I needed to compile a cuda source file (say, a.cu) into IR (a.bc), and then merge a.bc with another bitcode file (b.bc, compiled from b.cu). So I used llvm-link a.bc b.bc -o c.bc
However, I noticed that an internal function ’ _ZL21__nvvm_reflect_anchorv() ’ is defined in both a.bc & b.bc, and when merging these two files, one of the two definitions was renamed to ‘_ZL21__nvvm_reflect_anchorv.2()’, and written into c.bc.
Then I did llc c.bc -o c.s -march=nvptx ; ptxas c.s -o c.o
However, ptxas would give the following complaint:
ptxas c.s, line 171; error : Duplicate definition of function ‘_ZL21__nvvm_reflect_anchorv’
ptxas c.s, line 171; fatal : Parsing error near ‘.2’: syntax error
So I inspected c.s and found the issue above was caused by the following line:
.func (.param .b32 func_retval0) _ZL21__nvvm_reflect_anchorv.2() // @_ZL21__nvvm_reflect_anchorv.2
After I manually deleted the definition of this function in c.s, the compilation works file. I wonder how could I force llc to remove
**_ZL21__nvvm_reflect_anchorv.2**()? Or is that possible to prevent _ZL21__nvvm_reflect_anchorv**()** from being generated into a.bc & b.bc in the first place? Or is this possible to ask llvm-link to NOT rename _ZL21__nvvm_reflect_anchorv() into ZL21__nvvm_reflect_anchorv**.2****()**?
The actual problem here is that PTX appears to not allow ‘.’ in symbol names. We should probably just change our PTX emission to mangle dots somehow.
Yeah, ‘.’ is the direct reason for the ptxas failure here. I’m curious, however, about what the purpose of nvvm_reflect_anchorv() is here, and why does the front-end always generate this function? Since the current PTX emission doesn’t mangle dots, it would be a reasonable workaround for me to prevent the front-end from generating this function in the first place. Is there any magic option available to do so?
With David’s http://reviews.llvm.org/rL265060, do you think __nvvm_reflect_anchor is still necessary?
David’s change makes nvvm_reflect_anchor unnecessary. The issue with dots in names generated by llvm still needs to be fixed.
r267062 removed __nvvm_reflect_anchor() so you should no longer see this particular manifestation of illegal-character-in-PTX problem.
Awesome! Thanks for letting me know, Artem!