Stream: cranelift

Topic: invalid domtree verifier error not externally reproducible

view this post on Zulip bjorn3 (Jan 16 2021 at 10:28):

I got a few verifier errors with cg_clif saying that the domtree is not valid. I can't reproduce it with clif-util even when enabling the exact same flags.

thread 'rustc' panicked at 'called `Result::unwrap()` on an `Err` value: Compilation(Verifier(VerifierErrors([VerifierError { location: block3, context: None, message: "invalid domtree, expected idom(block3) = Some(inst11), got Some(inst10)" }])))', src/base.rs:133:14
thread 'rustc' panicked at 'called `Result::unwrap()` on an `Err` value: Compilation(Verifier(VerifierErrors([VerifierError { location: block4, context: None, message: "invalid domtree, expected idom(block4) = Some(inst4), got Some(inst5)" }])))', src/base.rs:133:14
thread 'rustc' panicked at 'called `Result::unwrap()` on an `Err` value: Compilation(Verifier(VerifierErrors([VerifierError { location: block4, context: None, message: "invalid domtree, expected idom(block4) = Some(inst15), got Some(inst14)" }])))', src/base.rs:133:14

view this post on Zulip bjorn3 (Jan 16 2021 at 10:29):

Any idea how I can debug this?

view this post on Zulip bjorn3 (Jan 16 2021 at 10:30):

The manual verifier pass just after the generation of clif ir doesn't give any error. Only once compilation happens does this verifier error get hit.

view this post on Zulip Benjamin Bouvier (Jan 16 2021 at 10:44):

Weird, do you have all the CLIF + same settings? And using the same backends in both cases?

view this post on Zulip Benjamin Bouvier (Jan 16 2021 at 10:44):

(same CPU flags too)

view this post on Zulip bjorn3 (Jan 16 2021 at 10:48):

Yes, I used the same cranelift commit, matched all flags and used the old backend in both cases.

view this post on Zulip bjorn3 (Jan 16 2021 at 10:48):

Currently testing if the latest cranelift version has the same problem.

view this post on Zulip bjorn3 (Jan 16 2021 at 10:54):

Updating doesn't help.

view this post on Zulip bjorn3 (Jan 16 2021 at 10:55):

If you want to reproduce, just use ./prepare.sh && ./test.sh --debug on the latest commit of cg_clif.

view this post on Zulip bjorn3 (Jan 16 2021 at 10:58):

Removing https://github.com/bjorn3/rustc_codegen_cranelift/blob/3ea8915d4a247b5b3c4cfb3424c230ccd2645b17/src/base.rs#L114-L119 doesn't help.

bjorn3/rustc_codegen_cranelift
Cranelift based backend for rustc. Contribute to bjorn3/rustc_codegen_cranelift development by creating an account on GitHub.

view this post on Zulip Benjamin Bouvier (Jan 16 2021 at 11:00):

One thing that would be interesting is the backtrace - the verifier runs several times during compilation, so knowing when it happens would be interesting.

view this post on Zulip Benjamin Bouvier (Jan 16 2021 at 11:00):

If you have the CLIF graph around, might be valuable too?

view this post on Zulip bjorn3 (Jan 16 2021 at 11:04):

This is the clif that gets compiled:

target x86_64-unknown-linux-gnu haswell

function u0:117(f32) -> i8 system_v {
; symbol _ZN4core3f3221_$LT$impl$u20$f32$GT$8classify17h334bc58d8063a2b5E
; instance Instance { def: Item(WithOptConstParam { did: DefId(0:125 ~ core[25e5]::f32::{impl#0}::classify), const_param_did: None }), substs: [] }

    ss0 = explicit_slot 1
    sig0 = (f32) -> i32 system_v
    fn0 = colocated u0:118 sig0

                                block0(v0: f32):
                                    v1 -> v0
                                    nop
                                    jump block1

                                block1:
                                    nop
@0002                               v2 = call fn0(v1)
                                    v3 -> v2
@0002                               jump block2

                                block2:
@0002                               nop
@0005                               v4 = iconst.i32 0x007f_ffff
@0005                               v5 = band.i32 v3, v4
                                    v11 -> v5
                                    v23 -> v5
@0008                               v6 = iconst.i32 0x7f80_0000
@0008                               v7 = band.i32 v3, v6
                                    v8 -> v7
@000c                               brz v5, block3
@000c                               jump block4(v7)

                                block3:
@000c                               nop
@000d                               brz.i32 v8, block7
@000d                               jump block4(v8)

                                block4(v9: i32):
@000d                               nop
@000e                               v10 = icmp_imm eq v9, 0x7f80_0000
@000e                               brnz v10, block5
@000e                               jump block12

                                block12:
@000e                               brz.i32 v9, block8
@000e                               jump block6

                                block5:
@000e                               nop
@000f                               brz.i32 v11, block9
@000f                               jump block10

                                block6:
@000f                               nop
@0010                               v12 = iconst.i8 4
@0010                               stack_store v12, ss0
@0011                               v13 = stack_load.i8 ss0
@0011                               return v13

                                block7:
@0011                               nop
@0012                               v14 = iconst.i8 2
@0012                               stack_store v14, ss0
@0011                               v15 = stack_load.i8 ss0
@0011                               return v15

                                block8:
@0011                               nop
@0013                               v16 = iconst.i8 3
@0013                               stack_store v16, ss0
@0011                               v17 = stack_load.i8 ss0
@0011                               return v17

                                block9:
@0011                               nop
@0014                               v18 = iconst.i8 1
@0014                               stack_store v18, ss0
@0011                               v19 = stack_load.i8 ss0
@0011                               return v19

                                block10:
@0011                               nop
@0015                               v20 = iconst.i8 0
@0015                               stack_store v20, ss0
@0011                               v21 = stack_load.i8 ss0
@0011                               return v21

                                block11:
@0011                               nop
@0017                               v22 = stack_load.i8 ss0
@0017                               return v22
}

view this post on Zulip Benjamin Bouvier (Jan 16 2021 at 11:12):

did cg_clif insert these nops?

view this post on Zulip Benjamin Bouvier (Jan 16 2021 at 11:14):

for instance for block3, the expected dominator value is wrong -- it's the actual value that's correct

view this post on Zulip bjorn3 (Jan 16 2021 at 11:19):

Yes, cg_clif inserts these nops at the start of the codegen of each mir block. I have code writing comments after specified instructions when printing the clif ir. The comment for the first mir statement is attached to this nop.

view this post on Zulip bjorn3 (Jan 16 2021 at 11:54):

When making Context::verify panic on verifier errors I got the following backtrace:

   0: rust_begin_unwind
             at /rustc/44e3daf5eee8263dfc3a2509e78ddd1f6f783a0e/library/std/src/panicking.rs:493:5
   1: std::panicking::begin_panic_fmt
             at /rustc/44e3daf5eee8263dfc3a2509e78ddd1f6f783a0e/library/std/src/panicking.rs:435:5
   2: cranelift_codegen::context::Context::verify
             at /home/bjorn/Documenten/wasmtime/cranelift/codegen/src/context.rs:293:13
   3: cranelift_codegen::context::Context::verify_if
             at /home/bjorn/Documenten/wasmtime/cranelift/codegen/src/context.rs:302:13
   4: cranelift_codegen::context::Context::preopt
             at /home/bjorn/Documenten/wasmtime/cranelift/codegen/src/context.rs:347:9
   5: cranelift_codegen::context::Context::compile
             at /home/bjorn/Documenten/wasmtime/cranelift/codegen/src/context.rs:173:13
   6: <cranelift_object::backend::ObjectModule as cranelift_module::module::Module>::define_function
             at /home/bjorn/Documenten/wasmtime/cranelift/object/src/backend.rs:263:13
   7: rustc_codegen_cranelift::base::codegen_fn::{{closure}}
             at /home/bjorn/Documenten/cg_clif/src/base.rs:126:9

The verifier error happens after preopt.

view this post on Zulip bjorn3 (Jan 16 2021 at 12:01):

context.domtree.clear() after the optimizations done by cg_clif itself fixes it.

view this post on Zulip bjorn3 (Jan 16 2021 at 12:02):

Some optimizations probably don't expect it to already be computed and thus don't invalidate it as necessary.

view this post on Zulip bjorn3 (Jan 16 2021 at 12:05):

Fixed in https://github.com/bjorn3/rustc_codegen_cranelift/commit/cfedad1f75bf22468fce59f754daf1501fa2827d

Clear domtree after cg_clif optimizations · bjorn3/rustc_codegen_cranelift@cfedad1
Cranelift based backend for rustc. Contribute to bjorn3/rustc_codegen_cranelift development by creating an account on GitHub.

view this post on Zulip Benjamin Bouvier (Jan 16 2021 at 12:34):

Nice to know there's a workaround, I'd be really interested in extracting a clif test case to reproduce. Maybe time to eat the frog and set up cg_clif :smile:

view this post on Zulip bjorn3 (Jan 16 2021 at 12:38):

I think the repro would be something like context.compute_cfg(); context.compute_domtree(); context.preopt(isa).unwrap(); with the verifier enabled on a function containing a nop instruction.

Last updated: Nov 22 2024 at 17:03 UTC