Output executable crashes with relocation-model=static

[Krysme]

The following code crashes with "relocation-model=static" on linux x86-64, reproducible.

#![feature(nll)]

fn main() {
    fn divide(numerator: f64, denominator: f64) -> Option<f64> {
        if denominator == 0.0 {
            None
        } else {
            Some(numerator / denominator)
        }
    }
    let result = divide(2.0, 3.0);

    match result {
        Some(x) => println!("Result: {}", x),
        None => println!("Cannot Divide by 0"),
    }
}

[mati865]

@lzutao you have to use relocation-model=static as mentioned in the issue.

LLDB backtrace

(lldb) r
Process 17646 launched: '/tmp/foo/t' (x86_64)
Process 17646 stopped
* thread #1, name = 't', stop reason = signal SIGSEGV: invalid address (fault address: 0x36a75)
    frame #0: 0x00000000004022ef t`t::main::h30886ce2ae13d6f9 + 127
t`t::main::h30886ce2ae13d6f9:
->  0x4022ef <+127>: movq   0x36a75, %rsi
    0x4022f7 <+135>: movsd  0x30(%rsp), %xmm0         ; xmm0 = mem[0],zero
    0x4022fd <+141>: movsd  %xmm0, 0x38(%rsp)
    0x402303 <+147>: leaq   0x38(%rsp), %rax
(lldb) bt
* thread #1, name = 't', stop reason = signal SIGSEGV: invalid address (fault address: 0x36a75)
  * frame #0: 0x00000000004022ef t`t::main::h30886ce2ae13d6f9 + 127
    frame #1: 0x0000000000402583 t`std::rt::lang_start::_$u7b$$u7b$closure$u7d$$u7d$::h819e5a92dacb5f12 + 3
    frame #2: 0x000000000040ad13 t`std::panicking::try::do_call::h37d577b51b92ce4b [inlined] std::rt::lang_start_internal::_$u7b$$u7b$closure$u7d$$u7d$::hc50706b6e8c189f3 at rt.rs:49:12
    frame #3: 0x000000000040ad07 t`std::panicking::try::do_call::h37d577b51b92ce4b at panicking.rs:294
    frame #4: 0x000000000040c76a t`__rust_maybe_catch_panic at lib.rs:82:7
    frame #5: 0x000000000040b7cd t`std::rt::lang_start_internal::h17a202bf41caec69 [inlined] std::panicking::try::hd2f41f2af573d056 at panicking.rs:273:12
    frame #6: 0x000000000040b78f t`std::rt::lang_start_internal::h17a202bf41caec69 [inlined] std::panic::catch_unwind::h1d1e6c4c3c3f0d9c at panic.rs:388
    frame #7: 0x000000000040b78f t`std::rt::lang_start_internal::h17a202bf41caec69 at rt.rs:48
    frame #8: 0x0000000000402568 t`std::rt::lang_start::h3d4b63434ee976c6 + 56
    frame #9: 0x000000000040240e t`main + 30
    frame #10: 0x00007ffff7d93b6b libc.so.6`__libc_start_main + 235
    frame #11: 0x000000000040216a t`_start + 42

@rustbot modify labels: +C-bug +O-linux

[nagisa]

Likely cause is that PIC libstd and libcore are used by the reporter – this results in assembly such as this:

_ZN4test4main17h21e0b94bad749a9eE:
  subq  $104, %rsp
  movq	_ZN4core3fmt5float52_$LT$impl$u20$core..fmt..Display$u20$for$u20$f64$GT$3fmt17he52c1166ee322c16E@GOTPCREL, %rsi
  <snip>

Which after assembling then results in machine code like this (note a null pointer dereference):

0000000000000000 <_ZN4test4main17h21e0b94bad749a9eE>:
   0:	48 83 ec 68          	sub    $0x68,%rsp
   4:	48 8b 34 25 00 00 00 	mov    0x0,%rsi
   b:	00 
   <snip>

which then linker relocates to, hiding the obvious null dereference:

Dump of assembler code for function _ZN4test4main17h21e0b94bad749a9eE:
   0x00000000004022f0 <+0>:	sub    $0x68,%rsp
   0x00000000004022f4 <+4>:	mov    0x32900,%rsi
   <snip>

From what I can tell this is very likely to be user error.

[Krysme]

@nagisa so, is there any way that I can get a version of libstd and libcore with relocation-model=static, or do I have to recompile it from scratch ? A link to the documentation would be helpful, thanks in advance.

[nagisa]

You can use xargo or whatever the current maintained version of that thing is to build libstd and libcore. cargo currently does not support that.

[jsgf]

I can reproduce this easily, but only with non-optimized builds. With any level of optimization this goes away. For example, this input:

fn main() {
  let _foo = format!("{}", 4);
}

will produce a crashing executable if compiled with rustc -Crelocation-mode=static --emit=asm,link --crate-type=bin reloc.rs, as it has the bogus

movq    _ZN4core3fmt3num3imp52_$LT$impl$u20$core..fmt..Display$u20$for$u20$i32$GT$3fmt17hd0b7b005436cd7b8E@GOTPCREL, %rsi

non-rip-relative GOTPCREL relocation.

But if I add -Copt-level=1, it generates a proper rip-relative reference:

movq    _ZN4core3fmt3num3imp52_$LT$impl$u20$core..fmt..Display$u20$for$u20$i32$GT$3fmt17hd0b7b005436cd7b8E@GOTPCREL(%rip), %rax

So I think there's more to it than just having static vs pic libstd.

