diff options
| author | Daniel Eklöf <daniel@ekloef.se> | 2020-11-22 15:11:01 +0000 |
|---|---|---|
| committer | Daniel Eklöf <daniel@ekloef.se> | 2020-11-22 18:29:44 +0000 |
| commit | af6d1aa5222762e68e47687a2d2ca2213811be4c (patch) | |
| tree | e561f96359dbd84082b8e66d001e1f6048c63d6a | |
| parent | pkgbuild: no need to run pgo input *that* many times (diff) | |
install: re-write description of recommended flags, and add size comparison
| -rw-r--r-- | INSTALL.md | 37 |
1 files changed, 27 insertions, 10 deletions
@@ -130,13 +130,20 @@ mkdir -p bld/release && cd bld/release export CFLAGS="$CFLAGS -O3 -march=native" meson --buildtype=release --prefix=/usr -Db_lto=true ../.. ``` -Both `-O3` and `-Db_lto=true` are **highly** recommended. -For performance reasons, I strongly recommend doing a +To optimize for performance, both `-O3`, `-Db_lto=true` and doing a [PGO](#profile-guided-optimization) (Profile Guided Optimization) -build. +build is recommended. -If you do not want this, just build: +If you instead want to optimize for size, use `-Os` instead of `-O3`. + +In general, `-Os` results in the smallest (and slowest) binary. A full +PGO build will be slightly larger than a `-Os` build, but smaller than +a `-O3` build. A partial PGO build will be somewhere in between a full +PGO build and a `-O3` build. See the [size +comparison](#size-comparison). + +If you are doing a non-PGO build, just build: ```sh ninja @@ -151,13 +158,8 @@ for example, uses `CFLAGS` to specify the default set of flags. Thus, we do `export CFLAGS+="..."` to at least not throw away those flags. -When packaging, you may want to use the default `CFLAGS` only, but -note this: foot is a performance critical application that relies on -compiler optimizations to perform well. - In particular, with GCC 10.1, it is **very** important `-O3` is used -(and not e.g. `-O2`) when doing a [PGO](#profile-guided-optimization) -build. +**when doing a [PGO](#profile-guided-optimization) build**. #### Profile Guided Optimization @@ -175,6 +177,12 @@ meson configure -Db_pgo=generate ninja ``` +Remainder: make sure `-O3` is being used. + +To get **maximum** performance, you can also add +`-fprofile-partial-training`, but do note that this will result in a +noticeably larger binary (see [size comparison](#size-comparison)) + Next, we need to actually generate the profiling data. There are two ways to do this: a [partial PGO build using a PGO @@ -266,6 +274,15 @@ ninja Continue reading in [Running the new build](#running-the-new-build) +##### Size comparison + +| PGO | default | `-fprofile-partial-training`| +|-----------------|--------:|----------------------------:| +| none (`-O3`) | 340K | - | +| none (`-Os`) | 260K | - | +| partial | 328K | 376K | +| full | 292K | 352K | + ### Debug build ```sh |
