digitalmars.D.learn - Issues using the in-line assembler
- solidstate1991 (26/26) Apr 04 2018 I have this code:
- solidstate1991 (1/1) Apr 04 2018 I forgot to tell, that xy0 ac, and bd are local to the class.
- Basile B. (27/53) Apr 04 2018 The "this" seems to be in R11, so you have to apply the asm
- Basile B. (4/12) Apr 04 2018 Ah sorry, the params must be removed ((new Foo).foo().writeln;)...
- solidstate1991 (39/65) Apr 05 2018 It seems that the compiler lets it through if I change it like
- solidstate1991 (37/37) Apr 05 2018 Seems I found a better solution hidden in the docs:
I have this code:
asm nogc{
movq XMM0, xy;
paddd XMM0, sXY; // xy + sXY
movq XMM3, xy0;
psubd XMM0, XMM3; // xy + sXY - x0y0
movq XMM1, ac;
movq XMM2, bd;
pmuludq XMM1, XMM0; // (ac * (xy + sXY - x0y0))
psrlq XMM1, 16; // (ac * (xy + sXY - x0y0))>>16
pmuludq XMM2, XMM0; // (bd * (xy + sXY - x0y0))
psrlq XMM2, 16; // (bd * (xy + sXY - x0y0))>>16
paddq XMM1, XMM2; // (bd * (xy + sXY - x0y0))>>16 * (ac * (xy +
sXY - x0y0))>>16
punpckldq XMM3, XMM7;
paddq XMM1, XMM3; // (bd * (xy + sXY - x0y0))>>16 * (ac * (xy +
sXY - x0y0))>>16 + x0y0
movups XMM2, XMM1; // Convert 64 bit vectors into 32 bit ones
psrldq XMM2, 4;
por XMM2, XMM1;
movq result, XMM2;
}
I'm getting "bad type/size of operand 'movq'" error on xy0, ac,
and bd when I try to compile it. All of the values are the type
of int[2], xy is function parameter, sXY is created locally. How
can I fix it?
Apr 04 2018
solidstate1991 <laszloszeremi outlook.com> I forgot to tell, that xy0 ac, and bd are local to the class.
Apr 04 2018
On Wednesday, 4 April 2018 at 21:00:44 UTC, solidstate1991 wrote:
I have this code:
asm nogc{
movq XMM0, xy;
paddd XMM0, sXY; // xy + sXY
movq XMM3, xy0;
psubd XMM0, XMM3; // xy + sXY - x0y0
movq XMM1, ac;
movq XMM2, bd;
pmuludq XMM1, XMM0; // (ac * (xy + sXY - x0y0))
psrlq XMM1, 16; // (ac * (xy + sXY - x0y0))>>16
pmuludq XMM2, XMM0; // (bd * (xy + sXY - x0y0))
psrlq XMM2, 16; // (bd * (xy + sXY - x0y0))>>16
paddq XMM1, XMM2; // (bd * (xy + sXY - x0y0))>>16 * (ac * (xy
+ sXY - x0y0))>>16
punpckldq XMM3, XMM7;
paddq XMM1, XMM3; // (bd * (xy + sXY - x0y0))>>16 * (ac * (xy
+ sXY - x0y0))>>16 + x0y0
movups XMM2, XMM1; // Convert 64 bit vectors into 32 bit ones
psrldq XMM2, 4;
por XMM2, XMM1;
movq result, XMM2;
}
I'm getting "bad type/size of operand 'movq'" error on xy0, ac,
and bd when I try to compile it. All of the values are the type
of int[2], xy is function parameter, sXY is created locally.
How can I fix it?
The "this" seems to be in R11, so you have to apply the asm
syntax for accessing the members using
<Type>.offsetof.<member>[R11], example:
```
class Foo
{
double a = 123456;
extern(D) double foo()
{
asm
{
naked;
movq XMM0, Foo.a.offsetof[R11];
ret;
}
}
}
void main()
{
import std.stdio;
(new Foo).foo(0,0).writeln;
}
```
However i cant find any specification saying that R11 is "this".
With a free function just pass the instance as param and replace
R11 by the register where the instance is passed.
Apr 04 2018
On Thursday, 5 April 2018 at 04:48:02 UTC, Basile B. wrote:
On Wednesday, 4 April 2018 at 21:00:44 UTC, solidstate1991
wrote:
void main()
{
import std.stdio;
(new Foo).foo(0,0).writeln;
}
```
Ah sorry, the params must be removed ((new Foo).foo().writeln;)...
