================
@@ -322,22 +322,26 @@ foreach( t ${LIBCLC_TARGETS_TO_BUILD} )
if ( ARCH STREQUAL spirv OR ARCH STREQUAL spirv64 )
set( build_flags -O0 -finline-hint-functions -DCLC_SPIRV )
- set( opt_flags )
+ set( clc_opt_flags )
+ # Inline CLC functions into OpenCL builtins
+ set( opencl_opt_flags "--passes=always-inline" )
----------------
frasercrmck wrote:
Do you mean for SPIR-V in the context of this PR (to stop the introduction of
new `always_inline` attributes), or generally across the whole of libclc?
I don't mind the former as, yeah, `always_inline` isn't ideal. But the
SPIR-V/Mesa targets seem quite explicit that they don't want any optimizations.
Limiting the optimizations to just the `always-inline` pass was my way of
adhering to that wish as best as I could. I don't know if SPIR-V/Mesa opt out
of optimizations because the SPIR-V translator couldn't handle optimized LLVM
IR, or produced worse SPIR-V, or crashed, or whether it was just for
compile-time reasons. Maybe it doesn't apply any more? I'm happy to do whatever
with those targets, as long as someone can vouch for them.
The latter would be a significant change as in practice _every_ function for
the AMDGPU/NVPTX targets - OpenCL or CLC - currently has the `always_inline`
attribute set, via `_CLC_DEF`. I'm not sure it's the best time to drop that
assumption.
https://github.com/llvm/llvm-project/pull/115699
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
