Improve LLVM IR Code generation

[Aatch]

This is a metabug to group together a bunch of related issues do with our current generation of LLVM IR.

The IR we produce is very verbose, this causes an inordinate amount of time to be spent during the optimization passes that we could trivially prevent by simply not outputting these sequences.

Instrumentation and testing issues

• 5073 : begin testing emitted LLVM

• #7456 : improved trans stats
• #7461 : LLVM codegen equivalence tests

  Too many copies of functions issues

(see http://people.mozilla.org/~graydon/symbols-by-name.txt)

• #8651 : rewrite type_use
• #2529 : normalize for monomorphization recursively
• #2537 : type_use should only consider fields before the selected field

• 7462 : match combinatorial codegen problems

• 7349 : multiply-instantiating monomorphic closures within polymorphic functions

• 8650 : reuse default methods

  Too much codegen issues

• 1980 : cross-crate monomorphic instance reuse

• 5016 : move semantics

• 6736 : tbaa.struct for skipping padding

• 8648 : intrinsics for trivial functions

  Incorrect or LLVM-frustrating codegen issues

• 7463 : LLVM lint-cleanness

• 4353 : fastisel vs. switch

• 3551 : fastisel vs. invoke

• 3298 : asm comments

  Overall use of LLVM issues

• #3519 : better LLVM optimizations

[graydon]

See -Z count-llvm-insns and pipe output to xdu.

[Aatch]

@graydon or, for people that do not have xdu, if you have xdiskusage instead, you can do rustc foo.rs -Z count-llvm-insns | xdiskusage

[farnoy]

Sorry for going off topic a bit, but this could improve readability of these kinds of metabugs: https://github.com/blog/1375-task-lists-in-gfm-issues-pulls-comments

[thestinger]

I think this kind of categorization is covered well by the usage of tags (I-slow, I-compiletime). It's not ever going to be possible to complete this metabug or mark it as a milestone so I don't see much value in leaving it open forever. The discussion here is out of date too.