================
@@ -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:
Thanks for letting me know, that's useful information to consider. I have seen
issues with LLVM optimizations generating invalid (or at least unwanted) SPIR-V
in the past, so it's good to be reminded of that.
It sounds as though you don't want any inlining, which is fair enough. Inlining
builtins would always be a danger if I make this change. I note that in the
current libclc library for SPIR-V (before any of my recent refactoring) we have
calls to "internal "__clc" functions such as `__clc_tanpi`, `__clc_rootn`, etc.
So perhaps it's fine if this and future PRs introduce more of those e.g. `sign`
calling `__clc_sign`?
I can undo the change which adds the `always_inline` attribute to CLC
functions, and undo the change that runs `opt --passes=always-inline` for
SPIR-V. Does that sound okay?
https://github.com/llvm/llvm-project/pull/115699
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
