value:[0-9]+]] wave_shr:1 row_mask:0xf bank_mask:0xf
+; GFX8MORE: v_mov_b32_dpp v[[wave_shr1:[0-9]+]], v{{[0-9]+}} wave_shr:1 row_mask:0xf bank_mask:0xf
 ; GFX8MORE: v_mov_b32_dpp v{{[0-9]+}}, v[[wave_shr1]] row_shr:1 row_mask:0xf bank_mask:0xf
 ; GFX8MORE: v_mov_b32_dpp v{{[0-9]+}}, v[[wave_shr1]] row_shr:2 row_mask:0xf bank_mask:0xf
 ; GFX8MORE: v_mov_b32_dpp v{{[0-9]+}}, v[[wave_shr1]] row_shr:3 row_mask:0xf bank_mask:0xf
 ; GFX8MORE: v_mov_b32_dpp v{{[0-9]+}}, v{{[0-9]+}} row_shr:4 row_mask:0xf bank_mask:0xe
 ; GFX8MORE: v_mov_b32_dpp v{{[0-9]+}}, v{{[0-9]+}} row_shr:8 row_mask:0xf bank_mask:0xc
 ; GFX8MORE: v_mov_b32_dpp v{{[0-9]+}}, v{{[0-9]+}} row_bcast:15 row_mask:0xa bank_mask:0xf
 ; GFX8MORE: v_mov_b32_dpp v{{[0-9]+}}, v{{[0-9]+}} row_bcast:31 row_mask:0xc bank_mask:0xf
-; GFX8MORE: v_sub_u32_e32 v[[sub_value]],{{( vcc,)?}} v[[sub_value]], v{{[0-9]+}}
-; GFX8MORE: v_readlane_b32 s[[scalar_value:[0-9]+]], v[[sub_value]], 63
+; GFX8MORE: v_readlane_b32 s[[scalar_value:[0-9]+]], v{{[0-9]+}}, 63
 ; GFX8MORE: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[scalar_value]]
 ; GFX8MORE: buffer_atomic_sub v[[value]]
 define amdgpu_kernel void @sub_i32_varying_vdata(i32 addrspace(1)* %out, <4 x i32> %inout) {
 entry:
   %lane = call i32 @llvm.amdgcn.workitem.id.x()
   %old = call i32 @llvm.amdgcn.buffer.atomic.sub(i32 %lane, <4 x i32> %inout, i32 0, i32 0, i1 0)
   store i32 %old, i32 addrspace(1)* %out
   ret void

llvm/trunk/test/CodeGen/AMDGPU/atomic_optimizations_global_pointer.ll

 ; RUN: llc -march=amdgcn -mtriple=amdgcn---amdgiz -amdgpu-atomic-optimizations=true -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX7LESS %s
-; RUN: llc -march=amdgcn -mtriple=amdgcn---amdgiz -mcpu=tonga -mattr=-flat-for-global -amdgpu-atomic-optimizations=true -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX8MORE %s
-; RUN: llc -march=amdgcn -mtriple=amdgcn---amdgiz -mcpu=gfx900 -mattr=-flat-for-global -amdgpu-atomic-optimizations=true -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX8MORE %s
+; RUN: llc -march=amdgcn -mtriple=amdgcn---amdgiz -mcpu=tonga -mattr=-flat-for-global -amdgpu-atomic-optimizations=true -amdgpu-dpp-combine=false -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX8MORE %s
+; RUN: llc -march=amdgcn -mtriple=amdgcn---amdgiz -mcpu=gfx900 -mattr=-flat-for-global -amdgpu-atomic-optimizations=true -amdgpu-dpp-combine=false -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX8MORE %s
 
 declare i32 @llvm.amdgcn.workitem.id.x()
 
 ; Show that what the atomic optimization pass will do for global pointers.
 
 ; GCN-LABEL: add_i32_constant:
 ; GCN: v_cmp_ne_u32_e64 s{{\[}}[[exec_lo:[0-9]+]]:[[exec_hi:[0-9]+]]{{\]}}, 1, 0
 ; GCN: v_mbcnt_lo_u32_b32{{(_e[0-9]+)?}} v[[mbcnt_lo:[0-9]+]], s[[exec_lo]], 0
entry:
   ret void
 }
 
 ; GCN-LABEL: sub_i32_varying:
 ; GFX7LESS-NOT: v_mbcnt_lo_u32_b32
 ; GFX7LESS-NOT: v_mbcnt_hi_u32_b32
 ; GFX7LESS-NOT: s_bcnt1_i32_b64
 ; GFX7LESS: buffer_atomic_sub v{{[0-9]+}}
-; GFX8MORE: v_mov_b32_dpp v{{[0-9]+}}, v[[sub_value:[0-9]+]] wave_shr:1 row_mask:0xf bank_mask:0xf
-; GFX8MORE: v_sub_u32_e32 v[[sub_value]],{{( vcc,)?}} v[[sub_value]], v{{[0-9]+}}
-; GFX8MORE: v_readlane_b32 s[[scalar_value:[0-9]+]], v[[sub_value]], 63
+; GFX8MORE: v_readlane_b32 s[[scalar_value:[0-9]+]], v{{[0-9]+}}, 63
 ; GFX8MORE: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[scalar_value]]
 ; GFX8MORE: buffer_atomic_sub v[[value]]
 define amdgpu_kernel void @sub_i32_varying(i32 addrspace(1)* %out, i32 addrspace(1)* %inout) {
 entry:
   %lane = call i32 @llvm.amdgcn.workitem.id.x()
   %old = atomicrmw sub i32 addrspace(1)* %inout, i32 %lane acq_rel
   store i32 %old, i32 addrspace(1)* %out
   ret void

