Operand 2 is promoted even when promote_operands = [0, 1]

[bangtianliu]

What happened?

When compiling a convolution operation with a tuning spec, the compilation fails because the generated code uses 89088 bytes of shared memory, which exceeds the hardware limit of 65536 bytes for the gfx942 target (AMD MI300X).

Steps to reproduce your issue

1. Compile the MLIR template with the following command:

iree-compile template.mlir -o=tmp.vmfb \
  --iree-codegen-tuning-spec-path=spec.mlir \
  --iree-hal-target-device=hip \
  --iree-hip-target=gfx942 \
  --compile-from=executable-sources

2. link to template.mlir and spec.mlir

What component(s) does this issue relate to?

No response

Version information

IREE commit: af241f9524 (Integrate LLVM @ 356479191ca0)

Additional context

No response