Compiler generated `main` symbol has unavoidable pointless stack check

[huonw]

#[no_split_stack]
#[start]
fn main(_: int, _: **u8) -> int {
    0
}

compiles to

    .text
    .file   "-.rs"
    .section    .text._ZN4main20he8340d7809f38c39eaa4v0.0E,"ax",@progbits
    .align  16, 0x90
    .type   _ZN4main20he8340d7809f38c39eaa4v0.0E,@function
_ZN4main20he8340d7809f38c39eaa4v0.0E:
    .cfi_startproc
    movabsq $0, %rax
    movq    %rdi, -8(%rsp)
    movq    %rsi, -16(%rsp)
    retq
.Ltmp0:
    .size   _ZN4main20he8340d7809f38c39eaa4v0.0E, .Ltmp0-_ZN4main20he8340d7809f38c39eaa4v0.0E
    .cfi_endproc

    .section    .text.main,"ax",@progbits
    .globl  main
    .align  16, 0x90
    .type   main,@function
main:
    .cfi_startproc
    cmpq    %fs:112, %rsp
    ja  .LBB1_0
    movabsq $8, %r10
    movabsq $0, %r11
    callq   __morestack
    retq
.LBB1_0:
    pushq   %rax
.Ltmp1:
    .cfi_def_cfa_offset 16
    callq   _ZN4main20he8340d7809f38c39eaa4v0.0E
    popq    %rdx
    retq
.Ltmp2:
    .size   main, .Ltmp2-main
    .cfi_endproc

    .section    ".note.GNU-stack","",@progbits

In particular the main function has a stack check, which can't be removed (there's no way to put a #[no_split_stack] on it: the compiler creates it automagically), and, (I think, I could easily be wrong) it's always pointless: when would the entry point of a program be overflowing the stack immediately?

[Florob]

I think that if the main function allocates a sufficient amount of data on the stack this might be necessary.
Being a bit annoyed by the unneeded stack checks I implemented avoiding them upstream a while ago: http://llvm.org/viewvc/llvm-project?view=revision&revision=209436. Once Rust's LLVM is updated to a newer revision we should get this. In particular it should get rid of the stack check in this specific case.

[alexcrichton]

Note that all platforms we run on have the initial stack limit as 0, which means that the initial thread has an "infinite stack". Fixing main is also not enough, one would have to fix all functions down to when the initial stack limit is set, and that's a lot of functions!

[mahkoh]

Still broken but you can disable all stack checks, including this one, with -C no-stack-check now.

[steveklabnik]

Updated code:

#![feature(start)]

#[no_stack_check]
#[start]
fn main(_: isize, _: *const *const u8) -> isize {
        0
}

I am bad at assembly, but

        .text
        .file   "hello.0.rs"
        .section        .text._ZN4main20h9021d633acfe0488eaaE,"ax",@progbits
        .align  16, 0x90
        .type   _ZN4main20h9021d633acfe0488eaaE,@function
_ZN4main20h9021d633acfe0488eaaE:
        .cfi_startproc
        xorl    %eax, %eax
        movq    %rdi, -8(%rsp)
        movq    %rsi, -16(%rsp)
        retq
.Ltmp0:
        .size   _ZN4main20h9021d633acfe0488eaaE, .Ltmp0-_ZN4main20h9021d633acfe0488eaaE
        .cfi_endproc

        .section        .text.main,"ax",@progbits
        .globl  main
        .align  16, 0x90
        .type   main,@function
main:
        .cfi_startproc
        cmpq    %fs:112, %rsp
        ja      .LBB1_0
        movabsq $8, %r10
        movabsq $0, %r11
        callq   __morestack
        retq
.LBB1_0:
        pushq   %rax
.Ltmp1:
        .cfi_def_cfa_offset 16
        callq   _ZN4main20h9021d633acfe0488eaaE
        popq    %rdx
        retq
.Ltmp2:
        .size   main, .Ltmp2-main
        .cfi_endproc


        .section        ".note.GNU-stack","",@progbits

[dotdash]

With the new LLVM version, that stack check gets removed in an optimized build. Can we close this?

    .text
    .file   "issue14912.0.rs"
    .section    .text.main,"ax",@progbits
    .globl  main
    .align  16, 0x90
    .type   main,@function
main:
    xorl    %eax, %eax
    retq
.Ltmp0:
    .size   main, .Ltmp0-main


    .section    ".note.GNU-stack","",@progbits

[alexcrichton]

great, thanks @dotdash!