Optimization application by llvm-gcc, opt

Project Infrastructure LLVM Dev List Archives
1

Hi all,

I wanted to clarify something about the application of optimizations by llvm-gcc, opt and llc.

Will the following two sequences generate the same optimized bitcode?
   a) $ llvm-gcc -O3 -c -emit-llvm foo.c -o foo.bc
   b) $ llvm-gcc -O0 -c -emit-llvm foo.c -o foo.bc
      $ opt -O3 foo.bc

2

Hi RaghuB,

I wanted to clarify something about the application of optimizations by llvm-gcc, opt and llc.

Will the following two sequences generate the same optimized bitcode?
    a) $ llvm-gcc -O3 -c -emit-llvm foo.c -o foo.bc
    b) $ llvm-gcc -O0 -c -emit-llvm foo.c -o foo.bc
       $ opt -O3 foo.bc

not necessarily, though usually it is the same. One reason for this is that
the per-function passes run by llvm-gcc are used to clean up functions as they
are output, while with opt they are run after the entire module has been
generated. The difference is that when a function is output some things like
global variables may have only been partially output (eg may not have an initial
value yet) resulting in less or different optimization.

Ciao, Duncan.

3

Thanks Duncan. Just to make sure I understand correctly, opt has more global information, due to which optimizations can have greater effect than in llvm-gcc, where optimizations are applied on a per-function basis. Is this correct ?

4

Hi Raghu,

Thanks Duncan. Just to make sure I understand correctly, opt has more global information, due to which optimizations can have greater effect than in llvm-gcc, where optimizations are applied on a per-function basis. Is this correct ?

if you look carefully at what "opt -O3" runs (using -debug-pass=Arguments) you
will see that it first runs a small list of passes (the "per function" passes),
and later runs a much more complete list of passes. As each function is output,
llvm-gcc runs these "per-function" passes on it. Once the entire module is
output, llvm-gcc runs the second list of passes on the whole module. Thus it
is not correct to say that optimizations are applied only on a per-function
basis in llvm-gcc. On the other hand opt applies both lists of passes to the
entire module. Your remark "opt has more global information, due to which
optimizations can have greater effect than in llvm-gcc" is correct, but only
applies to the first list of passes, since for the second list of passes
llvm-gcc also has the entire module available when applying them.

Ciao, Duncan.

5

Got it. Thanks for the detailed explanation Duncan.