Begin testing emitted LLVM IR

[bstrie]

<@pcwalton> strcat: it would be nice to have if foo == true just compile
            into mov eax,foo; test eax,eax; jne ... instead of mov
            eax,whatever; cmp eax,1; jne ...
<@pcwalton> err
<@pcwalton> mov eax,whatever; cmp al,1; jne ...
< strcat> yeah
< strcat> pcwalton: once the range asserts are there, LLVM will compile
          trunc to i1 and then a br to that
< strcat> that's what clang does
< strcat> if you don't put range asserts, it uses an AND for the trunc
<@pcwalton> eww
<@pcwalton> and al,feh I guess?
<@pcwalton> err
<@pcwalton> what am I on
<@pcwalton> and al,1
< strcat> pcwalton: yeah
< strcat> that's one of the things the range asserts will be good for
<@pcwalton> does and set the condition codes?
< strcat> pcwalton: not sure
< strcat> pcwalton: I'm just trying to copy clang as closely as possible
          and not worry about the details :P
< strcat> pcwalton: the LLVM optimizations seem to be very specialized
          for what clang outputs
<@pcwalton> not surprising
< strcat> pcwalton: because even with the range assert, I don't think it
          will avoid the cmp unless you do the trunc and br yourself
< strcat> they do a specific optimization for trunc when ranges are known
          (for clang)

It would be nice to begin testing the emitted LLVM IR to ensure that an expected optimization isn't suddenly deactivated under our noses, especially if LLVM bases its optimizations off of Clang-specific patterns.

[catamorphism]

I'm not sure whether this was ever nominated properly, but maturity #4 seems right to me.

[graydon]

We are now doing codegen tests against clang, with ratcheting turned on. #7461 / #7623 at least gives us those. Would you want to get more-specific than that? I suppose we can try to replicate the search-for-a-pattern directives that we have in the other compiletests, to scan the generated .ll file for a specific idiom? Or a specific instruction-count? Currently we're just ratcheting the rustc:clang LLVM code size ratio on each testcase.

[pnkfelix]

Need not block 1.0. Assigning P-low.

[sanxiyn]

Can this be closed now? We now run tests under src/test/codegen using LLVM's FileCheck tool.

[dotdash]

Yeah, since this says "begin testing", I think it's fair to close this. We'll probably want to extend tests to check for optimizations and/or assembly results later on, but that's no longer in the scope of this issue I think.