[AMDGPU] Add llvm.amdgcn.softwqm intrinsic

[critson]

Summary

Add llvm.amdgcn.softwqm intrinsic which behaves like llvm.amdgcn.wqm only if there is other WQM computation in the shader.

[arsenm]

lib/Target/AMDGPU/SIISelLowering.cpp

5955–5959 ↗

(On Diff #210619)

Is there some reason you can't just handle this with an instruction pattern?

[critson]

lib/Target/AMDGPU/SIISelLowering.cpp

5955–5959 ↗

(On Diff #210619)

For the same reason as llvm.amdgcn.wqm, we don't specify the input and output types.
Happy to be corrected, but I don't think there is a way to have a single instruction pattern covering all types.

[nhaehnle]

Have you checked that this actually fixes the reported CTS failure?

IIRC the CTS failure was essentially due to a shader of the form:

derivative calculation here
subgroup operation here

The derivative calculation enables WQM, but then we may leave WQM again for the subgroup operations which is unexpected (since helper lanes are expected to participate). So softwqm needs to seed WQM requirements, but only if there is at least one hard wqm requirement in the shader.

[arsenm]

lib/Target/AMDGPU/SIISelLowering.cpp

5955–5959 ↗

(On Diff #210619)

It's easier to directly select than to enumerate all the possible types. I would still expect all of these direct-to-machine-node intrinsics to be handled in AMDGPUISelDAGToDAG

[critson]

Add missing code in SI Fix SGPR copies.

[critson]

In D64935#1593333, @nhaehnle wrote:

Have you checked that this actually fixes the reported CTS failure?

Yes, with the associated (minimal) frontend changes this fixes the CTS failure.

In D64935#1593333, @nhaehnle wrote:

The derivative calculation enables WQM, but then we may leave WQM again for the subgroup operations which is unexpected (since helper lanes are expected to participate). So softwqm needs to seed WQM requirements, but only if there is at least one hard wqm requirement in the shader.

While my understanding of "seed requirements" means "for the whole shader", this code does what you expect.
If there are any hard WQM requirements for the shader, then all softwqm instructions (and their dependencies) are marked WQM.

[nhaehnle]

Okay thanks, I see the logic now.

lib/Target/AMDGPU/SIISelLowering.cpp

5955–5959 ↗

(On Diff #210619)

You mean adding an AMDGPUDAGToDAGISel::SelectINTRINSIC_WO_CHAIN and lowering the softwqm intrinsic there? That does make sense to me.

lib/Target/AMDGPU/SIInstructions.td

114

s/wcm/wqm/

[critson]

Move opcode selection to AMDGPUISelDAGToDAG.
Fix typo in comment.

[critson]

I've moved the selection to AMDGPUISelDAGToDAG.
If this code is appropriate I will submit a follow change to move the selection for llvm.amdgcn.wqm and llvm.amdgcn.wwm as well.

[nhaehnle]

LGTM. Followup for WQM/WWM sounds good to me as well.

include/llvm/IR/IntrinsicsAMDGPU.td

 
 
 // Copies the source value to the destination value, with the guarantee that
 // the source value is computed as if the entire program were executed in WQM.
 def int_amdgcn_wqm : Intrinsic<[llvm_any_ty],
   [LLVMMatchType<0>], [IntrNoMem, IntrSpeculatable]
 >;
 
+// Copies the source value to the destination value, such that the source
+// is computed as if the entire program were executed in WQM if any other
+// program code executes in WQM.
+def int_amdgcn_softwqm : Intrinsic<[llvm_any_ty],
+  [LLVMMatchType<0>], [IntrNoMem, IntrSpeculatable]
+>;
+
 // Return true if at least one thread within the pixel quad passes true into
 // the function.
 def int_amdgcn_wqm_vote : Intrinsic<[llvm_i1_ty],
   [llvm_i1_ty], [IntrNoMem, IntrConvergent]
 >;
 
