tree 8076b7d7d7b0d99bad26f6baacb897d36903266b
parent 35b2070788e70693b3c04ecfd96b282c03a4bf26
author Federico Bruzzone <federico.bruzzone.i@gmail.com> 1784875478 +0200
committer GitHub <noreply@github.com> 1784875478 +0200
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsFcBAABCAAQBQJqYwnWCRC1aQ7uu5UhlAAA/WEQAK56wGOaiTxgI78cKA8+aySV
 LGa34gLHuCfOlsv+Grd6iKnBUUoN55zCptxcWtUSZbpNilYFAyUEoQY3S9dHZKiq
 h4ZDVsehRG/9wq9Z2n0zfkrYKsKOJhiORLg11QfziPVT9HHrCTxjzrBklrwfk6nt
 7+M43r25IaIkT7AABKUxk0dKMtbp2T9opFQqe4jB04gMlrbtF+hAZzVgkFM3paIE
 5dcZV/oM9a222GF1U/jI2/gGrSFKKtQGws/1YGUPmCQAIqC5yHPeNMNNS2i1zDcX
 RKk7E1en6LAc8N9s/Jy4QwVrv9mF+ZUUcYdTI29UK6o4/u3zA6IaX5rh5M2nDW7N
 8lmWew9Gey7NmyCgznzpeL63DbCbGujwmMmUJxjxPpKfmXZMBqYI1FJ8INJ9RpfT
 i4URO3znsNGQ29RZOZE5pr5E6E/uXYlB6XvyHqddkjrWOb9IiLfy2O+eYo+FbgMT
 DZGJ2NSoBr67MbfgS0Z6fh3pNQBdR7SM3nwmQdAns5q1C6x+wwUVxxoV7qkgIB4w
 holUvz7W8XRmt925KhTT7d4FMQ2uQt0P62AiM4WMY8RxU0UuA28ezAE0ZwJDOiK1
 IkcQkAedNbBcZ9b5fox616g7XKt1edxC12cLGT6mh6QYPwkYqH4cIanl0GrbcQvh
 ovSl5KogtoxRWtG1bt5O
 =KnUx
 -----END PGP SIGNATURE-----
 

[LLVMCPU] Fix scalable vectorization fallback on SME-only (no `+sve`) targets (#24693)

**Stacked** on #24656.  Follow-up to #24661 / #24660.

This PR aims to removes the TODO introduced in #24656.

#24661 fixed the ArmSME lowering pipeline to not require classic `+sve`
when SME tiling is used. This PR fixes the same underlying problem for
the *fallback* path, which is what's used when the matmul falls back to
it (e.g., `--iree-llvmcpu-disable-arm-sme-tiling`, or op types SME
doesn't support): compiling a scalable-vectorized matmul on a
`+sme`-only target with SME tiling disabled crashed LLVM instruction
selection with `Cannot select: vscale`.

The fix is scoped to a single function:
`getMatmulVectorSizesUsingFillRegisterFileHeuristic` only checked
`hasAnySVEFeature`, so on a `+sme`-only target it always fell back to
non-scalable NEON tiles. It now also picks scalable tiles when `+sme` is
present *and* the user has explicitly passed
`--iree-llvmcpu-force-arm-streaming`.

Net effect: 
1. Non-matmul ops: always Peeling.
4. `linalg.matmul`, `+sme-only`, `disable-arm-sme-tiling=true`, and
`force-arm-streaming=true`: still reaches SSVE (that's the actual fix),
but now goes through `Peeling + masked tail` instead of masking-only (as
in #24661).

Unrelated to #24689

---------

Signed-off-by: Federico Bruzzone <federico.bruzzone.i@gmail.com>