[LLVMdev] Profiling in LLVM Patch Followup 1

[Andreas Neustifter]

Good Morning.

On 22.07.09 14:47), Andreas Neustifter wrote:
> Daniel Dunbar wrote:
>> I merged in my changes to your patch, which results in the attached
>> patch. There may be some missed merge errors. The main problem I have
>> with the rest of this patch is that it causes a regression in
>> llvm-prof's behavior. I tried running edge profiling on the
>> MultiSource/Applications/aha benchmark in the llvm test-suite, and
>> llvm-prof no longer prints any function information:
>> --
>> ddunbar at giles:aha$ opt -f Output/aha.linked.rbc -insert-edge-profiling -o foo.bc
>> ddunbar at giles:aha$ llc foo.bc -o - | gcc -x assembler - -x none
>> ~/llvm/Debug/lib/profile_rt.dylib
>> ddunbar at giles:aha$ ./a.out
>> ...
>> ddunbar at giles:aha$ ~/llvm/Release/bin/llvm-prof foo.bc llvmprof.out
>> [...]
> First of all, you have to use the original bytecode file
> (Output/aha.linked.bc) for use with llvm-prof like this:
> 
> $ ~/llvm/Release/bin/llvm-prof Output/aha.linked.bc llvmprof.out
> 
> Also you have to check that llvmprof.out is deleted between tests
> otherwise the figures are accumulated counts from several runs.
What I also use to make things easier is the utils/profile.pl script, it
automates the process of instrumenting running and profiling with
llvm-prof. I have attached a version that produces more output and
removes the llvmprof.out file in between runs automatically.

>> [...]
>>
>> The basic block counts also differ, sometimes very significantly, but
>> I'm not sure if this is to be expected. Can you investigate both of
>> these differences on some real example? I'd like to keep everything
>> working at least as good as it was before (modulo some other features,
>> like function counts or basic block traces, which we don't really
>> support).
> 
> I have to compare the revisions 75622 and 75625 to check what's going on
> there, but maybe its also due to undeleted llvmprof.out files?
I have compared the revisions now and apart from the passes being
executed faster (according to -time-passes) there is no difference in
the llvm-prof output.

Andi