tree ed8ff441568533010d3501ea65e5f07ab236a54e
parent 466e812764beb3eeafebba5e5a45324cd40b5341
author Ben Vanik <ben.vanik@gmail.com> 1698254217 -0700
committer GitHub <noreply@github.com> 1698254217 -0700
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJlOU2JCRBK7hj4Ov3rIwAAaa4IAG4nWEp83SXQPRJtbulVg/nW
 iObgSIp4SwimWEtiLiDudx9R6NZK4odtQ2zv+H5YKiKj9ZPSv973CXlRY82uLhTZ
 b6MC7aZ39PzMAPGe1f+mcWo4K4liGHo7pIQAgrJCbm+Ctq8DJ+/rDg87CPxTUxOs
 n1Ss5L5qjaLBhFfGlRFKGrfHx9j5RW4XFoxn2s6SPvHtacq+WiM2kehka2YYuuUR
 vhsLMhPaAZ8KC0sL0tsi3UmIII9C1gLvau1z19/cHJxDmRUAwWebjOHa/h0H+9rf
 SQo6oc9mSnUfsyHMongG7jJurI3fTbqJwZJjI59LD6FSZ+1VZqfHaA6pX+QtUII=
 =sPyQ
 -----END PGP SIGNATURE-----
 

Adding selection condition to hal.executable.variant. (#15284)

This allows for variants to declare host logic that determines whether
the variant should be selected for loading. When multiple variants are
available their declared conditions will be evaluated in op order along
with the existing executable format match.

Unfortunately the MLIR SymbolTable trait disallows multiple regions on
any op holding it so a new `hal.executable.condition` region op was
added that may be optionally present on any `hal.executable.variant`.
Ideally we clean this up and make it an optional region but that'll need
relaxing of upstream assertions like
https://sourcegraph.com/github.com/llvm/llvm-project/-/blob/mlir/lib/IR/SymbolTable.cpp?L122-123
(ideally either treating region 0 as the symbol table on ops or having
an interface override for selecting the region ala `getCallableRegion`
such as `getSymbolTableRegion`).

This removes the `hal.device.switch` op in favor of `scf.index_switch`.
When we start running const expr hoisting during the HAL pipeline this
should allow variant selection to be completely hoisted to
initialization time (or at least memoized per device). There's decent
low-hanging future work on optimizing the ranking/selection and
improving `scf.index_switch` hoisting/canonicalization to make things
better.