assorted performance optimizations

[thestinger]

Passing higher alignment values gives the optimization passes more freedom since it can copy in larger chunks. This change results in rustc outputting the same post-optimization IR as clang for swaps and most copies excluding the lack of information about padding.

Code snippet:

#[inline(never)]fnswap<T>(x:&mutT,y:&mutT){
util::swap(x, y);}

Original IR (for int):

defineinternalfastccvoid@_ZN9swap_283417_a71830ca3ed2d65d3_00E(i64*, i64*) #1 {
static_allocas:
%2 = icmpeqi64*%0, %1bri1%2, label%_ZN4util9swap_283717_a71830ca3ed2d65d3_00E.exit, label%3; <label>:3 ; preds = %static_allocas%4 = loadi64*%0, align1%5 = loadi64*%1, align1storei64%5, i64*%0, align1storei64%4, i64*%1, align1brlabel%_ZN4util9swap_283717_a71830ca3ed2d65d3_00E.exit
_ZN4util9swap_283717_a71830ca3ed2d65d3_00E.exit: ; preds = %3, %static_allocasretvoid
}

After #6710:

defineinternalfastccvoid@_ZN9swap_283017_a71830ca3ed2d65d3_00E(i64*nocapture, i64*nocapture) #1 {
static_allocas:
%2 = loadi64*%0, align1%3 = loadi64*%1, align1storei64%3, i64*%0, align1storei64%2, i64*%1, align1retvoid
}

After this change:

defineinternalfastccvoid@_ZN9swap_283017_a71830ca3ed2d65d3_00E(i64*nocapture, i64*nocapture) #1 {
static_allocas:
%2 = loadi64*%0, align8%3 = loadi64*%1, align8storei64%3, i64*%0, align8storei64%2, i64*%1, align8retvoid
}

Another example:

#[inline(never)]fnset<T>(x:&mutT,y:T){*x = y;}

Before, with (int, int) (align 1):

defineinternalfastccvoid@_ZN8set_282517_8fa972e3f9e451983_00E({ i64, i64 }* nocapture, { i64, i64 }* nocapture) #1 {
static_allocas:
%2 = bitcast { i64, i64 }* %1toi8*%3 = bitcast { i64, i64 }* %0toi8*tailcallvoid@llvm.memcpy.p0i8.p0i8.i64(i8*%3, i8*%2, i6416, i321, i1false)
retvoid
}

After, with (int, int) (align 8):

defineinternalfastccvoid@_ZN8set_282617_8fa972e3f9e451983_00E({ i64, i64 }* nocapture, { i64, i64 }* nocapture) #1 {
static_allocas:
%2 = bitcast { i64, i64 }* %1toi8*%3 = bitcast { i64, i64 }* %0toi8*tailcallvoid@llvm.memcpy.p0i8.p0i8.i64(i8*%3, i8*%2, i6416, i328, i1false)
retvoid
}

[thestinger]

Note that since #6742 landed, the parameters are also marked as noalias so the IR is strictly better than clang, but it results in pretty much the same assembly for these cases.

[graydon]

This is wonderful. Thanks.