On Nov 16, 2006, at 6:25 PM, Reid Spencer wrote:
> On Thu, 2006-11-16 at 17:44 -0800, Bill Wendling wrote:
>> Basically, this is to get rid of static c'tor/d'tors scattered all
>> over the source tree. (You get them whenever you #include
>> <iostream>.)
>> A smart compiler should be able to remove all of the empty functions
>> because they don't have any side-effects.
>
Hi Reid,
> I understand the motivation for the change and I commend it, but I'm
> concerned about getting it right. You're assuming a few things: a) the
> compiler is smart enough to get rid of calls that have no side
> effects,
> b) the compiler knows how to determine if a function has no side
> effects
> without help (operator<< isn't marked "pure"), and c) that there
> aren't
> any side effects. c) can probably be proven by your iostream
> implementation but a) and b) are a little tougher.
>
Of course, LLVM is self-hosted, so it should be smart enough to
handle it. :-)
> Can we build an optimized, with symbols (-g -O2) version of one of
> these
> modules and see if those symbols are getting removed ? If so, I'm
> fine
> with it. If not, we need a new strategy.
>
I hope I did this correctly. Anyway, here are the files:
t.h:
#include <ostream>
class llvm_ostream {
std::ostream* Stream;
public:
llvm_ostream() : Stream(0) {}
llvm_ostream(std::ostream& OStream) : Stream(&OStream) {}
template <typename Ty>
llvm_ostream& operator << (const Ty& Thing) {
#ifndef NDEBUG
if (Stream) *Stream << Thing;
#endif
return *this;
}
};
/// getErrorOutputStream - Returns the error output stream
(std::cerr). This
/// places the std::c* I/O streams into one .cpp file and relieves
the whole
/// program from having to have hundreds of static c'tor/d'tors for
them.
///
llvm_ostream getErrorOutputStream(const char *DebugType);
#ifdef NDEBUG
#define DOUT llvm_ostream()
#else
#define DOUT getErrorOutputStream(DEBUG_TYPE)
#endif
t.cpp:
#define DEBUG_TYPE "experiment"
#include "t.h"
int main()
{
DOUT << "Hello world\n";
}
foo.cpp:
#include "t.h"
#include <iostream>
llvm_ostream getErrorOutputStream(const char *DebugType) {
if (DebugType)
return llvm_ostream(std::cerr);
else
return llvm_ostream();
}
Compiled with -DNDEBUG -g -O2 gives:
$ nm -A foo.o
foo.o: 00000170 s EH_frame1
foo.o: 00000160 s __GLOBAL__I__Z20getErrorOutputStreamPKc
foo.o: 00000000 a __GLOBAL__I__Z20getErrorOutputStreamPKc.eh
foo.o: 00000000 T __Z20getErrorOutputStreamPKc
foo.o: 00000000 A __Z20getErrorOutputStreamPKc.eh
foo.o: 000000d0 s __Z41__static_initialization_and_destruction_0ii
foo.o: 0000018c s __Z41__static_initialization_and_destruction_0ii.eh
foo.o: U __ZNSt8ios_base4InitC1Ev
foo.o: U __ZNSt8ios_base4InitD1Ev
foo.o: U __ZSt4cerr
foo.o: 000001d0 b __ZSt8__ioinit
foo.o: U ___cxa_atexit
foo.o: U ___dso_handle
foo.o: U ___gxx_personality_v0
foo.o: 00000040 t ___tcf_0
foo.o: 00000000 a ___tcf_0.eh
foo.o: U dyld_stub_binding_helper
$ nm -A t.o
t.o: 00000000 T _main
t.o: 00000000 A _main.eh
Compiling with -g -O2 gives:
$ nm -A foo.o
foo.o: 00000170 s EH_frame1
foo.o: 00000160 s __GLOBAL__I__Z20getErrorOutputStreamPKc
foo.o: 00000000 a __GLOBAL__I__Z20getErrorOutputStreamPKc.eh
foo.o: 00000000 T __Z20getErrorOutputStreamPKc
foo.o: 00000000 A __Z20getErrorOutputStreamPKc.eh
foo.o: 000000d0 s __Z41__static_initialization_and_destruction_0ii
foo.o: 0000018c s __Z41__static_initialization_and_destruction_0ii.eh
foo.o: U __ZNSt8ios_base4InitC1Ev
foo.o: U __ZNSt8ios_base4InitD1Ev
foo.o: U __ZSt4cerr
foo.o: 000001d0 b __ZSt8__ioinit
foo.o: U ___cxa_atexit
foo.o: U ___dso_handle
foo.o: U ___gxx_personality_v0
foo.o: 00000040 t ___tcf_0
foo.o: 00000000 a ___tcf_0.eh
foo.o: U dyld_stub_binding_helper
$ nm -A t.o
t.o: 000000bc s EH_frame1
t.o: U __Z20getErrorOutputStreamPKc
t.o: U __ZStlsISt11char_traitsIcEERSt13basic_ostreamIcT_ES5_PKc
t.o: U ___gxx_personality_v0
t.o: 00000000 T _main
t.o: 000000d8 S _main.eh
t.o: U dyld_stub_binding_helper
So it seems like gcc is doing the right thing, I suppose. I'm going
to give it a try with the LLVM source in the CodeGen directory.
-bw
_______________________________________________
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
