Auto merge of #34189 - eddyb:mir-trans-imm, r=nagisa
trans: generalize immediate temporaries to all MIR locals. Added `Mir::local_index` which gives you an unified index for `Arg`, `Var`, `Temp` and `ReturnPointer`. Also available is `Mir::count_locals` which returns the total number of the above locals. This simplifies a lot of the code which can treat all of the local lvalues in the same manner. If we had `-> impl Iterator`, I could have added a bunch of useful `Ty` or `Lvalue` iterators for all locals. We could of course manually write such iterators as they are needed. The only place which currently takes advantage of unified locals is trans' alloca elision. Currently it's not as good as it could be, due to our usage of `llvm.dbg.declare` in debug mode. But passing some arguments and variables as immediates has some effect on release-mode `libsyntax`: Old trans: ``` time: 11.500; rss: 710MB translation time: 0.002; rss: 710MB assert dep graph time: 0.000; rss: 710MB serialize dep graph time: 4.410; rss: 628MB llvm function passes [0] time: 84.485; rss: 633MB llvm module passes [0] time: 23.898; rss: 634MB codegen passes [0] time: 0.002; rss: 634MB codegen passes [0] time: 113.408; rss: 634MB LLVM passes ``` `-Z orbit`, previously: ``` time: 12.588; rss: 723MB translation time: 0.002; rss: 723MB assert dep graph time: 0.000; rss: 723MB serialize dep graph time: 4.597; rss: 642MB llvm function passes [0] time: 77.347; rss: 646MB llvm module passes [0] time: 24.703; rss: 648MB codegen passes [0] time: 0.002; rss: 615MB codegen passes [0] time: 107.233; rss: 615MB LLVM passes ``` `-Z orbit`, after this PR: ``` time: 13.820; rss: 672MB translation time: 0.002; rss: 672MB assert dep graph time: 0.000; rss: 672MB serialize dep graph time: 3.969; rss: 591MB llvm function passes [0] time: 72.294; rss: 595MB llvm module passes [0] time: 24.610; rss: 597MB codegen passes [0] time: 0.002; rss: 597MB codegen passes [0] time: 101.439; rss: 597MB LLVM passes ```
Showing
想要评论请 注册 或 登录