llvm/trunk/test/CodeGen/AMDGPU/atomic_optimizations_local_pointer.ll

entry:
   ret void
 }
 
 ; GCN-LABEL: sub_i32_varying:
 ; GFX7LESS-NOT: v_mbcnt_lo_u32_b32
 ; GFX7LESS-NOT: v_mbcnt_hi_u32_b32
 ; GFX7LESS-NOT: s_bcnt1_i32_b64
 ; GFX7LESS: ds_sub_rtn_u32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
-; GFX8MORE: v_mov_b32_dpp v{{[0-9]+}}, v[[sub_value:[0-9]+]] wave_shr:1 row_mask:0xf bank_mask:0xf
-; GFX8MORE: v_sub_u32_e32 v[[sub_value]],{{( vcc,)?}} v[[sub_value]], v{{[0-9]+}}
-; GFX8MORE: v_readlane_b32 s[[scalar_value:[0-9]+]], v[[sub_value]], 63
+; GFX8MORE: v_readlane_b32 s[[scalar_value:[0-9]+]], v{{[0-9]+}}, 63
 ; GFX8MORE: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[scalar_value]]
 ; GFX8MORE: ds_sub_rtn_u32 v{{[0-9]+}}, v{{[0-9]+}}, v[[value]]
 define amdgpu_kernel void @sub_i32_varying(i32 addrspace(1)* %out) {
 entry:
   %lane = call i32 @llvm.amdgcn.workitem.id.x()
   %old = atomicrmw sub i32 addrspace(3)* @local_var32, i32 %lane acq_rel
   store i32 %old, i32 addrspace(1)* %out
   ret void

llvm/trunk/test/CodeGen/AMDGPU/atomic_optimizations_raw_buffer.ll

entry:
   ret void
 }
 
 ; GCN-LABEL: sub_i32_varying_vdata:
 ; GFX7LESS-NOT: v_mbcnt_lo_u32_b32
 ; GFX7LESS-NOT: v_mbcnt_hi_u32_b32
 ; GFX7LESS-NOT: s_bcnt1_i32_b64
 ; GFX7LESS: buffer_atomic_sub v{{[0-9]+}}
-; GFX8MORE: v_mov_b32_dpp v{{[0-9]+}}, v[[sub_value:[0-9]+]] wave_shr:1 row_mask:0xf bank_mask:0xf
-; GFX8MORE: v_sub_u32_e32 v[[sub_value]],{{( vcc,)?}} v[[sub_value]], v{{[0-9]+}}
-; GFX8MORE: v_readlane_b32 s[[scalar_value:[0-9]+]], v[[sub_value]], 63
+; GFX8MORE: v_readlane_b32 s[[scalar_value:[0-9]+]], v{{[0-9]+}}, 63
 ; GFX8MORE: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[scalar_value]]
 ; GFX8MORE: buffer_atomic_sub v[[value]]
 define amdgpu_kernel void @sub_i32_varying_vdata(i32 addrspace(1)* %out, <4 x i32> %inout) {
 entry:
   %lane = call i32 @llvm.amdgcn.workitem.id.x()
   %old = call i32 @llvm.amdgcn.raw.buffer.atomic.sub(i32 %lane, <4 x i32> %inout, i32 0, i32 0, i32 0)
   store i32 %old, i32 addrspace(1)* %out
   ret void

llvm/trunk/test/CodeGen/AMDGPU/atomic_optimizations_struct_buffer.ll

entry:
   ret void
 }
 
 ; GCN-LABEL: sub_i32_varying_vdata:
 ; GFX7LESS-NOT: v_mbcnt_lo_u32_b32
 ; GFX7LESS-NOT: v_mbcnt_hi_u32_b32
 ; GFX7LESS-NOT: s_bcnt1_i32_b64
 ; GFX7LESS: buffer_atomic_sub v{{[0-9]+}}
-; GFX8MORE: v_mov_b32_dpp v{{[0-9]+}}, v[[sub_value:[0-9]+]] wave_shr:1 row_mask:0xf bank_mask:0xf
-; GFX8MORE: v_sub_u32_e32 v[[sub_value]],{{( vcc,)?}} v[[sub_value]], v{{[0-9]+}}
-; GFX8MORE: v_readlane_b32 s[[scalar_value:[0-9]+]], v[[sub_value]], 63
+; GFX8MORE: v_readlane_b32 s[[scalar_value:[0-9]+]], v{{[0-9]+}}, 63
 ; GFX8MORE: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[scalar_value]]
 ; GFX8MORE: buffer_atomic_sub v[[value]]
 define amdgpu_kernel void @sub_i32_varying_vdata(i32 addrspace(1)* %out, <4 x i32> %inout) {
 entry:
   %lane = call i32 @llvm.amdgcn.workitem.id.x()
   %old = call i32 @llvm.amdgcn.struct.buffer.atomic.sub(i32 %lane, <4 x i32> %inout, i32 0, i32 0, i32 0, i32 0)
   store i32 %old, i32 addrspace(1)* %out
   ret void

