Compiler profiles for size optimization #2973

TeroJaasko · 2016-10-10T15:41:36Z

Description

This pull request adds two new compiler profiles, one for size optimization for release builds and one for minimally optimized debug build.

Status

READY/IN DEVELOPMENT/HOLD

Migrations

If this PR changes any APIs or behaviors, give a short description of what API users should do when this PR is merged.

NO

Related PRs

List related PRs against other branches:

branch	PR
other_pr_production	link
other_pr_master	link

Todos

Tests
Documentation
Review by @theotherjimmy
Review by @screamerbg

Deploy notes

Steps to test or reproduce

mbed compile -m K64F -t ARM --profile mbed-os/tools/profiles/size.json
mbed compile -m K64F -t GCC_ARM --profile mbed-os/tools/profiles/debug-size.json

The default profile in mbed-os optimizes for size on GCC byt for speed on ARMCC and IAR. Add a profile which optimizes for size on all compilers. This file is just a copy of default.json with the different optimization targets. Before (ARMCC, default.json): Total Static RAM memory (data + bss): 72112 bytes Total RAM memory (data + bss + heap + stack): 72112 bytes Total Flash memory (text + data + misc): 358112 bytes After (ARMCC, size.json): Total Static RAM memory (data + bss): 72112 bytes Total RAM memory (data + bss + heap + stack): 72112 bytes Total Flash memory (text + data + misc): 314359 bytes Before (IAR, default.json): Total Static RAM memory (data + bss): 70088 bytes Total RAM memory (data + bss + heap + stack): 168392 bytes Total Flash memory (text + data + misc): 309739 bytes After (IAR, size.json) Total Static RAM memory (data + bss): 70092 bytes Total RAM memory (data + bss + heap + stack): 168396 bytes Total Flash memory (text + data + misc): 290523 bytes

…bled On some cases the mbed-os's debug profile is unusable, as it disables all optimizations to get best possible debugging environment which may make the code too big to actually fit into system. This new profile is based on the debug.json, but it enables the lowest level of compiler optimizations along with the debug flag. Debugging with this may not be optimal, but on some cases it is actually possible at all. Before (GCC 5.4.1, debug.json): Total Static RAM memory (data + bss): 74144 bytes Total RAM memory (data + bss + heap + stack): 139684 bytes Total Flash memory (text + data + misc): 598554 bytes After (GCC 5.4.1, debug_size.json): Total Static RAM memory (data + bss): 73624 bytes Total RAM memory (data + bss + heap + stack): 139164 bytes Total Flash memory (text + data + misc): 392197 bytes Before (ARMCC 5.06u3, debug.json): Total Static RAM memory (data + bss): 72100 bytes Total RAM memory (data + bss + heap + stack): 72100 bytes Total Flash memory (text + data + misc): 412428 bytes After (ARMCC 5.06u3, debug_size.json): Static RAM memory (data + bss): 72112 bytes Total RAM memory (data + bss + heap + stack): 72112 bytes Total Flash memory (text + data + misc): 334792 bytes Before (IAR, debug.json): Total Static RAM memory (data + bss): 70115 bytes Total RAM memory (data + bss + heap + stack): 168419 bytes Total Flash memory (text + data + misc): 407543 bytes After (IAR, debug_size.json): Total Static RAM memory (data + bss): 70115 bytes Total RAM memory (data + bss + heap + stack): 168419 bytes Total Flash memory (text + data + misc): 377093 bytes

TeroJaasko · 2016-10-10T17:19:31Z

As a side note, the debug build compiled with ARMCC for K64F fails to load properly by gdb, but binary still does work without debugger. The same happens with stock debug.json and the cloned debug-size.json. Gdb shows this warning when it starts:
--8<---
Reading symbols from BUILD/K64F/ARM/mbed-cloud-client-example.elf...
warning: Loadable section "RW_m_data" outside of ELF segments
warning: Loadable section "RW_m_data_2" outside of ELF segments
done.
--8<---

0xc0170 · 2016-10-11T09:49:46Z

cc @sg- @screamerbg @bridadan

bridadan

It seems fine to me.

The only question I have is the default setting for GCC_ARM already optimizes for size. Does anyone have concerns about this causing confusion?

sg- · 2016-10-11T16:21:02Z

This is a good start. We need some docs about the design principles behind each one and a second set of eyes that the right settings are chosen across the toolchains. Dont want to see people flipping these back and forth. If needed local profiles can be used rather than a collection of similar but different ones. @c1728p9

c1728p9 · 2016-10-12T23:49:59Z

Something like debug-size.json might be better left as a local profile. It would be good to add a size profile though. I created #3005 which bases the small profile on the size profile here. I tried to boil it down to the 3 most common use cases. Let me know what you think.

TeroJaasko · 2016-10-13T07:14:31Z

The debug-size kind of profile is something that we need every time a debug session is needed. IMHO, it truely should be part of mbed os, as the problem is same on many real world applications which do their best to add as much as features as ROM allows.

