Bug 15432 – Win64: bad code offset in debug line number info

This program: ////////////// import std.stdio; enum AnotherColor { Red, } enum Color { red, } int main(string[] argv) { Color c = Color.red; writeln(AnotherColor.Red.stringof); // line 10 return 0; } ////////////// compiled with "dmd -m64 -g main.d" generates this disassembly for main (dumpbin): _Dmain: 0000000000000000: 55 push rbp 0000000000000001: 48 8B EC mov rbp,rsp 0000000000000004: 48 83 EC 28 sub rsp,28h 0000000000000008: 53 push rbx 0000000000000009: 48 89 4D 10 mov qword ptr [rbp+10h],rcx 000000000000000D: 31 C0 xor eax,eax 000000000000000F: 89 45 E0 mov dword ptr [rbp-20h],eax 0000000000000012: 48 8D 0D 00 00 00 lea rcx,[_TMP0] 00 0000000000000019: BB 03 00 00 00 mov ebx,3 000000000000001E: 48 89 5D F0 mov qword ptr [rbp-10h],rbx 0000000000000022: 48 89 4D F8 mov qword ptr [rbp-8],rcx 0000000000000026: 48 8D 4D F0 lea rcx,[rbp-10h] 000000000000002A: 48 83 EC 20 sub rsp,20h 000000000000002E: E8 00 00 00 00 call _D3std5stdio16__T7writelnTAyaZ7writelnFNfAyaZv 0000000000000033: 48 83 C4 20 add rsp,20h 0000000000000037: 31 C0 xor eax,eax 0000000000000039: 5B pop rbx 000000000000003A: 48 8D 65 00 lea rsp,[rbp] 000000000000003E: 5D pop rbp 000000000000003F: C3 ret and these line number infos ("c:\Program Files (x86)\VisualD\cv2pdb\dumplines"): Sym: _Dmain File: main.d Off 0x0: Line 7 Off 0xd: Line 9 Off 0x12: Line 10 Off 0x36: Line 12 Please note, that the code offset for line 12 is not at the beginning of an address, so setting a breakpoint there corrupts execution (the debugger changes the respective code byte to 0xcc).

I understand that you are talking about the debug symbols, right? If so, the following is not the cause of this bug. This bug reminded me of a change I made in the stack trace addresses generation (https://github.com/D-Programming-Language/druntime/commit/8822115bc8d52fa61c15cef38fe77349f18747b9) that I did not test the effect of on Windows. I made this change so that the addresses coming from the stack trace point before the return address (i.e. to point at the call instruction even though it most probably won't point at the start of it). I don't know if the address to line handler on Windows produces correct mappings for such addresses.

Yes, I'm referring to the debug information generated by dmd, so libraries cannot cause this.

I reproduced this with 2.074.1. Detailed steps to reproduce (without assuming prior knowledge): 1. You need: - dumpbin.exe (Visual Studio or Windows SDK) - dumplines.exe (build from source from https://github.com/rainers/cv2pdb, install Visual D from https://github.com/dlang/visuald/releases, or download stand-alone binary I compiled from https://dump.thecybershadow.net/3d28f947dd7aeeb30325935bc2463d5d/dumplines.exe) 2. Save code to test.d 3. Run: dmd -g -m64 -c test.d 4. Run: dumpbin.exe /DISASM test.obj > test.asm 5. Run: dumplines.exe test.obj > test-lines.txt 6. In test-lines.txt, note the offsets (Off 0x### numbers) for the _Dmain function (should be the first in the file) 7. In test.asm, note the offsets (first column of hex numbers) for the instructions for the _Dmain function (should be the first in the file) If there are any offsets in test-lines.txt's _Dmain which do not appear in test.asm's _Dmain, then it is an instance of this bug. A self-contained example would be nice. Would it be possible to reproduce this issue e.g. by throwing an exception and looking at the stack trace for an address not mapped to a line number OSLT?

Here's a further reduction: /////////////////// void call15432(string col) {} int test15432() // line 8 { call15432(null); return 0; } /////////////////// dumpbin: _D7testpdb9test15432FZi: 0000000000000000: 55 push rbp 0000000000000001: 48 8B EC mov rbp,rsp 0000000000000004: 48 83 EC 10 sub rsp,10h 0000000000000008: 48 C7 45 F0 00 00 mov qword ptr [rbp-10h],0 00 00 0000000000000010: 48 C7 45 F8 00 00 mov qword ptr [rbp-8],0 00 00 0000000000000018: 48 8D 4D F0 lea rcx,[rbp-10h] 000000000000001C: 48 83 EC 20 sub rsp,20h 0000000000000020: E8 00 00 00 00 call _D7testpdb9call15432FAyaZv 0000000000000025: 48 83 C4 20 add rsp,20h 0000000000000029: 31 C0 xor eax,eax 000000000000002B: 48 8B E5 mov rsp,rbp 000000000000002E: 5D pop rbp 000000000000002F: C3 ret cvdump (https://github.com/Microsoft/microsoft-pdb/blob/master/cvdump/cvdump.exe): *** LINES 8 00000000 10 00000008 11 00000029 12 0000002A You could get the offset to the line number before a stack address, but the return adress is still in the preceding line because there is some cleanup code after the call. In this code, the bad adress is even after the line with "return". Enumerating the line number info is possible in test/runnable/testpdb.d, but a test will have to check specific code to be generated. At least, I don't have an idea what to test for instead.

Commits pushed to master at https://github.com/dlang/dmd https://github.com/dlang/dmd/commit/399b12d8d6feaf43ced88c2559404f29ab70c496 fix issue 15432: Win64: bad code offset in debug line number info pinhole optimization should not yield different results if run twice https://github.com/dlang/dmd/commit/a7867ca72bf51a6db9e3ed44428e142a21d3de75 Merge pull request #6979 from rainers/issue15432 fix issue 15432: Win64: bad code offset in debug line number info merged-on-behalf-of: Rainer Schuetze <[email protected]>

Commits pushed to newCTFE at https://github.com/dlang/dmd https://github.com/dlang/dmd/commit/399b12d8d6feaf43ced88c2559404f29ab70c496 fix issue 15432: Win64: bad code offset in debug line number info https://github.com/dlang/dmd/commit/a7867ca72bf51a6db9e3ed44428e142a21d3de75 Merge pull request #6979 from rainers/issue15432

Commits pushed to stable at https://github.com/dlang/dmd https://github.com/dlang/dmd/commit/399b12d8d6feaf43ced88c2559404f29ab70c496 fix issue 15432: Win64: bad code offset in debug line number info https://github.com/dlang/dmd/commit/a7867ca72bf51a6db9e3ed44428e142a21d3de75 Merge pull request #6979 from rainers/issue15432