llvm/trunk/test/CodeGen/AMDGPU/fix-wwm-liveness.mir

-# RUN: llc -march=amdgcn -verify-machineinstrs -run-pass si-fix-wwm-liveness -o -  %s | FileCheck %s
-
-# Test a then phi value.
-#CHECK: test_wwm_liveness_then_phi
-#CHECK: %21:vgpr_32 = V_MOV_B32_e32 1, implicit $exec, implicit %21
-
----
-name:            test_wwm_liveness_then_phi
-alignment:       0
-exposesReturnsTwice: false
-legalized:       false
-regBankSelected: false
-selected:        false
-tracksRegLiveness: true
-registers:
-  - { id: 0, class: sreg_64, preferred-register: '' }
-  - { id: 1, class: sgpr_32, preferred-register: '' }
-  - { id: 2, class: sgpr_32, preferred-register: '' }
-  - { id: 3, class: vgpr_32, preferred-register: '' }
-  - { id: 4, class: vgpr_32, preferred-register: '' }
-  - { id: 5, class: vgpr_32, preferred-register: '' }
-  - { id: 6, class: vgpr_32, preferred-register: '' }
-  - { id: 7, class: vgpr_32, preferred-register: '' }
-  - { id: 8, class: sreg_64, preferred-register: '$vcc' }
-  - { id: 9, class: sreg_64, preferred-register: '' }
-  - { id: 10, class: sreg_32_xm0, preferred-register: '' }
-  - { id: 11, class: sreg_64, preferred-register: '' }
-  - { id: 12, class: sreg_32_xm0, preferred-register: '' }
-  - { id: 13, class: sreg_32_xm0, preferred-register: '' }
-  - { id: 14, class: sreg_32_xm0, preferred-register: '' }
-  - { id: 15, class: sreg_128, preferred-register: '' }
-  - { id: 16, class: vgpr_32, preferred-register: '' }
-  - { id: 17, class: vgpr_32, preferred-register: '' }
-  - { id: 18, class: vgpr_32, preferred-register: '' }
-  - { id: 19, class: sreg_64, preferred-register: '' }
-  - { id: 20, class: sreg_64, preferred-register: '' }
-  - { id: 21, class: vgpr_32, preferred-register: '' }
-  - { id: 22, class: sreg_64, preferred-register: '' }
-  - { id: 23, class: sreg_64, preferred-register: '' }
-liveins:
-body:             |
-  bb.0:
-    successors: %bb.1(0x40000000), %bb.2(0x40000000)
-
-    %21 = V_MOV_B32_e32 0, implicit $exec
-    %5 = V_MBCNT_LO_U32_B32_e64 -1, 0, implicit $exec
-    %6 = V_MBCNT_HI_U32_B32_e32 -1, killed %5, implicit $exec
-    %8 = V_CMP_GT_U32_e64 32, killed %6, implicit $exec
-    %22 = COPY $exec, implicit-def $exec
-    %23 = S_AND_B64 %22, %8, implicit-def dead $scc
-    %0 = S_XOR_B64 %23, %22, implicit-def dead $scc
-    $exec = S_MOV_B64_term killed %23
-    SI_MASK_BRANCH %bb.2, implicit $exec
-    S_BRANCH %bb.1
-
-  bb.1:
-    successors: %bb.2(0x80000000)
-
-    %13 = S_MOV_B32 61440
-    %14 = S_MOV_B32 -1
-    %15 = REG_SEQUENCE undef %12, 1, undef %10, 2, killed %14, 3, killed %13, 4
-    %19 = COPY $exec
-    $exec = S_MOV_B64 -1
-    %16 = BUFFER_LOAD_DWORD_OFFSET %15, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 4)
-    %17 = V_ADD_F32_e32 1065353216, killed %16, implicit $exec
-    $exec = EXIT_WWM killed %19
-    %21 = V_MOV_B32_e32 1, implicit $exec
-    early-clobber %18 = WWM killed %17, implicit $exec
-    BUFFER_STORE_DWORD_OFFSET killed %18, killed %15, 0, 0, 0, 0, 0, implicit $exec :: (store 4)
-
-  bb.2:
-    $exec = S_OR_B64 $exec, killed %0, implicit-def $scc
-    $vgpr0 = COPY killed %21
-    SI_RETURN_TO_EPILOG killed $vgpr0
-
-...
-
-# Test a loop with a loop exit value and a loop phi.
-#CHECK: test_wwm_liveness_loop
-#CHECK: %4:vgpr_32 = IMPLICIT_DEF
-#CHECK: bb.1:
-#CHECK: %4:vgpr_32 = FLAT_LOAD_DWORD{{.*}}, implicit %4
-#CHECK: %27:vgpr_32 = COPY killed %21, implicit %27
-
----
-name:            test_wwm_liveness_loop
-alignment:       0
-exposesReturnsTwice: false
-legalized:       false
-regBankSelected: false
-selected:        false
-failedISel:      false
-tracksRegLiveness: true
-registers:
-  - { id: 0, class: vgpr_32, preferred-register: '' }
-  - { id: 1, class: sreg_32_xm0, preferred-register: '' }
