`-flto` on clang/gcc?

[indutny]

I wonder if we should give a try to the Link Time Optimizations flag in clang. It appears to be used when building v8 (though, as far as I can tell no on mac), and may be useful for building OpenSSL and/or core itself.

Thoughts?

cc @nodejs/collaborators

[indutny]

cc @bnoordhuis

[indutny]

Here are tls-throughput result comparison on my mac (10 runs):

tls/throughput.js dur="5" type="buf" size="2": ./out/Release/node-flto: 6.1295 ./out/Release/node: 5.6876 ....... 7.77%
tls/throughput.js dur="5" type="buf" size="1024": ./out/Release/node-flto: 1567 ./out/Release/node: 1455 ........ 7.70%
tls/throughput.js dur="5" type="buf" size="1048576": ./out/Release/node-flto: 4167.8 ./out/Release/node: 4026.7 . 3.50%
tls/throughput.js dur="5" type="asc" size="2": ./out/Release/node-flto: 5.5664 ./out/Release/node: 5.0363 ...... 10.52%
tls/throughput.js dur="5" type="asc" size="1024": ./out/Release/node-flto: 1462.2 ./out/Release/node: 1336.8 .... 9.38%
tls/throughput.js dur="5" type="asc" size="1048576": ./out/Release/node-flto: 3947.5 ./out/Release/node: 3847.4 . 2.60%
tls/throughput.js dur="5" type="utf" size="2": ./out/Release/node-flto: 5.5544 ./out/Release/node: 5.0786 ....... 9.37%
tls/throughput.js dur="5" type="utf" size="1024": ./out/Release/node-flto: 1328.5 ./out/Release/node: 1255.7 .... 5.80%
tls/throughput.js dur="5" type="utf" size="1048576": ./out/Release/node-flto: 3051.1 ./out/Release/node: 2985.1 . 2.21%

[indutny]

I suspect that it will also slightly improve start time and overall performance.

[mscdex]

Out of curiosity, was this prompted by news of ThinLTO?

[bnoordhuis]

@indutny Did you build with CXX=clang++? If yes, can you pass LINK=clang++ as well? GYP links with g++ by default and in that case you won't get cross-compilation unit LTO.

[indutny]

@mscdex yup

@bnoordhuis yep, and also I am on OS X, so gcc is not an option anyway.

[indutny]

@bnoordhuis btw, g++ has -flto option too. Though, it accepts number of threads there.

[indutny]

Something to consider too: http://clang.llvm.org/docs/UsersManual.html#cmdoption-fwhole-program-vtables

[bnoordhuis]

btw, g++ has -flto option too.

What I mean is that when you compile with clang++ but link with g++, then g++ won't know how to do LTO across files because the .o files don't contain GIMPLE (because they contain LLVM bitcode.)

[indutny]

@bnoordhuis oh, of course. Do you think we need to do something about it?