Switch translate() to the header file

[tannewt]

This allows the compile stage to optimize most of the translate()
function away and saves a ton of space (~40k on ESP). However, it
requires us to wait for the qstr output before we compile the rest
of our .o files. (Only qstr.o used to wait.)

This isn't as good as the current setup with LTO though. Trinket M0
loses <1k with this setup.

So, we should probably conditionalize this along with LTO.

[d-c-d]

What caused the Trinket M0
To grow by 1k with this setup? ( duplication of strings? )
Haven't studied the changes yet

[dhalbert]

I can experiment with this on other builds. Definitely conditionalize on LTO.

[dhalbert]

If you compiled translate.c with -O3 or something like that, would it help the LTO builds (or the non-LTO builds)?

[tannewt]

What caused the Trinket M0 To grow by 1k with this setup? ( duplication of strings? ) Haven't studied the changes yet

I suspect LTO is duplicating copies of the compressed data because each compilation unit (.o) now has it's own copy. So if the same error occurs in two files, there will now be two copies of it. I haven't proven this though.

[tannewt]

Ok, @dhalbert. This finally built. There is a lot of diff noise due to the header move. Let me know if you want me to clean it up. I considered folding the translate.h include into py/runtime.h but the include-what-you-use philosophy would have you list it I think.

[dhalbert]

This is great work! One formatting oddity, and one change to improve the compile times.

[dhalbert]

The LTO builds are taking about the same amount of time, but the non-LTO builds are now much longer, and each translate build takes about 3x the time of before. Basically the non-LTO builds are now catching up to the LTO builds in slowness. But it does save a lot of space.

So the entire latest PR build was about 132 minutes, compared with about 75 minutes before.

atmel-samd (LTO), before and after:

espressif, before and after:

[dhalbert]

I am thinking about two ways around the long build times:

1. On a PR, do only representative builds: English, and a few large builds, like de, ja, ru or even fewer. So kind of like the windows-builds. Do the complete set only on merge.
2. More radical: give up on using real strings in translate(), and use message ids, e.g. translate(MSG_INVALID_Q_PIN) or whatever. I don't mean qstr symbol nameparsing, they would be too long, but a central message table file. It could still be compiled per build, or if we can use id numbers, maybe it can just be a library to link in, with a separate section or whatever for each message, so there are no unused compressed message strings in a build.

[tannewt]

I think the simplest thing would be to have the TRANSLATE_OBJECT version on by default and only do the header thing when we need the space (like small S3 builds.)

[dhalbert]

I reran the latest build when nothing else was queued, and got a total runtime of 1h 16m 18s, which is only 10 minutes more than what I was getting in #6436. Thanks for spending the time on this: the space savings is great!

[dhalbert]

The single failure was a transient CI issue.