• Target / CPU: amdgcn-amd-amdhsa / gfx1250
• Toolchain:
  AMD clang version 23.0.0git (https://github.com/ROCm/llvm-project.git 43215c7+PATCHED:d17c5aa0e3ea29cde402f58f27e39b6034effa27)
  Target: x86_64-unknown-linux-gnu
  Thread model: posix
  InstalledDir: /opt/rocm-versions/gfx1250-7.14.0a20260521/lib/llvm/bin

Issue

gfx1250-assembler-bug-mx-scales.zip

For v_wmma_scale_f32_16x16x128_f8f6f4, the ISA constrains the legal
(matrix_a_fmt, matrix_a_scale_fmt, matrix_b_fmt, matrix_b_scale_fmt)
tuples (see table-valid-combinations.txt): matrix-format classes FP8/BF8
and FP6/BF6 require scale E8; class FP4 allows E8 / E5M3 / E4M3;
when both sides are class FP4, the two scales must match.

The integrated assembler ignores these joint constraints on the
matrix_*_fmt / matrix_*_scale_fmt modifiers and accepts arbitrary
combinations, emitting encodings that are not legal per the ISA. Out of
the 225 (A fmt, A scale, B fmt, B scale) tuples, only 43 are valid;
the other 182 are all silently accepted today.

Steps to reproduce

The reproducer is a single self-contained script with no arguments
required. From the directory containing it:

python3 enumerate.py --clean

enumerate.py enumerates all 225 tuples, runs each through llvm-mc
(/opt/rocm/llvm/bin/llvm-mc by default; override with --llvm-mc PATH
or $LLVM_MC), splits the cases into four buckets under results/
(valid-accepted/, valid-rejected/, invalid-accepted/,
invalid-rejected/) as individual .s files, and prints a summary.

Reproduce any single failing case (each .s is self-contained and headed):

/opt/rocm/llvm/bin/llvm-mc -triple=amdgcn-amd-amdhsa -mcpu=gfx1250 \
    -filetype=obj -o /dev/null \
    results/invalid-accepted/A-FP8-E8__B-FP8-E5M3.s ; echo $?
# 0   (expected: non-zero with a diagnostic on matrix_b_scale_fmt)

CI gate that flips when any fix lands:

N=$(ls results/invalid-accepted/*.s | wc -l)
fail=0
for f in results/invalid-accepted/*.s; do
  /opt/rocm/llvm/bin/llvm-mc -triple=amdgcn-amd-amdhsa -mcpu=gfx1250 \
      -filetype=obj -o /dev/null "$f" 2>/dev/null || fail=$((fail + 1))
done
[ "$fail" -eq 0 ] \
  && echo "STILL BUGGY: $N/$N invalid combinations accepted" \
  || echo "PROGRESS: $fail/$N now rejected"

Results / summary

enumerate.py output on the toolchain above:

============================================================
Summary  (total tested: 225)
============================================================
  valid-accepted    :  43   (expected behavior)
  invalid-rejected  :   0   (expected behavior)
  invalid-accepted  : 182   (BUG: assembler should reject)
  valid-rejected    :   0   (BUG: assembler should accept)
------------------------------------------------------------
  correct           :  43 / 225
  incorrect (bugs)  : 182 / 225

• 182 / 225 invalid (A fmt, A scale, B fmt, B scale) tuples are
  accepted; 0 / 225 are correctly rejected.
• 43 / 225 valid tuples are accepted (no false negatives on the
  valid side).
• All 182 bug repros are persisted as standalone .s files in
  results/invalid-accepted/; the CI gate above stays red until the
  assembler starts rejecting any of them.