test classify-runtime-const for f16

[folkertdev]

View all comments

tracking issue: #116909

r? tgross35

[rustbot]

tgross35 is currently at their maximum review capacity.
They may take a while to respond.

[tgross35]

Tbh I feel like this and similar tests should be folded up into https://github.com/rust-lang/rust/blob/d27207d3d64a1399f90344d6ef3c5743204b3fd3/library/coretests/tests/floats/mod.rs, and we should run that specific file both with and without optimizations. But no reason to exclude f16 here for now.

@bors r+ rollup

View changes since this review

[rust-bors]

📌 Commit c716a5c has been approved by tgross35

It is now in the queue for this repository.

[Zalathar]

Failed in rollup: #153932 (comment)

@bors r-

[rust-bors]

This pull request was unapproved.

This PR was contained in a rollup (#153932), which was unapproved.

[folkertdev]

Maybe this is an LLVM issue? Otherwise mingw is apparently buggy

https://triage.rust-lang.org/gha-logs/rust-lang/rust/67158354905

@bors try jobs=x86_64-mingw-1

[rust-bors]

💔 Test for 3201c11 failed: CI. Failed job:

• try - x86_64-mingw-1 (web logs, enhanced plaintext logs)

[folkertdev]

Well, that is unfortunate, LLVM miscompiles. With wine I can reproduce this

#![feature(f16)]

use std::hint::black_box;

fn main() {
    dbg!(black_box(0.0f16) / black_box(0.0f16));
    dbg!(0.0f16 / 0.0f16);
}

emits

[src/main.rs:6:5] black_box(0.0f16) / black_box(0.0f16) = 0x0000
[src/main.rs:7:5] 0.0f16 / 0.0f16 = 0x7e00

I believe this is an ABI problem, e.g. in https://godbolt.org/z/8xGePaqPn the argument to __extendhfsf2 is passed via xmm, but looking at the disassembly of that method it appears to expect the argument via a GPR:

