AMDGPU/SI: Implement GroupStaticSize Intrinsic for Dynamic LDS

[cfang]

Summary

Static LDS size is saved in MachineFunctionInfo::LDSSize,
We define a pseudo instruction with usesCustomInserter bit set. Then, in EmitInstrWithCustomInserter,
we replace this pseudo instruction with a mov of MachineFunctionInfo::LDSSize.

[• tstellarAMD]

lib/Target/AMDGPU/SIISelLowering.cpp

1064

I don't think this will work if we call this intrinsic in the entry block and then the LDS variables are used in different blocks. Can you add a test case for this and then fix it if it doesn't work.

[arsenm]

include/llvm/IR/IntrinsicsAMDGPU.td

219–220

I think the get should be dropped from the name

[arsenm]

lib/Target/AMDGPU/SIISelLowering.cpp

1064

This is ok because this is called by the ExpandISelPseudos pass which runs on the completed MachineFunction

[cfang]

updated based on Tom and Matt's comments:

1. drop "get_" in the intrinsic name;
2. add new test to test multiple block case.

[• tstellarAMD]

LGTM.

include/llvm/IR/IntrinsicsAMDGPU.td

   GCCBuiltin<"__builtin_amdgcn_mbcnt_hi">,
   Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
 
+def int_amdgcn_groupstaticsize :
+  GCCBuiltin<"__builtin_amdgcn_groupstaticsize">,
+  Intrinsic<[llvm_i32_ty], [], [IntrNoMem]>;
+
 //===----------------------------------------------------------------------===//
 // CI+ Intrinsics
 //===----------------------------------------------------------------------===//

lib/Target/AMDGPU/SIISelLowering.cpp

     MachineInstr * MI, MachineBasicBlock * BB) const {
 
   switch (MI->getOpcode()) {
-  default:
-    return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
   case AMDGPU::BRANCH:
     return BB;
+  case AMDGPU::GET_GROUPSTATICSIZE: {
+    const SIInstrInfo *TII =
+      static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
+    MachineFunction *MF = BB->getParent();
+    SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
+    DebugLoc DL = MI->getDebugLoc();
+    BuildMI (*BB, MI, DL, TII->get(AMDGPU::S_MOVK_I32))
+      .addOperand(MI->getOperand(0))
+      .addImm(MFI->LDSSize);
+    MI->eraseFromParent();
+    return BB;
+  }
+  default:
+    return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
   }
   return BB;
 }

lib/Target/AMDGPU/SIInstructions.td

 def SGPR_USE : InstSI <(outs),(ins), "", []>;
 }
 
+let usesCustomInserter = 1, SALU = 1 in {
+def GET_GROUPSTATICSIZE : InstSI <(outs SReg_32:$sdst), (ins), "",
+  [(set SReg_32:$sdst, (int_amdgcn_groupstaticsize))]>;
+} // End let usesCustomInserter = 1, SALU = 1
+
 // SI pseudo instructions. These are used by the CFG structurizer pass
 // and should be lowered to ISA instructions prior to codegen.
 

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

+; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck %s
+; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=kaveri -verify-machineinstrs < %s | FileCheck %s
+; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=fiji -verify-machineinstrs < %s | FileCheck %s
+
+
+@lds0 = addrspace(3) global [512 x float] undef, align 4
+@lds1 = addrspace(3) global [256 x float] undef, align 4
+
+; FUNC-LABEL: {{^}}groupstaticsize_test0:
+; CHECK: s_movk_i32 s{{[0-9]+}}, 0x800
+define void @get_groupstaticsize_test0(float addrspace(1)* %out, i32 addrspace(1)* %lds_size) #0 {
+  %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
+  %idx.0 = add nsw i32 %tid.x, 64
+  %static_lds_size = call i32 @llvm.amdgcn.groupstaticsize() #1
+  store i32 %static_lds_size, i32 addrspace(1)* %lds_size, align 4
+  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds0, i32 0, i32 %idx.0
+  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
+  store float %val0, float addrspace(1)* %out, align 4
+
+  ret void
+}
+
+
+; FUNC-LABEL: {{^}}groupstaticsize_test1:
+; CHECK: s_movk_i32 s{{[0-9]+}}, 0xc00
+define void @groupstaticsize_test1(float addrspace(1)* %out, i32 %cond, i32 addrspace(1)* %lds_size) {
+entry:
+  %static_lds_size = call i32 @llvm.amdgcn.groupstaticsize() #1
+  store i32 %static_lds_size, i32 addrspace(1)* %lds_size, align 4
+  %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
+  %idx.0 = add nsw i32 %tid.x, 64
+  %tmp = icmp eq i32 %cond, 0
+  br i1 %tmp, label %if, label %else
+
+if:                                               ; preds = %entry
+  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds0, i32 0, i32 %idx.0
+  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
+  store float %val0, float addrspace(1)* %out, align 4
+  br label %endif
+
+else:                                             ; preds = %entry
+  %arrayidx1 = getelementptr inbounds [256 x float], [256 x float] addrspace(3)* @lds1, i32 0, i32 %idx.0
+  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
+  store float %val1, float addrspace(1)* %out, align 4
+  br label %endif
+
+endif:                                            ; preds = %else, %if
+  ret void
+}
+
+
+declare i32 @llvm.amdgcn.groupstaticsize() #1
+declare i32 @llvm.amdgcn.workitem.id.x() #1
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind readnone }