AMDGPU: Fix legalization of MUBUF instructions in shaders

Summary:
The addr64-based legalization is incorrect for MUBUF instructions with idxen
set as well as for BUFFER_LOAD/STORE_FORMAT_* instructions.  This affects
e.g.  shaders that access buffer textures.

Since we never actually need the addr64-legalization in shaders, this patch
takes the easy route and keys off the calling convention.  If this ever
affects (non-OpenGL) compute, the type of legalization needs to be chosen
based on some TSFlag.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98664

Reviewers: arsenm, tstellarAMD

Subscribers: kzhuravl, wdng, yaxunl, tony-tye, llvm-commits

Differential Revision: https://reviews.llvm.org/D26747

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@287339 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/AMDGPU/SIInstrInfo.cpp

@@ -2580,7 +2580,8 @@ void SIInstrInfo::legalizeGenericOperand(MachineBasicBlock &InsertMBB,
 }
 
 void SIInstrInfo::legalizeOperands(MachineInstr &MI) const {
-  MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
+  MachineFunction &MF = *MI.getParent()->getParent();
+  MachineRegisterInfo &MRI = MF.getRegInfo();
 
   // Legalize VOP2
   if (isVOP2(MI) || isVOPC(MI)) {
@@ -2690,8 +2691,14 @@ void SIInstrInfo::legalizeOperands(MachineInstr &MI) const {
     return;
   }
 
-  // Legalize MIMG
-  if (isMIMG(MI)) {
+  // Legalize MIMG and MUBUF/MTBUF for shaders.
+  //
+  // Shaders only generate MUBUF/MTBUF instructions via intrinsics or via
+  // scratch memory access. In both cases, the legalization never involves
+  // conversion to the addr64 form.
+  if (isMIMG(MI) ||
+      (AMDGPU::isShader(MF.getFunction()->getCallingConv()) &&
+       (isMUBUF(MI) || isMTBUF(MI)))) {
     MachineOperand *SRsrc = getNamedOperand(MI, AMDGPU::OpName::srsrc);
     if (SRsrc && !RI.isSGPRClass(MRI.getRegClass(SRsrc->getReg()))) {
       unsigned SGPR = readlaneVGPRToSGPR(SRsrc->getReg(), MI, MRI);
@@ -2706,9 +2713,10 @@ void SIInstrInfo::legalizeOperands(MachineInstr &MI) const {
     return;
   }
 
-  // Legalize MUBUF* instructions
+  // Legalize MUBUF* instructions by converting to addr64 form.
   // FIXME: If we start using the non-addr64 instructions for compute, we
-  // may need to legalize them here.
+  // may need to legalize them as above. This especially applies to the
+  // buffer_load_format_* variants and variants with idxen (or bothen).
   int SRsrcIdx =
       AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::srsrc);
   if (SRsrcIdx != -1) {

test/CodeGen/AMDGPU/mubuf-shader-vgpr.ll

@@ -0,0 +1,49 @@
+;RUN: llc < %s -march=amdgcn -mcpu=verde -verify-machineinstrs | FileCheck %s -check-prefix=CHECK
+;RUN: llc < %s -march=amdgcn -mcpu=tonga -verify-machineinstrs | FileCheck %s -check-prefix=CHECK
+
+; Test that buffer_load_format with VGPR resource descriptor is properly
+; legalized.
+
+; CHECK-LABEL: {{^}}test_none:
+; CHECK: buffer_load_format_x v0, off, {{s\[[0-9]+:[0-9]+\]}}, 0{{$}}
+define amdgpu_vs float @test_none(<4 x i32> addrspace(2)* inreg %base, i32 %i) {
+main_body:
+  %ptr = getelementptr <4 x i32>, <4 x i32> addrspace(2)* %base, i32 %i
+  %tmp2 = load <4 x i32>, <4 x i32> addrspace(2)* %ptr, align 32
+  %tmp7 = call float @llvm.amdgcn.buffer.load.format.f32(<4 x i32> %tmp2, i32 0, i32 0, i1 0, i1 0)
+  ret float %tmp7
+}
+
+; CHECK-LABEL: {{^}}test_idxen:
+; CHECK: buffer_load_format_x v0, {{v[0-9]+}}, {{s\[[0-9]+:[0-9]+\]}}, 0 idxen{{$}}
+define amdgpu_vs float @test_idxen(<4 x i32> addrspace(2)* inreg %base, i32 %i) {
+main_body:
+  %ptr = getelementptr <4 x i32>, <4 x i32> addrspace(2)* %base, i32 %i
+  %tmp2 = load <4 x i32>, <4 x i32> addrspace(2)* %ptr, align 32
+  %tmp7 = call float @llvm.amdgcn.buffer.load.format.f32(<4 x i32> %tmp2, i32 undef, i32 0, i1 0, i1 0)
+  ret float %tmp7
+}
+
+; CHECK-LABEL: {{^}}test_offen:
+; CHECK: buffer_load_format_x v0, {{v[0-9]+}}, {{s\[[0-9]+:[0-9]+\]}}, 0 offen{{$}}
+define amdgpu_vs float @test_offen(<4 x i32> addrspace(2)* inreg %base, i32 %i) {
+main_body:
+  %ptr = getelementptr <4 x i32>, <4 x i32> addrspace(2)* %base, i32 %i
+  %tmp2 = load <4 x i32>, <4 x i32> addrspace(2)* %ptr, align 32
+  %tmp7 = call float @llvm.amdgcn.buffer.load.format.f32(<4 x i32> %tmp2, i32 0, i32 undef, i1 0, i1 0)
+  ret float %tmp7
+}
+
+; CHECK-LABEL: {{^}}test_both:
+; CHECK: buffer_load_format_x v0, {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 idxen offen{{$}}
+define amdgpu_vs float @test_both(<4 x i32> addrspace(2)* inreg %base, i32 %i) {
+main_body:
+  %ptr = getelementptr <4 x i32>, <4 x i32> addrspace(2)* %base, i32 %i
+  %tmp2 = load <4 x i32>, <4 x i32> addrspace(2)* %ptr, align 32
+  %tmp7 = call float @llvm.amdgcn.buffer.load.format.f32(<4 x i32> %tmp2, i32 undef, i32 undef, i1 0, i1 0)
+  ret float %tmp7
+}
+
+declare float @llvm.amdgcn.buffer.load.format.f32(<4 x i32>, i32, i32, i1, i1) nounwind readonly
+
+attributes #0 = { nounwind readnone }