[PATCH] D88081: [AMDGPU] Move WQM Pass after MI Scheduler

Fri Nov 6 05:25:43 PST 2020

mceier added inline comments.

================
Comment at: llvm/lib/Target/AMDGPU/SIWholeQuadMode.cpp:653-657
+      MachineInstr *EndMI = LIS->getInstructionFromIndex(S->end.getBaseIndex());
+      assert(EndMI && "Segment does not end on valid instruction");
+      auto NextI = std::next(EndMI->getIterator());
+      SlotIndex Next = LIS->getInstructionIndex(*NextI);
       if (Next > LastIdx)
----------------
critson wrote:
> mceier wrote:
> > mceier wrote:
> > > mceier wrote:
> > > > critson wrote:
> > > > > mceier wrote:
> > > > > > Is this really correct ? CS:GO crashes with git version of llvm at line 657 (on Radeon 5700XT):
> > > > > > 
> > > > > > 
> > > > > > ```
> > > > > > #0  0x00007fad33ab57b0 in llvm::IndexListEntry::getIndex (this=0x0) at /var/tmp/portage/sys-devel/llvm-12.0.0.9999-r1/work/llvm/include/llvm/CodeGen/SlotIndexes.h:58
> > > > > > #1  0x00007fad33ab581d in llvm::SlotIndex::getIndex (this=0x7fad2ffa2960) at /var/tmp/portage/sys-devel/llvm-12.0.0.9999-r1/work/llvm/include/llvm/CodeGen/SlotIndexes.h:125
> > > > > > #2  0x00007fad33ab591d in llvm::SlotIndex::operator> (this=0x7fad2ffa2960, other=...) at /var/tmp/portage/sys-devel/llvm-12.0.0.9999-r1/work/llvm/include/llvm/CodeGen/SlotIndexes.h:187
> > > > > > #3  0x00007fad362e842c in (anonymous namespace)::SIWholeQuadMode::prepareInsertion (this=0x1460000, MBB=..., First=..., Last=..., PreferLast=false, SaveSCC=true)
> > > > > >     at /var/tmp/portage/sys-devel/llvm-12.0.0.9999-r1/work/llvm/lib/Target/AMDGPU/SIWholeQuadMode.cpp:657
> > > > > > #4  0x00007fad362e923c in (anonymous namespace)::SIWholeQuadMode::processBlock (this=0x1460000, MBB=..., LiveMaskReg=2147484277, isEntry=true)
> > > > > >     at /var/tmp/portage/sys-devel/llvm-12.0.0.9999-r1/work/llvm/lib/Target/AMDGPU/SIWholeQuadMode.cpp:859
> > > > > > #5  0x00007fad362ea140 in (anonymous namespace)::SIWholeQuadMode::runOnMachineFunction (this=0x1460000, MF=...)
> > > > > >     at /var/tmp/portage/sys-devel/llvm-12.0.0.9999-r1/work/llvm/lib/Target/AMDGPU/SIWholeQuadMode.cpp:1055
> > > > > > #6  0x00007fad33d15048 in llvm::MachineFunctionPass::runOnFunction (this=0x1460000, F=...) at /var/tmp/portage/sys-devel/llvm-12.0.0.9999-r1/work/llvm/lib/CodeGen/MachineFunctionPass.cpp:73
> > > > > > #7  0x00007fad3391406a in llvm::FPPassManager::runOnFunction (this=0x148ba80, F=...) at /var/tmp/portage/sys-devel/llvm-12.0.0.9999-r1/work/llvm/lib/IR/LegacyPassManager.cpp:1519
> > > > > > #8  0x00007fad352ab96b in (anonymous namespace)::CGPassManager::RunPassOnSCC (this=0x148bc40, P=0x148ba80, CurSCC=..., CG=..., CallGraphUpToDate=@0x7fad2ffa2d8d: true, DevirtualizedCall=@0x7fad2ffa2e30: false)
> > > > > >     at /var/tmp/portage/sys-devel/llvm-12.0.0.9999-r1/work/llvm/lib/Analysis/CallGraphSCCPass.cpp:178
> > > > > > #9  0x00007fad352ac4ed in (anonymous namespace)::CGPassManager::RunAllPassesOnSCC (this=0x148bc40, CurSCC=..., CG=..., DevirtualizedCall=@0x7fad2ffa2e30: false)
> > > > > >     at /var/tmp/portage/sys-devel/llvm-12.0.0.9999-r1/work/llvm/lib/Analysis/CallGraphSCCPass.cpp:476
> > > > > > #10 0x00007fad352ac7e4 in (anonymous namespace)::CGPassManager::runOnModule (this=0x148bc40, M=...) at /var/tmp/portage/sys-devel/llvm-12.0.0.9999-r1/work/llvm/lib/Analysis/CallGraphSCCPass.cpp:541
> > > > > > #11 0x00007fad339147f6 in (anonymous namespace)::MPPassManager::runOnModule (this=0x1412800, M=...) at /var/tmp/portage/sys-devel/llvm-12.0.0.9999-r1/work/llvm/lib/IR/LegacyPassManager.cpp:1634
> > > > > > #12 0x00007fad3390f6ce in llvm::legacy::PassManagerImpl::run (this=0x1430200, M=...) at /var/tmp/portage/sys-devel/llvm-12.0.0.9999-r1/work/llvm/lib/IR/LegacyPassManager.cpp:615
> > > > > > #13 0x00007fad33915087 in llvm::legacy::PassManager::run (this=0x1427bb8, M=...) at /var/tmp/portage/sys-devel/llvm-12.0.0.9999-r1/work/llvm/lib/IR/LegacyPassManager.cpp:1761
> > > > > > #14 0x00007fad3abc7145 in ac_compile_module_to_elf (p=0x1427b60, module=0x35af9300, pelf_buffer=0x111a5ad0, pelf_size=0x111a5ad8) at ../mesa-9999/src/amd/llvm/ac_llvm_helper.cpp:259
> > > > > > #15 0x00007fad3aad8501 in si_compile_llvm (sscreen=0xc23400, binary=0x111a5ad0, conf=0x111a5ae8, compiler=0xc23cb0, ac=0x7fad2ffa3490, debug=0x2577a030, stage=MESA_SHADER_FRAGMENT,
> > > > > >     name=0x7fad3ad7582a "Pixel Shader", less_optimized=false) at ../mesa-9999/src/gallium/drivers/radeonsi/si_shader_llvm.c:104
> > > > > > #16 0x00007fad3aad61d2 in si_llvm_compile_shader (sscreen=0xc23400, compiler=0xc23cb0, shader=0x111a5a00, debug=0x2577a030, nir=0xc5a6450, free_nir=false)
> > > > > >     at ../mesa-9999/src/gallium/drivers/radeonsi/si_shader.c:1891
> > > > > > #17 0x00007fad3aad634d in si_compile_shader (sscreen=0xc23400, compiler=0xc23cb0, shader=0x111a5a00, debug=0x2577a030) at ../mesa-9999/src/gallium/drivers/radeonsi/si_shader.c:1927
> > > > > > #18 0x00007fad3ab109c3 in si_init_shader_selector_async (job=0x2577a000, thread_index=0) at ../mesa-9999/src/gallium/drivers/radeonsi/si_state_shaders.c:2492
> > > > > > #19 0x00007fad3a73196b in util_queue_thread_func (input=0xc34ed0) at ../mesa-9999/src/util/u_queue.c:304
> > > > > > #20 0x00007fad3a730acb in impl_thrd_routine (p=0xc34ec0) at ../mesa-9999/include/c11/threads_posix.h:87
> > > > > > #21 0x00007fad3fbcbf9e in start_thread (arg=0x7fad2ffa5640) at pthread_create.c:463
> > > > > > #22 0x00007fad3ff4865f in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:95
> > > > > > ```
> > > > > Thanks for bringing this to my attention.  Sorry for the slow response.  I am currently investigating.  Do you happen to have any more details for reproduction?
> > > > The crash happens at CS:GO startup and unfortunately besides this and the stacktrace I don't have more details about it; I don't know how to extract the shader that triggers this crash (not even sure if it will help with anything). So far only CS:GO triggers it.
> > > > 
> > > > I wonder if the move of "insertPass" from line 1000 to 993 might be the culprit ? If so I could rebuild the llvm (I didn't do it yet because it takes a lot of time) and verify this.
> > > I tried changing the order of passes and CS:GO still crashed. Also I was able to dump LLVM IR of the shader (by setting AMD_DEBUG=ps,vs): {F13704476}
> > > 
> > > It's always shader 387 that crashes. Hopefully it will help you debug this.
> > > 
> > After commenting out top 2 lines, llc produces the same stacktrace as CS:GO.
> Thanks for this shader.
> Can you share you llc command line options as I was unable to get a crash feeding the shader to: llc -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs
Well, I just renamed shader4.txt to shader4.ll, commented out 2 top lines and ran llc shader4.ll and llc crashed. I didn't use any llc options.