-  - { id: 2, class: sreg_64, preferred-register: '' }
-  - { id: 3, class: sreg_32_xm0, preferred-register: '' }
-  - { id: 4, class: vgpr_32, preferred-register: '' }
-  - { id: 5, class: sreg_32_xm0, preferred-register: '' }
-  - { id: 6, class: sreg_64, preferred-register: '' }
-  - { id: 7, class: sreg_64, preferred-register: '' }
-  - { id: 8, class: sreg_64, preferred-register: '' }
-  - { id: 9, class: vreg_64, preferred-register: '' }
-  - { id: 10, class: vgpr_32, preferred-register: '' }
-  - { id: 11, class: vgpr_32, preferred-register: '' }
-  - { id: 12, class: vgpr_32, preferred-register: '' }
-  - { id: 13, class: sreg_64, preferred-register: '' }
-  - { id: 14, class: vreg_64, preferred-register: '' }
-  - { id: 15, class: sreg_32_xm0, preferred-register: '' }
-  - { id: 16, class: vgpr_32, preferred-register: '' }
-  - { id: 17, class: sreg_64, preferred-register: '$vcc' }
-  - { id: 18, class: vgpr_32, preferred-register: '' }
-  - { id: 19, class: vgpr_32, preferred-register: '' }
-  - { id: 20, class: vgpr_32, preferred-register: '' }
-  - { id: 21, class: vgpr_32, preferred-register: '' }
-  - { id: 22, class: vgpr_32, preferred-register: '' }
-  - { id: 23, class: sreg_64, preferred-register: '' }
-  - { id: 24, class: sreg_64, preferred-register: '' }
-  - { id: 25, class: sreg_64, preferred-register: '' }
-  - { id: 26, class: sreg_64, preferred-register: '' }
-  - { id: 27, class: vgpr_32, preferred-register: '' }
-liveins:
-frameInfo:
-  isFrameAddressTaken: false
-  isReturnAddressTaken: false
-  hasStackMap:     false
-  hasPatchPoint:   false
-  stackSize:       0
-  offsetAdjustment: 0
-  maxAlignment:    0
-  adjustsStack:    false
-  hasCalls:        false
-  stackProtector:  ''
-  maxCallFrameSize: 4294967295
-  hasOpaqueSPAdjustment: false
-  hasVAStart:      false
-  hasMustTailInVarArgFunc: false
-  localFrameSize:  0
-  savePoint:       ''
-  restorePoint:    ''
-fixedStack:
-stack:
-constants:
-body:             |
-  bb.0:
-    successors: %bb.1(0x80000000)
-
-    %25:sreg_64 = S_OR_SAVEEXEC_B64 -1, implicit-def $exec, implicit-def dead $scc, implicit $exec
-    %0:vgpr_32 = FLAT_LOAD_DWORD undef %9:vreg_64, 0, 0, 0, implicit $exec, implicit $flat_scr :: (volatile load 4 from `float addrspace(1)* undef`, addrspace 1)
-    $exec = EXIT_WWM killed %25
-    %12:vgpr_32 = V_MBCNT_LO_U32_B32_e64 -1, 0, implicit $exec
-    %7:sreg_64 = S_MOV_B64 0
-    %26:sreg_64 = COPY killed %7
-    %27:vgpr_32 = COPY killed %12
-
-  bb.1:
-    successors: %bb.2(0x04000000), %bb.1(0x7c000000)
-
-    %24:sreg_64 = S_OR_SAVEEXEC_B64 -1, implicit-def $exec, implicit-def dead $scc, implicit $exec
-    %20:vgpr_32 = COPY killed %27
-    %2:sreg_64 = COPY killed %26
-    %4:vgpr_32 = FLAT_LOAD_DWORD undef %14:vreg_64, 0, 0, 0, implicit $exec, implicit $flat_scr :: (volatile load 4 from `float addrspace(1)* undef`, addrspace 1)
-    $exec = EXIT_WWM killed %24
-    %22:vgpr_32 = V_ADD_I32_e32 -1, killed %20, implicit-def dead $vcc, implicit $exec
-    %17:sreg_64 = V_CMP_EQ_U32_e64 0, %22, implicit $exec
-    %6:sreg_64 = S_OR_B64 killed %17, killed %2, implicit-def $scc
-    %21:vgpr_32 = COPY killed %22
-    %26:sreg_64 = COPY %6
-    %27:vgpr_32 = COPY killed %21
-    $exec = S_ANDN2_B64_term $exec, %6, implicit-def $scc
-    S_CBRANCH_EXECNZ %bb.1, implicit $exec
-    S_BRANCH %bb.2
-
-  bb.2:
-    $exec = S_OR_B64 $exec, killed %6, implicit-def $scc
-    %23:sreg_64 = S_OR_SAVEEXEC_B64 -1, implicit-def $exec, implicit-def dead $scc, implicit $exec
-    %18:vgpr_32 = V_ADD_F32_e32 killed %0, killed %4, implicit $exec
-    $exec = EXIT_WWM killed %23
-    early-clobber %19:vgpr_32 = COPY killed %18, implicit $exec
-    $vgpr0 = COPY killed %19
-    SI_RETURN_TO_EPILOG killed $vgpr0
-
-...
-

