Hi,
I was using "clang -O3 -S -emit-llvm" got some very optimized output.
Then I did "clang -S -emit-llvm" (without optimization) and wanted to optimized the code in a
separate pass. The llvm program "opt" did not do anything.
How can I invoke the optimizer on some un-optimized program, possibly showing the output of each optimizer stage?
I would like to get a deeper understanding of the optimization pass.
Best,
Patrick
Hi Patrick,
You have to tell opt which optimizations you want it to do. The
-std-compile-opts option for opt may be what you're looking for.
--Sam
In addition to Sam's advice, I want to point out that clang's IR generator doesn't necessarily emit the same code when compiling for optimization and not. Most obviously, we never emit available_externally function definitions at -O0 because we assume that would just waste compile time.
John.
OK, I am looking for "LTO"/global optimization. So the function definition will remain "somewhere else" (externally), and the optimizer will find in some other module to possibly inline it later on.
I am planning to concat all the *.ll (eg "link" the files) and pass them to the "global" optimizer, as "size" is a very important optimization criterium to me. After that, the back-end will be invoked.
Is that a good approach?
Best, Patrick
How about just add -O4 on clang (or llvm-gcc) command line to get LTO optimizations ?
See http://llvm.org/docs/LinkTimeOptimization.html
http://llvm.org/docs/GoldPlugin.html
On Mac OS X, it is a checkbox in Xcode.
My backend doesnt supply lto 
I am trying to emulate lto like this:
clang *.c -S -emit-llvm
cat *.ll > very_big.ll
opt -O3 -extra-things...
llc very_big.ll.optimized
('Ease of use' is not so important 
Best, Patrick
Nice to see it being support on OSX now - updating XCode right now...
OK, I am looking for “LTO”/global optimization. So the function definition will remain “somewhere else” (externally), and the optimizer will find in some other module to possibly inline it later on.
I am planning to concat all the *.ll (eg “link” the files) and pass them to the “global” optimizer, as “size” is a very important optimization criterium to me. After that, the back-end will be invoked.
Is that a good approach?
How about just add -O4 on clang (or llvm-gcc) command line to get LTO optimizations ?
On Mac OS X, it is a checkbox in Xcode.
Devang
My backend doesnt supply lto
The backend (or code generator) does not do LTO in this case also.
clang *.c -S -emit-llvm
cat *.ll > very_big.ll
opt -O3 -extra-things…
llc very_big.ll.optimized
Try replacing “cat *.ll > very_big.ll” with “llvm-ld *.ll -o big.bc”. ‘llvm-ld’ is another llvm tool, just like ‘opt’
Sry for spam... I already have XCode 3.2.5... what is the lto option called, or is it only available in XCode 4 preview?
OK, I am looking for “LTO”/global optimization. So the function definition will remain “somewhere else” (externally), and the optimizer will find in some other module to possibly inline it later on.
I am planning to concat all the *.ll (eg “link” the files) and pass them to the “global” optimizer, as “size” is a very important optimization criterium to me. After that, the back-end will be invoked.
Is that a good approach?
How about just add -O4 on clang (or llvm-gcc) command line to get LTO optimizations ?
On Mac OS X, it is a checkbox in Xcode.
Devang
My backend doesnt supply lto
The backend (or code generator) does not do LTO in this case also.
Roger, but the effect is equivalent to lto, isnt it?
As I understand, lto-enabled GCC does the same (save as compiled code and IR) and the linker will optimize over that IR (if present) and then invoke its backend.
I was playing with gcc-lto on linux last year, but it didnt work very well for me at that time. So when I saw llvm, that is what came to my mind first: emulated lto due to the IR!
clang *.c -S -emit-llvm
cat *.ll > very_big.ll
opt -O3 -extra-things…
llc very_big.ll.optimizedTry replacing “cat *.ll > very_big.ll” with “llvm-ld *.ll -o big.bc”. ‘llvm-ld’ is another llvm tool, just like ‘opt’
OK! I will be able to have a look at any intermediate file with llvm-dis I suppose in order to verify the transformations done…
Best, Patrick
This build setting is LLVM_LTO show as "Link Time Optimization" in the build settings panel.
It is available only with clang and llvm-gcc AFAIK.
-- Jean-Daniel