PGO cannot be used via cargo because of how `RUSTFLAGS` are handled.

[michaelwoerister]

Problem

PGO works in two phases:

1. Compile an instrumented binary of your program that enables collecting profile data. This is done via RUSTFLAGS=-Cprofile-generate.
2. Compile an optimized binary that makes use of the profile data collected in step (1). This is done via RUSTFLAGS=-Cprofile-use.

In order for this to work, the symbol names within the two program versions must match up because LLVM generates profiling data in terms of symbol names.

However, since #6503, RUSTFLAGS are fed into the -Cmetadata argument to rustc. This causes symbol names to differ because one version has -Cprofile-generate and the other has -Cprofile-use in their RUSTFLAGS.

I think I would prefer not feeding RUSTFLAGS into -C metadata at all. It does not seem right to me that symbol names are affected by random command line arguments to the compiler.

[alexcrichton]

I've proposed a revert of #6503 at #7417 to fix this.

@michaelwoerister out of curiosity was this the cause of why y'all weren't seeing great wins for PGO in Firefox?

[michaelwoerister]

@alexcrichton I have confirmed that this issue prevents Firefox from getting any PGO wins, yes. However, I don't know yet if it is the only issue (e.g. GNU ld seems to have a bug that sometimes causes instrumented builds to not record anything 😱).

I have some microbenchmarks that show a 10-15% win with PGO but have not been able to replicate that with the regex benchmark suite, which is my next bigger test case.

[alexcrichton]

Oh dear that must have been absolutely gnarly to figure this out, sorry about that!

[michaelwoerister]

It's exactly the kind of issue I expected to encounter :)