 // If false, set EXEC=0 for the current thread until the end of program.
 def int_amdgcn_kill : Intrinsic<[], [llvm_i1_ty], []>;

lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

private:
   void SelectS_BFE(SDNode *N);
   bool isCBranchSCC(const SDNode *N) const;
   void SelectBRCOND(SDNode *N);
   void SelectFMAD_FMA(SDNode *N);
   void SelectATOMIC_CMP_SWAP(SDNode *N);
   void SelectDSAppendConsume(SDNode *N, unsigned IntrID);
   void SelectDS_GWS(SDNode *N, unsigned IntrID);
   void SelectINTRINSIC_W_CHAIN(SDNode *N);
+  void SelectINTRINSIC_WO_CHAIN(SDNode *N);
   void SelectINTRINSIC_VOID(SDNode *N);
 
 protected:
   // Include the pieces autogenerated from the target description.
 #include "AMDGPUGenDAGISel.inc"
 };
 
 class R600DAGToDAGISel : public AMDGPUDAGToDAGISel {
case AMDGPUISD::CVT_PK_I16_I32: {
     }
 
     break;
   }
   case ISD::INTRINSIC_W_CHAIN: {
     SelectINTRINSIC_W_CHAIN(N);
     return;
   }
+  case ISD::INTRINSIC_WO_CHAIN: {
+    SelectINTRINSIC_WO_CHAIN(N);
+    return;
+  }
   case ISD::INTRINSIC_VOID: {
     SelectINTRINSIC_VOID(N);
     return;
   }
   }
 
   SelectCode(N);
 }
case Intrinsic::amdgcn_ds_consume: {
     SelectDSAppendConsume(N, IntrID);
     return;
   }
   }
 
   SelectCode(N);
 }
 
+void AMDGPUDAGToDAGISel::SelectINTRINSIC_WO_CHAIN(SDNode *N) {
+  unsigned IntrID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
+  unsigned Opcode;
+  switch (IntrID) {
+  case Intrinsic::amdgcn_softwqm:
+    Opcode = AMDGPU::SOFT_WQM;
+    break;
+  default:
+    SelectCode(N);
+    return;
+  }
+
+  SDValue Src = N->getOperand(1);
+  CurDAG->SelectNodeTo(N, Opcode, N->getVTList(), {Src});
+}
+
 void AMDGPUDAGToDAGISel::SelectINTRINSIC_VOID(SDNode *N) {
   unsigned IntrID = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
   switch (IntrID) {
   case Intrinsic::amdgcn_ds_gws_init:
   case Intrinsic::amdgcn_ds_gws_barrier:
   case Intrinsic::amdgcn_ds_gws_sema_v:
   case Intrinsic::amdgcn_ds_gws_sema_br:
   case Intrinsic::amdgcn_ds_gws_sema_p:

lib/Target/AMDGPU/SIFixSGPRCopies.cpp

for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
          I != E; ++I) {
       MachineInstr &MI = *I;
 
       switch (MI.getOpcode()) {
       default:
         continue;
       case AMDGPU::COPY:
       case AMDGPU::WQM:
+      case AMDGPU::SOFT_WQM:
       case AMDGPU::WWM: {
         // If the destination register is a physical register there isn't really
         // much we can do to fix this.
         if (!TargetRegisterInfo::isVirtualRegister(MI.getOperand(0).getReg()))
           continue;
 
         const TargetRegisterClass *SrcRC, *DstRC;
         std::tie(SrcRC, DstRC) = getCopyRegClasses(MI, *TRI, MRI);

lib/Target/AMDGPU/SIInstrInfo.cpp

 unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) const {
   switch (MI.getOpcode()) {
   default: return AMDGPU::INSTRUCTION_LIST_END;
   case AMDGPU::REG_SEQUENCE: return AMDGPU::REG_SEQUENCE;
   case AMDGPU::COPY: return AMDGPU::COPY;
   case AMDGPU::PHI: return AMDGPU::PHI;
   case AMDGPU::INSERT_SUBREG: return AMDGPU::INSERT_SUBREG;
   case AMDGPU::WQM: return AMDGPU::WQM;
+  case AMDGPU::SOFT_WQM: return AMDGPU::SOFT_WQM;
   case AMDGPU::WWM: return AMDGPU::WWM;
   case AMDGPU::S_MOV_B32: {
     const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
     return MI.getOperand(1).isReg() ||
            RI.isAGPR(MRI, MI.getOperand(0).getReg()) ?
            AMDGPU::COPY : AMDGPU::V_MOV_B32_e32;
   }
   case AMDGPU::S_ADD_I32:
for (MachineRegisterInfo::use_iterator I = MRI.use_begin(DstReg),
          E = MRI.use_end(); I != E;) {
     MachineInstr &UseMI = *I->getParent();
 
