size optimization for "V4" firmware (-flto)

[softhack007]

This is a continuation (and replacement) for

• Enable Link Time Optimisation #3824 by @deece

Thanks to the contribution made by @deece to platform-espressif32, link time optimization is now possible in esp32 framework version 6.6.0 and above without hacking build scripts (8266 - no way, sorry).

Overview and background

https://gcc.gnu.org/onlinedocs/gcc/Optimize-Options.html#index-flto

• LTO generally reduces firmware size by post-optimizing the binary when the linker runs, considering all pieces of the code instead of looking at each ".cpp.o" in isolation.
• In my first tests, there was no additional speedup, however sometimes the stack usage is more than previously.
• LTO makes it very hard to read exception decoder dumps, so I've added "-g3 -ggdb" which is recommended to improve the situation
• esp32-C3: not possible to apply LTO, due to compiler bugs
• Currently we can only use LTO on "V4" builds. I've enabled it for the 4MB binaries which benefit the most (see table below).

size comparison

firmware	0_15 without -flto	0_15 with -flto
esp32_wrover 83.6kB saved	RAM: 4.6% (used 60156 bytes); Flash: 92.3% (used 1572241 bytes)	RAM: 4.6% (used 60140 bytes) Flash: 87.4% (used 1488605 bytes)
esp32dev_8M 80.5kB saved	RAM: 18.3% (used 60084 bytes); Flash: 73.0% (used 1530305 bytes)	RAM: 18.3% (used 60052 bytes); Flash: 69.1% (used 1449817 bytes)
esp32s3_4M_qspi 71.7kB saved	RAM: 18.4% (used 60288 bytes); Flash: 91.4% (used 1437945 bytes)	RAM: 18.4% (used 60264 bytes); Flash: 86.9% (used 1366221 bytes)
lolin_s2_mini 72.3kB saved	RAM: 20.2% (used 66100 bytes); Flash: 92.7% (used 1458094 bytes)	RAM: 20.2% (used 66100 bytes); Flash: 88.1% (used 1385770 bytes)

[softhack007]

Update: omitting "-g3 -ggdb" saves another 26 kB on the wrover build.

Flash: 85.8% (used 1461869 bytes)

[blazoncek]

This is beyond my skill level but IMO if we can resolve flash use increase in latest platforms it would be best if we upgraded ESP32 to use newer platform in general. Platform 3.5.0 is ancient to say the least.

[softhack007]

This is beyond my skill level but IMO if we can resolve flash use increase in latest platforms it would be best if we upgraded ESP32 to use newer platform in general. Platform 3.5.0 is ancient to say the least.

yes, this is what i was thinking, too. 80-100Kb might be enough to recover from the core size increase in the new framework.

The only problem that remains - as explained previously:

in userland most people flash their devices once from install.wled.me, then update by OTA. Unfortunately, OTA cannot update the bootloader. Going to the new framework means that we'll have lots of "V3 bootloader and V4 firmware" setups. In the past this caused very unresponsive devices (suspected case LittleFs). Not sure if this is still happening. We definitely need more testing to safeguard the "OTA to V4" upgrade path on esp32.

So we might go for this change after 0.15.0 is released - to have some more time for testing. We also need to check how development is affected. My first experience with LTO was "WTF" when i got my first exception decoder dump, and it was completely gibberish (all functions renamed to "zaXXXXX").

[softhack007]

Another option to decrease flash size is creating a custom arduino-esp32 where unneeded features (rainmaker, insights, bluetooth, etc) are removed. I've never tried it, but @Aircoookie did something similar for the old framework. And @troyhacks has some experiences with this for the new arduino framework. Plan B: use a Tasmota framework, as the tasmota team is already having their own custom arduino framework that saves around 150Kb program size.

[troyhacks]

@softhack007 - yes, absolutely can reliably build the v4 and v5 frameworks now without the overhead (or at least test that out)

My v5.3 for the ESP32-P4 has all of the extras removed. No Rainmaker, Insights, NN, etc.