llvm/trunk/test/CodeGen/AMDGPU/indirect-addressing-term.ll

define amdgpu_kernel void @extract_w_offset_vgpr(i32 addrspace(1)* %out) {
   ; GCN:   SI_SPILL_S128_SAVE killed $sgpr8_sgpr9_sgpr10_sgpr11, %stack.1, implicit $exec, implicit $sgpr96_sgpr97_sgpr98_sgpr99, implicit $sgpr3, implicit-def dead $m0 :: (store 16 into %stack.1, align 4, addrspace 5)
   ; GCN:   SI_SPILL_V512_SAVE killed $vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32, %stack.2, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr3, 0, implicit $exec :: (store 64 into %stack.2, align 4, addrspace 5)
   ; GCN:   SI_SPILL_S64_SAVE killed $sgpr0_sgpr1, %stack.3, implicit $exec, implicit $sgpr96_sgpr97_sgpr98_sgpr99, implicit $sgpr3, implicit-def dead $m0 :: (store 8 into %stack.3, align 4, addrspace 5)
   ; GCN:   SI_SPILL_V32_SAVE killed $vgpr1, %stack.4, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr3, 0, implicit $exec :: (store 4 into %stack.4, addrspace 5)
   ; GCN:   SI_SPILL_S64_SAVE killed $sgpr24_sgpr25, %stack.5, implicit $exec, implicit $sgpr96_sgpr97_sgpr98_sgpr99, implicit $sgpr3, implicit-def dead $m0 :: (store 8 into %stack.5, align 4, addrspace 5)
   ; GCN: bb.1:
   ; GCN:   successors: %bb.1(0x40000000), %bb.2(0x40000000)
   ; GCN:   $sgpr0_sgpr1 = SI_SPILL_S64_RESTORE %stack.5, implicit $exec, implicit $sgpr96_sgpr97_sgpr98_sgpr99, implicit $sgpr3, implicit-def dead $m0 :: (load 8 from %stack.5, align 4, addrspace 5)
-  ; GCN:   $vgpr0 = SI_SPILL_V32_RESTORE %stack.4, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr3, 0, implicit $exec :: (load 4 from %stack.4, addrspace 5)
   ; GCN:   $vgpr1 = SI_SPILL_V32_RESTORE %stack.0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr3, 0, implicit $exec :: (load 4 from %stack.0, addrspace 5)
   ; GCN:   renamable $sgpr2 = V_READFIRSTLANE_B32 $vgpr1, implicit $exec
   ; GCN:   renamable $sgpr4_sgpr5 = V_CMP_EQ_U32_e64 $sgpr2, killed $vgpr1, implicit $exec
   ; GCN:   renamable $sgpr4_sgpr5 = S_AND_SAVEEXEC_B64 killed renamable $sgpr4_sgpr5, implicit-def $exec, implicit-def $scc, implicit $exec
   ; GCN:   S_SET_GPR_IDX_ON killed renamable $sgpr2, 1, implicit-def $m0, implicit undef $m0
   ; GCN:   $vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17 = SI_SPILL_V512_RESTORE %stack.2, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr3, 0, implicit $exec :: (load 64 from %stack.2, align 4, addrspace 5)
   ; GCN:   renamable $vgpr18 = V_MOV_B32_e32 undef $vgpr3, implicit $exec, implicit killed $vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17, implicit $m0
   ; GCN:   S_SET_GPR_IDX_OFF
   ; GCN:   renamable $vgpr19 = COPY renamable $vgpr18
   ; GCN:   renamable $sgpr6_sgpr7 = COPY renamable $sgpr4_sgpr5
   ; GCN:   SI_SPILL_S64_SAVE killed $sgpr6_sgpr7, %stack.5, implicit $exec, implicit $sgpr96_sgpr97_sgpr98_sgpr99, implicit $sgpr3, implicit-def dead $m0 :: (store 8 into %stack.5, align 4, addrspace 5)
-  ; GCN:   SI_SPILL_S64_SAVE killed $sgpr0_sgpr1, %stack.6, implicit $exec, implicit $sgpr96_sgpr97_sgpr98_sgpr99, implicit $sgpr3, implicit-def dead $m0 :: (store 8 into %stack.6, align 4, addrspace 5)
   ; GCN:   SI_SPILL_V32_SAVE killed $vgpr19, %stack.4, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr3, 0, implicit $exec :: (store 4 into %stack.4, addrspace 5)
-  ; GCN:   SI_SPILL_V32_SAVE killed $vgpr0, %stack.7, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr3, 0, implicit $exec :: (store 4 into %stack.7, addrspace 5)
   ; GCN:   SI_SPILL_V32_SAVE killed $vgpr18, %stack.8, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr3, 0, implicit $exec :: (store 4 into %stack.8, addrspace 5)
   ; GCN:   $exec = S_XOR_B64_term $exec, killed renamable $sgpr4_sgpr5, implicit-def $scc
   ; GCN:   S_CBRANCH_EXECNZ %bb.1, implicit $exec
   ; GCN: bb.2:
   ; GCN:   $sgpr0_sgpr1 = SI_SPILL_S64_RESTORE %stack.3, implicit $exec, implicit $sgpr96_sgpr97_sgpr98_sgpr99, implicit $sgpr3, implicit-def dead $m0 :: (load 8 from %stack.3, align 4, addrspace 5)
   ; GCN:   $exec = S_MOV_B64 killed renamable $sgpr0_sgpr1
   ; GCN:   $vgpr0 = SI_SPILL_V32_RESTORE %stack.8, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr3, 0, implicit $exec :: (load 4 from %stack.8, addrspace 5)
   ; GCN:   $sgpr4_sgpr5_sgpr6_sgpr7 = SI_SPILL_S128_RESTORE %stack.1, implicit $exec, implicit $sgpr96_sgpr97_sgpr98_sgpr99, implicit $sgpr3, implicit-def dead $m0 :: (load 16 from %stack.1, align 4, addrspace 5)

