Reduce stage1 peak memory usage to within 3.5 GiB

[ghost]

No description provided.

[tiehuis]

I've just checked this on the current latest master to get some more actionable results:

In summary the following tests were the most memory intensive. I did not hit anywhere near 4Gb, but with other things running on the system at the same time it isn't unlikely to hit that.

• test-std 880Mb peak
• test-compiler-rt 1.76Gb peak

time -v

1539.64Mb

Command being timed: "zig build --build-file ../build.zig test"
	User time (seconds): 451.28
	System time (seconds): 31.24
	Percent of CPU this job got: 98%
	Elapsed (wall clock) time (h:mm:ss or m:ss): 8:12.10
	Average shared text size (kbytes): 0
	Average unshared data size (kbytes): 0
	Average stack size (kbytes): 0
	Average total size (kbytes): 0
	Maximum resident set size (kbytes): 1576592
	Average resident set size (kbytes): 0
	Major (requiring I/O) page faults: 3159
	Minor (reclaiming a frame) page faults: 21121360
	Voluntary context switches: 35057
	Involuntary context switches: 132143
	Swaps: 0
	File system inputs: 507560
	File system outputs: 1051144
	Socket messages sent: 0
	Socket messages received: 0
	Signals delivered: 0
	Page size (bytes): 4096
	Exit status: 0

Massif

This seems incorrect and reporting too low.

valgrind --tool=massif --pages-as-heap=yes --massif-out-file=massif.out zig build --build-file ../build.zig test

Manual \w Htop

Just to confirm I watched a run using htop.

Start: 2.66Gb
Peak: 4.36Gb

Notable peaks were 3.54Gb during test-std and 4.36Gb during test-compiler-rt so I would target those segments. The rest of the tests memory consumption were negligible in comparison.

[andrewrk]

There are a lot of memory inefficiencies in the c++ compiler which I never really expect to fix. However I do want to keep the self hosted compiler memory efficient. I'm also considering having a feature where you can upper bound the memory and it would utilize disk space beyond that.

[andrewrk]

I added -DZIG_ENABLE_MEM_PROFILE option and -fmem-report flag in #3482. @mikdusan is doing some work on this in #3502.

Here is what I observe with zig-0.5.0+e98e5d, with a fresh build (nothing cached):

• test-std: 4.3 GiB peak
• test-compiler-rt: 2.3 GiB peak

I'm going to clarify the requirements of stage1 here: the goal is to get stage1 memory usage down to 3.5 GiB peak for the entire test suite. The reasoning for this is straightforward: 4 GiB is how much RAM we want zig to require a CI server to have. We leave 0.5 GiB for the base system and 3.5 peak for all the tests - including building the self-hosted compiler. This number is significant because it is the memory requirement of bootstrapping.

Once this goal is accomplished we could use something like time -v -o time.txt zig build test and then look at the results, failing the build if the memory usage exceeded 3.5 GiB peak.

[andrewrk]

Duplicate of #6485