pw_bloat#
The bloat module provides tools and helpers around using Bloaty McBloatface including generating generate size report cards for output binaries through Pigweed’s GN build system.
Bloat report cards allow tracking the memory usage of a system over time as code changes are made and provide a breakdown of which parts of the code have the largest size impact.
Defining size reports#
Size reports are defined using the GN template pw_size_diff. The template
requires at least two executable targets on which to perform a size diff. The
base for the size diff can be specified either globally through the top-level
base argument, or individually per-binary within the binaries list.
Arguments
title: Title for the report card.base: Optional default base target for all listed binaries.binaries: List of binaries to size diff. Each binary specifies a target, a label for the diff, and optionally a base target that overrides the default base.source_filter: Optional regex to filter labels in the diff output.full_report: Boolean flag indicating whether to output a full report of all symbols in the binary, or a summary of the segment size changes. Default false.
import("$dir_pw_bloat/bloat.gni")
executable("empty_base") {
sources = [ "empty_main.cc" ]
}
executable("hello_world_printf") {
sources = [ "hello_printf.cc" ]
}
executable("hello_world_iostream") {
sources = [ "hello_iostream.cc" ]
}
pw_size_diff("my_size_report") {
title = "Hello world program using printf vs. iostream"
base = ":empty_base"
binaries = [
{
target = ":hello_world_printf"
label = "Hello world using printf"
},
{
target = ":hello_world_iostream"
label = "Hello world using iostream"
},
]
}
Single Binary Sizes Reports#
Argument
target: Binary target to run size report on.
import("$dir_pw_bloat/bloat.gni")
executable("hello_world_iostream") {
sources = [ "hello_iostream.cc" ]
}
pw_size_report("hello_world_iostream_size_report") {
target = ":hello_iostream"
}
ASCII Table Generated
+-------------+-----------------------------------------------+------+
|segment_names| fullsymbols | sizes|
+=============+===============================================+======+
|FLASH | |20,416|
| |[147 Others] | 6,976|
| |_dtoa_r | 3,036|
| |[section .code] | 2,985|
| |_printf_float | 1,132|
| |__adddf3 | 630|
| |_vfiprintf_r | 608|
| |__aeabi_dmul | 596|
| |_printf_i | 588|
| |_svfprintf_r | 512|
| |__aeabi_ddiv | 464|
| |pw_assert_basic_HandleFailure | 388|
| |__multiply | 340|
| |quorem | 278|
| |__mdiff | 276|
| |__sflush_r | 268|
| |_ctype_ | 257|
| |__lshift | 224|
| |__swsetup_r | 220|
| |pw::allocator::FreeListHeap::Free(void*) | 220|
| |_printf_common | 218|
| |__mprec_tens | 200|
+-------------+-----------------------------------------------+------+
|RAM | | 672|
| |__global_locale | 364|
| |(anonymous namespace)::buf | 104|
| |impure_data | 96|
| |kCrashBanner | 48|
| |object.0 | 24|
| |[section .zero_init_ram] | 9|
| |[section .static_init_ram] | 4|
| |_impure_ptr | 4|
| |errno | 4|
| |pw::log_basic::(anonymous namespace)::write_log| 4|
| |pw_freelist_heap | 4|
| |unoptimizable | 4|
| |__lock___sfp_recursive_mutex | 1|
| |__lock___sinit_recursive_mutex | 1|
+-------------+-----------------------------------------------+------+
|Total | |21,088|
+-------------+-----------------------------------------------+------+
Size reports are typically included in ReST documentation, as described in
Documentation integration. Size reports may also be printed in the build
output if desired. To enable this in the GN build, set the
pw_bloat_SHOW_SIZE_REPORTS build arg to true.
Documentation integration#
Bloat reports are easy to add to documentation files. All pw_size_diff
targets output a file containing a tabular report card. This file can be
imported directly into a ReST documentation file using the include
directive.
For example, the simple_bloat_loop and simple_bloat_function size
reports under //pw_bloat/examples are imported into this file as follows:
Simple bloat loop example
^^^^^^^^^^^^^^^^^^^^^^^^^
.. include:: examples/simple_bloat_loop
Simple bloat function example
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
.. include:: examples/simple_bloat_function
Resulting in this output:
Simple bloat loop example#
Warning
The pw_size_report_toolchains build variable is empty for this target.
Size reports will not be generated.
See Defining size reports for details on how to set up size reports.
Simple bloat function example#
Warning
The pw_size_report_toolchains build variable is empty for this target.
Size reports will not be generated.
See Defining size reports for details on how to set up size reports.
Additional Bloaty data sources#
Bloaty McBloatface by itself cannot help
answer some questions which embedded developers frequently face such as
understanding how much space is left. To address this, Pigweed provides Python
tooling (pw_bloat.bloaty_config) to generate bloaty configuration files
based on the final ELF files through small tweaks in the linker scripts to
expose extra information.
See the sections below on how to enable the additional data sections through modifications in your linker script(s).