0xc0170 · 2016-10-13T10:57:00Z

@TeroJaasko Can you review referenced PR by @c1728p9 please? And agree if that one makes sense and this one will be an addition ? Or ?

screamerbg · 2016-10-14T11:08:51Z

The profiles look good to me. Please address @c1728p9. I'm up for the size profile. Note that the more profiles we have, the more confusing it would get.
@sg- @theotherjimmy

teetak01 · 2016-10-18T11:45:22Z

This looks ready. Can we get this merged ?

0xc0170 · 2016-10-18T12:14:24Z

This looks ready. Can we get this merged ?

Final review by @c1728p9 and @sg- please

For me size profile +1, not certain about debug-size. A user can tweak it, we got debug already, and 3 other profiles, plus this size. So each of them should get debug?

sg- · 2016-10-19T11:35:53Z

@TeroJaasko @c1728p9 @teetak01

Please work together to get this and #3005 unified. Less profiles are better and local profiles should be encouraged. Profiles without docs and description statement are not helpful. Next we'll see samgro01-profile.json checked in and so on.

Also, need verification that the flags are profiles are best picked across each toolchain. Report of build output for each would be nice.

c1728p9 · 2016-10-19T16:15:14Z

tools/profiles/size.json

@@ -0,0 +1,44 @@
+{


Size doesn't really imply big or small. Maybe rename this file to small.json or something which implies that this is smaller than the default?

These profiles need docs. Feel free to pull the DOCs from pr #3005.

c1728p9 · 2016-10-19T16:19:13Z

tools/profiles/debug-size.json

+                   "-fmessage-length=0", "-fno-exceptions", "-fno-builtin",
+                   "-ffunction-sections", "-fdata-sections", "-funsigned-char",
+                   "-MMD", "-fno-delete-null-pointer-checks",
+                   "-fomit-frame-pointer", "-O1", "-g"],


With -O1 the optimization level is changed from the size.json configuration. This is one more configuration to verify. I would recommend sticking with just one size profile.

c1728p9 · 2016-10-19T16:28:24Z

tools/profiles/debug-size.json

+                   "-fmessage-length=0", "-fno-exceptions", "-fno-builtin",
+                   "-ffunction-sections", "-fdata-sections", "-funsigned-char",
+                   "-MMD", "-fno-delete-null-pointer-checks",
+                   "-fomit-frame-pointer", "-O1", "-g"],


I believe the reason debug symbols aren't turn on in all configurations is because this increases build sizes too much for testing and the online IDE. @screamerbg, @bridadan or @0xc0170 can you confirm this? If debug symbols could be turned on/off independent of the profile the each of the profiles wouldn't need have a corresponding "debug-*" profile to add symbols. This would be especially useful if the binary output doesn't change (which it shouldn't).

@c1728p9 That's correct. Good catch!

sg- · 2016-10-25T22:52:37Z

@TeroJaasko can you please work with @c1728p9 to get comments resolved or close this out and make requests against #3005

TeroJaasko added 2 commits October 10, 2016 18:12

TeroJaasko mentioned this pull request Oct 10, 2016

Developer should have a way to select the ARMCC compiler optimization target #2600

Closed

theotherjimmy approved these changes Oct 10, 2016

View reviewed changes

bridadan approved these changes Oct 11, 2016

View reviewed changes

sg- added the needs: review label Oct 11, 2016

sg- added needs: work and removed needs: review labels Oct 19, 2016

c1728p9 suggested changes Oct 19, 2016

View reviewed changes

sg- closed this Nov 1, 2016

kjbracey mentioned this pull request Jun 12, 2019

Make ARMC5 and IAR develop profile also size optimized #10813

Merged

Compiler profiles for size optimization #2973

Compiler profiles for size optimization #2973

Uh oh!

Conversation

TeroJaasko commented Oct 10, 2016 • edited by screamerbg Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Status

Migrations

Related PRs

Todos

Deploy notes

Steps to test or reproduce

Uh oh!

TeroJaasko commented Oct 10, 2016

Uh oh!

0xc0170 commented Oct 11, 2016

Uh oh!

bridadan left a comment

Choose a reason for hiding this comment

Uh oh!

sg- commented Oct 11, 2016

Uh oh!

c1728p9 commented Oct 12, 2016

Uh oh!

TeroJaasko commented Oct 13, 2016

Uh oh!

0xc0170 commented Oct 13, 2016 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

screamerbg commented Oct 14, 2016

Uh oh!

teetak01 commented Oct 18, 2016

Uh oh!

0xc0170 commented Oct 18, 2016

Uh oh!

sg- commented Oct 19, 2016

Uh oh!

c1728p9 Oct 19, 2016

Choose a reason for hiding this comment

Uh oh!

c1728p9 Oct 19, 2016

Choose a reason for hiding this comment

Uh oh!

c1728p9 Oct 19, 2016

Choose a reason for hiding this comment

Uh oh!

c1728p9 Oct 19, 2016

Choose a reason for hiding this comment

Uh oh!

screamerbg Oct 19, 2016

Choose a reason for hiding this comment

Uh oh!

sg- commented Oct 25, 2016

Uh oh!

Uh oh!

TeroJaasko commented Oct 10, 2016 •

edited by screamerbg

Loading

0xc0170 commented Oct 13, 2016 •

edited

Loading