================
@@ -420,12 +420,37 @@ foreach( t ${LIBCLC_TARGETS_TO_BUILD} )
-D${CLC_TARGET_DEFINE}
# All libclc builtin libraries see CLC headers
-I${CMAKE_CURRENT_SOURCE_DIR}/clc/include
+ # Error on undefined macros
+ -Werror=undef
)
if( NOT "${cpu}" STREQUAL "" )
list( APPEND build_flags -mcpu=${cpu} )
endif()
+ # Generic address space support.
+ # Note: when declaring builtins, we must consider that even if a target
+ # formally/nominally supports the generic address space, in practice that
+ # target may map it to the same target address space as another address
+ # space (often the private one). In such cases we must be careful not to
+ # multiply-define a builtin in a single target address space, as it would
+ # result in a mangling clash.
+ # For this reason we must consider the target support of the generic
+ # address space separately from the *implementation* decision about whether
+ # to declare certain builtins in that address space.
+ # FIXME: Shouldn't clang automatically enable this extension based on the
+ # target?
----------------
frasercrmck wrote:
With #137940 we can remove this CMake logic, thanks all!
One result of #137636 is that (I believe) because of the default AMDGPU devices
we compile libclc for, the generic address space isn't being enabled by default
for any AMDGPU target in this PR. Should we perhaps be building libclc for a
newer AMDGCN device? It should be at least GFX700 for generic address space
support.
https://github.com/llvm/llvm-project/pull/137183
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