As an example to generate the helper configuration which enables additional data
sources for example.elf if you’ve updated your linker script(s) accordingly,
simply run
python -m pw_bloaty.bloaty_config example.elf > example.bloaty. The
example.bloaty can then be used with bloaty using the -c flag, for
example
bloaty -c example.bloaty example.elf --domain vm -d memoryregions,utilization
which may return something like:
84.2% 1023Ki FLASH
94.2% 963Ki Free space
5.8% 59.6Ki Used space
15.8% 192Ki RAM
100.0% 192Ki Used space
0.0% 512 VECTOR_TABLE
96.9% 496 Free space
3.1% 16 Used space
0.0% 0 Not resident in memory
NAN% 0 Used space
utilization data source#
The most common question many embedded developers face when using bloaty is
how much space you are using and how much space is left. To correctly answer
this, section sizes must be used in order to correctly account for section
alignment requirements.
The generated utilization data source will work with any ELF file, where
Used Space is reported for the sum of virtual memory size of all sections.
In order for Free Space to be reported, your linker scripts must include
properly aligned sections which span the unused remaining space for the relevant
memory region with the unused_space string anywhere in their name. This
typically means creating a trailing section which is pinned to span to the end
of the memory region.
For example imagine this partial example GNU LD linker script:
MEMORY
{
FLASH(rx) : \
ORIGIN = PW_BOOT_FLASH_BEGIN, \
LENGTH = PW_BOOT_FLASH_SIZE
RAM(rwx) : \
ORIGIN = PW_BOOT_RAM_BEGIN, \
LENGTH = PW_BOOT_RAM_SIZE
}
SECTIONS
{
/* Main executable code. */
.code : ALIGN(8)
{
/* Application code. */
*(.text)
*(.text*)
KEEP(*(.init))
KEEP(*(.fini))
. = ALIGN(8);
/* Constants.*/
*(.rodata)
*(.rodata*)
} >FLASH
/* Explicitly initialized global and static data. (.data)*/
.static_init_ram : ALIGN(8)
{
*(.data)
*(.data*)
. = ALIGN(8);
} >RAM AT> FLASH
/* Zero initialized global/static data. (.bss) */
.zero_init_ram : ALIGN(8)
{
*(.bss)
*(.bss*)
*(COMMON)
. = ALIGN(8);
} >RAM
}
Could be modified as follows enable Free Space reporting:
MEMORY
{
FLASH(rx) : ORIGIN = PW_BOOT_FLASH_BEGIN, LENGTH = PW_BOOT_FLASH_SIZE
RAM(rwx) : ORIGIN = PW_BOOT_RAM_BEGIN, LENGTH = PW_BOOT_RAM_SIZE
}
SECTIONS
{
/* Main executable code. */
.code : ALIGN(8)
{
/* Application code. */
*(.text)
*(.text*)
KEEP(*(.init))
KEEP(*(.fini))
. = ALIGN(8);
/* Constants.*/
*(.rodata)
*(.rodata*)
} >FLASH
/* Explicitly initialized global and static data. (.data)*/
.static_init_ram : ALIGN(8)
{
*(.data)
*(.data*)
. = ALIGN(8);
} >RAM AT> FLASH
/* Zero initialized global/static data. (.bss). */
.zero_init_ram : ALIGN(8)
{
*(.bss)
*(.bss*)
*(COMMON)
. = ALIGN(8);
} >RAM
/*
* Do not declare any output sections after this comment. This area is
* reserved only for declaring unused sections of memory. These sections are
* used by pw_bloat.bloaty_config to create the utilization data source for
* bloaty.
*/
.FLASH.unused_space (NOLOAD) : ALIGN(8)
{
. = ABSOLUTE(ORIGIN(FLASH) + LENGTH(FLASH));
} >FLASH
.RAM.unused_space (NOLOAD) : ALIGN(8)
{
. = ABSOLUTE(ORIGIN(RAM) + LENGTH(RAM));
} >RAM
}
memoryregions data source#
Understanding how symbols, sections, and other data sources can be attributed
back to the memory regions defined in your linker script is another common
problem area. Unfortunately the ELF format does not include the original memory
regions, meaning bloaty can not do this today by itself. In addition, it’s
relatively common that there are multiple memory regions which alias to the same
memory but through different buses which could make attribution difficult.
Instead of taking the less portable and brittle approach to parse *.map
files, pw_bloat.bloaty_config consumes symbols which are defined in the
linker script with a special format to extract this information from the ELF
file: pw_bloat_config_memory_region_NAME_{start,end}{_N,}.
These symbols are then used to determine how to map segments to these memory regions. Note that segments must be used in order to account for inter-section padding which are not attributed against any sections.
As an example, if you have a single view in the single memory region named
FLASH, then you should produce the following two symbols in your linker
script:
pw_bloat_config_memory_region_FLASH_start = ORIGIN(FLASH);
pw_bloat_config_memory_region_FLASH_end = ORIGIN(FLASH) + LENGTH(FLASH);
As another example, if you have two aliased memory regions (DCTM and
ITCM) into the same effective memory named you’d like to call RAM, then
you should produce the following four symbols in your linker script:
pw_bloat_config_memory_region_RAM_start_0 = ORIGIN(ITCM);
pw_bloat_config_memory_region_RAM_end_0 = ORIGIN(ITCM) + LENGTH(ITCM);
pw_bloat_config_memory_region_RAM_start_1 = ORIGIN(DTCM);
pw_bloat_config_memory_region_RAM_end_1 = ORIGIN(DTCM) + LENGTH(DTCM);