It's still a bit unclear what exactly causes the bug. I was able to bisect an error in Higgs unittests down to this change. https://github.com/D-Programming-Language/dmd/pull/3885#issuecomment-69158463
Please provide a reproducible example.
(In reply to Walter Bright from comment #1) > Please provide a reproducible example. I already spend almost a day on this and couldn't reduce it to less than compiling Higgs and running the tests. Will try again, but it's definitely cause by the mentioned commit.
After spending yet another day on this :(, I finally found it. It's a bug in Higgs that was triggered by the change to the struct literal code. https://github.com/higgsjs/Higgs/pull/177 I'm wondering why this worked before. Was the struct previously fully initialized before calling the ctor, or did this just work by accidentally? Anyhow, I filed bug 14107 because this should never have compiled.
struct X86Imm { ulong imm; } struct X86Opnd { union { X86Reg reg; X86Imm imm; } ubyte tag; this(X86Reg r) { reg = r; } } struct X86Reg { /// Register type ubyte type; /// Register index number ubyte regNo; /// Size in bits ushort size; } X86Opnd opnd(X86Reg reg) { return X86Opnd(reg); } Here is the asm that for the opnd function with 2.066.1 _D4bug24opndFS4bug26X86RegZS4bug27X86Opnd PROC push rbp ; 0000 _ 55 mov rbp, rsp ; 0001 _ 48: 8B. EC sub rsp, 32 ; 0004 _ 48: 83. EC, 20 lea rax, [rbp-20H] ; 0008 _ 48: 8D. 45, E0 xor rcx, rcx ; 000C _ 48: 31. C9 mov qword ptr [rax], rcx ; 000F _ 48: 89. 08 mov qword ptr [rax+8H], rcx ; 0012 _ 48: 89. 48, 08 mov rsi, rdi ; 0016 _ 48: 89. FE mov rdi, rax ; 0019 _ 48: 89. C7 call _D4bug27X86Opnd6__ctorMFNcS4bug26X86RegZS4bug27X86Opnd; 001C _ E8, 00000000(rel) mov rdx, qword ptr [rax+8H] ; 0021 _ 48: 8B. 50, 08 mov rax, qword ptr [rax] ; 0025 _ 48: 8B. 00 mov rsp, rbp ; 0028 _ 48: 8B. E5 pop rbp ; 002B _ 5D ret ; 002C _ C3 _D4bug24opndFS4bug26X86RegZS4bug27X86Opnd ENDP That's the same function now. _D4bug24opndFS4bug26X86RegZS4bug27X86Opnd PROC push rbp ; 0000 _ 55 mov rbp, rsp ; 0001 _ 48: 8B. EC sub rsp, 32 ; 0004 _ 48: 83. EC, 20 xor eax, eax ; 0008 _ 31. C0 mov dword ptr [rbp-18H], eax ; 000A _ 89. 45, E8 mov dword ptr [rbp-14H], eax ; 000D _ 89. 45, EC xor ecx, ecx ; 0010 _ 31. C9 mov byte ptr [rbp-10H], cl ; 0012 _ 88. 4D, F0 mov byte ptr [rbp-0FH], cl ; 0015 _ 88. 4D, F1 mov word ptr [rbp-0EH], ax ; 0018 _ 66: 89. 45, F2 mov edx, dword ptr [rbp-10H] ; 001C _ 8B. 55, F0 mov dword ptr [rbp-20H], edx ; 001F _ 89. 55, E0 mov byte ptr [rbp-18H], cl ; 0022 _ 88. 4D, E8 mov rsi, rdi ; 0025 _ 48: 89. FE lea rdi, [rbp-20H] ; 0028 _ 48: 8D. 7D, E0 call _D4bug27X86Opnd6__ctorMFNcS4bug26X86RegZS4bug27X86Opnd; 002C _ E8, 00000000(rel) mov rdx, qword ptr [rax+8H] ; 0031 _ 48: 8B. 50, 08 mov rax, qword ptr [rax] ; 0035 _ 48: 8B. 00 mov rsp, rbp ; 0038 _ 48: 8B. E5 pop rbp ; 003B _ 5D ret ; 003C _ C3 _D4bug24opndFS4bug26X86RegZS4bug27X86Opnd ENDP Just focusing on the X86Opnd init code lea rax, [rbp-20H] xor rcx, rcx mov qword ptr [rax], rcx mov qword ptr [rax+8H], rcx In the new code the 4 bytes at [rbp-1CH] are never initialized. xor eax, eax mov dword ptr [rbp-18H], eax mov dword ptr [rbp-14H], eax xor ecx, ecx mov byte ptr [rbp-10H], cl mov byte ptr [rbp-0FH], cl mov word ptr [rbp-0EH], ax mov edx, dword ptr [rbp-10H] mov dword ptr [rbp-20H], edx mov byte ptr [rbp-18H], cl mov rsi, rdi lea rdi, [rbp-20H] Without a constructor it's the same for both dmd versions and also only initializes the first union field of X86Opnd. xor eax, eax mov dword ptr [rbp-18H], eax mov dword ptr [rbp-14H], eax mov dword ptr [rbp-20H], edi So this isn't really a regression and it was a bug in Higgs to rely on memory compare. I filed https://issues.dlang.org/show_bug.cgi?id=14107 to disallow default comparison for structs with unions. The amount of code generated for the constructor is still worrisome. Especially the part that initializes X86Reg, because it initializes every field individually, it constructs a temporary on the stack only to end up loading zero into edx. xor ecx, ecx mov byte ptr [rbp-10H], cl mov byte ptr [rbp-0FH], cl mov word ptr [rbp-0EH], ax mov edx, dword ptr [rbp-10H] So we should still consider to revert that part of the change.
Changed to reflect your last comment.
(In reply to Walter Bright from comment #5) > Changed to reflect your last comment. No, this is still a regression issue around the struct constructor call. https://github.com/D-Programming-Language/dmd/pull/4387
(In reply to Kenji Hara from comment #6) > (In reply to Walter Bright from comment #5) > > Changed to reflect your last comment. > > No, this is still a regression issue around the struct constructor call. > > https://github.com/D-Programming-Language/dmd/pull/4387 I opened a new performance issue: https://issues.dlang.org/show_bug.cgi?id=14133
Commits pushed to master at https://github.com/D-Programming-Language/dmd https://github.com/D-Programming-Language/dmd/commit/e5f21beb82acade73900aa50a0e092b3935e980b fix Issue 13952 - change in struct ctor lowering triggers codegen bug https://github.com/D-Programming-Language/dmd/commit/2ebb60303208768adfa29d29f11f7fc8c95fffa7 Merge pull request #4387 from 9rnsr/fix13952 [REG2.067a] Issue 13952 - change in struct ctor lowering triggers codegen bug
Commits pushed to https://github.com/D-Programming-Language/dmd https://github.com/D-Programming-Language/dmd/commit/e5f21beb82acade73900aa50a0e092b3935e980b fix Issue 13952 - change in struct ctor lowering triggers codegen bug https://github.com/D-Programming-Language/dmd/commit/2ebb60303208768adfa29d29f11f7fc8c95fffa7 Merge pull request #4387 from 9rnsr/fix13952