     unsigned OpNo = 0;
 
     switch (UseMI.getOpcode()) {
     case AMDGPU::COPY:
     case AMDGPU::WQM:
+    case AMDGPU::SOFT_WQM:
     case AMDGPU::WWM:
     case AMDGPU::REG_SEQUENCE:
     case AMDGPU::PHI:
     case AMDGPU::INSERT_SUBREG:
       break;
     default:
       OpNo = I.getOperandNo();
       break;
const TargetRegisterClass *SIInstrInfo::getDestEquivalentVGPRClass(
   // For target instructions, getOpRegClass just returns the virtual register
   // class associated with the operand, so we need to find an equivalent VGPR
   // register class in order to move the instruction to the VALU.
   case AMDGPU::COPY:
   case AMDGPU::PHI:
   case AMDGPU::REG_SEQUENCE:
   case AMDGPU::INSERT_SUBREG:
   case AMDGPU::WQM:
+  case AMDGPU::SOFT_WQM:
   case AMDGPU::WWM: {
     const TargetRegisterClass *SrcRC = getOpRegClass(Inst, 1);
     if (RI.hasAGPRs(SrcRC)) {
       if (RI.hasAGPRs(NewDstRC))
         return nullptr;
 
       NewDstRC = RI.getEquivalentAGPRClass(NewDstRC);
       if (!NewDstRC)

lib/Target/AMDGPU/SIInstructions.td

 // SIFoldOperands pass to enable folding of inline immediates.
 def V_MOV_B64_PSEUDO : VPseudoInstSI <(outs VReg_64:$vdst),
                                       (ins VSrc_b64:$src0)>;
 
 // Pseudoinstruction for @llvm.amdgcn.wqm. It is turned into a copy after the
 // WQM pass processes it.
 def WQM : PseudoInstSI <(outs unknown:$vdst), (ins unknown:$src0)>;
 
+// Pseudoinstruction for @llvm.amdgcn.softwqm. Like @llvm.amdgcn.wqm it is
+// turned into a copy by WQM pass, but does not seed WQM requirements.
+def SOFT_WQM : PseudoInstSI <(outs unknown:$vdst), (ins unknown:$src0)>;
+
 // Pseudoinstruction for @llvm.amdgcn.wwm. It is turned into a copy post-RA, so
 // that the @earlyclobber is respected. The @earlyclobber is to make sure that
 // 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)>;
 }
 

lib/Target/AMDGPU/SIWholeQuadMode.cpp

 
 // Scan instructions to determine which ones require an Exact execmask and
 // which ones seed WQM requirements.
 char SIWholeQuadMode::scanInstructions(MachineFunction &MF,
                                        std::vector<WorkItem> &Worklist) {
   char GlobalFlags = 0;
   bool WQMOutputs = MF.getFunction().hasFnAttribute("amdgpu-ps-wqm-outputs");
   SmallVector<MachineInstr *, 4> SetInactiveInstrs;
+  SmallVector<MachineInstr *, 4> SoftWQMInstrs;
 
   // We need to visit the basic blocks in reverse post-order so that we visit
   // defs before uses, in particular so that we don't accidentally mark an
   // instruction as needing e.g. WQM before visiting it and realizing it needs
   // WQM disabled.
   ReversePostOrderTraversal<MachineFunction *> RPOT(&MF);
   for (auto BI = RPOT.begin(), BE = RPOT.end(); BI != BE; ++BI) {
     MachineBasicBlock &MBB = **BI;
for (auto II = MBB.begin(), IE = MBB.end(); II != IE; ++II) {
         markInstructionUses(MI, StateWQM, Worklist);
         GlobalFlags |= StateWQM;
         continue;
       } else if (Opcode == AMDGPU::WQM) {
         // The WQM intrinsic requires its output to have all the helper lanes
         // correct, so we need it to be in WQM.
         Flags = StateWQM;
         LowerToCopyInstrs.push_back(&MI);
+      } else if (Opcode == AMDGPU::SOFT_WQM) {
+        LowerToCopyInstrs.push_back(&MI);
+        SoftWQMInstrs.push_back(&MI);
+        continue;
       } else if (Opcode == AMDGPU::WWM) {
         // The WWM intrinsic doesn't make the same guarantee, and plus it needs
         // to be executed in WQM or Exact so that its copy doesn't clobber
         // inactive lanes.
         markInstructionUses(MI, StateWWM, Worklist);
         GlobalFlags |= StateWWM;
         LowerToCopyInstrs.push_back(&MI);
         continue;
for (auto II = MBB.begin(), IE = MBB.end(); II != IE; ++II) {
       markInstruction(MI, Flags, Worklist);
       GlobalFlags |= Flags;
     }
   }
 
   // Mark sure that any SET_INACTIVE instructions are computed in WQM if WQM is
   // ever used anywhere in the function. This implements the corresponding
   // semantics of @llvm.amdgcn.set.inactive.
+  // Similarly for SOFT_WQM instructions, implementing @llvm.amdgcn.softwqm.
   if (GlobalFlags & StateWQM) {
     for (MachineInstr *MI : SetInactiveInstrs)
       markInstruction(*MI, StateWQM, Worklist);
+    for (MachineInstr *MI : SoftWQMInstrs)
+      markInstruction(*MI, StateWQM, Worklist);
   }
 
   return GlobalFlags;
 }
 
 void SIWholeQuadMode::propagateInstruction(MachineInstr &MI,
                                            std::vector<WorkItem>& Worklist) {
   MachineBasicBlock *MBB = MI.getParent();
bool SIWholeQuadMode::runOnMachineFunction(MachineFunction &MF) {
   MRI = &MF.getRegInfo();
   LIS = &getAnalysis<LiveIntervals>();
 
   char GlobalFlags = analyzeFunction(MF);
   unsigned LiveMaskReg = 0;
   unsigned Exec = ST->isWave32() ? AMDGPU::EXEC_LO : AMDGPU::EXEC;
   if (!(GlobalFlags & StateWQM)) {
     lowerLiveMaskQueries(Exec);
-    if (!(GlobalFlags & StateWWM))
+    if (!(GlobalFlags & StateWWM) && LowerToCopyInstrs.empty())
       return !LiveMaskQueries.empty();
   } else {
     // Store a copy of the original live mask when required
     MachineBasicBlock &Entry = MF.front();
     MachineBasicBlock::iterator EntryMI = Entry.getFirstNonPHI();
 
     if (GlobalFlags & StateExact || !LiveMaskQueries.empty()) {
       LiveMaskReg = MRI->createVirtualRegister(TRI->getBoolRC());

