assert_instr fails when using lld as the linker

[GabrielMajeri]

Tried running ./ci/run.sh, everything ran smoothly until:

failures:

---- api::ops::vector_rotates::u64x8::assert_rotate_left_vpro stdout ----
disassembly for verify::api::ops::vector_rotates::u64x8::assert_rotate_left_vpro::rotate_left_shim:
         0: lea -0x3f307(%rip),%rax # 1000 <anon.c3685bac7e5bc7b153e587724ab9a96a.7.llvm.2311252202702433396+0x80>
         1: mov 0xbacda(%rip),%rcx # fafe8 <_DYNAMIC+0x248>
         2: mov %rax,(%rcx)
         3: movq $0x27,0x8(%rcx)
         4:
         5: vprolq $0xc,%zmm0,%zmm0
         6: retq
         7: int3
         8: int3
         9: int3
        10: int3
        11: int3
        12: int3
        13: int3
        14: int3
        15: int3
        16: int3
        17: int3
        18: int3
        19: int3
        20: int3
        21: int3
thread 'api::ops::vector_rotates::u64x8::assert_rotate_left_vpro' panicked at 'instruction found, but the disassembly contains too many instructions: #instructions = 22 >= 22 (limit)', /home/gabriel/.cargo/git/checkouts/stdsimd-64c3bc0ac1e47113/b3bd07b/crates/stdsimd-test/src/lib.rs:467:9
note: Run with `RUST_BACKTRACE=1` for a backtrace.

---- api::ops::vector_rotates::u64x8::assert_rotate_right_vpro stdout ----
disassembly for verify::api::ops::vector_rotates::u64x8::assert_rotate_right_vpro::rotate_right_shim:
         0: lea -0x3f71d(%rip),%rax # bba <anon.c3685bac7e5bc7b153e587724ab9a96a.5.llvm.2311252202702433396+0x11>
         1: mov 0xbad0a(%rip),%rcx # fafe8 <_DYNAMIC+0x248>
         2: mov %rax,(%rcx)
         3: movq $0x29,0x8(%rcx)
         4:
         5: vprolq $0x34,%zmm0,%zmm0
         6: retq
         7: int3
         8: int3
         9: int3
        10: int3
        11: int3
        12: int3
        13: int3
        14: int3
        15: int3
        16: int3
        17: int3
        18: int3
        19: int3
        20: int3
        21: int3
thread 'api::ops::vector_rotates::u64x8::assert_rotate_right_vpro' panicked at 'instruction found, but the disassembly contains too many instructions: #instructions = 22 >= 22 (limit)', /home/gabriel/.cargo/git/checkouts/stdsimd-64c3bc0ac1e47113/b3bd07b/crates/stdsimd-test/src/lib.rs:467:9


failures:
    api::ops::vector_rotates::u64x8::assert_rotate_left_vpro
    api::ops::vector_rotates::u64x8::assert_rotate_right_vpro

test result: FAILED. 3 passed; 2 failed; 0 ignored; 0 measured; 0 filtered out

error: test failed, to rerun pass '--lib'
+ return 

${TARGET}: x86_64-unknown-linux-gnu rustc: rustc 1.30.0-nightly (6e0f1cc15 2018-09-05)

[gnzlbg]

That's weird and wrong, the disassembly should have been truncated after the retq instruction.

Can you post exactly which system are you on? Which linux, which version of objdump, etc. ?

Do you have any RUSTFLAGS or .cargo/config ?

The int3 instruction is 1 byte long so it could be used to align the function but... I've never run into this :/

cc @alexcrichton

[GabrielMajeri]

@gnzlbg AFAIK int $3 is simply the default padding between functions, will trigger a breakpoint when running under a debugger.

Found the issue: I had enabled LLD as my linker globally

linker = "clang"
rustflags = [ "-C", "link-arg=-fuse-ld=lld" ]

Using ld instead doesn't break the tests.

LLD version: LLD 6.0.1 (compatible with GNU linkers), installed from the ArchLinux package repository objdump version: GNU objdump (GNU Binutils) 2.31.1

[gnzlbg]

I think we should try to fix assert_instr to work with lld. Enabling cross-language inlining for sleef will require using lld as a linker so this is a problem we'll have to solve anyways. I'm renaming this issue.

[alexcrichton]

Odd! It may be the case that ld doesn't emit informating saying the padding is part of the function whereas LLD does emit that information? This would also be reasonably easy enough to prune in the assert-instr crate as well