"llc -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs shader4.ll" also doesn't crash here.

LLVM commit is f738aee0bbf39d11b9f0104e094c7893ffca040c

llc --version shows:

LLVM (http://llvm.org/):
  LLVM version 12.0.0git
  Optimized build.
  Default target: x86_64-pc-linux-gnu
  Host CPU: skylake

  Registered Targets:
    aarch64    - AArch64 (little endian)
    aarch64_32 - AArch64 (little endian ILP32)
    aarch64_be - AArch64 (big endian)
    amdgcn     - AMD GCN GPUs
    arm        - ARM
    arm64      - ARM64 (little endian)
    arm64_32   - ARM64 (little endian ILP32)
    armeb      - ARM (big endian)
    bpf        - BPF (host endian)
    bpfeb      - BPF (big endian)
    bpfel      - BPF (little endian)
    nvptx      - NVIDIA PTX 32-bit
    nvptx64    - NVIDIA PTX 64-bit
    r600       - AMD GPUs HD2XXX-HD6XXX
    riscv32    - 32-bit RISC-V
    riscv64    - 64-bit RISC-V
    thumb      - Thumb
    thumbeb    - Thumb (big endian)
    wasm32     - WebAssembly 32-bit
    wasm64     - WebAssembly 64-bit
    x86        - 32-bit X86: Pentium-Pro and above
    x86-64     - 64-bit X86: EM64T and AMD64

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D88081/new/

https://reviews.llvm.org/D88081