test/CodeGen/AMDGPU/llvm.amdgcn.softwqm.ll

+; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=CHECK %s
+
+; Check that WQM is not triggered by the softwqm intrinsic alone.
+;
+;CHECK-LABEL: {{^}}test1:
+;CHECK-NOT: s_wqm_b64 exec, exec
+;CHECK: buffer_load_dword
+;CHECK: buffer_load_dword
+;CHECK: v_add_f32_e32
+define amdgpu_ps float @test1(i32 inreg %idx0, i32 inreg %idx1) {
+main_body:
+  %src0 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx0, i32 0, i1 0, i1 0)
+  %src1 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx1, i32 0, i1 0, i1 0)
+  %out = fadd float %src0, %src1
+  %out.0 = call float @llvm.amdgcn.softwqm.f32(float %out)
+  ret float %out.0
+}
+
+; Check that the softwqm intrinsic works correctly for integers.
+;
+;CHECK-LABEL: {{^}}test2:
+;CHECK-NOT: s_wqm_b64 exec, exec
+;CHECK: buffer_load_dword
+;CHECK: buffer_load_dword
+;CHECK: v_add_f32_e32
+define amdgpu_ps float @test2(i32 inreg %idx0, i32 inreg %idx1) {
+main_body:
+  %src0 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx0, i32 0, i1 0, i1 0)
+  %src1 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx1, i32 0, i1 0, i1 0)
+  %out = fadd float %src0, %src1
+  %out.0 = bitcast float %out to i32
+  %out.1 = call i32 @llvm.amdgcn.softwqm.i32(i32 %out.0)
+  %out.2 = bitcast i32 %out.1 to float
+  ret float %out.2
+}
+
+; Make sure the transition from WQM to Exact to softwqm does not trigger WQM.
+;
+;CHECK-LABEL: {{^}}test_softwqm1:
+;CHECK-NOT: s_wqm_b64 exec, exec
+;CHECK: buffer_load_dword
+;CHECK: buffer_load_dword
+;CHECK: buffer_store_dword
+;CHECK-NOT; s_wqm_b64 exec, exec
+;CHECK: v_add_f32_e32
+define amdgpu_ps float @test_softwqm1(i32 inreg %idx0, i32 inreg %idx1) {
+main_body:
+  %src0 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx0, i32 0, i1 0, i1 0)
+  %src1 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx1, i32 0, i1 0, i1 0)
+  %temp = fadd float %src0, %src1
+  call void @llvm.amdgcn.buffer.store.f32(float %temp, <4 x i32> undef, i32 %idx0, i32 0, i1 0, i1 0)
+  %out = fadd float %temp, %temp
+  %out.0 = call float @llvm.amdgcn.softwqm.f32(float %out)
+  ret float %out.0
+}
+
+; Make sure the transition from WQM to Exact to softwqm does trigger WQM.
+;
+;CHECK-LABEL: {{^}}test_softwqm2:
+;CHECK: s_mov_b64 [[ORIG:s\[[0-9]+:[0-9]+\]]], exec
+;CHECK: s_wqm_b64 exec, exec
+;CHECK: buffer_load_dword
+;CHECK: buffer_load_dword
+;CHECK: s_and_b64 exec, exec, [[ORIG]]
+;CHECK: buffer_store_dword
+;CHECK; s_wqm_b64 exec, exec
+;CHECK: v_add_f32_e32
+define amdgpu_ps float @test_softwqm2(i32 inreg %idx0, i32 inreg %idx1) {
+main_body:
+  %src0 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx0, i32 0, i1 0, i1 0)
+  %src1 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx1, i32 0, i1 0, i1 0)
+  %temp = fadd float %src0, %src1
+  %temp.0 = call float @llvm.amdgcn.wqm.f32(float %temp)
+  call void @llvm.amdgcn.buffer.store.f32(float %temp.0, <4 x i32> undef, i32 %idx0, i32 0, i1 0, i1 0)
+  %out = fadd float %temp, %temp
+  %out.0 = call float @llvm.amdgcn.softwqm.f32(float %out)
+  ret float %out.0
+}
+
+; Make sure the transition from Exact to WWM then softwqm does not trigger WQM.
+;
+;CHECK-LABEL: {{^}}test_wwm1:
+;CHECK: buffer_load_dword
+;CHECK: buffer_store_dword
+;CHECK: s_or_saveexec_b64 [[ORIG:s\[[0-9]+:[0-9]+\]]], -1
+;CHECK: buffer_load_dword
+;CHECK: v_add_f32_e32
+;CHECK: s_mov_b64 exec, [[ORIG]]
+;CHECK-NOT: s_wqm_b64
+define amdgpu_ps float @test_wwm1(i32 inreg %idx0, i32 inreg %idx1) {
+main_body:
+  %src0 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx0, i32 0, i1 0, i1 0)
+  call void @llvm.amdgcn.buffer.store.f32(float %src0, <4 x i32> undef, i32 %idx0, i32 0, i1 0, i1 0)
+  %src1 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx1, i32 0, i1 0, i1 0)
