"BOLT" post-link optimizer gives 15% speed boost to Clang
welkam
wwwelkam at gmail.com
Wed Oct 24 14:07:57 UTC 2018
On Wednesday, 24 October 2018 at 01:57:38 UTC, Walter Bright
wrote:
> Yup, it's reinvention of the same thing.
The authors never claimed that they invented anything. What they
claim is that doing reordering after linking has benefits that
compiler cant easily do. Bolt improves performance on top of
PGO+LTO by:
1. Using more accurate profile data from intel`s last branch
record(LBR)
2. Inlining function some times changes hotness of code blocks
based on where it was inlined so doing profiling after inlining
gives better decision making for Bolt.
The negative of Bolt`s approach is that is limited on what
allowed transformations it can do so people should use
PGO+LTO+Bolt for maximum performance. Second negative is that it
works with ELF format so windows users wont benefit from it.
More information about the Digitalmars-d
mailing list