[ELF] Rework debug line parsing to use llvm::Error and callbacks (LLD-side)

[jhenderson]

Summary

Feedback on D44382 suggested to change the interface to use a callback instead of the structures used. As I didn't want to effectively overwrite that diff, I created D44560 as an alternative implementation using the suggested approach. This is the corresponding patch required for LLD assuming D44560 were to be applied.

As with D44382, D44560 changes the debug line parser interface to report LLVM errors in an interface that different executables can use, rather than always being printed directly as warnings to stderr. This change allows LLD to make use of the new interface and call its own warning methods to report problems.

To test this, I have extended the bad-debug undefined symbol message case to show that a corresponding warning is printed, if the debug line cannot be parsed. In addition, I have also added tests for LLD attempting to parse a non-existent/empty debug line section, showing that the new warning for attempting to do this is not emitted.

[grimar]

ELF/InputFiles.cpp

128 ↗

(On Diff #138691)

I would move this comment above isValidOffset(0) and update it as you are
doing something with offset 0 right there already.

136 ↗

(On Diff #138691)

Is it useful to warn here? I think we only call this method on linkage error,
so anyways going to terminate the link.
With that, it seems easier to always error out here to simplify code/logic.

[jhenderson]

ELF/InputFiles.cpp

128 ↗

(On Diff #138691)

Actually, the comment (and code) is technically wrong - it is possible for there to be multiple CUs in a single object input. I'm thinking objects built via LTO or -r links. I'll add a FIXME and file a bug, along with updating the comment.

136 ↗

(On Diff #138691)

I'm currently preserving behaviour (or fixing intended behaviour, at least), though I certainly see where you're coming from. In general we try to emit as many errors/warnings as we can before stopping, so I still think it's useful - it might help the user identify some other problem, for example.

I'm not bothered either way though, so if the consensus is to not emit the warning, I can easily change that.

[grimar]

ELF/InputFiles.cpp

136 ↗

(On Diff #138691)

I meant we always can error here instead of warn for all cases I think.

[grimar]

ELF/InputFiles.cpp

136 ↗

(On Diff #138691)

As at this point, we are already in error state.

[jhenderson]

Update comment.

ELF/InputFiles.cpp

136 ↗

(On Diff #138691)

Right, but that would suggest to the user that fixing the malformed debug line is a requirement to get a working link, which it isn't. Perhaps another way to think about it is if we wanted to introduce a new warning that could use source information. We'd want to use this function too, but we shouldn't stop the link succeeding if we can't parse the line table.

I could actually make a case for making the second one a warning too, but I made that an error, because we don't actually ever expect it to be reported currently.

[• espindola]

test/ELF/undef.s

37 ↗

(On Diff #138892)

I think I agree with George about producing an error on corrupted output. In practice we expect to never see it, so we may as well make the code simpler.

[• espindola]

ELF/InputFiles.cpp

130 ↗

(On Diff #138892)

When do we expect to have LineData that doesn't include the current CU?

[• rafael]

David Blaikie via llvm-commits <[e-mail]> writes:

[jhenderson]

@rafael on the mailing list:

Now that I think of it, the fact that lld could not parse the debug info to provide a better error message should always be a note, not an error, regardless of how broken the section happens to be.

To be clear, are you saying you're happy with the warnings for now, or would you prefer to drop the severity to just use "message()"? Also, what do you think about the generic error type which currently emits an error? I'd be happy to change the severity of that one too. It's currently an error because we don't expect an Error of that type ever to be reported.

ELF/InputFiles.cpp

130 ↗

(On Diff #138892)

LineData contains the contents of the .debug_line section. It is possible to have an empty section, or a missing section, in which case there will be no valid contents in LineData. Previously, when LLD requested parsing of this section, the parser would return false immediately, because the offset was invalid. With the changes in D44560, an Error is now returned saying the offset is invalid.

In addition, since an object file can contain multiple CUs (e.g. via LTO or -r links), some of which might be missing debug data and others not, we can't know for certain using this method that offset 0 is the offset of the CU for the corresponding symbol (see the reproducible in PR36793).

[• espindola]

In D44562#1042990, @jhenderson wrote:

@rafael on the mailing list:

Now that I think of it, the fact that lld could not parse the debug info to provide a better error message should always be a note, not an error, regardless of how broken the section happens to be.

To be clear, are you saying you're happy with the warnings for now, or would you prefer to drop the severity to just use "message()"? Also, what do you think about the generic error type which currently emits an error? I'd be happy to change the severity of that one too. It's currently an error because we don't expect an Error of that type ever to be reported.

Using message() in all cases seems better. We know the link is failing for unrelated reasons and we are just letting the user know why we are not printing better error messages (could not parse the debug info).

[jhenderson]

In D44562#1043904, @espindola wrote:

Using message() in all cases seems better. We know the link is failing for unrelated reasons and we are just letting the user know why we are not printing better error messages (could not parse the debug info).

message() prints to stdout, which I hadn't realised earlier. This clearly should be printed to stderr, for which there is no mechanism currently, if we don't want a warning or error. Thinking about it more, the rest of LLVM has a concept of "remarks" which are "lower" severity messages than errors or warnings (and can't be promoted to errors, as I understand it). How about I add this to LLD? Alternatively, I'll need to add a note() function, or just print to errs() directly.

[jhenderson]

Add a "remark" method to the LLD ErrorHandler, and use that instead of warnings to report problems with the debug line. Also added another case to the testing, to demonstrate failures via the callback in contrast to failures via the return value.

[• espindola]

Tentative LGTM pending the llvm change.

[jhenderson]

Thanks @espindola. I'll still need to rebase this once I've figured out what to do following rLLD328284 (see also my comments in D44560). I don't anticipate this changing from using remark() though. It's just how that function gets called.

[jhenderson]

Rebase following rLLD328284. I was able to drop one of the test inputs, now that LLD can handle multiple CUs in the same input file. Also added a comment to the test explaining the purpose of the zed6/zed7 test cases.

[jhenderson]

@espindola/@ruiu, are you happy with the latest update?

[ruiu]

Common/ErrorHandler.cpp

86–91 ↗

(On Diff #140069)

Honestly I don't think we want to add a new level of error messages, as I think error() for errors, warning() for warnings and message() for non-error messages is enough. If something needs to be fixed, it should be an error or a warning. If something doesn't have to be fixed, lld shouldn't print out anything. A situation like "verbose messages are printed out but you can ignore them" isn't actionable and thus not desirable.

So, can you choose either (1) show it as a warning or (2) don't show anything?

[jhenderson]

Common/ErrorHandler.cpp

86–91 ↗

(On Diff #140069)

Okay, I have no problem making it a warning. I think this might be a safer course than not printing anything. If undefined symbol messages are incomplete, due to a problem in the input, it might be possible for users to fix it, depending on the nature of the problem (e.g. bogus hand-written debug_line assembly), and therefore allowing them to see the useful error information, allowing them to more quickly identify the problem. I could also see us potentially using this function in the future to produce other informational messages, or implement other features, without always emitting an error, so a warning would be appropriate here too.

[jhenderson]

Change back to using warnings.

[• espindola]

LGTM. I will review the LLVM side too.

lld/trunk/ELF/InputFiles.cpp

 template <class ELFT> void ObjFile<ELFT>::initializeDwarf() {
   Dwarf = llvm::make_unique<DWARFContext>(make_unique<LLDDwarfObj<ELFT>>(this));
   const DWARFObject &Obj = Dwarf->getDWARFObj();
   DwarfLine.reset(new DWARFDebugLine);
   DWARFDataExtractor LineData(Obj, Obj.getLineSection(), Config->IsLE,
                               Config->Wordsize);
 
   for (std::unique_ptr<DWARFCompileUnit> &CU : Dwarf->compile_units()) {
-    const DWARFDebugLine::LineTable *LT = Dwarf->getLineTableForUnit(CU.get());
+    Expected<const DWARFDebugLine::LineTable *> ExpectedLT =
+        Dwarf->getLineTableForUnit(CU.get(), warn);
+    const DWARFDebugLine::LineTable *LT = nullptr;
+    if (ExpectedLT)
+      LT = *ExpectedLT;
+    else
+      handleAllErrors(ExpectedLT.takeError(),
+                      [](ErrorInfoBase &Err) { warn(Err.message()); });
     if (!LT)
       continue;
     LineTables.push_back(LT);
 
     // Loop over variable records and insert them to VariableLoc.
     for (const auto &Entry : CU->dies()) {
       DWARFDie Die(CU.get(), &Entry);
       // Skip all tags that are not variables.

