AMDGPU: Fix an out of date assert in addressing FrameIndex

[cfang]

Summary

Should use StachPtrOffsetReg instead of FrameOffsetReg to address FrameIndex.

[arsenm]

test/CodeGen/AMDGPU/stack-pointer-offset-relative-frameindex.ll

5 ↗

(On Diff #220176)

You should be able to reduce this. You don't need all of these types defined

24 ↗

(On Diff #220176)

This isn't a useful check

[cfang]

test/CodeGen/AMDGPU/stack-pointer-offset-relative-frameindex.ll

5 ↗

(On Diff #220176)

But if I replace these defined types of struct with type of float or int for example, I can not reproduce the assert.

24 ↗

(On Diff #220176)

What do you expect to check for a case of assert? I just wanted to avoid irrelevant checks.

[arsenm]

test/CodeGen/AMDGPU/stack-pointer-offset-relative-frameindex.ll

5 ↗

(On Diff #220176)

You can always avoid the struct types. You may also switch to a MIR test

24 ↗

(On Diff #220176)

You can check the stack accesses or generate the checks

[cfang]

1. Remove the use of structs to simplify the test
2. generate llc checks with utils/update_llc_test_checks.py

[arsenm]

test/CodeGen/AMDGPU/stack-pointer-offset-relative-frameindex.ll

45 ↗

(On Diff #222017)

Probably should make this defined

57 ↗

(On Diff #222017)

Probably should not use unreachable

[cfang]

update the test based on the comment:

1. use a defined comparison
2. remove "unreachable", and make it branch to exit.

[arsenm]

LGTM

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

for (const MachineOperand &MO: MI.operands()) {
     }
   }
 #endif
 
   MachineOperand *FIOp = TII->getNamedOperand(MI, AMDGPU::OpName::vaddr);
   assert(FIOp && FIOp->isFI() && "frame index must be address operand");
   assert(TII->isMUBUF(MI));
   assert(TII->getNamedOperand(MI, AMDGPU::OpName::soffset)->getReg() ==
-         MF->getInfo<SIMachineFunctionInfo>()->getFrameOffsetReg() &&
-         "should only be seeing frame offset relative FrameIndex");
-
+         MF->getInfo<SIMachineFunctionInfo>()->getStackPtrOffsetReg() &&
+         "should only be seeing stack pointer offset relative FrameIndex");
 
   MachineOperand *OffsetOp = TII->getNamedOperand(MI, AMDGPU::OpName::offset);
   int64_t NewOffset = OffsetOp->getImm() + Offset;
   assert(isUInt<12>(NewOffset) && "offset should be legal");
 
   FIOp->ChangeToRegister(BaseReg, false);
   OffsetOp->setImm(NewOffset);
 }

llvm/trunk/test/CodeGen/AMDGPU/stack-pointer-offset-relative-frameindex.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs | FileCheck -check-prefix=GCN %s
+
+; An assert was hit when frame offset register was used to address FrameIndex.
+define amdgpu_kernel void @kernel_background_evaluate(float addrspace(5)* %kg, <4 x i32> addrspace(1)* %input, <4 x float> addrspace(1)* %output, i32 %i) {
+; GCN-LABEL: kernel_background_evaluate:
+; GCN:       ; %bb.0: ; %entry
+; GCN-NEXT:    s_load_dword s6, s[0:1], 0x24
+; GCN-NEXT:    s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GCN-NEXT:    s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GCN-NEXT:    s_mov_b32 s38, -1
+; GCN-NEXT:    s_mov_b32 s39, 0x31c16000
+; GCN-NEXT:    s_mov_b32 s33, s3
+; GCN-NEXT:    s_mov_b64 s[0:1], s[36:37]
+; GCN-NEXT:    v_mov_b32_e32 v1, 0x2000
+; GCN-NEXT:    v_mov_b32_e32 v2, 0x4000
+; GCN-NEXT:    s_mov_b64 s[2:3], s[38:39]
+; GCN-NEXT:    v_mov_b32_e32 v3, 0
+; GCN-NEXT:    v_mov_b32_e32 v4, 0x400000
+; GCN-NEXT:    s_add_u32 s32, s33, 0xc0000
+; GCN-NEXT:    v_add_nc_u32_e64 v32, 4, 0x4000
+; GCN-NEXT:    ; implicit-def: $vcc_hi
+; GCN-NEXT:    s_getpc_b64 s[4:5]
+; GCN-NEXT:    s_add_u32 s4, s4, svm_eval_nodes@rel32@lo+4
+; GCN-NEXT:    s_addc_u32 s5, s5, svm_eval_nodes@rel32@hi+4
+; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    v_mov_b32_e32 v0, s6
+; GCN-NEXT:    s_swappc_b64 s[30:31], s[4:5]
+; GCN-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GCN-NEXT:    s_and_saveexec_b32 s0, vcc_lo
+; GCN-NEXT:    ; mask branch BB0_2
+; GCN-NEXT:    s_cbranch_execz BB0_2
+; GCN-NEXT:  BB0_1: ; %if.then4.i
+; GCN-NEXT:    buffer_load_dword v0, v32, s[36:39], s32 offen
+; GCN-NEXT:    buffer_load_dword v1, v32, s[36:39], s32 offen offset:4
+; GCN-NEXT:    s_waitcnt vmcnt(0)
+; GCN-NEXT:    v_add_nc_u32_e32 v0, v1, v0
+; GCN-NEXT:    v_mul_lo_u32 v0, 0x41c64e6d, v0
+; GCN-NEXT:    v_add_nc_u32_e32 v0, 0x3039, v0
+; GCN-NEXT:    buffer_store_dword v0, v0, s[36:39], s33 offen
+; GCN-NEXT:  BB0_2: ; %shader_eval_surface.exit
+; GCN-NEXT:    s_endpgm
+entry:
+  %sd = alloca < 1339 x i32>, align 16, addrspace(5)
+  %state = alloca <4 x i32>, align 4, addrspace(5)
+  %rslt = call i32 @svm_eval_nodes(float addrspace(5)* %kg, <1339 x i32> addrspace(5)* %sd, <4 x i32> addrspace(5)* %state, i32 0, i32 4194304)
+  %cmp = icmp eq i32 %rslt, 0
+  br i1 %cmp, label %shader_eval_surface.exit, label %if.then4.i
+
+if.then4.i:                                       ; preds = %entry
+  %rng_hash.i.i = getelementptr inbounds < 4 x i32>, <4 x i32> addrspace(5)* %state, i32 0, i32 1
+  %tmp0 = load i32, i32 addrspace(5)* %rng_hash.i.i, align 4
+  %rng_offset.i.i = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(5)* %state, i32 0, i32 2
+  %tmp1 = load i32, i32 addrspace(5)* %rng_offset.i.i, align 4
+  %add.i.i = add i32 %tmp1, %tmp0
+  %add1.i.i = add i32 %add.i.i, 0
+  %mul.i.i.i.i = mul i32 %add1.i.i, 1103515245
+  %add.i.i.i.i = add i32 %mul.i.i.i.i, 12345
+  store i32 %add.i.i.i.i, i32 addrspace(5)* undef, align 16
+  br label %shader_eval_surface.exit
+
+shader_eval_surface.exit:                         ; preds = %entry
+  ret void
+}
+
+declare hidden i32 @svm_eval_nodes(float addrspace(5)*, <1339 x i32> addrspace(5)*, <4 x i32> addrspace(5)*, i32, i32) local_unnamed_addr