I built clang twice, once with:

cmake -GNinja -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=ON -DLLVM_ENABLE_TERMINFO=OFF -DLLVM_USE_CRT_RELEASE=MT -DCLANG_PLUGIN_SUPPORT=OFF -DLLVM_ENABLE_LIBXML2=FORCE_ON -DLLVM_ENABLE_THREADS=OFF -DLLVM_ENABLE_PIC=OFF ../llvm-rw/
ninja clang llvm-ar lld

and once with:

cmake -GNinja -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=ON -DLLVM_ENABLE_TERMINFO=OFF -DLLVM_USE_CRT_RELEASE=MT -DCLANG_PLUGIN_SUPPORT=OFF -DLLVM_ENABLE_LIBXML2=FORCE_ON -DLLVM_ENABLE_THREADS=OFF ../llvm-rw/
ninja clang llvm-ar lld

The only difference is that the first has -DLLVM_ENABLE_PIC=OFF and the second doesn't. I checked that one build.ninja file passes -fPIC to Sema.cpp.o and the other doesn't.

I then built target 'base' in chrome three times (actually four times, but I discarded the first result as cache warmup) with 

symbol_level = 1
is_component_build = true
is_debug = false
clang_base_path = getenv("HOME") + "/src/llvm-build-pic"
clang_use_chrome_plugins 

by doing

thakis@thakis:~/src/chrome/src$ gn clean out/gn
thakis@thakis:~/src/chrome/src$ gn gen out/gn
Done. Made 12210 targets from 1873 files in 3059ms
thakis@thakis:~/src/chrome/src$ time ninja -C out/gn base

I got:

real	0m27.230s
user	13m24.791s
sys	0m32.916s

real	0m27.012s
user	13m26.341s
sys	0m32.099s

real	0m27.156s
user	13m25.614s
sys	0m32.425s


Then I did the same with
clang_base_path = getenv("HOME") + "/src/llvm-build-nopic"
and the rest the same. Here I got:

real	0m26.668s
user	13m19.867s
sys	0m32.466s

real	0m26.976s
user	13m20.559s
sys	0m32.160s

real	0m26.678s
user	13m19.674s
sys	0m32.012s

So it looks like this is in fact measuably faster, but the min-of-3 is only 1.27% faster. Still, it's basically free now that we build with the plugins compiled in. And the binary is 400kB smaller (again, not much, but free.)



(This was all on my linux desktop box. I think mac maybe always uses pic? not sure, need to check. LLVM's cmake file has some comment about -mdynamic-no-pic in Xcode, but that might be obsolete.)