lld/trunk/test/ELF/Inputs/undef-bad-debug.s

 .section .text,"ax"
 sym:
     .quad zed6
-    
+sym2:
+    .quad zed7
+
+.section .debug_line,"",@progbits
+.Lunit:
+    .long .Lunit_end - .Lunit_start # unit length
+.Lunit_start:
+    .short 4                        # version
+    .long .Lprologue_end - .Lprologue_start # prologue length
+.Lprologue_start:
+    .byte 1                         # minimum instruction length
+    .byte 1                         # maximum operatiosn per instruction
+    .byte 1                         # default is_stmt
+    .byte -5                        # line base
+    .byte 14                        # line range
+    .byte 13                        # opcode base
+    .byte 0, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 1 # standard opcode lengths
+    .asciz "dir"                    # include directories
+    .byte 0
+    .asciz "undef-bad-debug.s"      # file names
+    .byte 1, 0, 0
+    .byte 0
+    .byte 0                         # extraneous byte
+.Lprologue_end:
+    .byte 0, 9, 2                   # DW_LNE_set_address
+    .quad sym
+    .byte 3                         # DW_LNS_advance_line
+    .byte 10
+    .byte 1                         # DW_LNS_copy
+    .byte 2                         # DW_LNS_advance_pc
+    .byte 8
+    .byte 0, 1, 1                   # DW_LNE_end_sequence
+.Lunit_end:
+
+.Lunit2:
+    .long .Lunit2_end - .Lunit2_start # unit length
+.Lunit2_start:
+    .short 4                        # version
+    .long .Lprologue2_end - .Lprologue2_start # prologue length
+.Lprologue2_start:
+    .byte 1                         # minimum instruction length
+    .byte 1                         # maximum operatiosn per instruction
+    .byte 1                         # default is_stmt
+    .byte -5                        # line base
+    .byte 14                        # line range
+    .byte 13                        # opcode base
+    .byte 0, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 1 # standard opcode lengths
+    .asciz "dir2"                   # include directories
+    .byte 0
+    .asciz "undef-bad-debug2.s"     # file names
+    .byte 1, 0, 0
+    .byte 0
+.Lprologue2_end:
+    .byte 0, 9, 2                   # DW_LNE_set_address
+    .quad sym2
+    .byte 3                         # DW_LNS_advance_line
+    .byte 10
+    .byte 1                         # DW_LNS_copy
+    .byte 2                         # DW_LNS_advance_pc
+    .byte 8
+    .byte 0, 1, 1                   # DW_LNE_end_sequence
+    .byte 0, 9, 2                   # DW_LNE_set_address
+    .quad 0x0badbeef
+    .byte 3                         # DW_LNS_advance_line
+    .byte 99
+    .byte 1                         # DW_LNS_copy
+    .byte 99                        # DW_LNS_advance_pc
+    .byte 119
+    # Missing end of sequence.
+.Lunit2_end:
 .section .debug_info,"",@progbits
     .long   .Lcu_end - .Lcu_start   # Length of Unit
 .Lcu_start:
     .short  4                       # DWARF version number
     .long   .Lsection_abbrev        # Offset Into Abbrev. Section
     .byte   8                       # Address Size (in bytes)
     .byte   1                       # Abbrev [1] 0xb:0x79 DW_TAG_compile_unit
+    .long   .Lunit                  # DW_AT_stmt_list
     .byte   2                       # Abbrev [2] 0x2a:0x15 DW_TAG_variable
     .long   .Linfo_string           # DW_AT_name
                                         # DW_AT_external
     .byte   1                       # DW_AT_decl_file
     .byte   3                       # DW_AT_decl_line
     .byte   0                       # End Of Children Mark
 .Lcu_end:
 
+    .long   .Lcu2_end - .Lcu2_start # Length of Unit
+.Lcu2_start:
+    .short  4                       # DWARF version number
+    .long   .Lsection_abbrev        # Offset Into Abbrev. Section
+    .byte   8                       # Address Size (in bytes)
+    .byte   1                       # Abbrev [1] 0xb:0x79 DW_TAG_compile_unit
+    .long   .Lunit2                 # DW_AT_stmt_list
+    .byte   2                       # Abbrev [2] 0x2a:0x15 DW_TAG_variable
+    .long   .Linfo2_string          # DW_AT_name
+                                        # DW_AT_external
+    .byte   1                       # DW_AT_decl_file
+    .byte   3                       # DW_AT_decl_line
+    .byte   0                       # End Of Children Mark
+.Lcu2_end:
+
 .section .debug_abbrev,"",@progbits
 .Lsection_abbrev:
     .byte   1                       # Abbreviation Code
     .byte   17                      # DW_TAG_compile_unit
     .byte   1                       # DW_CHILDREN_yes
+    .byte   16                      # DW_AT_stmt_list
+    .byte   23                      # DW_FORM_sec_offset
     .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   14                      # DW_FORM_strp
     .byte   63                      # DW_AT_external
     .byte   25                      # DW_FORM_flag_present
     .byte   58                      # DW_AT_decl_file
     .byte   11                      # DW_FORM_data1
     .byte   59                      # DW_AT_decl_line
     .byte   11                      # DW_FORM_data1
     .byte   0                       # EOM(1)
     .byte   0                       # EOM(2)
     .byte   0                       # EOM(3)
 
