Move PTLS load emission from codegen to late-gc-lowering

[pchintalapudi]

PTLS load instructions can interfere with optimizing allocations due to being unhoistable and inseparable from their corresponding gc_alloc_obj call. By moving their generation to late-gc-lowering, we can optimize allocation function calls directly without worry about the PTLS load.

[vchuravy]

@pchintalapudi thanks for opening this PR.

IIRC you had a couple of testcases that we were looking at? Could you add them to test/llvmpasses?

[vchuravy]

@nanosoldier runtests(ALL, vs = ":master")

[nanosoldier]

Your package evaluation job has completed - possible new issues were detected. A full report can be found here.

[vchuravy]

@vtjnash it is my understanding that currently we need to always redo the ptls load from the current_task. @pchintalapudi is looking at enabling hoisting of allocations from loops and with the ptls load the allocation call is no longer invariant.

Do you see any reason for why emitting the load earlier might be benefitial? (Outside starting to add invariant range annotations.)

[vtjnash]

I'd suggest passing pptls there, since the GEP computation adds unnecessary cost, but otherwise SGTM

[vtjnash]

Actually, I guess later optimization passes are likely to fold that away regardless

[vchuravy]

@pchintalapudi the last thing you will need to fix is the llvmpasses failure https://buildkite.com/julialang/julia-master/builds/4661#78e454c5-f38e-4605-b5ed-0346286be2f0

[DilumAluthge]

@pchintalapudi Can you rebase this PR on the latest master? This should fix the Downloads tests.