tree 43905505e4cabcd4f19dabb9308c64e06a92b8bf
parent b48a5955575cf8924ffc191f3b0804976e5738b9
author Ege Beysel <beysel@roofline.ai> 1787425577 +0200
committer GitHub <noreply@github.com> 1787425577 +0200
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsFcBAABCAAQBQJqifMpCRC1aQ7uu5UhlAAAgJYQAKq/0NuejdtEynI8Q7E/Lgui
 c5m0KzDDjgBrg0clk9qFDLKAFCBxx/bnbBfCX+1Ij0zB+hHCJm88cOA3frhDuNKI
 qFcDgGwv8o8MM9csDLQegGYIkuI1xuJzJFCCHqmOhDvdyGHWvnjO64wLPlgER0vH
 mTTopFXbHfCLOBgkghTG5J3YcfOfXu0Pl1Z8Lu4PNsDYyzdt7EVMq6Q9b9jPgmx1
 VW5ffoDIcm6Xrfg+GlZ6zhM8wLiRFBGSZ67m2osWI07p0WAX1qb5Gb4Xa7a5XjEU
 nFHOmCpdNdALCEkbz5YQgg/bUB/BlGXfCZjMsjOAaFong9UZnfd81uTWcyXpTHD9
 qWjbnO3uBne/lGxSofrCYAwxjTkmoggouvVJnI2CMjK3pwFZ4DhrwP3NiJrg1Wqf
 1dhnYaksWCDrDhIeAqpfgTgIZGnHm5C9eHQGza6NDD0pMNo9xihwOkvLSQpZQS9B
 klLatwtNs4gH7mmlFL5JlN53sYp5jnbJEWo/cRX1vPqfPLa4FI0zSaAlWpTOvoQu
 BK6Dl09hrRhA7IJkR4iTw5CtsLmbaHQPTcWjGEcX94e2kL6muAb2X/n3rJ2Eyz0+
 1SOQlQCxVa9QC08IwVciFU/Fl+9M6ki2SoqtftqupmrX+De1nB3IYC+rL9G8cfsX
 Q7IZvLdyr5/Et3NrANSk
 =DzLK
 -----END PGP SIGNATURE-----
 

[LLVMCPU] strip per-function target attributes to allow inlining ukernels (#24833)

The ukernel bitcode carries per-function
target-cpu/tune-cpu/target-features attributes that conflict with the
attribute-less dispatch (which inherits the module-level TargetMachine)
and stop LLVM from inlining the ukernel. Strip them (and vscale_range,
so RVV bitcode inlines) at link time.

Fixes #24755.

Assisted-by: Claude Code

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