AMDGPU: Move v_readlane lane select from VGPR to SGPR

[nhaehnle]

Summary

Fix a compiler bug when the lane select happens to end up in a VGPR.

Clarify the semantic of the corresponding intrinsic to be that of
the corresponding GLSL: the lane select must be uniform across a
wave front, otherwise results are undefined.

[arsenm]

lib/Target/AMDGPU/SIInstrInfo.cpp

2639

Also mention the source is assumed to be uniform?

2642

SReg32_XM0

2643

const reference

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

26

Can you add a GEP on workitem ID to ensure the scalar load optimization won't ever trigger on this

[nhaehnle]

Address review comments.

[arsenm]

LGTM

include/llvm/IR/IntrinsicsAMDGPU.td

 def int_amdgcn_fcmp :
   Intrinsic<[llvm_i64_ty], [llvm_anyfloat_ty, LLVMMatchType<0>, llvm_i32_ty],
             [IntrNoMem, IntrConvergent]>;
 
 def int_amdgcn_readfirstlane :
   GCCBuiltin<"__builtin_amdgcn_readfirstlane">,
   Intrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoMem, IntrConvergent]>;
 
+// The lane argument must be uniform across the currently active threads of the
+// current wave. Otherwise, the result is undefined.
 def int_amdgcn_readlane :
   GCCBuiltin<"__builtin_amdgcn_readlane">,
   Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem, IntrConvergent]>;
 
 //===----------------------------------------------------------------------===//
 // CI+ Intrinsics
 //===----------------------------------------------------------------------===//
 

lib/Target/AMDGPU/SIInstrInfo.cpp

if (Src0.isReg() && RI.isSGPRReg(MRI, Src0.getReg()))
       legalizeOpWithMove(MI, Src0Idx);
   }
 
   // VOP2 src0 instructions support all operand types, so we don't need to check
   // their legality. If src1 is already legal, we don't need to do anything.
   if (isLegalRegOperand(MRI, InstrDesc.OpInfo[Src1Idx], Src1))
     return;
 
+  // Special case: V_READLANE_B32 accepts only immediate or SGPR operands for
+  // lane select. Fix up using V_READFIRSTLANE, since we assume that the lane
+  // select is uniform.
+  if (Opc == AMDGPU::V_READLANE_B32 && Src1.isReg() &&
+      RI.isVGPR(MRI, Src1.getReg())) {
+    unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
+    const DebugLoc &DL = MI.getDebugLoc();
+    BuildMI(*MI.getParent(), MI, DL, get(AMDGPU::V_READFIRSTLANE_B32), Reg)
+        .add(Src1);
+    Src1.ChangeToRegister(Reg, false);
+    return;
+  }
+
   // We do not use commuteInstruction here because it is too aggressive and will
   // commute if it is possible. We only want to commute here if it improves
   // legality. This can be called a fairly large number of times so don't waste
   // compile time pointlessly swapping and checking legality again.
   if (HasImplicitSGPR || !MI.isCommutable()) {
     legalizeOpWithMove(MI, Src1Idx);
     return;
   }

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

 ; CHECK: v_mov_b32_e32 [[VVAL:v[0-9]]], 32
 ; CHECK: v_readlane_b32 s{{[0-9]+}}, [[VVAL]], s{{[0-9]+}}
 define amdgpu_kernel void @test_readlane_imm_sreg(i32 addrspace(1)* %out, i32 %src1) #1 {
   %readlane = call i32 @llvm.amdgcn.readlane(i32 32, i32 %src1)
   store i32 %readlane, i32 addrspace(1)* %out, align 4
   ret void
 }
 
+; CHECK-LABEL: {{^}}test_readlane_vregs:
+; CHECK: v_readfirstlane_b32 [[LANE:s[0-9]+]], v{{[0-9]+}}
+; CHECK: v_readlane_b32 s{{[0-9]+}}, v{{[0-9]+}}, [[LANE]]
+define amdgpu_kernel void @test_readlane_vregs(i32 addrspace(1)* %out, <2 x i32> addrspace(1)* %in) #1 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep.in = getelementptr <2 x i32>, <2 x i32> addrspace(1)* %in, i32 %tid
+  %args = load <2 x i32>, <2 x i32> addrspace(1)* %gep.in
+  %value = extractelement <2 x i32> %args, i32 0
+  %lane = extractelement <2 x i32> %args, i32 1
+  %readlane = call i32 @llvm.amdgcn.readlane(i32 %value, i32 %lane)
+  store i32 %readlane, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
 ; TODO: m0 should be folded.
 ; CHECK-LABEL: {{^}}test_readlane_m0_sreg:
 ; CHECK: s_mov_b32 m0, -1
 ; CHECK: s_mov_b32 [[COPY_M0:s[0-9]+]], m0
 ; CHECK: v_mov_b32_e32 [[VVAL:v[0-9]]], [[COPY_M0]]
 ; CHECK: v_readlane_b32 s{{[0-9]+}}, [[VVAL]], s{{[0-9]+}}
 define amdgpu_kernel void @test_readlane_m0_sreg(i32 addrspace(1)* %out, i32 %src1) #1 {
   %m0 = call i32 asm "s_mov_b32 m0, -1", "={M0}"()
   %readlane = call i32 @llvm.amdgcn.readlane(i32 %m0, i32 %src1)
   store i32 %readlane, i32 addrspace(1)* %out, align 4
   ret void
 }
 
 ; CHECK-LABEL: {{^}}test_readlane_imm:
 ; CHECK: v_readlane_b32 s{{[0-9]+}}, v{{[0-9]+}}, 32
 define amdgpu_kernel void @test_readlane_imm(i32 addrspace(1)* %out, i32 %src0) #1 {
   %readlane = call i32 @llvm.amdgcn.readlane(i32 %src0, i32 32) #0
   store i32 %readlane, i32 addrspace(1)* %out, align 4
   ret void
 }
 
+declare i32 @llvm.amdgcn.workitem.id.x() #2
+
 attributes #0 = { nounwind readnone convergent }
 attributes #1 = { nounwind }
+attributes #2 = { nounwind readnone }