I was actually trying to play with params and extern linkage to
see R11 always work...
Apr 04 2018
On Thursday, 5 April 2018 at 04:48:02 UTC, Basile B. wrote:
The "this" seems to be in R11, so you have to apply the asm
syntax for accessing the members using
<Type>.offsetof.<member>[R11], example:
```
class Foo
{
double a = 123456;
extern(D) double foo()
{
asm
{
naked;
movq XMM0, Foo.a.offsetof[R11];
ret;
}
}
}
void main()
{
import std.stdio;
(new Foo).foo(0,0).writeln;
}
```
However i cant find any specification saying that R11 is "this".
With a free function just pass the instance as param and
replace R11 by the register where the instance is passed.
It seems that the compiler lets it through if I change it like
this:
asm nogc{
naked;
movd XMM1, dword ptr sX[EBP];
pslldq XMM1, 4;
movss XMM1, dword ptr sY[EBP];
movq XMM0, xy;
paddd XMM0, XMM1; // [x,y] + [sX,sY]
movq XMM3, qword ptr xy0[EBP];
psubd XMM0, XMM3; // ([x,y] + [sX,sY] - [x_0,y_0])
movq XMM1, qword ptr ac[EBP];
movq XMM2, qword ptr bd[EBP];
pmuludq XMM1, XMM0; // [A,0,C,0] * ([x,y] + [sX,sY] - [x_0,y_0])
psrlq XMM1, 16; // ([A,0,C,0] * ([x,y] + [sX,sY] -
[x_0,y_0]))>>16
movups XMM4, XMM0;
psrldq XMM4, 4;
pslldq XMM0, 4;
por XMM4, XMM0;
pmuludq XMM2, XMM4; // [0,B,0,D] * ([x,y] + [sX,sY] - [x_0,y_0])
psrlq XMM2, 16; // ([0,B,0,D] * ([x,y] + [sX,sY] -
[x_0,y_0]))>>16
paddq XMM1, XMM2; // ([A,B,C,D] * ([x,y] + [sX,sY] -
[x_0,y_0]))>>16
punpckldq XMM3, XMM7;
paddq XMM1, XMM3; // ([A,B,C,D] * ([x,y] + [sX,sY] -
[x_0,y_0]))>>16 + [x_0,y_0]
movups XMM0, XMM1; // Convert 64 bit vectors into 32 bit ones
psrldq XMM0, 4;
por XMM0, XMM1;
ret ;
}
I wonder if I can return an int[2] in XMM0. I can do some
modifications to either move the result to the stack first, or
add an import to core.simd (which needs to be refactored
completely) and instead make the return type int4 on SIMD enabled
CPUs.
Apr 05 2018
Seems I found a better solution hidden in the docs:
nogc protected int[2] transformFunc(int[2] xy){
version(X86){
asm nogc{
naked;
mov EBX, this;
movd XMM1, sX[EBX];
pslldq XMM1, 4;
movss XMM1, sY[EBX];
movq XMM0, xy;
paddd XMM0, XMM1; // [x,y] + [sX,sY]
movq XMM3, xy0[EBX];
psubd XMM0, XMM3; // ([x,y] + [sX,sY] - [x_0,y_0])
movq XMM1, ac[EBX];
movq XMM2, bd[EBX];
pmuludq XMM1, XMM0; // [A,0,C,0] * ([x,y] + [sX,sY] - [x_0,y_0])
psrlq XMM1, 16; // ([A,0,C,0] * ([x,y] + [sX,sY] -
[x_0,y_0]))>>16
movups XMM4, XMM0;
psrldq XMM4, 4;
pslldq XMM0, 4;
por XMM4, XMM0;
pmuludq XMM2, XMM4; // [0,B,0,D] * ([x,y] + [sX,sY] - [x_0,y_0])
psrlq XMM2, 16; // ([0,B,0,D] * ([x,y] + [sX,sY] -
[x_0,y_0]))>>16
paddq XMM1, XMM2; // ([A,B,C,D] * ([x,y] + [sX,sY] -
[x_0,y_0]))>>16
punpckldq XMM3, XMM7;
paddq XMM1, XMM3; // ([A,B,C,D] * ([x,y] + [sX,sY] -
[x_0,y_0]))>>16 + [x_0,y_0]
movups XMM0, XMM1; // Convert 64 bit vectors into 32 bit ones
psrldq XMM0, 4;
por XMM0, XMM1;
ret ;
}
}(...)
}
Apr 05 2018