dissassembly of `__extendhfsf2`

``` 000000014009f200 <__extendhfsf2>: 14009f200: 53 push %rbx 14009f201: 48 83 ec 30 sub $0x30,%rsp 14009f205: 0f ae 5c 24 2c stmxcsr 0x2c(%rsp) 14009f20a: 0f b7 c1 movzwl %cx,%eax 14009f20d: 66 81 e1 ff 03 and $0x3ff,%cx 14009f212: 89 c2 mov %eax,%edx 14009f214: c1 e8 0f shr $0xf,%eax 14009f217: c1 ea 0a shr $0xa,%edx 14009f21a: 41 89 c0 mov %eax,%r8d 14009f21d: 83 e2 1f and $0x1f,%edx 14009f220: 41 83 e0 01 and $0x1,%r8d 14009f224: 48 8d 42 01 lea 0x1(%rdx),%rax 14009f228: a8 1e test $0x1e,%al 14009f22a: 74 24 je 14009f250 <__extendhfsf2+0x50> 14009f22c: 48 8d 5a 70 lea 0x70(%rdx),%rbx 14009f230: 81 e1 ff 03 00 00 and $0x3ff,%ecx 14009f236: 41 c1 e0 1f shl $0x1f,%r8d 14009f23a: c1 e3 17 shl $0x17,%ebx 14009f23d: c1 e1 0d shl $0xd,%ecx 14009f240: 09 cb or %ecx,%ebx 14009f242: 44 09 c3 or %r8d,%ebx 14009f245: 66 0f 6e c3 movd %ebx,%xmm0 14009f249: 48 83 c4 30 add $0x30,%rsp 14009f24d: 5b pop %rbx 14009f24e: c3 ret 14009f24f: 90 nop 14009f250: 0f b7 c1 movzwl %cx,%eax 14009f253: 48 85 d2 test %rdx,%rdx 14009f256: 75 48 jne 14009f2a0 <__extendhfsf2+0xa0> 14009f258: 48 85 c0 test %rax,%rax 14009f25b: 74 73 je 14009f2d0 <__extendhfsf2+0xd0> 14009f25d: 4c 0f bd c8 bsr %rax,%r9 14009f261: ba a6 00 00 00 mov $0xa6,%edx 14009f266: 44 89 c3 mov %r8d,%ebx 14009f269: 49 83 f1 3f xor $0x3f,%r9 14009f26d: c1 e3 1f shl $0x1f,%ebx 14009f270: 41 8d 49 d8 lea -0x28(%r9),%ecx 14009f274: 44 29 ca sub %r9d,%edx 14009f277: 48 d3 e0 shl %cl,%rax 14009f27a: c1 e2 17 shl $0x17,%edx 14009f27d: b9 02 00 00 00 mov $0x2,%ecx 14009f282: 25 ff ff 7f 00 and $0x7fffff,%eax 14009f287: 09 d0 or %edx,%eax 14009f289: 09 c3 or %eax,%ebx 14009f28b: e8 30 07 00 00 call 14009f9c0 <__sfp_handle_exceptions> 14009f290: 66 0f 6e c3 movd %ebx,%xmm0 14009f294: 48 83 c4 30 add $0x30,%rsp 14009f298: 5b pop %rbx 14009f299: c3 ret 14009f29a: 66 0f 1f 44 00 00 nopw 0x0(%rax,%rax,1) 14009f2a0: 48 85 c0 test %rax,%rax 14009f2a3: 74 43 je 14009f2e8 <__extendhfsf2+0xe8> 14009f2a5: 48 c1 e0 0d shl $0xd,%rax 14009f2a9: 41 c1 e0 1f shl $0x1f,%r8d 14009f2ad: 48 89 c3 mov %rax,%rbx 14009f2b0: 81 cb 00 00 c0 7f or $0x7fc00000,%ebx 14009f2b6: 44 09 c3 or %r8d,%ebx 14009f2b9: 66 81 f9 ff 01 cmp $0x1ff,%cx 14009f2be: 77 85 ja 14009f245 <__extendhfsf2+0x45> 14009f2c0: b9 01 00 00 00 mov $0x1,%ecx 14009f2c5: eb c4 jmp 14009f28b <__extendhfsf2+0x8b> 14009f2c7: 66 0f 1f 84 00 00 00 nopw 0x0(%rax,%rax,1) 14009f2ce: 00 00 14009f2d0: 44 89 c0 mov %r8d,%eax 14009f2d3: c1 e0 1f shl $0x1f,%eax 14009f2d6: 89 c3 mov %eax,%ebx 14009f2d8: 66 0f 6e c3 movd %ebx,%xmm0 14009f2dc: 48 83 c4 30 add $0x30,%rsp 14009f2e0: 5b pop %rbx 14009f2e1: c3 ret 14009f2e2: 66 0f 1f 44 00 00 nopw 0x0(%rax,%rax,1) 14009f2e8: 44 89 c3 mov %r8d,%ebx 14009f2eb: c1 e3 1f shl $0x1f,%ebx 14009f2ee: 81 cb 00 00 80 7f or $0x7f800000,%ebx 14009f2f4: 66 0f 6e c3 movd %ebx,%xmm0 14009f2f8: 48 83 c4 30 add $0x30,%rsp 14009f2fc: 5b pop %rbx 14009f2fd: c3 ret 14009f2fe: 90 nop 14009f2ff: 90 nop ```

This seems like it's an LLVM issue? Though I'm honestly not sure how any of our tests pass on windows at all with this being broken.

[folkertdev]

I guess this is just gcc-mirror/gcc@8b6a18e missing actually. Should we just skip windows altogether? Or use target_has_reliable_f16?

[tgross35]

Ah yeah, the f16 parts should probably be gated by target_has_reliable_f16.

[folkertdev]

@bors try jobs=x86_64-mingw-1

[rust-bors]

💔 Test for 0aa6729 failed: CI. Failed job:

• try - x86_64-mingw-1 (web logs, enhanced plaintext logs)

[folkertdev]

@bors try jobs=x86_64-mingw-1

[rust-bors]

☀️ Try build successful (CI)
Build commit: 8e51988 (8e5198828147a496f8152ad2db49afecb8cbbaf5, parent: 6bdc342ddb05b520a1f741aabf002319b44d2475)

[folkertdev]

@bors r=tgross35

[rust-bors]

📌 Commit 19e9ec7 has been approved by tgross35

It is now in the queue for this repository.