www.digitalmars.com         C & C++   DMDScript  

digitalmars.D.learn - Using inline assembler

reply Etienne <etcimon gmail.com> writes:
I'm a bit new to the inline assembler, I'm trying to use the `movdqu` 
operation to move a 128 bit double quadword from a pointer location into 
another location like this:

align(16) union __m128i { ubyte[16] data };

void store(__m128i* src, __m128i* dst) {
	asm { movdqu [dst], src; }
}


The compiler complains about a "bad type/size of operands 'movdqu'", but 

register? Is there something I'm missing here?
Oct 09 2014
next sibling parent reply "anonymous" <anonymous example.com> writes:
On Thursday, 9 October 2014 at 12:37:20 UTC, Etienne wrote:

 I'm a bit new to the inline assembler, I'm trying to use the 
 `movdqu` operation to move a 128 bit double quadword from a 
 pointer location into another location like this:

 align(16) union __m128i { ubyte[16] data };

 void store(__m128i* src, __m128i* dst) {
 	asm { movdqu [dst], src; }
 }


 The compiler complains about a "bad type/size of operands 
 'movdqu'", but these two data segments are 16 byte align so 

 missing here?


I know virtually nothing about SSE, but you can't move directly
from memory to memory, can you? You need go through a register,
no?

This compiles:

align(16) union __m128i { ubyte[16] data; } /* note the position
of the semicolon */

void store(__m128i* src, __m128i* dst) {
      asm
      {
          movdqu XMM0, [src]; /* note: [src] */
          movdqu [dst], XMM0;
      }
}
Oct 09 2014
parent reply Etienne <etcimon gmail.com> writes:
On 2014-10-09 8:54 AM, anonymous wrote:

 This compiles:

 align(16) union __m128i { ubyte[16] data; } /* note the position
 of the semicolon */

 void store(__m128i* src, __m128i* dst) {
       asm
       {
           movdqu XMM0, [src]; /* note: [src] */
           movdqu [dst], XMM0;
       }
 }


Yes, this does compile, but the value from src never ends up stored in dst.

void main() {
	__m128i src;
	src.data[0] = 255;
	__m128i dst;
	writeln(src.data); // shows 255 at offset 0
	store(&src, &dst);
	writeln(dst.data); // remains set as the initial array
}

http://x86.renejeschke.de/html/file_module_x86_id_184.html

Is this how it's meant to be used?
Oct 09 2014
parent reply "anonymous" <anonymous example.com> writes:
On Thursday, 9 October 2014 at 13:29:27 UTC, Etienne wrote:

 On 2014-10-09 8:54 AM, anonymous wrote:

 This compiles:

 align(16) union __m128i { ubyte[16] data; } /* note the 
 position
 of the semicolon */

 void store(__m128i* src, __m128i* dst) {
      asm
      {
          movdqu XMM0, [src]; /* note: [src] */
          movdqu [dst], XMM0;
      }
 }


 Yes, this does compile, but the value from src never ends up 
 stored in dst.

 void main() {
 	__m128i src;
 	src.data[0] = 255;
 	__m128i dst;
 	writeln(src.data); // shows 255 at offset 0
 	store(&src, &dst);
 	writeln(dst.data); // remains set as the initial array
 }

 http://x86.renejeschke.de/html/file_module_x86_id_184.html

 Is this how it's meant to be used?


I'm out of my knowledge zone here, but it seems to work when you
move the pointers to registers first:

void store(__m128i* src, __m128i* dst) {
      asm
      {
          mov RAX, src;
          mov RBX, dst;
          movdqu XMM0, [RAX];
          movdqu [RBX], XMM0;
      }
}
Oct 09 2014
parent Etienne <etcimon gmail.com> writes:
On 2014-10-09 9:46 AM, anonymous wrote:

 I'm out of my knowledge zone here, but it seems to work when you
 move the pointers to registers first:

 void store(__m128i* src, __m128i* dst) {
       asm
       {
           mov RAX, src;
           mov RBX, dst;
           movdqu XMM0, [RAX];
           movdqu [RBX], XMM0;
       }
 }


Absolutely incredible! My first useful working assembler code. You save 
the day. Now I can probably write a whole SIMD library ;)
Oct 09 2014
prev sibling parent Etienne <etcimon gmail.com> writes:
Maybe someone can help with the more specific problem. I'm translating a 
crypto engine here:

https://github.com/etcimon/botan/blob/master/source/botan/block/aes_ni/aes_ni.d

But I need this to work on DMD, LDC and GDC. I decided to write the 
assembler code directly for the functions in this module:

https://github.com/etcimon/botan/blob/master/source/botan/utils/simd/xmmintrin.d

If there's anything someone can tell me about this, I'd be thankful. I'm 
very experienced in every aspect of programming, but still at my first 
baby steps in assembler.
Oct 09 2014