Improve handling of immediate values

[Aatch]

The primary improvement here is to handle function calls that return
imemdiate values such that it doesn't force the use of a stack slot. It
also avoids loads from slots that have an store to it in the same block.

This especially improves the codegen for intrinsics, as most of them
produce immediate values.

cc @rust-lang/compiler

[rust-highfive]

r? @jroesch

(rust_highfive has picked a reviewer for you, use r? to override)

[dotdash]

I'd prefer to add that check after the if-else block, so it can be merged with the same one in the else-arm.

[dotdash]

👍 I wanted to get rid of the store/load combo for function calls for sooooo long.

I'm not sure about the load elision thing. How often / in which situations does that come into play?

[brson]

Neat. Do compile times improve?

[Aatch]

Updated to address comments.

@brson I haven't tested, but my previous attempt at this suggested, no. At least, nothing that isn't covered by noise. The main reason presumably being that let a = foo() and similar expressions still compile to about the same code, which probably accounts for a fair amount of function calls.

[jroesch]

@Aatch read the patch over and it looks good to me, not sure if you want someone who spends more time in trans to look it over instead, otherwise r=me.

[dotdash]

@bors r+ 8905343

[Manishearth]

might be causing http://buildbot.rust-lang.org/builders/auto-mac-64-nopt-t/builds/6075

[bors]

⌛ Testing commit 8905343 with merge 924bd8b...

[bors]

💔 Test failed - auto-mac-64-nopt-t

[bors]

☔ The latest upstream changes (presumably #27169) made this pull request unmergeable. Please resolve the merge conflicts.

[jroesch]

@Aatch ping, whats the status on this?

[dotdash]

@Aatch Are you ok with me finishing this in a new PR?

[alexcrichton]

Closing due to inactivity, but feel free to resubmit with a rebase! (also looks like @jroesch or @dotdash may be willing to help out on this)