Our armv7 builds are currently built with -mthumb but our armv6 builds are not. We should build armv6 with -mthumb as well (and ensure the resulting code is smaller)

I started a build with the attached mozconfig (basically a copy of my usual mozconfig but for armv6 and with thumb enabled) and it died while building some of the JS code. I'll attach a build log output file.

Right, the main reason we didn't do this at the time was because we actually have a bunch of code that doesn't build as thumb-1.

Also, just for reference, I built fennec for ARMv7 with and without thumb enabled. The thumb APKs was only ~407 KiB smaller, but once I extracted the libraries inside there was a much larger size difference (with the lion's share going to a 16 MiB savings in libxul.so): kats@kgupta-pc thumb$ (find . -type f | while read file; do SIZE=$(cat $file | wc -c); NFILE=$(echo $file | sed -e 's/-v7a//'); NSIZE=$(cat ../nothumb/$NFILE | wc -c); if [ $SIZE -ne $NSIZE ]; then echo "$((NSIZE - SIZE)) $file"; fi; done) | sort -rn 16434464 ./libxul.so 516092 ./libnss3.so 344064 ./libmozsqlite3.so 135168 ./libfreebl3.so 126664 ./lib/armeabi-v7a/libmozglue.so 103008 ./libssl3.so 102400 ./libsoftokn3.so 88324 ./libnspr4.so 65532 ./libsmime3.so 45056 ./libnssckbi.so 42344 ./libnssutil3.so 5976 ./lib/armeabi-v7a/libomxplugin.so 4440 ./libplc4.so 2800 ./libplds4.so 1828 ./libmozalloc.so 1760 ./lib/armeabi-v7a/libplugin-container.so 1144 ./libxpcom.so 5 ./omni.ja -12 ./META-INF/MANIFEST.MF -12 ./META-INF/CERT.SF

It's hard to draw comparisons there, since thumb1 and thumb2 are not the same. We'd have to invest some effort to make things compile with thumb1 to even test.

All ARMv7 support thumb-2, but not all ARMv6 support thumb-2. Random example: the HTC Magic has an ARM1136 ARMv6, which doesn't support thumb-2. These CPUs however do support thumb-1, and it has different size patterns compared to thumb-2. Also, the additional instructions available in armv7 and thumb-2 make for more compact code.

(In reply to Ted Mielczarek [:ted.mielczarek] from comment #5) > It's hard to draw comparisons there, since thumb1 and thumb2 are not the > same. We'd have to invest some effort to make things compile with thumb1 to > even test. Fair enough. I'll take a crack at doing this. I'll probably just disable stuff that doesn't build as thumb1 for a first cut. I assume there isn't so much of that stuff that it will make a significant difference.

I'm not even sure there are phones with ARM1156, supporting thumb-2.

Another note, on older cores thumb-1 may actually be slower than arm instructions.

(In reply to Mike Hommey [:glandium] from comment #9) > Another note, on older cores thumb-1 may actually be slower than arm > instructions. Not just older cores, thumb-1 is slower in most cases AFAIK. Still worth the investigation though.

I built with the attached patch with thumb and non-thumb. Diffs show that libxul is ~8.7 megs smaller with thumb. kats@kgupta-pc armv6-thumb$ (find . -type f | while read file; do SIZE=$(cat $file | wc -c); NFILE=$(echo $file | sed -e 's/-v7a//'); NSIZE=$(cat ../armv6-nothumb/$NFILE | wc -c); if [ $SIZE -ne $NSIZE ]; then echo "$((NSIZE - SIZE)) $file"; fi; done) | sort -rn 8794032 ./libxul.so 315380 ./libnss3.so 209776 ./libmozsqlite3.so 61872 ./libssl3.so 61432 ./libsoftokn3.so 58728 ./lib/armeabi-v7a/libmozglue.so 50100 ./libnspr4.so 36844 ./libfreebl3.so 33220 ./libsmime3.so 28668 ./libnssckbi.so 23268 ./libnssutil3.so 2872 ./lib/armeabi-v7a/libomxplugin.so 2164 ./libplc4.so 1384 ./libmozalloc.so 1376 ./lib/armeabi-v7a/libplugin-container.so 1248 ./libplds4.so 1156 ./libxpcom.so -10 ./omni.ja -12 ./META-INF/MANIFEST.MF -12 ./META-INF/CERT.SF

Here is a better patch that (I think) should provide reasonable fallback code paths for all of the thumb2 stuff. Compiles but completely untested.

Adding a see-also bug I split off for one of the issues I found while working on this, as it has some relevant discussion that I don't want to lose a pointer to.

We're dropping support for armv6.