Looking at AWSY: https://areweslimyet.com/ We see ~15MB regression in RSS after TP5 which seems to come completely from JS. Here is the relevant pushlog: https://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?fromchange=2629257557ff&tochange=73055a66d95a Its been suggested that this might be due to bug 984537.

Bug 984537 does seem like a likely culprit. Looking at the first data point before and the first after the regression: - js-main-runtime/runtime went from 12.6 to 25.77 - explict/js-non-window/runtime/code went from to 1280 KiB to 14.38 MiB, with the following details: Before: > - 1,280.00 KiB ( 9.92%) code > - 996.52 KiB (77.85%) unused > - 200.48 KiB (15.66%) baseline > - 46.93 KiB ( 3.67%) other > - 30.66 KiB ( 2.40%) ion > - 5.41 KiB ( 0.42%) regexp After > - 14.38 MiB (55.79%) code > - 12.91 MiB (89.82%) unused > - 1,399.57 KiB ( 9.51%) baseline > - 50.38 KiB ( 0.34%) ion > - 46.33 KiB ( 0.31%) other > - 2,040.00 B ( 0.01%) regexp So it's mostly unused code, though baseline code also jumped a lot. The minimum code chunk size is 64 KiB, IIRC. I wonder if this is fragmentation, i.e. many of these chunks are mostly empty, or if we're somehow holding onto completely empty chunks.

Tracking this for 32 and 31 for now.

Backouts for both branches have landed, so we just need to wait for AWSY to update across the backout to confirm the fix.

AWSY has dropped back down. Bug 984537 comment 32 has the details.