More defailbloating

[brson]

With this, all the failure lang item calls, and all begin_unwind (fail!) calls are optimized to take either one or two pointers, at least one of which is a constant, with the exception that the bounds check lang item also takes index and len. Code gen for all that is gorgeous.

Divide-by-zero before:

        leaq    "str\"str\"(1762)"(%rip), %rax
        movq    %rax, 16(%rsp)
        movq    $27, 24(%rsp)
        leaq    "str\"str\"(1542)"(%rip), %rax
        movq    %rax, (%rsp)
        movq    $19, 8(%rsp)
        leaq    16(%rsp), %rdi
        leaq    (%rsp), %rsi
        movl    $32, %edx
        callq   _ZN7failure5fail_20hc04408f955ce60aaqWjE@PLT

After:

        leaq    .Lconst(%rip), %rdi
        callq   _ZN7failure5fail_20haf918a97c8f7f2bfqWjE@PLT

Bounds check before:

        leaq    "str\"str\"(1542)"(%rip), %rax
        movq    %rax, 8(%rsp)
        movq    $19, 16(%rsp)
        leaq    8(%rsp), %rdi
        movl    $38, %esi
        movl    $1, %edx
        movl    $1, %ecx
        callq   _ZN7failure17fail_bounds_check20hf4bc3c69e96caf41RXjE@PLT

Bounds check after:

        leaq    .Lconst2(%rip), %rdi
        movl    $1, %esi
        movl    $1, %edx
        callq   _ZN7failure17fail_bounds_check20h5267276a537a7de22XjE@PLT

Size before:

21277995 librustc-4e7c5e5c.s

text       data
12554881   6089335

Size after:

21247617 librustc-4e7c5e5c.so

text       data
12518497   6095748

The next two steps might be to change the amount of information we generate based on a compiler flag, and maybe debug levels; and to investigate improving codegen for fmt::Arguments, which appears in many fail! invocations.

r? @pcwalton

[brson]

Note that this time the total rustc binary size actually shrank! That doesn't seem to be true of every binary though.

[alexcrichton]

🍠