ldc-developers/ldc

Add a build target for bolt-optimized ldc?

JohanEngelen opened this issue · 4 comments

JohanEngelen commented

The following optimization of ldc2 with llvm-bolt results in a 3%-4% faster build time (-O0) at Weka.

# Instrument LDC
/usr/local/bin/llvm-bolt bin/ldc2 -instrument -o bin/ldc2_instr -instrumentation-file=bolt_profile.fdata
# Run workload which generates a profile
bin/ldc2_instr -c -of=dummy.o include/d/std/*.d include/d/std/regex/*.d include/d/std/regex/internal/*.d  -d-version=StdUnittest -unittest -dip1000
# Use the profile to let llvm-bolt optimize our binary
/usr/local/bin/llvm-bolt bin/ldc2 -o bin/ldc2.bolt -data=bolt_profile.fdata \
        -reorder-blocks=cache+ -reorder-functions=hfsort+ -split-functions=3 -split-all-cold -icf -split-eh -dyno-stats
kinke commented

That sounds like it might be worth it. Which ldc2 executable did you use exactly? One from an official release, already using LTO+PGO?

JohanEngelen commented

That sounds like it might be worth it. Which ldc2 executable did you use exactly? One from an official release, already using LTO+PGO?

Forgot to mention, it is on self-built binary with LTO+PGO. (ThinLTO in fact, including LLVM LTO libraries)

kinke commented

including LLVM LTO libraries

Oh wow! :D - I tried that once, took ages and resulted in no measurable performance gains IIRC.