/cc @dr2chase @cherrymui @jeremyfaller
Maybe this test failure is real and the linker is failing to eliminate data it used to?

Friendly ping.

This issue currently has the label Soon, which is for regressions, serious bugs, and outages. This doesn't seem to qualify for that, so I'll remove the Soon label.

Sorry 'bout that. Probably got lost in the runup to 1.14 freeze. I'll take a look.

No problem, and thank you Jeremy. I just wanted to update our list of Soon issues.

My guess is that this is due to alignment padding between segments. It seems the failures are all on architectures with large alignment (usually 0x10000 i.e. 64K). Worst case the alignment padding can increase the binary size by 64K, which can shadow the 50K difference we're looking for.

If the test is checking if the table is dropped by the linker, maybe use the nm command to look for particular symbols?

Or it can add the size of sections without the paddings. But then it needs to be OS dependent: will need codepaths for ELF, Mach-O, xcoff, etc..

Or the test can use linker's -R flag to reduce the alignment. This might cause problems at run time, but we don't run the resulted binaries.

Change https://golang.org/cl/221104 mentions this issue: unicode/norm: look for symbols in TestLinking