mirror of
https://github.com/espressif/esp-idf.git
synced 2026-09-22 13:01:16 +03:00
4052daa8a6
ESP-IDF compiles every component with -ffunction-sections/-fdata-sections and
links with -Wl,--gc-sections, so that code and data nothing references does not
reach the image. With CONFIG_COMPILER_LTO_LINKTIME the object files carry GIMPLE
instead of machine code and code generation is deferred to the link, where those
two options no longer apply. The LTO partition is therefore emitted as a single
.text/.rodata, and since --gc-sections works at section granularity it can only
keep or drop that section as a whole. Some function in it is always live, so
nothing is dropped.
The options are recorded in the object file (in the .gnu.lto_.opts section), but
they do not reach the code generator. Code generation at link time runs as a
separate compiler invocation (LTRANS) whose command line lto-wrapper
reconstructs from those recorded options, and append_compiler_options() in
gcc/lto-wrapper.cc forwards only CL_TARGET options plus a small hard-coded list:
-fPIC/-fpic/-fPIE/-fpie, -fcommon, -fgnu-tm, -fopenmp/-fopenacc, -fcf-protection=,
(-fasynchronous-)unwind-tables, -g, -O/-Os/-Og/-Ofast/-Oz and the diagnostics
formatting options. Everything else hits the default arm and is dropped.
-ffunction-sections/-fdata-sections are neither. They are plain Common options in
common.opt without the Optimization marker, so they are not part of the
per-function state that is streamed with each function (which is why -O2 and -Os
do survive per translation unit under LTO), and they are not target options
either. They reach LTRANS only if they are repeated on the link command line.
Pass them next to -flto=auto, in both build systems.
This can be verified on any LTO build by adding -save-temps to the link options
and inspecting the generated <output>.ltrans.mk, which contains the literal
LTRANS command line:
grep -o -- "-ffunction-sections\|-fdata-sections" build/*.ltrans.mk
Measured on esp32c3 with -Os and CONFIG_COMPILER_LTO_COMPILETIME, application
binary size in bytes:
no LTO LTO LTO+fix fix saves
hello_world 116112 119792 115808 -3984 (-3.33%)
wifi/getting_started 734720 792960 732976 -59984 (-7.56%)
Without this change, enabling LTO produces a larger image than not using LTO at
all; with it, LTO is size neutral to slightly positive.
Two effects contribute. The dead code inside the partition itself stays, which
scales with how much code was compiled with LTO. On top of that, every function
retained this way keeps whatever it references alive as well, transitively and
across object boundaries: constant data, other functions, and sections of objects
that were not compiled with LTO at all. That second effect is not bounded by the
size of the LTO partition and can dominate. In the wifi/getting_started case a
single retained function, esp_crt_bundle_attach(), is the only referrer of the
mbedTLS certificate bundle, so a 55 KB .rodata blob stayed in an application that
never uses TLS. A plain non-LTO build collects all of it, which is what this
change restores.
Signed-off-by: Frantisek Hrbata <frantisek.hrbata@espressif.com>