LLVM 5 fails debuginfo/by-value-self-argument-in-trait-impl.rs

[cuviper]

rustc compiled with LLVM 5 fails debuginfo/by-value-self-argument-in-trait-impl.rs on x86_64-unknown-linux-gnu. GDB fails to read self for the following impl in the test:

impl Trait for (f64, isize, isize, f64) {
    fn method(self) -> (f64, isize, isize, f64) {
        zzz(); // #break
        self
    }
}

Cannot access memory at address 0x40b15c8000000000

It's called with (4444.5, 5555, 6666, 7777.5).method(), and it just so happens that 4444.5 transmuted to a u64 is 0x40b15c8000000000, so there seems to be an indirection/deref problem in the debuginfo.

Full test output:

---- [debuginfo-gdb] debuginfo/by-value-self-argument-in-trait-impl.rs stdout ----
        NOTE: compiletest thinks it is using GDB with native rust support
NOTE: compiletest thinks it is using GDB version 8000001

error: line not found in debugger output: $3 = (4444.5, 5555, 6666, 7777.5)
status: exit code: 0
command: "/usr/bin/gdb" "-quiet" "-batch" "-nx" "-command=/home/jistone/rust/rust/build/x86_64-unknown-linux-gnu/test/debuginfo/by-value-self-argument-in-trait-impl.debugger.script"
stdout:
------------------------------------------
GNU gdb (GDB) Fedora 8.0.1-33.fc27
Copyright (C) 2017 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law.  Type "show copying"
and "show warranty" for details.
This GDB was configured as "x86_64-redhat-linux-gnu".
Type "show configuration" for configuration details.
For bug reporting instructions, please see:
<http://www.gnu.org/software/gdb/bugs/>.
Find the GDB manual and other documentation resources online at:
<http://www.gnu.org/software/gdb/documentation/>.
For help, type "help".
Type "apropos word" to search for commands related to "word".
Breakpoint 1 at 0xf9c: file /home/jistone/rust/rust/src/test/debuginfo/by-value-self-argument-in-trait-impl.rs, line 59.
Breakpoint 2 at 0xfc0: file /home/jistone/rust/rust/src/test/debuginfo/by-value-self-argument-in-trait-impl.rs, line 71.
Breakpoint 3 at 0xfeb: file /home/jistone/rust/rust/src/test/debuginfo/by-value-self-argument-in-trait-impl.rs, line 78.
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib64/libthread_db.so.1".

Breakpoint 1, <isize as by_value_self_argument_in_trait_impl::Trait>::method (self=1111) at /home/jistone/rust/rust/src/test/debuginfo/by-value-self-argument-in-trait-impl.rs:59
59              zzz(); // #break
$1 = 1111

Breakpoint 2, <by_value_self_argument_in_trait_impl::Struct as by_value_self_argument_in_trait_impl::Trait>::method (self=...) at /home/jistone/rust/rust/src/test/debuginfo/by-value-self-argument-in-trait-impl.rs:71
71              zzz(); // #break
$2 = by_value_self_argument_in_trait_impl::Struct {x: 2222, y: 3333}

Breakpoint 3, <(f64, isize, isize, f64) as by_value_self_argument_in_trait_impl::Trait>::method (self=<error reading variable: Cannot access memory at address 0x40b15c8000000000>) at /home/jistone/rust/rust/src/test/debuginfo/by-value-self-argument-in-trait-impl.rs:78
78              zzz(); // #break

------------------------------------------
stderr:
------------------------------------------
/home/jistone/rust/rust/build/x86_64-unknown-linux-gnu/test/debuginfo/by-value-self-argument-in-trait-impl.debugger.script:16: Error in sourced command file:
Cannot access memory at address 0x40b15c8000000000

------------------------------------------

thread '[debuginfo-gdb] debuginfo/by-value-self-argument-in-trait-impl.rs' panicked at 'explicit panic', tools/compiletest/src/runtest.rs:2884:9

[cuviper]

Here's the .debug_loc in the working version compiled with LLVM 4:

    Offset   Begin            End              Expression
    00000000 0000000000000fe0 0000000000000ff3 (DW_OP_breg4 (rsi): 0)
    00000014 0000000000000ff3 000000000000104c (DW_OP_breg6 (rbp): -48; DW_OP_deref)
    00000029 <End of list>

And here it is compiled with LLVM 5 -- note the extra DW_OP_deref:

    Offset   Begin            End              Expression
    00000000 0000000000000fe0 0000000000000ff3 (DW_OP_breg4 (rsi): 0; DW_OP_deref)
    00000015 0000000000000ff3 000000000000104c (DW_OP_breg6 (rbp): -48; DW_OP_deref; DW_OP_deref)
    0000002b <End of list>

[alexcrichton]

Nice investigation @cuviper! Grepping the LLVM history may mean that llvm-mirror/llvm@ca18763 is a point of interest.

[cuviper]

@alexcrichton AFAICS that commit is after LLVM 5, present in the release_60 and master branches -- so are you looking at that commit that it might have fixed this?

In release_40..release_50, this one looks suspicious to me: llvm-mirror/llvm@b560ea7

I had to fix the semantics of dbg.declare for describing function
arguments. After this patch a dbg.declare always takes the address
of a variable as the first argument, even if the argument is not an
alloca.

Which might mean that our own deref logic is out of date.

[alexcrichton]

Oh heh, good point!

I actually saw that LLVM 6 was tagged with an rc today so I was kicking the tires tonight, and unfortunately this failure is still here after updating to LLVM 6. (that's how I saw it)

In that case you're probably right! @eddyb or @michaelwoerister would y'all be able to help out with this?

[cuviper]

To test, I tried simply masking that manual deref op:

diff --git a/src/librustc_trans/mir/mod.rs b/src/librustc_trans/mir/mod.rs
index b367eb6548d0..56e545cf8c3a 100644
--- a/src/librustc_trans/mir/mod.rs
+++ b/src/librustc_trans/mir/mod.rs
@@ -492,7 +492,7 @@ fn arg_local_refs<'a, 'tcx>(bx: &Builder<'a, 'tcx>,
                 };

                 if let PassMode::Indirect(ref attrs) = arg.mode {
-                    if !attrs.contains(ArgAttribute::ByVal) {
+                    if !attrs.contains(ArgAttribute::ByVal) && false {
                         variable_access = VariableAccess::IndirectVariable {
                             alloca: place.llval,
                             address_operations: &deref_op,

This passes all debuginfo tests with LLVM 5! But (predictably) it fails on LLVM 4 and 3.9. So I guess we could add a conditional flag from rustllvm to tell us whether a manual deref should be added, depending on LLVM version. Or is there a better way to accomplish this?

[alexcrichton]

@cuviper heh I was just reading over the commit you linked (which seems like the correct diagnosis) and the rustc code and concluded to test the same thing, glad it actually works!

I'd personally be fine with a "do this on LLVM 4 and below but not 5 and up" conditional (with a comment), and @michaelwoerister could do the final sign off to make sure it's not completely crazy.