[Issue 10226] New: core.simd bad codegen

[d-bugmail at puremagic.com]

http://d.puremagic.com/issues/show_bug.cgi?id=10226

           Summary: core.simd bad codegen
           Product: D
           Version: D2
          Platform: All
        OS/Version: All
            Status: NEW
          Severity: normal
          Priority: P2
         Component: DMD
        AssignedTo: nobody at puremagic.com
        ReportedBy: code at benjamin-thaut.de


--- Comment #0 from Benjamin Thaut <code at benjamin-thaut.de> 2013-06-01 03:12:12 PDT ---
The following testcode:

import core.simd;
import std.stdio;

void main(string[] args)
{
    float[] value1 = [1.0f, 2.0f, 3.0f, 4.0f];
    float4 result = __simd(XMM.LODAPS, *cast(float4*)value1.ptr);
    result = __simd(XMM.ADDPS, result, result);
    __simd_sto(XMM.STOAPS, *cast(float4*)value1.ptr, result);
    writefln("%s", value1);
}

Will produce the following assembly

1 mov         rax,qword ptr [rbp-68h]  
2 movaps      xmm0,xmmword ptr [rax]  
3 movaps      xmmword ptr [rbp-60h],xmm0  
4 movdqa      xmm0,xmmword ptr [rbp-60h]  
5 addps       xmm0,xmmword ptr [rbp-60h]  
6 movaps      xmmword ptr [rbp-60h],xmm0  
7 movdqa      xmm0,xmmword ptr [rbp-60h]  
8 mov         rax,qword ptr [rbp-68h]  
9 movaps      xmmword ptr [rax],xmm0  

The instructions 3 and 4 are completely useless, as well as the instructions 6
and 7. Instruction 8 has no effect because RAX already contains that value.
Ideally the assembly should look as follows:

1 mov         rax,qword ptr [rbp-68h]  
2 movaps      xmm0,xmmword ptr [rax]  
5 addps       xmm0,xmmword ptr [rbp-60h]  
9 movaps      xmmword ptr [rax],xmm0  

This is a huge problem because SIMD does only start to get effective if you
stay within the xmm registers as long as possible.

tested with dmd 2.063

-- 
Configure issuemail: http://d.puremagic.com/issues/userprefs.cgi?tab=email
------- You are receiving this mail because: -------