llvm/trunk/test/CodeGen/AMDGPU/wqm.mir

 # RUN: llc -march=amdgcn -verify-machineinstrs -run-pass si-wqm -o -  %s | FileCheck %s
 
 ---
 # Check for awareness that s_or_saveexec_b64 clobbers SCC
 #
-#CHECK: S_OR_SAVEEXEC_B64
+#CHECK: ENTER_WWM
 #CHECK: S_CMP_LT_I32
 #CHECK: S_CSELECT_B32
 name:            test_wwm_scc
 alignment:       0
 exposesReturnsTwice: false
 legalized:       false
 regBankSelected: false
 selected:        false

llvm/trunk/test/CodeGen/AMDGPU/wwm-reserved.ll

+; RUN: llc -O0 -march=amdgcn -mcpu=gfx900 -amdgpu-dpp-combine=false -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9,GFX9-O0 %s
+; RUN: llc -march=amdgcn -mcpu=gfx900 -amdgpu-dpp-combine=false -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9,GFX9-O3 %s
+
+define amdgpu_cs void @no_cfg(<4 x i32> inreg %tmp14) {
+  %tmp100 = call <2 x float> @llvm.amdgcn.raw.buffer.load.v2f32(<4 x i32> %tmp14, i32 0, i32 0, i32 0)
+  %tmp101 = bitcast <2 x float> %tmp100 to <2 x i32>
+  %tmp102 = extractelement <2 x i32> %tmp101, i32 0
+  %tmp103 = extractelement <2 x i32> %tmp101, i32 1
+  %tmp105 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp102, i32 0)
+  %tmp107 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp103, i32 0)
+
+; GFX9: v_mov_b32_dpp v[[FIRST_MOV:[0-9]+]], v{{[0-9]+}} row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9: v_add_u32_e32 v[[FIRST_ADD:[0-9]+]], v{{[0-9]+}}, v[[FIRST_MOV]]
+; GFX9: v_mov_b32_e32 v[[FIRST:[0-9]+]], v[[FIRST_ADD]]
+  %tmp120 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp105, i32 323, i32 12, i32 15, i1 false)
+  %tmp121 = add i32 %tmp105, %tmp120
+  %tmp122 = tail call i32 @llvm.amdgcn.wwm.i32(i32 %tmp121)
+
+; GFX9: v_mov_b32_dpp v[[SECOND_MOV:[0-9]+]], v{{[0-9]+}} row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9: v_add_u32_e32 v[[SECOND_ADD:[0-9]+]], v{{[0-9]+}}, v[[SECOND_MOV]]
+; GFX9: v_mov_b32_e32 v[[SECOND:[0-9]+]], v[[SECOND_ADD]]
+  %tmp135 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp107, i32 323, i32 12, i32 15, i1 false)
+  %tmp136 = add i32 %tmp107, %tmp135
+  %tmp137 = tail call i32 @llvm.amdgcn.wwm.i32(i32 %tmp136)
+
+; GFX9-O3: v_cmp_eq_u32_e32 vcc, v[[FIRST]], v[[SECOND]]
+; GFX9-O0: v_cmp_eq_u32_e64 s{{\[}}{{[0-9]+}}:{{[0-9]+}}{{\]}}, v[[FIRST]], v[[SECOND]]
+  %tmp138 = icmp eq i32 %tmp122, %tmp137
+  %tmp139 = sext i1 %tmp138 to i32
+  %tmp140 = shl nsw i32 %tmp139, 1
+  %tmp141 = and i32 %tmp140, 2
+  %tmp145 = bitcast i32 %tmp141 to float
+  call void @llvm.amdgcn.raw.buffer.store.f32(float %tmp145, <4 x i32> %tmp14, i32 4, i32 0, i32 0)
+  ret void
+}
+
+define amdgpu_cs void @cfg(<4 x i32> inreg %tmp14, i32 %arg) {
+entry:
+  %tmp100 = call <2 x float> @llvm.amdgcn.raw.buffer.load.v2f32(<4 x i32> %tmp14, i32 0, i32 0, i32 0)
+  %tmp101 = bitcast <2 x float> %tmp100 to <2 x i32>
+  %tmp102 = extractelement <2 x i32> %tmp101, i32 0
+  %tmp105 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp102, i32 0)
+
+; GFX9: v_mov_b32_dpp v[[FIRST_MOV:[0-9]+]], v{{[0-9]+}} row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9: v_add_u32_e32 v[[FIRST_ADD:[0-9]+]], v{{[0-9]+}}, v[[FIRST_MOV]]
+; GFX9: v_mov_b32_e32 v[[FIRST:[0-9]+]], v[[FIRST_ADD]]
+; GFX9-O0: buffer_store_dword v[[FIRST]], off, s{{\[}}{{[0-9]+}}:{{[0-9]+}}{{\]}}, s[[FIRST_SGPR_OFFSET:[0-9]+]] offset:[[FIRST_IMM_OFFSET:[0-9]+]]
+  %tmp120 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp105, i32 323, i32 12, i32 15, i1 false)
+  %tmp121 = add i32 %tmp105, %tmp120
+  %tmp122 = tail call i32 @llvm.amdgcn.wwm.i32(i32 %tmp121)
+
+  %cond = icmp eq i32 %arg, 0
+  br i1 %cond, label %if, label %merge
+if:
+  %tmp103 = extractelement <2 x i32> %tmp101, i32 1
+  %tmp107 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp103, i32 0)
+
+; GFX9: v_mov_b32_dpp v[[SECOND_MOV:[0-9]+]], v{{[0-9]+}} row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9: v_add_u32_e32 v[[SECOND_ADD:[0-9]+]], v{{[0-9]+}}, v[[SECOND_MOV]]
+; GFX9: v_mov_b32_e32 v[[SECOND:[0-9]+]], v[[SECOND_ADD]]
+; GFX9-O0: buffer_store_dword v[[SECOND]], off, s{{\[}}{{[0-9]+}}:{{[0-9]+}}{{\]}}, s[[SECOND_SGPR_OFFSET:[0-9]+]] offset:[[SECOND_IMM_OFFSET:[0-9]+]]
+  %tmp135 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp107, i32 323, i32 12, i32 15, i1 false)
+  %tmp136 = add i32 %tmp107, %tmp135
+  %tmp137 = tail call i32 @llvm.amdgcn.wwm.i32(i32 %tmp136)
+  br label %merge
+
+merge:
+  %merge_value = phi i32 [ 0, %entry ], [%tmp137, %if ]
+; GFX9-O3: v_cmp_eq_u32_e32 vcc, v[[FIRST]], v[[SECOND]]
+; GFX9-O0: buffer_load_dword v[[SECOND:[0-9]+]], off, s{{\[}}{{[0-9]+}}:{{[0-9]+}}{{\]}}, s[[SECOND_SGPR_OFFSET]] offset:[[SECOND_IMM_OFFSET]]
+; GFX9-O0: buffer_load_dword v[[FIRST:[0-9]+]], off, s{{\[}}{{[0-9]+}}:{{[0-9]+}}{{\]}}, s[[FIRST_SGPR_OFFSET]] offset:[[FIRST_IMM_OFFSET]]
+; GFX9-O0: v_cmp_eq_u32_e64 s{{\[}}{{[0-9]+}}:{{[0-9]+}}{{\]}}, v[[FIRST]], v[[SECOND]]
+  %tmp138 = icmp eq i32 %tmp122, %merge_value
+  %tmp139 = sext i1 %tmp138 to i32
+  %tmp140 = shl nsw i32 %tmp139, 1
+  %tmp141 = and i32 %tmp140, 2
+  %tmp145 = bitcast i32 %tmp141 to float
+  call void @llvm.amdgcn.raw.buffer.store.f32(float %tmp145, <4 x i32> %tmp14, i32 4, i32 0, i32 0)
+  ret void
+}
+
+define i32 @called(i32 %a) noinline {
+; GFX9: v_add_u32_e32 v1, v0, v0
+  %add = add i32 %a, %a
+; GFX9: v_mul_lo_i32 v0, v1, v0
+  %mul = mul i32 %add, %a
+; GFX9: v_sub_u32_e32 v0, v0, v1
+  %sub = sub i32 %mul, %add
+  ret i32 %sub
+}
+
+define amdgpu_kernel void @call(<4 x i32> inreg %tmp14, i32 inreg %arg) {
+; GFX9-O0: v_mov_b32_e32 v0, s2
+; GFX9-O3: v_mov_b32_e32 v2, s0
+; GFX9-NEXT: s_not_b64 exec, exec
+; GFX9-O0-NEXT: v_mov_b32_e32 v0, s3
+; GFX9-O3-NEXT: v_mov_b32_e32 v2, 0
+; GFX9-NEXT: s_not_b64 exec, exec
+  %tmp107 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %arg, i32 0)
+; GFX9-O0: buffer_store_dword v0
+; GFX9-O3: v_mov_b32_e32 v0, v2
+; GFX9: s_swappc_b64
+  %tmp134 = call i32 @called(i32 %tmp107)
+; GFX9-O0: buffer_load_dword v1
+; GFX9-O3: v_mov_b32_e32 v1, v0
+; GFX9-O0: v_add_u32_e32 v0, v0, v1
+; GFX9-O3: v_add_u32_e32 v1, v1, v2
+  %tmp136 = add i32 %tmp134, %tmp107
+  %tmp137 = tail call i32 @llvm.amdgcn.wwm.i32(i32 %tmp136)
+; GFX9-O0: buffer_store_dword v2
+; GFX9-O3: buffer_store_dword v0
+  call void @llvm.amdgcn.raw.buffer.store.i32(i32 %tmp137, <4 x i32> %tmp14, i32 4, i32 0, i32 0)
+  ret void
+}
+
+define i64 @called_i64(i64 %a) noinline {
+  %add = add i64 %a, %a
+  %mul = mul i64 %add, %a
+  %sub = sub i64 %mul, %add
+  ret i64 %sub
+}
+
+define amdgpu_kernel void @call_i64(<4 x i32> inreg %tmp14, i64 inreg %arg) {
+; GFX9-O0: v_mov_b32_e32 v0, s0
+; GFX9-O0: v_mov_b32_e32 v1, s1
+; GFX9-O3: v_mov_b32_e32 v7, s1
+; GFX9-O3: v_mov_b32_e32 v6, s0
+; GFX9-NEXT: s_not_b64 exec, exec