 .section .debug_str,"MS",@progbits,1
 .Linfo_string:
     .asciz "sym"
+.Linfo2_string:
+    .asciz "sym2"

lld/trunk/test/ELF/no-line-parser-errors-if-empty-section.s

+# REQUIRES: x86
+
+# LLD uses the debug data to get information for error messages, if possible.
+# However, if the debug line section is empty, we should not attempt to parse
+# it, as that would result in errors from the parser.
+
+# RUN: llvm-mc -filetype=obj -triple=x86_64-pc-linux %s -o %t.o
+# RUN: not ld.lld %t.o -o %t.elf 2>&1 | FileCheck %s
+
+# CHECK-NOT: warning:
+# CHECK-NOT: error:
+# CHECK: error: undefined symbol: undefined
+# CHECK-NEXT: {{.*}}.o:(.text+0x1)
+# CHECK-NOT: warning:
+# CHECK-NOT: error:
+
+.globl _start
+_start:
+    callq undefined
+
+.section .debug_line,"",@progbits

lld/trunk/test/ELF/no-line-parser-errors-if-no-section.s

+# REQUIRES: x86
+
+# LLD uses the debug data to get information for error messages, if possible.
+# However, if there is no debug line section, we should not attempt to parse
+# it, as that would result in errors from the parser.
+
+# RUN: llvm-mc -filetype=obj -triple=x86_64-pc-linux %s -o %t.o
+# RUN: not ld.lld %t.o -o %t.elf 2>&1 | FileCheck %s
+
+# CHECK-NOT: warning:
+# CHECK-NOT: error:
+# CHECK: error: undefined symbol: undefined
+# CHECK-NEXT: {{.*}}.o:(.text+0x1)
+# CHECK-NOT: warning:
+# CHECK-NOT: error:
+
+.globl _start
+_start:
+    callq undefined

lld/trunk/test/ELF/undef.s

 # CHECK: error: undefined symbol: zed4
 # CHECK: >>> referenced by undef-debug.s:7 (dir{{/|\\}}undef-debug.s:7)
 # CHECK: >>>               {{.*}}.o:(.text.1+0x0)
 
 # CHECK: error: undefined symbol: zed5
 # CHECK: >>> referenced by undef-debug.s:11 (dir{{/|\\}}undef-debug.s:11)
 # CHECK: >>>               {{.*}}.o:(.text.2+0x0)
 
+# Show that all line table problems are mentioned as soon as the object's line information
+# is requested, even if that particular part of the line information is not currently required.
+# CHECK: warning: parsing line table prologue at 0x00000000 should have ended at 0x00000038 but it ended at 0x00000037
+# CHECK: warning: last sequence in debug line table is not terminated!
 # CHECK: error: undefined symbol: zed6
 # CHECK: >>> referenced by {{.*}}tmp4.o:(.text+0x0)
 
+# Show that a problem with one line table's information doesn't affect getting information from
+# a different one in the same object.
+# CHECK: error: undefined symbol: zed7
+# CHECK: >>> referenced by undef-bad-debug2.s:11 (dir2{{/|\\}}undef-bad-debug2.s:11)
+# CHECK: >>>               {{.*}}tmp4.o:(.text+0x8)
+
 # RUN: not ld.lld %t.o %t2.a -o %t.exe -no-demangle 2>&1 | \
 # RUN:   FileCheck -check-prefix=NO-DEMANGLE %s
 # NO-DEMANGLE: error: undefined symbol: _Z3fooi
 
 .file "undef.s"
 
   .globl _start
 _start:
   call foo
   call bar
   call zed1
   call _Z3fooi