-O0 flag broken

[michel2323]

using AMDGPU
kernel() = (AMDGPU.trap(); nothing)
AMDGPU.@device_code_llvm wait(@roc kernel())

This runs fine with default julia flags but segfaults using no optimization -O0. @device_code_llvm output is the same in both cases:

# default flags and -O0
; CompilerJob of kernel kernel() for GPUCompiler.GCNCompilerTarget
;  @ /gpfs/alpine/csc359/scratch/mschanen/julia_depot_crusher/dev/AMDGPU/mwe.jl:4 within `kernel`
define amdgpu_kernel void @_Z16julia_kernel_968() local_unnamed_addr #1 {
entry:
; ┌ @ /gpfs/alpine/csc359/scratch/mschanen/julia_depot_crusher/dev/AMDGPU/src/device/llvm.jl:3 within `trap`
   call void @llvm.trap()
; └
  unreachable
}

This is on branch ms/julia-1.8 and

Julia Version 1.8.0-DEV.1490
Commit 94bbfd45f9 (2022-02-10 14:02 UTC)
Platform Info:
  OS: Linux (x86_64-suse-linux)
  CPU: AMD EPYC 7A53 64-Core Processor
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-13.0.0 (ORCJIT, znver3)
Environment:
  LD_LIBRARY_PATH = /opt/rocm-4.5.0/lib64:/opt/rocm-4.5.0/lib:/opt/rocm-4.5.0/hsa/lib:/opt/rocm-4.5.0/llvm/lib:/opt/cray/pe/papi/6.0.0.12/lib64:/opt/cray/libfabric/1.13.0.0/lib64
  JULIA_FOLDER = /gpfs/alpine/scratch/mschanen/csc359/julia
  JULIA_AMDGPU_DISABLE_ARTIFACTS = 1
  JULIA_DEPOT_PATH = /gpfs/alpine/scratch/mschanen/csc359/julia_depot_crusher
  JULIA_MPI_PATH = /opt/cray/pe/mpich/8.1.12/ofi/amd/4.4
Project AMDGPU v0.2.17
Status `/gpfs/alpine/csc359/scratch/mschanen/julia_depot_crusher/dev/AMDGPU/Project.toml`
  [621f4979] AbstractFFTs v1.1.0
  [79e6a3ab] Adapt v3.3.3
  [b99e7846] BinaryProvider v0.5.10
  [fa961155] CEnum v0.4.1
  [e2ba6199] ExprTools v0.1.8
  [0c68f7d7] GPUArrays v8.2.1
  [61eb1bfa] GPUCompiler v0.13.11
  [929cbde3] LLVM v4.7.1
  [1914dd2f] MacroTools v0.5.9
  [ae029012] Requires v1.3.0
⌅ [efcf1570] Setfield v0.7.1
  [276daf66] SpecialFunctions v2.1.2
  [2696aab5] HIP_jll v4.2.0+0
  [873c0968] ROCmDeviceLibs_jll v4.2.0+0
  [dd59ff1a] hsa_rocr_jll v4.2.0+1
  [a6151927] rocRAND_jll v4.2.0+0
  [8f399da3] Libdl
  [37e2e46d] LinearAlgebra
  [44cfe95a] Pkg v1.8.0
  [de0858da] Printf
  [9a3f8284] Random
  [10745b16] Statistics
Info Packages marked with ⌅ have new versions available but cannot be upgraded. To see why use `status --outdated`

[jpsamaroo]

I can reproduce with 1.7.2 and jps/julia-1.7; the generated GCN seems fine, and running with --check-bounds=yes --inline=no doesn't produce any more meaningful info.

[jpsamaroo]

This is due to TLS access in AMDGPU.queue_error_handler, which is eliminated on a non--O0 configuration. We can possibly prevent this function from reaching @cfunction with JuliaLang/julia#43747, but it won't fix the fact that we need to eliminate the TLS access.

[jpsamaroo]

@vchuravy emission and non-elision of the "error in type bounds" check appears to be causing this.

[jpsamaroo]

Nope, sorry, it's a typeassert