Add x86_64 builtins #40
Conversation
| // I don't believe that this behavior is guaranteed, and a program that uses | ||
| // only __alloca could have ___chkstk removed by --gc-sections. Call | ||
| // ___chkstk here to guarantee that neither of those happen. | ||
| ___chkstk(); |
There was a problem hiding this comment.
Since __chkstk uses a custom calling convention, you should probably jump to it using a jmp instruction instead of calling it from Rust.
There was a problem hiding this comment.
I had thought about that, but couldn't __chkstk possibly be removed by the linker in that case? Now that I think about it again, the linker could probably see the use of the ___chkstk symbol in the jmp and leave it in, just the same as it would for a call instruction.
👍 Awesome.
A good sanity check is to manually check that the output assembly of this crate is similar/equal to the disassemble of the libcompiler-rt.a shipped with Rust. (I'm not sure what's the Windows'
This what the disassembly of Rust's libcompiler-rt.a looks like on my Linux machine: 0000000000000000 <__alloca>:
0: 48 89 c8 mov %rcx,%rax
0000000000000003 <___chkstk>:
3: 51 push %rcx
4: 48 3d 00 10 00 00 cmp $0x1000,%rax
a: 48 8d 4c 24 10 lea 0x10(%rsp),%rcx
f: 72 18 jb 29 <___chkstk+0x26>
11: 48 81 e9 00 10 00 00 sub $0x1000,%rcx
18: 48 85 09 test %rcx,(%rcx)
1b: 48 2d 00 10 00 00 sub $0x1000,%rax
21: 48 3d 00 10 00 00 cmp $0x1000,%rax
27: 77 e8 ja 11 <___chkstk+0xe>
29: 48 29 c1 sub %rax,%rcx
2c: 48 85 09 test %rcx,(%rcx)
2f: 48 8d 44 24 08 lea 0x8(%rsp),%rax
34: 48 89 cc mov %rcx,%rsp
37: 48 8b 48 f8 mov -0x8(%rax),%rcx
3b: ff 30 pushq (%rax)
3d: 48 29 e0 sub %rsp,%rax
40: c3 retqI'm not sure if a function call would generate equivalent assembly. I can't really read x86 assembly but I think that __alloca is not supposed to trigger an extra function call over __chkstk.
Makes sense to me. |
|
The generated assembly looks much better without the function call, and I've convinced myself that |
The appveyor build failed because the test build has name mangling so |
|
The problem here is that we only specify |
This prevents linker errors in test builds due to the `jmp` instruction in __alloca
|
We can probably make all of them unconditionally |
[ci ignore]
|
Thank you @mattico! |
This is a fairly straightforward port of the original assembly files. I'm hoping to help port a whole bunch of these builtins over the next few days, so let me know if you have any style or other concerns. I don't believe that these functions have any tests in compiler-rt and they would be quite difficult functions to test. Nonetheless, I'm going to attempt to find a windows binary that exercises these and swap them in to see if anything breaks.
Two notes:
__allocafalls through to___chkstkin the original assembly. I think having a function call there is the best way to implement this reliably, other than duplicating the assembly.