https://issues.dlang.org/show_bug.cgi?id=10226

Benjamin Thaut <code@benjamin-thaut.de> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |SIMD

--

https://issues.dlang.org/show_bug.cgi?id=10226

Walter Bright <bugzilla@digitalmars.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |bugzilla@digitalmars.com

--- Comment #9 from Walter Bright <bugzilla@digitalmars.com> ---
(In reply to Benjamin Thaut from comment #8)
> Building with debug symbols gives a ICE ;-)

It works when I compile with -g.

--

https://issues.dlang.org/show_bug.cgi?id=10226

Walter Bright <bugzilla@digitalmars.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement

--- Comment #10 from Walter Bright <bugzilla@digitalmars.com> ---
It isn't 'bad' codegen, it's just inefficient codegen. I reclassified this as an enhancement request.

--

https://issues.dlang.org/show_bug.cgi?id=10226

Walter Bright <bugzilla@digitalmars.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|core.simd bad codegen       |core.simd inefficient
                   |                            |codegen

--

https://issues.dlang.org/show_bug.cgi?id=10226

Martin Nowak <code@dawg.eu> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |code@dawg.eu

--- Comment #11 from Martin Nowak <code@dawg.eu> ---
cat > bug.d << CODE
import core.simd;

alias vec = __vector(double[2]);

void storeUnaligned(ref vec a, in ref vec val)
{
    __simd_sto(XMM.STOUPD, a, val);
}

void test(ref vec a, in vec b)
{
    __simd_sto(XMM.STOUPD, a, b * b); // works
    immutable tmp = b * b;
    __simd_sto(XMM.STOUPD, a, tmp); // dips stack
    storeUnaligned(a, tmp); // dips stack
}
CODE
dmd -c -O -release -inline bug
----
        mulpd   xmm1, xmm3                              ; 000C _ 66: 0F 59. CB
        movapd  xmmword ptr [rsp], xmm1                 ; 0010 _ 66: 0F 29. 0C
24
        movdqa  xmm2, xmmword ptr [rsp]                 ; 0015 _ 66: 0F 6F. 14
24
        movupd  xmmword ptr [rdi], xmm2                 ; 001A _ 66: 0F 11. 17
----
Turns out that dmd's codegen is almost prohibitively inefficient, the intermediate value unnecessarily get pushed to the stack.

--

https://issues.dlang.org/show_bug.cgi?id=10226

--- Comment #12 from Martin Nowak <code@dawg.eu> ---
The cause for the example in comment 11 seems to be the void16 parameters of __simd_sto which breaks the

  (tyfloating(em->Ety) != 0) == (tyfloating(e->Ety) != 0)

condition in localizer.

https://github.com/dlang/dmd/blob/0089ae06db7c7b4bebe4d11bfcf02eab69936d81/src/ddmd/backend/glocal.c#L344


The assignment of value

{
(__vector(float[4])* p = &a[0];) , ((__vector(float[4]) value = b - c;));
__simd_sto(cast(XMM)3857, cast(__vector(void[16]))*p,
cast(__vector(void[16]))value);
}

is optimized to this el

el:0x9003b0 cnt=0 cs=0 *  TYsigned char[16] 0x900350
 el:0x900350 cnt=0 cs=0 var  TY*  __a_8
el:0x8ffc94 cnt=0 cs=0 -  TYsigned char[16] 0x8ffbd4 0x8ffc34
 el:0x8ffbd4 cnt=0 cs=0 var  TYfloat[4]  b
 el:0x8ffc34 cnt=0 cs=0 var  TYfloat[4]  c

which assigns a float[4] vector to a byte[16] vector.

--

https://issues.dlang.org/show_bug.cgi?id=10226

--- Comment #13 from github-bugzilla@puremagic.com ---
Commits pushed to master at https://github.com/dlang/dmd

https://github.com/dlang/dmd/commit/194fa0a92aec4a2b4fa5a6e28feeedca6daceb76 partly resolve Issue 10226 - inefficient core.simd codegen

- fix localization of float vector expressions passed to
  void16 parameters of simd intrinsics

https://github.com/dlang/dmd/commit/0d4eb1f06b938599293249eaecc2130fa6c89c81 Merge pull request #6626 from MartinNowak/issue10226

partly resolve Issue 10226 - inefficient core.simd codegen

--

https://issues.dlang.org/show_bug.cgi?id=10226

--- Comment #14 from github-bugzilla@puremagic.com ---
Commits pushed to newCTFE at https://github.com/dlang/dmd

https://github.com/dlang/dmd/commit/194fa0a92aec4a2b4fa5a6e28feeedca6daceb76 partly resolve Issue 10226 - inefficient core.simd codegen

https://github.com/dlang/dmd/commit/0d4eb1f06b938599293249eaecc2130fa6c89c81 Merge pull request #6626 from MartinNowak/issue10226

--

https://issues.dlang.org/show_bug.cgi?id=10226

Iain Buclaw <ibuclaw@gdcproject.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P2                          |P4

--