[jesboat]

I think this might be in LLVM. I made a self-contained (i.e. doesn't pull in any library stuff like format!) example, the tldr of which "call a function where an argument is a function pointer to another crate":

extern crate a;
  
pub fn takep(arg: fn() -> ()) -> () {}  

pub fn example1() {
  takep(a::voidfn);
} 

The last LLIR before the instruction selection phases is

define void @_ZN4main8example117h149bc290ecddb5b8E() unnamed_addr #0 {
start:
  call void @_ZN4main5takep17h851b97fbf6407c1bE(void ()* nonnull @_ZN1a6voidfn17h7d790e74ed3ff940E)
  br label %bb1

bb1:                                              ; preds = %start
  ret void
}

and immediately after the instruction selection phases, the call is lowered to:

  %0:gr64 = MOV64rm $noreg, 1, $noreg, target-flags(x86-gotpcrel) @_ZN1a6voidfn17h7d790e74ed3ff940E, $noreg
  ADJCALLSTACKDOWN64 0, 0, 0, implicit-def $rsp, implicit-def $eflags, implicit-def $ssp, implicit $rsp, implicit $ssp
  $rdi = COPY %0:gr64  
  CALL64pcrel32 @_ZN4main5takep17h851b97fbf6407c1bE, <regmask $bh $bl $bp $bph $bpl $bx $ebp $ebx $hbp $hbx $rbp $rbx $r12 $r13 $r14 $r15 $r12b   $r13b $r14b $r15b $r12bh $r13bh $r14bh $r15bh $r12d $r13d $r14d $r15d $r12w $r13w $r14w $r15w $r12wh and 3 more...>, implicit $rsp, implicit      $ssp, implicit $rdi
  ADJCALLSTACKUP64 0, 0, implicit-def $rsp, implicit-def $eflags, implicit-def $ssp, implicit $rsp, implicit $ssp

Note how the MOV64rn has $noreg.

In contrast, when the example is modified slightly, so that it would need to unwind if takep throws:

pub fn example2() {
  let _v = Droppable{};
  takep(a::voidfn);
}

The LLIR right before ISel (excluding basic blocks which are only called if unwinding)

define void @_ZN4main8example217h4f443cbc5e17bbddE() unnamed_addr #0 personality i32 (i32, i32, i64, %"unwind::libunwind::_Unwind_Exception"*,    %"unwind::libunwind::_Unwind_Context"*)* @rust_eh_personality {
start:
  %personalityslot = alloca { i8*, i32 }, align 8
  %_v = alloca %Droppable, align 1
  invoke void @_ZN4main5takep17h851b97fbf6407c1bE(void ()* nonnull @_ZN1a6voidfn17h7d790e74ed3ff940E)
          to label %bb2 unwind label %cleanup

bb2:                                              ; preds = %start
  call void @_ZN4core3ptr18real_drop_in_place17h6782efda6a962288E(%Droppable* nonnull align 1 %_v)
  br label %bb4

bb4:                                              ; preds = %bb2
  ret void

The LLVM language reference says about invoke

This instruction is designed to operate as a standard ‘call’ instruction in most regards. The primary difference is that it establishes an association with a label, which is used by the runtime library to unwind the stack.

However, here, the lowering is

  EH_LABEL <mcsymbol .Ltmp0>
  ADJCALLSTACKDOWN64 0, 0, 0, implicit-def dead $rsp, implicit-def dead $eflags, implicit-def dead $ssp, implicit $rsp, implicit $ssp
  %0:gr64 = MOV64rm $rip, 1, $noreg, target-flags(x86-gotpcrel) @_ZN1a6voidfn17h7d790e74ed3ff940E, $noreg :: (load 8 from got)
  $rdi = COPY %0:gr64
  CALL64pcrel32 @_ZN4main5takep17h851b97fbf6407c1bE, <regmask $bh $bl $bp $bph $bpl $bx $ebp $ebx $hbp $hbx $rbp $rbx $r12 $r13 $r14 $r15 $r12b   $r13b $r14b $r15b $r12bh $r13bh $r14bh $r15bh $r12d $r13d $r14d $r15d $r12w $r13w $r14w $r15w $r12wh and 3 more...>, implicit $rsp, implicit      $ssp, implicit $rdi, implicit-def $rsp, implicit-def $ssp
  ADJCALLSTACKUP64 0, 0, implicit-def dead $rsp, implicit-def dead $eflags, implicit-def dead $ssp, implicit $rsp, implicit $ssp
  EH_LABEL <mcsymbol .Ltmp1>
  JMP_1 %bb.2

where the MOV64rn has $rip.

A not-particularly-educated guess would be that FastISel has a fast path for call which has a bug, and that bug is not present in the slow path for invoke.

The full example, including source and unabridged LLIR and MIR is https://gist.github.com/jesboat/a88aa28f0a80c876872425901bdd9279

[jesboat]

I am reasonably confident at this point that this is a bug in LLVM's FastIsel. This LLIR

define void @_ZN4main8example117h149bc290ecddb5b8E() unnamed_addr #0 {
start:
  %ptr = alloca void ()*
  store void ()* @_ZN1a6voidfn17h7d790e74ed3ff940E, void ()** %ptr
  ret void
}

is sufficient to produce a broken binary on trunk LLVM with llc -O0 --relocation-model=static. Adding -fast-isel=false to the options produces a non-broken load.

https://godbolt.org/z/Ga7aJg

[Krysme]

Thanks for all the replies, and I hope that LLVM developers will resolve this one as soon as possible.

[yshui]

I just hit this problem. Did anyone report this bug to LLVM?