Skip to content

AMDGPU DS cmpxhg fails machine verifier with AGPR inputs #155777

@arsenm

Description

@arsenm

The backend does not ensure that both data operands are AGPR or VGPR for DS atomic cmpxchg

# After SI Fold Operands
# Machine code for function ds_atomic_cmpxchg_b32_ret_a_a__a: IsSSA, TracksLiveness
Function Live Ins: $vgpr0 in %8

bb.0 (%ir-block.0):
  liveins: $vgpr0
  %8:vgpr_32 = COPY $vgpr0
  INLINEASM &"; def $0" [attdialect], $0:[regdef:AGPR_32], def %9:agpr_32
  INLINEASM &"; def $0" [attdialect], $0:[regdef:AGPR_32], def %10:agpr_32
  %13:vgpr_32 = COPY %9:agpr_32
  %12:vgpr_32 = DS_CMPST_RTN_B32_gfx9 %8:vgpr_32, %13:vgpr_32, %10:agpr_32, 40, 0, implicit $exec :: (load store seq_cst monotonic (s32) on %ir.gep.0, addrspace 3)
  %11:agpr_32 = COPY %12:vgpr_32
  INLINEASM &"; use $0" [attdialect], $0:[reguse:AGPR_32], %11:agpr_32
  SI_RETURN

# End machine code for function ds_atomic_cmpxchg_b32_ret_a_a__a.

*** Bad machine code: Invalid register class: both data operands should be VGPR or AGPR ***
- function:    ds_atomic_cmpxchg_b32_ret_a_a__a
- basic block: %bb.0  (0x15b85f5a8)
- instruction: %12:vgpr_32 = DS_CMPST_RTN_B32_gfx9 %8:vgpr_32, %13:vgpr_32, %10:agpr_32, 40, 0, implicit $exec :: (load store seq_cst monotonic (s32) on %ir.gep.0, addrspace 3)

; RUN: llc -mtriple=amdgcn -mcpu=gfx90a -verify-machineinstrs < %s | FileCheck %s

define void @ds_atomic_cmpxchg_b32_ret_av_av__av(ptr addrspace(3) %ptr) #0 {
  %gep.0 = getelementptr inbounds [512 x float], ptr addrspace(3) %ptr, i32 0, i32 10
  %data0 = call i32 asm "; def $0", "=^VA"()
  %data1 = call i32 asm "; def $0", "=^VA"()
  %pair = cmpxchg ptr addrspace(3) %gep.0, i32 %data0, i32 %data1 seq_cst monotonic
  %result = extractvalue { i32, i1 } %pair, 0
  call void asm "; use $0", "^VA"(i32 %result)
  ret void
}

define void @ds_atomic_cmpxchg_b32_ret_av_av__v(ptr addrspace(3) %ptr) #0 {
  %gep.0 = getelementptr inbounds [512 x float], ptr addrspace(3) %ptr, i32 0, i32 10
  %data0 = call i32 asm "; def $0", "=^VA"()
  %data1 = call i32 asm "; def $0", "=^VA"()
  %pair = cmpxchg ptr addrspace(3) %gep.0, i32 %data0, i32 %data1 seq_cst monotonic
  %result = extractvalue { i32, i1 } %pair, 0
  call void asm "; use $0", "v"(i32 %result)
  ret void
}

define void @ds_atomic_cmpxchg_b32_ret_av_av__a(ptr addrspace(3) %ptr) #0 {
  %gep.0 = getelementptr inbounds [512 x float], ptr addrspace(3) %ptr, i32 0, i32 10
  %data0 = call i32 asm "; def $0", "=^VA"()
  %data1 = call i32 asm "; def $0", "=^VA"()
  %pair = cmpxchg ptr addrspace(3) %gep.0, i32 %data0, i32 %data1 seq_cst monotonic
  %result = extractvalue { i32, i1 } %pair, 0
  call void asm "; use $0", "a"(i32 %result)
  ret void
}

; FIXME: Broken
define void @ds_atomic_cmpxchg_b32_ret_a_a__a(ptr addrspace(3) %ptr) #0 {
  %gep.0 = getelementptr inbounds [512 x float], ptr addrspace(3) %ptr, i32 0, i32 10
  %data0 = call i32 asm "; def $0", "=a"()
  %data1 = call i32 asm "; def $0", "=a"()
  %pair = cmpxchg ptr addrspace(3) %gep.0, i32 %data0, i32 %data1 seq_cst monotonic
  %result = extractvalue { i32, i1 } %pair, 0
  call void asm "; use $0", "a"(i32 %result)
  ret void
}

; FIXME: Broken
define void @ds_atomic_cmpxchg_b32_ret_a_a__v(ptr addrspace(3) %ptr) #0 {
  %gep.0 = getelementptr inbounds [512 x float], ptr addrspace(3) %ptr, i32 0, i32 10
  %data0 = call i32 asm "; def $0", "=a"()
  %data1 = call i32 asm "; def $0", "=a"()
  %pair = cmpxchg ptr addrspace(3) %gep.0, i32 %data0, i32 %data1 seq_cst monotonic
  %result = extractvalue { i32, i1 } %pair, 0
  call void asm "; use $0", "v"(i32 %result)
  ret void
}

