Add a -Z flag for LLVM align attributes on arguments

LLVM seems to still put the assume calls in when inlining, so this probably isn't in a place where it can be turned on by default, but it's interesting to experiment with.

For example, this makes `swap::<u64x8>` be 8x `vmovaps	ymm` instead of 16x `vmovups	xmm`, on my cpu.

src/librustc/session/config.rs

@@ -1208,6 +1208,8 @@ options! {DebuggingOptions, DebuggingSetter, basic_debugging_options,
           "set the MIR optimization level (0-3, default: 1)"),
     mutable_noalias: bool = (false, parse_bool, [UNTRACKED],
           "emit noalias metadata for mutable references"),
+    arg_align_attributes: bool = (false, parse_bool, [UNTRACKED],
+          "emit align metadata for reference arguments"),
     dump_mir: Option<String> = (None, parse_opt_string, [UNTRACKED],
           "dump MIR state at various points in translation"),
     dump_mir_dir: String = (String::from("mir_dump"), parse_string, [UNTRACKED],

src/librustc_trans/abi.rs

@@ -778,7 +778,7 @@ impl<'a, 'tcx> FnType<'tcx> {
 
                     // HACK(eddyb) LLVM inserts `llvm.assume` calls when inlining functions
                     // with align attributes, and those calls later block optimizations.
-                    if !is_return {
+                    if !is_return && !cx.tcx.sess.opts.debugging_opts.arg_align_attributes {
                         attrs.pointee_align = None;
                     }