tree c5f3edfc7d6f325910bdc1d5a4112aa1240853bf
parent 9080be216769444461f5f2c3a7d76a733a1e2e09
author Federico Bruzzone <federico.bruzzone.i@gmail.com> 1783424688 +0200
committer GitHub <noreply@github.com> 1783424688 +0000
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsFcBAABCAAQBQJqTOawCRC1aQ7uu5UhlAAAnVMQAESKpByy/9XKvq6wgdw1gKOE
 U5Llfa/GBjjUFrMaLtAKuqVsW4JRtTZ+GH9Oaa+GUa1/Y7P/QfvmIR31ODCAycmQ
 KmdLwWONI5XCR+Y3g05gm4Ozer/+SUFIW8itrsWi2GQdFTpn5qff1eT9Xae+P6Kq
 eo5rDIwXHiRN0xL7xbfUCkyoiLeFk9sP8rhqrIiQ7Wtw15It3NERUergTUy4HZNq
 yK4vGsiTysozhIOV4CttI36Vhu29oDq3vBs22RHO+yVLxcoMMJLqVVLI8qX8GEyA
 ju2mxZEID2V2YTm38NPna9mUNtqKgViXwuFLC5rsApmjfuH5vb93/P5nWDuJ62MB
 RCoxvRfGbCOrMvzfAtRhAi9QkNusPSuW0B7o5xKH30PcnwJ1LKeZhaRdx870L517
 iZqM39oZm3ahSfxEeYzn4Sdex+meVNs3sa5Ee7bR16MezcKE2RCA8LrgImR5+Q/L
 5EjB2OGkC6sL7U0A7HhBX49KgcazkE6ETfJOqM5RvhWLkuKhQUq+XnpnUX4FUEpj
 pvwtuoHWAnwN2zvZZlz2GqO+XwZGhrXyrdtoRie3LkBYe6zdM/2ApUg43MAktso4
 4NEZQC+Hs74zUygXyINjCTpxu6QDw7zAVThLLa/vlQbPENxq5lVAQn9vIUZ19L1V
 ysV8LreTrA8md5QHrJjh
 =zWHH
 -----END PGP SIGNATURE-----
 

[LLVMCPU] Fix ArmSME requiring classic SVE, breaking SME-only targets (e.g., Apple Silicon) (#24661)

`enableAArch64SME` required `hasAnySVEFeature` in addition to
`hasSMEFeature`, even though SME runs in Streaming SVE mode and does not
architecturally need classic (non-streaming) SVE. On SME-only targets
(e.g. Apple Silicon, which implements `FEAT_SME`/`FEAT_SME2` but not
`FEAT_SVE`), this meant:

- **`+sme` alone**: `enableAArch64SME` was `false`, so the ArmSME
lowering pipeline never ran. Vectorization silently fell back to scalar
code, but a leftover `vscale` was never inside a function marked
"streaming", so it had no legal instruction-selection pattern:
`iree-compile` crashed with `LLVM ERROR: Cannot select: vscale`.
- **`+sve,+sme` together** (the only combination the existing SME e2e
tests use): compiles fine, but a separate, Darwin-specific mechanism in
`AArch64FrameLowering`/`AArch64PrologueEpilogue` is gated on `hasSVE()`.
With `+sve` set, LLVM emits `cntd` directly in the function prologue
*before* `smstart`, which is illegal on hardware without non-streaming
SVE and crashes at runtime with `SIGILL`.

Neither combination worked; does that mean that, prior to this PR, it
wasn't possible to use IREE's SME support on this class of hardware? If
so, it implies that, until now, IREE couldn't leverage SME on Apple M4
Pro processors.
But I feel like I might be missing something...

Fixes #24660.

See: https://github.com/llvm/llvm-project/issues/204853


@banach-space @egebeysel @AGindinson @maxbartel

---------

Signed-off-by: Federico Bruzzone <federico.bruzzone.i@gmail.com>
Co-authored-by: Ege Beysel <beyselege@gmail.com>