+; GFX9-O0-NEXT: v_mov_b32_e32 v0, s2
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s3
+; GFX9-O3-NEXT: v_mov_b32_e32 v6, 0
+; GFX9-O3-NEXT: v_mov_b32_e32 v7, 0
+; GFX9-NEXT: s_not_b64 exec, exec
+  %tmp107 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %arg, i64 0)
+; GFX9-O0: buffer_store_dword v0
+; GFX9-O0: buffer_store_dword v1
+; GFX9: s_swappc_b64
+  %tmp134 = call i64 @called_i64(i64 %tmp107)
+; GFX9-O0: buffer_load_dword v3
+; GFX9-O0: buffer_load_dword v4
+  %tmp136 = add i64 %tmp134, %tmp107
+  %tmp137 = tail call i64 @llvm.amdgcn.wwm.i64(i64 %tmp136)
+  %tmp138 = bitcast i64 %tmp137 to <2 x i32>
+  call void @llvm.amdgcn.raw.buffer.store.v2i32(<2 x i32> %tmp138, <4 x i32> %tmp14, i32 4, i32 0, i32 0)
+  ret void
+}
+
+define amdgpu_cs void @_amdgpu_cs_main(<4 x i32> inreg %desc, i32 %index) {
+  %tmp17 = shl i32 %index, 5
+; GFX9: buffer_load_dwordx4
+  %tmp18 = tail call <4 x i32> @llvm.amdgcn.s.buffer.load.v4i32(<4 x i32> %desc, i32 %tmp17, i32 0)
+  %.i0.upto1.bc = bitcast <4 x i32> %tmp18 to <2 x i64>
+  %tmp19 = or i32 %tmp17, 16
+; GFX9: buffer_load_dwordx2
+  %tmp20 = tail call <2 x i32> @llvm.amdgcn.s.buffer.load.v2i32(<4 x i32> %desc, i32 %tmp19, i32 0)
+  %.i0.upto1.extract = extractelement <2 x i64> %.i0.upto1.bc, i32 0
+  %tmp22 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %.i0.upto1.extract, i64 9223372036854775807)
+  %tmp97 = tail call i64 @llvm.amdgcn.wwm.i64(i64 %tmp22)
+  %.i1.upto1.extract = extractelement <2 x i64> %.i0.upto1.bc, i32 1
+  %tmp99 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %.i1.upto1.extract, i64 9223372036854775807)
+  %tmp174 = tail call i64 @llvm.amdgcn.wwm.i64(i64 %tmp99)
+  %.i25 = bitcast <2 x i32> %tmp20 to i64
+  %tmp176 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %.i25, i64 9223372036854775807)
+  %tmp251 = tail call i64 @llvm.amdgcn.wwm.i64(i64 %tmp176)
+  %.cast = bitcast i64 %tmp97 to <2 x float>
+  %.cast6 = bitcast i64 %tmp174 to <2 x float>
+  %.cast7 = bitcast i64 %tmp251 to <2 x float>
+  %tmp254 = shufflevector <2 x float> %.cast, <2 x float> %.cast6, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; GFX9: buffer_store_dwordx4
+  tail call void @llvm.amdgcn.raw.buffer.store.v4f32(<4 x float> %tmp254, <4 x i32> %desc, i32 %tmp17, i32 0, i32 0)
+  ; GFX9: buffer_store_dwordx2
+  tail call void @llvm.amdgcn.raw.buffer.store.v2f32(<2 x float> %.cast7, <4 x i32> %desc, i32 %tmp19, i32 0, i32 0)
+  ret void
+}
+
+declare i32 @llvm.amdgcn.wwm.i32(i32)
+declare i64 @llvm.amdgcn.wwm.i64(i64)
+declare i32 @llvm.amdgcn.set.inactive.i32(i32, i32)
+declare i64 @llvm.amdgcn.set.inactive.i64(i64, i64)
+declare i32 @llvm.amdgcn.update.dpp.i32(i32, i32, i32, i32, i32, i1)
+declare <2 x float> @llvm.amdgcn.raw.buffer.load.v2f32(<4 x i32>, i32, i32, i32)
+declare void @llvm.amdgcn.raw.buffer.store.f32(float, <4 x i32>, i32, i32, i32)
+declare void @llvm.amdgcn.raw.buffer.store.i32(i32, <4 x i32>, i32, i32, i32)
+declare void @llvm.amdgcn.raw.buffer.store.v2i32(<2 x i32>, <4 x i32>, i32, i32, i32)
+declare void @llvm.amdgcn.raw.buffer.store.v2f32(<2 x float>, <4 x i32>, i32, i32, i32)
+declare void @llvm.amdgcn.raw.buffer.store.v4f32(<4 x float>, <4 x i32>, i32, i32, i32) 
+declare <2 x i32> @llvm.amdgcn.s.buffer.load.v2i32(<4 x i32>, i32, i32)
+declare <4 x i32> @llvm.amdgcn.s.buffer.load.v4i32(<4 x i32>, i32, i32)