https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97106
Bug ID: 97106
Summary: [nvptx] Issues with weak aliases introduced by C++
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Keywords: openmp
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: burnus at gcc dot gnu.org
CC: vries at gcc dot gnu.org
Depends on: 97102
Target Milestone: ---
Created attachment 49240
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=49240&action=edit
C++ test case - note: matches mostly uncommitted
testsuite/libgomp.c++/pr96390.C of PR96390
Related PR 97102: nvptx + aliases
Related PR 96390: alias-related: automatically mark functions called in omp
target as declare target (that's where the initial testcase came from)
For the attached testcase, GCC calls the "complete object constructor" C1:
_ZN1VILi1EEC1ImvEET_
but generates the function definition for "base object constructor" C2:
_ZN1VILi1EEC2ImvEET_
And it generates a weak alias between the two.
Using .alias should work – the aliasee should not be weak but I believe in our
case only the alias itself is weak. – Hence, it should work.
Caveat: .alias is ISA version 6.3 = CUDA 10.0, cf. PR 97102.
I wonder whether this could be solved at LTO time – after all, except for
linked libraries (libgomp, libm) it should be know.
Caveat: currently, no real LTO is done on the offloading side.
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97102
[Bug 97102] [nvptx] PTX JIT compilation failed when using aliases
