Skip to content

NOTES.md

Latest commit

 

History

History
110 lines (83 loc) · 3.87 KB

NOTES.md

File metadata and controls

110 lines (83 loc) · 3.87 KB
Raw

m114: AMDGPUPromoteKernelArguments unconditionally addrspacecasts flat kernarg pointers to global, silently miscompiling LDS/private-aperture inputs

Discovery method: code inspection. Sibling shape to m088/m091 -- a mid-level pass making an unwarranted assumption about kernarg semantics.

The bug

amdgpu/third_party/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUPromoteKernelArguments.cpp:105-128 (promoteFlatArgument):

// For every kernarg with type ptr in FLAT_ADDRESS:
auto *Cast0 = new AddrSpaceCastInst(P, GlobalPtrTy, "", BB->getFirstInsertionPt());
auto *Cast1 = new AddrSpaceCastInst(Cast0, FlatPtrTy, "", BB->getFirstInsertionPt());
P->replaceUsesWithIf(Cast1, [&](Use &U) { ... });

The pass wraps every FLAT_ADDRESS kernel-arg pointer in addrspacecast(addrspacecast(p to ptr addrspace(1)) to ptr) so InferAddressSpaces (run next in the pipeline) will rewrite the downstream memops to global_*.

There is NO check that the flat pointer is actually in the global aperture. A flat kernarg can legitimately carry a pointer whose aperture is LOCAL (AS 3) or PRIVATE (AS 5) -- e.g. the host sets addrspacecast(@LDS to ptr) and passes that into the kernel as a flat ptr.

Per LangRef:

If the source value's address space and the destination address space are not in the same address space hierarchy, the resulting pointer value is undefined behavior.

AMDGPU's flat aperture is a single virtual range; casting flat -> global strips the aperture base, so the resulting "global" address is garbage. The subsequent global_store then hits arbitrary global memory instead of the LDS / scratch slot.

Reproducer

reduced.ll:

define amdgpu_kernel void @k(ptr addrspace(1) readonly %pp, ptr %dst) {
  store i32 7, ptr %dst, align 4
  ret void
}

opt -S -passes='amdgpu-promote-kernel-arguments,infer-address-spaces' reduced.ll:

%0 = addrspacecast ptr %dst to ptr addrspace(1)
store i32 7, ptr addrspace(1) %0, align 4

Then llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950:

global_store_dword v0, v1, s[0:1]    ; original would have been flat_store

The original flat_store_dword would have correctly inspected the aperture and dispatched to LDS / scratch / global. The promoted global_store_dword bypasses the flat aperture entirely.

The companion case repro_local.ll (kernarg slot holds an LDS-aperture flat ptr loaded from another global) exhibits the same promotion -- also miscompiled.

Why this matters for default pipeline

AMDGPUPromoteKernelArguments runs at -O2 for the SDAG path on every AMDGPU target (registered in AMDGPUTargetMachine.cpp). Any client that passes flat pointers into a kernel with a non-global aperture is affected.

Suggested fix

Skip the cast unless we can prove the loaded/argument pointer is in the global aperture. Options:

  1. Restrict to kernargs that have an explicit aperture-asserting attribute (e.g. !amdgpu.global metadata or a new addrspace(1) constraint).
  2. Use InferAddressSpaces's own analysis: only promote when the flat pointer's aperture can be proven GLOBAL via the existing flat-AS resolver.
  3. Conservative: gate on a target opt-in flag for clients that guarantee all flat kernargs are global.

Toolchain Results

Toolchain Result
LLVM HEAD with the local PR patches (build/llvm-fuzzer) Reproduces.
ROCm 7.1.1 (/opt/rocm-7.1.1/lib/llvm/bin/opt) Same pass, same bug.

Why the fuzzer hasn't caught it

  • The current IR fuzzer does not emit kernargs that are flat-typed pointers pointing into LDS / private. All kernarg pointers are ptr addrspace(1) (global) directly.
  • The interpreter oracle assumes flat kernargs are global.
  • Per MEMORY.md (Prefer-random-over-idioms), the right hook is to emit kernargs typed as ptr (flat) with a runtime addrspacecast from @LDS symbols so the pass sees aperture diversity.