; FIXME: Broken
define void @ds_atomic_cmpxchg_b32_ret_v_a__v(ptr addrspace(3) %ptr) #0 {
  %gep.0 = getelementptr inbounds [512 x float], ptr addrspace(3) %ptr, i32 0, i32 10
  %data0 = call i32 asm "; def $0", "=v"()
  %data1 = call i32 asm "; def $0", "=a"()
  %pair = cmpxchg ptr addrspace(3) %gep.0, i32 %data0, i32 %data1 seq_cst monotonic
  %result = extractvalue { i32, i1 } %pair, 0
  call void asm "; use $0", "v"(i32 %result)
  ret void
}

; FIXME: Broken
define void @ds_atomic_cmpxchg_b32_ret_a_v__v(ptr addrspace(3) %ptr) #0 {
  %gep.0 = getelementptr inbounds [512 x float], ptr addrspace(3) %ptr, i32 0, i32 10
  %data0 = call i32 asm "; def $0", "=a"()
  %data1 = call i32 asm "; def $0", "=v"()
  %pair = cmpxchg ptr addrspace(3) %gep.0, i32 %data0, i32 %data1 seq_cst monotonic
  %result = extractvalue { i32, i1 } %pair, 0
  call void asm "; use $0", "v"(i32 %result)
  ret void
}

define void @ds_atomic_cmpxchg_b32_ret_v_v__a(ptr addrspace(3) %ptr) #0 {
  %gep.0 = getelementptr inbounds [512 x float], ptr addrspace(3) %ptr, i32 0, i32 10
  %data0 = call i32 asm "; def $0", "=v"()
  %data1 = call i32 asm "; def $0", "=v"()
  %pair = cmpxchg ptr addrspace(3) %gep.0, i32 %data0, i32 %data1 seq_cst monotonic
  %result = extractvalue { i32, i1 } %pair, 0
  call void asm "; use $0", "a"(i32 %result)
  ret void
}

define void @ds_atomic_cmpxchg_b32_ret_av_v__av(ptr addrspace(3) %ptr) #0 {
  %gep.0 = getelementptr inbounds [512 x float], ptr addrspace(3) %ptr, i32 0, i32 10
  %data0 = call i32 asm "; def $0", "=^VA"()
  %data1 = call i32 asm "; def $0", "=v"()
  %pair = cmpxchg ptr addrspace(3) %gep.0, i32 %data0, i32 %data1 seq_cst monotonic
  %result = extractvalue { i32, i1 } %pair, 0
  call void asm "; use $0", "^VA"(i32 %result)
  ret void
}

define void @ds_atomic_cmpxchg_b32_ret_v_av__av(ptr addrspace(3) %ptr) #0 {
  %gep.0 = getelementptr inbounds [512 x float], ptr addrspace(3) %ptr, i32 0, i32 10
  %data0 = call i32 asm "; def $0", "=v"()
  %data1 = call i32 asm "; def $0", "=^VA"()
  %pair = cmpxchg ptr addrspace(3) %gep.0, i32 %data0, i32 %data1 seq_cst monotonic
  %result = extractvalue { i32, i1 } %pair, 0
  call void asm "; use $0", "^VA"(i32 %result)
  ret void
}

; FIXME: Broken
define void @ds_atomic_cmpxchg_b32_ret_av_a__av(ptr addrspace(3) %ptr) #0 {
  %gep.0 = getelementptr inbounds [512 x float], ptr addrspace(3) %ptr, i32 0, i32 10
  %data0 = call i32 asm "; def $0", "=^VA"()
  %data1 = call i32 asm "; def $0", "=a"()
  %pair = cmpxchg ptr addrspace(3) %gep.0, i32 %data0, i32 %data1 seq_cst monotonic
  %result = extractvalue { i32, i1 } %pair, 0
  call void asm "; use $0", "^VA"(i32 %result)
  ret void
}

define void @ds_atomic_cmpxchg_b32_ret_a_av__av(ptr addrspace(3) %ptr) #0 {
  %gep.0 = getelementptr inbounds [512 x float], ptr addrspace(3) %ptr, i32 0, i32 10
  %data0 = call i32 asm "; def $0", "=a"()
  %data1 = call i32 asm "; def $0", "=^VA"()
  %pair = cmpxchg ptr addrspace(3) %gep.0, i32 %data0, i32 %data1 seq_cst monotonic
  %result = extractvalue { i32, i1 } %pair, 0
  call void asm "; use $0", "^VA"(i32 %result)
  ret void
}

attributes #0 = { nounwind "amdgpu-waves-per-eu"="10,10" }

Metadata

Metadata

Assignees

No one assigned

    Type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions