| Issue |
244346
|
| Summary |
[DWARFVerifier] `--verify` requires a .debug_names entry for a local whose pieces are a constant address value and a stack slot
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
clackary
|
## Summary
This is a false positive in the DWARF verifier (`llvm-dwarfdump --verify` / `dsymutil --verify-dwarf`). It's neither a crash nor a miscompile: the generated DWARF is valid, but the verifier rejects it, and that fails dsymutil's output verification.
Since #228574 ("[DWARF] Don't index variables whose address is only their value"), `isVariableIndexable` treats a variable as static storage when both of these are true:
1. `_expression_.isMemoryLocation()`, meaning any piece of the location is a memory location, and
2. the _expression_ contains `DW_OP_addr` (or a TLS op) anywhere in the _expression_.
The two checks are evaluated independently, so they can be satisfied by different pieces. A local whose location is
```
DW_OP_addr X, DW_OP_stack_value, DW_OP_piece 8, <- constant value (an address used as a value)
DW_OP_breg29 -152, DW_OP_piece 8 <- memory, but a stack slot
```
has no piece at a static address, but the verifier still requires it to be in `.debug_names`. DWARFLinker (correctly) doesn't index it, so `dsymutil`'s output verification fails:
```
error: Name Index @ 0x0: Entry for DIE @ 0x788ed (DW_TAG_variable) with name expectedMultiplied missing.
error: Name Index @ 0x0: Entry for DIE @ 0x78943 (DW_TAG_variable) with name expectedDivided missing.
error: Name Index @ 0x0: Entry for DIE @ 0x784fb (DW_TAG_variable) with name expectedMultiplied missing.
error: Name Index @ 0x0: Entry for DIE @ 0x78551 (DW_TAG_variable) with name expectedDivided missing.
error: output verification failed for arm64
```
The comment on the existing `var_mixed_piece` test case states the intended rule: "a piece that does live at an address does [call for an index entry]". Every existing test case has either no memory piece or a memory piece that *is* the `DW_OP_addr`. None has a non-static memory piece next to an `addr; stack_value` piece, so this case isn't covered.
## Reproducer
This needs `llvm-mc` and `llvm-dwarfdump` from a build that includes #228574, such as current `main`. Older versions use the previous rule and also report `value_and_register`. `repro.s` (below) has two locals. Each is made of two 8-byte pieces, and the first piece of each is a constant address used as a value (`DW_OP_addr; DW_OP_stack_value`). The second piece is a stack slot in one variable and a register in the other. Neither variable has static storage, so neither needs a `.debug_names` entry.
```sh
llvm-mc -triple x86_64-pc-linux -filetype=obj repro.s -o repro.o
llvm-dwarfdump --verify repro.o
```
Expected: no errors. Only the usual `Name Index @ 0x0 does not contain a hash table` warning should appear.
Actual: the variable with the stack-slot piece is reported, and the one with the register piece isn't:
```
Verifying .debug_names...
warning: Name Index @ 0x0 does not contain a hash table.
error: Name Index @ 0x0: Entry for DIE @ 0xc (DW_TAG_variable) with name value_and_stack_slot missing.
error: Aggregated error counts:
error: Name Index DIE entry missing name occurred 1 time(s).
Errors detected.
```
This output is from `llvm-mc` and `llvm-dwarfdump` built from swiftlang/llvm-project at the tag `swift-DEVELOPMENT-SNAPSHOT-2026-11-01-a`, on arm64 macOS. The verifier code involved (`isVariableIndexable`, which uses `DWARFExpression::isMemoryLocation`) is identical on llvm/llvm-project `main`.
<details>
<summary><code>repro.s</code></summary>
```asm
# Two locals, each made of two 8-byte pieces. The first piece of each is a
# constant address used as a value (DW_OP_addr; DW_OP_stack_value). The second
# piece is a stack slot in one variable and a register in the other. Neither
# variable has static storage, so neither needs a .debug_names entry.
#
# llvm-mc -triple x86_64-pc-linux -filetype=obj repro.s -o repro.o
# llvm-dwarfdump --verify repro.o
.section .debug_abbrev,"",@progbits
.byte 1 # Abbreviation Code
.byte 17 # DW_TAG_compile_unit
.byte 1 # DW_CHILDREN_yes
.byte 0 # EOM(1)
.byte 0 # EOM(2)
.byte 2 # Abbreviation Code
.byte 52 # DW_TAG_variable
.byte 0 # DW_CHILDREN_no
.byte 3 # DW_AT_name
.byte 8 # DW_FORM_string
.byte 2 # DW_AT_location
.byte 24 # DW_FORM_exprloc
.byte 0 # EOM(1)
.byte 0 # EOM(2)
.byte 0 # EOM(3)
.section .debug_info,"",@progbits
.Lcu_begin0:
.long .Lcu_end0-.Lcu_start0 # Length of Unit
.Lcu_start0:
.short 4 # DWARF version number
.long .debug_abbrev # Offset Into Abbrev. Section
.byte 8 # Address Size (in bytes)
.byte 1 # Abbrev [1] DW_TAG_compile_unit
.byte 2 # Abbrev [2] DW_TAG_variable
.asciz "value_and_stack_slot" # DW_AT_name
.byte .Lloc1_end-.Lloc1 # DW_AT_location
.Lloc1:
.byte 0x03 # DW_OP_addr
.quad 0x47
.byte 0x9f # DW_OP_stack_value
.byte 0x93, 8 # DW_OP_piece 8
.byte 0x77, 0x78 # DW_OP_breg7 RSP-8
.byte 0x93, 8 # DW_OP_piece 8
.Lloc1_end:
.byte 2 # Abbrev [2] DW_TAG_variable
.asciz "value_and_register" # DW_AT_name
.byte .Lloc2_end-.Lloc2 # DW_AT_location
.Lloc2:
.byte 0x03 # DW_OP_addr
.quad 0x47
.byte 0x9f # DW_OP_stack_value
.byte 0x93, 8 # DW_OP_piece 8
.byte 0x50 # DW_OP_reg0 RAX
.byte 0x93, 8 # DW_OP_piece 8
.Lloc2_end:
.byte 0 # End Of Children Mark
.Lcu_end0:
.section .debug_names,"",@progbits
.long .Lnames_end0-.Lnames_start0 # Header: contribution length
.Lnames_start0:
.short 5 # Header: version
.short 0 # Header: padding
.long 1 # Header: compilation unit count
.long 0 # Header: local type unit count
.long 0 # Header: foreign type unit count
.long 0 # Header: bucket count
.long 0 # Header: name count
.long .Lnames_abbrev_end0-.Lnames_abbrev_start0 # Header: abbreviation table size
.long 0 # Header: augmentation length
.long .Lcu_begin0 # Compilation unit 0
.Lnames_abbrev_start0:
.byte 0 # End of abbrev list
.Lnames_abbrev_end0:
.Lnames_end0:
```
</details>
### Real-world (Swift, dsymutil)
Found with `swift-DEVELOPMENT-SNAPSHOT-2026-11-01-a` building a package's tests with `swift build --build-tests -c release` on arm64 macOS (Xcode 27.1 SDK). Release only; debug builds are fine.
The test code holds the result of a derivative function, a `(value: T, pullback: (T) -> (T, T))` tuple, in a local. The pullback closure captures two values, and at `-O` its context gets saved to the stack. The local's location list in the linked dSYM:
```
0x000784fb: DW_TAG_variable
DW_AT_location (0x00000535:
[0xbef8, 0xbf04): DW_OP_piece 0x10, DW_OP_reg0 W0, DW_OP_piece 0x8
[0xbf04, 0xbf20): DW_OP_constu 0xbff2000000000000, DW_OP_stack_value, DW_OP_piece 0x8, DW_OP_addr 0xe124, DW_OP_stack_value, DW_OP_piece 0x8, DW_OP_reg0 W0, DW_OP_piece 0x8
[0xbf20, 0xc068): DW_OP_constu 0xbff2000000000000, DW_OP_stack_value, DW_OP_piece 0x8, DW_OP_addr 0xe124, DW_OP_stack_value, DW_OP_piece 0x8, DW_OP_reg24 W24, DW_OP_piece 0x8
[0xc068, 0xc6ec): DW_OP_constu 0xbff2000000000000, DW_OP_stack_value, DW_OP_piece 0x8, DW_OP_addr 0xe124, DW_OP_stack_value, DW_OP_piece 0x8, DW_OP_breg29 W29-152, DW_OP_piece 0x8)
DW_AT_name ("expectedMultiplied")
```
`0xe124` is a partial-apply forwarder (`$sS3dIegnrr_S3dIegydd_TRTA`), i.e. the closure's function pointer, used as a value. Only the last range has a memory piece (`breg29`), and that range is enough to trigger the error.
Similar locals in the same function confirm the pattern; none of them is flagged:
- Closures that capture nothing have a `DW_OP_lit0` context piece.
- Closures that capture one value keep the context in a register (`reg23`/`reg24`) for the whole range.
Exactly the four DIEs that have an `addr; stack_value` piece next to a `breg` piece are flagged.
## Workaround for users
Build without full debug info for affected targets (`-gline-tables-only`, or skip dSYM generation).
## AI assistance
I found the original failure in a real project build, and used the robot to distill a representative reproducer, which I ran myself to validate. It was used to generate this issue, which I have reviewed and curated for relevance. I am more familiar with the swiftlang dev process, but I have reviewed LLVM's bug report and AI policy docs before submitting.
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs