Issue 9200 - Wrong SIMD code generated
Summary: Wrong SIMD code generated
Status: RESOLVED FIXED
Alias: None
Product: D
Classification: Unclassified
Component: dmd (show other issues)
Version: D2
Hardware: All All
: P2 critical
Assignee: No Owner
URL:
Keywords: SIMD, wrong-code
Depends on:
Blocks:
 
Reported: 2012-12-23 10:38 UTC by jerro.public
Modified: 2013-01-14 12:44 UTC (History)
2 users (show)

See Also:

Attachments
Add an attachment (proposed patch, testcase, etc.)

Note You need to log in before you can comment on or make changes to this issue.
Description jerro.public 2012-12-23 10:38:26 UTC 
When compiled with no flags, the following program gives wrong results:

import std.stdio;
import core.simd;

double2 * v(double* a)
{
    return cast(double2*)a;
}

void main()
{
    double2 a;
    auto p = cast(double*) &a;
    p[0] = 1;
    p[1] = 2;

    double2 b = v(p)[0];
    v(p)[0] = b;

    writeln(p[0 .. 2]); // prints [1, 0]
}

Disassembly of the relevant part of the code:


call   426344 <_D3tmp1vFPdZPNhG2d>
movapd xmm0,XMMWORD PTR [rax]
movapd XMMWORD PTR [rbp-0x10],xmm0
movapd xmm1,XMMWORD PTR [rbp-0x10]
movsd  QWORD PTR [rbp-0x40],xmm1        ; should be movapd
mov    rdi,QWORD PTR [rbp-0x20]
call   426344 <_D3tmp1vFPdZPNhG2d>
movsd  xmm1,QWORD PTR [rbp-0x40]        ; should be movapd
movapd XMMWORD PTR [rax],xmm1

This happens with both DMD 2.060 and the latest version of 2.061 from github. It doesn't happen if I use either -O flag or -inline. It doesn't happen with LDC or GDC.

I have only tested this on linux.
Comment 1 jerro.public 2012-12-23 10:49:42 UTC 
I managed to reduce it a bit further:

import std.stdio;
import core.simd;

double2 * v(double2* a)
{
    return a;
}

void main()
{
    double2 a = [1, 2];

    *v(&a) = a;

    writeln(a.array);
}

And the disassembly:


movsd  QWORD PTR [rbp-0x20],xmm1
lea    rdi,[rbp-0x10]
call   4263f4 <_D3tmp1vFPNhG2dZPNhG2d>
movsd  xmm1,QWORD PTR [rbp-0x20]
movapd XMMWORD PTR [rax],xmm1
Comment 2 Walter Bright 2012-12-23 21:10:51 UTC 
This is happening in cod3.c REGSAVE::save() and REGSAVE::restore(). Unfortunately, just changing the opcodes doesn't work because MOVAPD requires 16 bit alignment of the operands. Fixing that exposes further problems.

Essentially, it'll have to wait a bit.
Comment 3 jerro.public 2012-12-24 16:24:04 UTC 
(In reply to comment #2)
> This is happening in cod3.c REGSAVE::save() and REGSAVE::restore().
> Unfortunately, just changing the opcodes doesn't work because MOVAPD requires
> 16 bit alignment of the operands. Fixing that exposes further problems.
> 
> Essentially, it'll have to wait a bit.

I know nothing about the DMD back end, so this may be an obviously bad idea, but if alignment is the main problem, wouldn't using MOVUPD work in the meantime?
Comment 4 Walter Bright 2012-12-24 19:00:01 UTC 
MOVUPD is terribly slow.
Comment 5 yebblies 2013-01-14 03:08:27 UTC 
(In reply to comment #4)
> MOVUPD is terribly slow.

Terribly slow is still much better than wrong-code.