[PATCH] D40573: [NVPTX] Assign valid global names

[Artem Belevich via Phabricator via llvm-commits]

tra added inline comments.


================
Comment at: lib/IR/ValueSymbolTable.cpp:54-59
+      // MSVC name mangling but is fine for PTX.
+      if (M && Triple(M->getTargetTriple()).isNVPTX())
+        S << "$";
+      else
+        S << ".";
+    }
----------------
Hahnfeld wrote:
> tra wrote:
> > This patch addresses "we can't compile generated PTX because LLVM uses illegal characters", but exposes another issue -- having potentially different names on host and device is a problem for CUDA. For some objects host side may need to know what it's called on device side. We need it in order to access it from host (eg cudaMemcpyToSymbol(), or initializing static variables) and we currently assume that the names are the same. If such symbol gets different names on host and device, compilation will succeed, but we'll have problems at runtime.
> > 
> > Does "." have any special meaning? Can we skip the unique delimiter altogether? 
> > 
> > If we can't find a suitable way to guarantee identical naming, we'll need a way to have a reliable way to determine the name used on the other side of the compilation.
> > 
> > 
> So the interesting question is: When will this code ever be hit? Most programming languages (including C and C++) obviously don't allow multiple variables of the same name - how would the compiler say which symbol you meant. The use case I've mostly seen is for compiler generated function, `omp_outlined` for example. These can be generated multiple times in the same translation unit and have to get unique names. Do you have another example where this could happen?
> 
> I'm not really sure '.' has a special meaning. Maybe @rafael can help because one of his old commits (https://reviews.llvm.org/rL253804) says `For globals it is important to use "foo.1" to help C++ name demangling.`
I vaguely recall that '.' was an indication for demangler that it should not proceed further. I.e. a sort-of-special character to indicate the end of the C++-mangled part of the symbol name.

If name mangling can't be made identical (and it looks like it may be the case), we can probably work around it. I.e. for symbols that must have identical names on both sides we can generate a unique alias that's identical on both sides and use it instead when CUDA needs it.



https://reviews.llvm.org/D40573