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
Comment #1 by code — 2013-06-01T03:13:40Z
Small correction, ideal assembly should look like this:
1 mov rax,qword ptr [rbp-68h]
2 movaps xmm0,xmmword ptr [rax]
5 addps xmm0,xmm0
9 movaps xmmword ptr [rax],xmm0
Instruction 5 should use a xmm register as well and not add from memory.
Comment #2 by turkeyman — 2013-06-03T03:43:34Z
(In reply to comment #1)
> Small correction, ideal assembly should look like this:
>
> 1 mov rax,qword ptr [rbp-68h]
> 2 movaps xmm0,xmmword ptr [rax]
> 5 addps xmm0,xmm0
> 9 movaps xmmword ptr [rax],xmm0
>
> Instruction 5 should use a xmm register as well and not add from memory.
Interesting. I haven't scrutinised DMD's codegen as much as GDC yet.
I've been working on the std.simd API mostly against GDC, and once I'm happy with that, I'll be logging codegen bugs in DMD accordingly.
What do you get if you do:
float4 result = [1,2,3,4];
result = __simd(XMM.ADDPS, result, result);
writefln("%s", result.array);
Why do you need to issue the loads and stores explicitly?
Comment #3 by code — 2013-06-03T05:37:15Z
(In reply to comment #2)
> (In reply to comment #1)
> > Small correction, ideal assembly should look like this:
> >
> > 1 mov rax,qword ptr [rbp-68h]
> > 2 movaps xmm0,xmmword ptr [rax]
> > 5 addps xmm0,xmm0
> > 9 movaps xmmword ptr [rax],xmm0
> >
> > Instruction 5 should use a xmm register as well and not add from memory.
>
> Interesting. I haven't scrutinised DMD's codegen as much as GDC yet.
> I've been working on the std.simd API mostly against GDC, and once I'm happy
> with that, I'll be logging codegen bugs in DMD accordingly.
>
> What do you get if you do:
>
> float4 result = [1,2,3,4];
> result = __simd(XMM.ADDPS, result, result);
> writefln("%s", result.array);
>
> Why do you need to issue the loads and stores explicitly?
On modern processors unaligned loads come at almost no penalty, and it is a lot easier to use unaligned loads within a highly optimized functions compared to making all the data in the whole project 16 byte aligned and use aligned loads.
Comment #4 by turkeyman — 2013-06-03T08:42:51Z
(In reply to comment #3)
> (In reply to comment #2)
> > (In reply to comment #1)
> > > Small correction, ideal assembly should look like this:
> > >
> > > 1 mov rax,qword ptr [rbp-68h]
> > > 2 movaps xmm0,xmmword ptr [rax]
> > > 5 addps xmm0,xmm0
> > > 9 movaps xmmword ptr [rax],xmm0
> > >
> > > Instruction 5 should use a xmm register as well and not add from memory.
> >
> > Interesting. I haven't scrutinised DMD's codegen as much as GDC yet.
> > I've been working on the std.simd API mostly against GDC, and once I'm happy
> > with that, I'll be logging codegen bugs in DMD accordingly.
> >
> > What do you get if you do:
> >
> > float4 result = [1,2,3,4];
> > result = __simd(XMM.ADDPS, result, result);
> > writefln("%s", result.array);
> >
> > Why do you need to issue the loads and stores explicitly?
>
> On modern processors unaligned loads come at almost no penalty, and it is a lot
> easier to use unaligned loads within a highly optimized functions compared to
> making all the data in the whole project 16 byte aligned and use aligned loads.
Umm, What?
This has nothing to do with alignment... where did that come from?
Your issue is the codegen right? You expect the codegen to be what you pasted below, and I agree. I'm suggesting you remove the explicit load/store and let the compiler generate them. What do you get using the code I suggest?
Comment #5 by bearophile_hugs — 2013-06-03T09:46:24Z
(In reply to comment #3)
> On modern processors unaligned loads come at almost no penalty, and it is a lot
> easier to use unaligned loads within a highly optimized functions compared to
> making all the data in the whole project 16 byte aligned and use aligned loads.
If this is true then is it worth reducing the smallest allocation size of the D garbage collector from 16 bytes to 8 bytes and reduce the alignment to 8 bytes? This saves some memory when you allocate many small objects, like small tree nodes.
Comment #6 by code — 2013-06-03T10:07:12Z
Reducing alignment on all allocations to 8 byte is not a good idea. When I said recent processors I meant i7 ivy bridge. On older processors this still has a significant performance impact. Initially the manual load and store operation were unaligned, but as of this bug
http://d.puremagic.com/issues/show_bug.cgi?id=10225
I had to replace them with aligned loads and stores for the time beeing. But this doesn't help that when the unaligned store gets fixed, the generated code will be equally bad as when using aligned operations.
Comment #7 by turkeyman — 2013-06-03T16:21:54Z
(In reply to comment #6)
> Reducing alignment on all allocations to 8 byte is not a good idea. When I said
> recent processors I meant i7 ivy bridge. On older processors this still has a
> significant performance impact. Initially the manual load and store operation
> were unaligned, but as of this bug
> http://d.puremagic.com/issues/show_bug.cgi?id=10225
> I had to replace them with aligned loads and stores for the time beeing. But
> this doesn't help that when the unaligned store gets fixed, the generated code
> will be equally bad as when using aligned operations.
Okay, I see. It occurred to me that maybe DMD promoted that movups in your other bug to a movaps because it was able to detect that the source was always aligned? (I have no hard reason to believe that, but just a thought. Walter?)
Yeah it looks kinda like the compiler is treating the mov*ps opcodes like a typical add or something, and the compiler is generating typical register allocation around the opcodes which performs explicit register allocation ;)
My guess is the compiler doesn't actually understand what the mov*ps (and friends) opcodes do, and therefore erroneously tries to helpfully configure a register state to execute them ;)
What flags are you building with? If it's building un-optimised or in some debug mode, the extra mov's could be to try and keep stack synchronisation.
Comment #8 by code — 2013-06-03T22:36:41Z
Building with debug symbols gives a ICE ;-)
I'm building with -release -O -inline
Comment #9 by bugzilla — 2016-04-25T07:00:47Z
(In reply to Benjamin Thaut from comment #8)
> Building with debug symbols gives a ICE ;-)
It works when I compile with -g.
Comment #10 by bugzilla — 2016-04-25T07:01:33Z
It isn't 'bad' codegen, it's just inefficient codegen. I reclassified this as an enhancement request.
Comment #11 by code — 2017-03-13T19:41:51Z
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.
Comment #12 by code — 2017-03-14T01:03:42Z
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.
Comment #13 by github-bugzilla — 2017-03-15T04:11:37Z