Bug 11310 – Alignment failure in test42.d

Here's the code that fails in test/runnable/test42.d: ----------------------------- align(16) struct S41 { int[4] a; } shared int x41; shared S41 s41; void test41() { printf("&x = %p\n", &x41); printf("&s = %p\n", &s41); assert((cast(int)&s41 & 0xF) == 0); } ---------------------------- when compiled on Linux with: -O -fPIC Curiously, it prints: &x = 0x7fef00 &s = 0x7fef10 which shows it is aligned correctly - so why does the assert fail?

I noticed importing random stuff (e.g. std.string, std.stdio, std.algorithm) made it run without error, then I looked at the assembly code: Disassembly of section .text._D6_113106test41FZv: 0000000000000000 <_D6_113106test41FZv>: 0: 55 push %rbp 1: 48 8b ec mov %rsp,%rbp 4: f6 05 00 00 00 00 0f testb $0xf,0x0(%rip) b: 74 0a je 17 <_D6_113106test41FZv+0x17> d: bf 11 00 00 00 mov $0x11,%edi 12: e8 00 00 00 00 callq 17 <_D6_113106test41FZv+0x17> 17: 5d pop %rbp 18: c3 retq 19: 0f 1f 80 00 00 00 00 nopl 0x0(%rax) Looks like it is and'ing the instruction pointer with 0xf.

That is the assembly without printf's, btw.

Created attachment 1273 Edb screenshot for second printf

Created attachment 1274 Edb screenshot for assert test

As you can see in the first screen shot, the address for s41 is calculated properly, whereas in the second screenshot the is a mistake. rip for the printf calculation: 0x419c84 rip for the assert calculation: 0x419c99 difference: 0x15 offset for the printf calculation: 0x21e35c offset for the assert calculation: 0x21e344 difference: 0x18 Looking at the relocation section of the executable, printf seems to be giving the correct location (0x63cdd0.)

Looking at the non-optimized version of the code, the test instruction operates on dwords, which makes the constant 3 bytes longer (i.e. the difference noted in comment 5,) the constant off of rip is the same as before.

Created attachment 1279 Edb screenshot for unoptimized version Just to make it more clear what the issue is (I was quite sick when I wrote the other comment): The code responsible for calculating the offsets from the IP to the offset table is calculating an offset that is off by 3 bytes. This offset is appropriate for the instruction present in the unoptimized version which operates on dword.

Created attachment 1280 This hack makes the program run properly for me I don't know if there are cases where this will produce invalid code, I'll pull out the 64 bit architecture manual tonight and figure it out.

After checking manuals, I've concluded that this is the correct fix and I've submitted a pull request: https://github.com/D-Programming-Language/dmd/pull/2691

Commits pushed to master at https://github.com/D-Programming-Language/dmd https://github.com/D-Programming-Language/dmd/commit/6ec2800443a088b42722b1580a3388111ed9c2ab fix Issue 11310 - Fix RIP relative offset calculation for TEST instruction. https://github.com/D-Programming-Language/dmd/commit/5138bc4a1aa09e935875c7343adeee3def5ca21c Merge pull request #2692 from Safety0ff/fix11310m fix Issue 11310 - Fix RIP relative offset calculation for TEST instruction

Commit pushed to 2.064 at https://github.com/D-Programming-Language/dmd https://github.com/D-Programming-Language/dmd/commit/6edfd4cdd96889dd539f4b888fcf7e1bd70e09a1 Merge pull request #2692 from Safety0ff/fix11310m fix Issue 11310 - Fix RIP relative offset calculation for TEST instruction