Excessive Memory Usage on Linux when Compiling Large Arrays

[NeedleInAJayStack]

Describe the bug
When compiling large array literals in a Linux environment, Swift uses an extremely large amount of memory. For example, memory usage exceeds 500MB when compiling a 1000 element array literal composed of simple structs. If the array is declared as an existential, memory usage balloons to 2GB.

Steps To Reproduce
Steps to reproduce the behavior:

1. Clone this reproducer repo: https://github.com/NeedleInAJayStack/CompileMemUsageReproducer
2. Follow the workflow in the readme

Watching the memory usage via docker stats is definitely not ideal, but I'm not familiar with a better approach to extract peak memory usage or diagnostics from the Swift compiler. I'm definitely open to suggestions here. When collecting the observed peak usage numbers, the swift build command was run multiple times in a row and the maximum number observed in docker stats was recorded.

Expected behavior
Large array literals are able to be compiled with a much smaller amount of memory than currently observed. At runtime these arrays are significantly less than 1MB, yet take many times that amount to compile.

Environment (please fill out the following information)

• OS: Ubuntu 20.04
• Swift: Nightly toolchain (August 10 at time of filing)
• Xcode Version/Tag/Branch: N/A

Additional context
This may be related to #58380

[Austinpayne]

cc: @asl

[asl]

The issue could be reproduced on MacOS as well. Besides the memory consumption, the running time also depends on # of elements. Here is the profiler trace for LargeExistentialArray.swift:

LargeObjectArray.swift has the peak memory consumption / time also in constraint solver.