+  %temp = fadd float %src0, %src1
+  %temp.0 = call float @llvm.amdgcn.wwm.f32(float %temp)
+  %out = fadd float %temp.0, %temp.0
+  %out.0 = call float @llvm.amdgcn.softwqm.f32(float %out)
+  ret float %out.0
+}
+
+; Check that softwqm on one case of branch does not trigger WQM for shader.
+;
+;CHECK-LABEL: {{^}}test_control_flow_0:
+;CHECK-NEXT: ; %main_body
+;CHECK-NOT: s_wqm_b64 exec, exec
+;CHECK: %ELSE
+;CHECK: store
+;CHECK: %IF
+;CHECK: buffer_load
+;CHECK: buffer_load
+define amdgpu_ps float @test_control_flow_0(<8 x i32> inreg %rsrc, <4 x i32> inreg %sampler, i32 inreg %idx0, i32 inreg %idx1, i32 %c, i32 %z, float %data) {
+main_body:
+  %cmp = icmp eq i32 %z, 0
+  br i1 %cmp, label %IF, label %ELSE
+
+IF:
+  %src0 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx0, i32 0, i1 0, i1 0)
+  %src1 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx1, i32 0, i1 0, i1 0)
+  %out = fadd float %src0, %src1
+  %data.if = call float @llvm.amdgcn.softwqm.f32(float %out)
+  br label %END
+
+ELSE:
+  call void @llvm.amdgcn.buffer.store.f32(float %data, <4 x i32> undef, i32 %c, i32 0, i1 0, i1 0)
+  br label %END
+
+END:
+  %r = phi float [ %data.if, %IF ], [ %data, %ELSE ]
+  ret float %r
+}
+
+; Check that softwqm on one case of branch is treated as WQM in WQM shader.
+;
+;CHECK-LABEL: {{^}}test_control_flow_1:
+;CHECK-NEXT: ; %main_body
+;CHECK-NEXT: s_mov_b64 [[ORIG:s\[[0-9]+:[0-9]+\]]], exec
+;CHECK-NEXT: s_wqm_b64 exec, exec
+;CHECK: %ELSE
+;CHECK: s_and_saveexec_b64 [[SAVED:s\[[0-9]+:[0-9]+\]]], [[ORIG]]
+;CHECK: store
+;CHECK: s_mov_b64 exec, [[SAVED]]
+;CHECK: %IF
+;CHECK-NOT: s_and_saveexec_b64
+;CHECK-NOT: s_and_b64 exec
+;CHECK: buffer_load
+;CHECK: buffer_load
+define amdgpu_ps float @test_control_flow_1(<8 x i32> inreg %rsrc, <4 x i32> inreg %sampler, i32 inreg %idx0, i32 inreg %idx1, i32 %c, i32 %z, float %data) {
+main_body:
+  %c.bc = bitcast i32 %c to float
+  %tex = call <4 x float> @llvm.amdgcn.image.sample.1d.v4f32.f32(i32 15, float %c.bc, <8 x i32> %rsrc, <4 x i32> %sampler, i1 0, i32 0, i32 0) #0
+  %tex0 = extractelement <4 x float> %tex, i32 0
+  %dtex = call <4 x float> @llvm.amdgcn.image.sample.1d.v4f32.f32(i32 15, float %tex0, <8 x i32> %rsrc, <4 x i32> %sampler, i1 0, i32 0, i32 0) #0
+  %data.sample = extractelement <4 x float> %dtex, i32 0
+
+  %cmp = icmp eq i32 %z, 0
+  br i1 %cmp, label %IF, label %ELSE
+
+IF:
+  %src0 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx0, i32 0, i1 0, i1 0)
+  %src1 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx1, i32 0, i1 0, i1 0)
+  %out = fadd float %src0, %src1
+  %data.if = call float @llvm.amdgcn.softwqm.f32(float %out)
+  br label %END
+
+ELSE:
+  call void @llvm.amdgcn.buffer.store.f32(float %data.sample, <4 x i32> undef, i32 %c, i32 0, i1 0, i1 0)
+  br label %END
+
+END:
+  %r = phi float [ %data.if, %IF ], [ %data, %ELSE ]
+  ret float %r
+}
+
+declare void @llvm.amdgcn.buffer.store.f32(float, <4 x i32>, i32, i32, i1, i1) #2
+declare void @llvm.amdgcn.buffer.store.v4f32(<4 x float>, <4 x i32>, i32, i32, i1, i1) #2
+declare float @llvm.amdgcn.buffer.load.f32(<4 x i32>, i32, i32, i1, i1) #3
+declare <4 x float> @llvm.amdgcn.image.sample.1d.v4f32.f32(i32, float, <8 x i32>, <4 x i32>, i1, i32, i32) #3
+declare <4 x float> @llvm.amdgcn.image.sample.2d.v4f32.f32(i32, float, float, <8 x i32>, <4 x i32>, i1, i32, i32) #3
+declare void @llvm.amdgcn.kill(i1) #1
+declare float @llvm.amdgcn.wqm.f32(float) #3
+declare float @llvm.amdgcn.softwqm.f32(float) #3
+declare i32 @llvm.amdgcn.softwqm.i32(i32) #3
+declare float @llvm.amdgcn.wwm.f32(float) #3
+
+attributes #1 = { nounwind }
+attributes #2 = { nounwind readonly }
+attributes #3 = { nounwind readnone }