Compiling with clang++
Now that clang++ 3.9.0 is installed on titan (/opt/llvm-3.9.0/bin/clang++) it is possible to do some compilation test with this compiler.
Modify your LocalConfig.cmake to use clang++ and rebuild everything.
A full Ctest on Xtest have been done and give almost same results as g++.
First remark regarding CMAKEFILE, clang give the following message:
clang-3.9: warning: optimization flag '-frounding-math' is not supported
This mean that we have put this flag every-where or almost but it is compiler dependent !!! If we pass to modern cmake we will be able to use $<COMPILER_ID:GNU>,$<COMPILER_ID:Clang>
stuff to appropriately use -frounding-math or not
When using MUMPS or PASTIX clang need to have an explicit -lpthread. Those libraries where compiled with g++. Maybe with this compiler the -lpthread is automatically set. Like above if we pass to modern cmake we may encapsulate MUMPS or PASTIX target dependencies and fix cleanly those dependencies with $<COMPILER_ID:GNU>,$<COMPILER_ID:Clang>
stuff.
What is more interesting is that clang have note exactly the same behavior as g++ regarding errors and warnings. For example with -Wconversion clang check implicit conversion miss formed (this is not possible with g++):
#include <iostream>
#include <vector>
using namespace std;
int main ()
{
size_t h=1234567891123456Lu;
int hi;
short hs;
hi=h;
hs=h;
cout<<"h "<<h<<endl;
cout<<"hi "<<hi<<endl;
cout<<"hs "<<hs<<endl;
hi=-1;
h=hi;
cout<<"h "<<h<<endl;
cout<<"hi "<<hi<<endl;
vector<int> v(3,1);
int l=v.size();
v[l-1]=4;
for (auto val : v)
cout<<val<<endl;
return 0;
}
With this code clang produce following messages:
clang++ -std=c++11 -Wconversion main.cc
main.cc:13:8: warning: implicit conversion loses integer precision: 'size_t' (aka 'unsigned long') to 'short' [-Wconversion]
hs=h;
~^
main.cc:20:7: warning: implicit conversion changes signedness: 'int' to 'size_t' (aka 'unsigned long') [-Wsign-conversion]
h=hi;
~^~
main.cc:27:8: warning: implicit conversion changes signedness: 'int' to 'size_type' (aka 'unsigned long') [-Wsign-conversion]
v[l-1]=4;
~ ~^~
main.cc:12:8: warning: implicit conversion loses integer precision: 'size_t' (aka 'unsigned long') to 'int' [-Wshorten-64-to-32]
hi=h;
~^
main.cc:26:13: warning: implicit conversion loses integer precision: 'size_type' (aka 'unsigned long') to 'int' [-Wshorten-64-to-32]
int l=v.size();
~ ~~^~~~~~
5 warnings generated.
With eXlibris many messages of those types are issued ... For many it is only a matter of using int
instead of size_t
when dealing with STL container. But unfortunately this is masking more interesting ones that may be mistakes or error !