I need to access the x86_64 RDTSCP assembly instruction from D.

I found this for C++:
http://stackoverflow.com/questions/14783782/which-inline-assembly-code-is-correct-for-rdtscp

Does anyone here know how (if?) I can do this from D?

On 30/08/2016 2:04 PM, kookman wrote:
> I need to access the x86_64 RDTSCP assembly instruction from D.
>
> I found this for C++:
> http://stackoverflow.com/questions/14783782/which-inline-assembly-code-is-correct-for-rdtscp
>
>
> Does anyone here know how (if?) I can do this from D?

It appears dmd's inline assembler does not support it.

So a workaround:

asm {
	// rdtscp
	db 0x0F;
	db 0x01;
	db 0xF9;
}

http://dlang.org/spec/iasm.html

On Tuesday, 30 August 2016 at 02:04:55 UTC, kookman wrote:
> I need to access the x86_64 RDTSCP assembly instruction from D.
>
> I found this for C++:
> http://stackoverflow.com/questions/14783782/which-inline-assembly-code-is-correct-for-rdtscp
>
> Does anyone here know how (if?) I can do this from D?

ALternatively to Rikki K's solution, you can do this to mimic the rdtscp behavior:

asm
{
  cpuid;
  rdtsc;
  // store time in locals
}
// bench
{
  rdtsc;
  // store time in locals
}
// compute delta


explanations here:
- http://www.intel.com/content/dam/www/public/us/en/documents/white-papers/ia-32-ia-64-benchmark-code-execution-paper.pdf
- http://stackoverflow.com/a/14214220

But according to the first link, this solution, while better than rdtsc alone, is not as good as rdtscp.

On Tuesday, 30 August 2016 at 09:04:41 UTC, Basile B. wrote:
> ALternatively to Rikki K's solution, you can do this to mimic the rdtscp behavior:
>
> asm
> {
>   cpuid;
>   rdtsc;
>   // store time in locals
> }
> // bench
> {
>   rdtsc;
>   // store time in locals
> }
> // compute delta
>
>
> explanations here:
> - http://www.intel.com/content/dam/www/public/us/en/documents/white-papers/ia-32-ia-64-benchmark-code-execution-paper.pdf
> - http://stackoverflow.com/a/14214220
>
> But according to the first link, this solution, while better than rdtsc alone, is not as good as rdtscp.

Indeed, I want to use rdtscp to get access to the core ID it makes available (to at least know when tsc is from different cores/packages). I guess in the meantime I can use Rikki's solution.

Sorry I'm a DMD asm newbie: DO i need to worry about which registers I clobber here? I couldn't find any info about this on the Inline Assembly doco page on dlang.org.

On 31/08/2016 7:34 PM, kookman wrote:
> On Tuesday, 30 August 2016 at 09:04:41 UTC, Basile B. wrote:
>> ALternatively to Rikki K's solution, you can do this to mimic the
>> rdtscp behavior:
>>
>> asm
>> {
>>   cpuid;
>>   rdtsc;
>>   // store time in locals
>> }
>> // bench
>> {
>>   rdtsc;
>>   // store time in locals
>> }
>> // compute delta
>>
>>
>> explanations here:
>> -
>> http://www.intel.com/content/dam/www/public/us/en/documents/white-papers/ia-32-ia-64-benchmark-code-execution-paper.pdf
>>
>> - http://stackoverflow.com/a/14214220
>>
>> But according to the first link, this solution, while better than
>> rdtsc alone, is not as good as rdtscp.
>
> Indeed, I want to use rdtscp to get access to the core ID it makes
> available (to at least know when tsc is from different cores/packages).
> I guess in the meantime I can use Rikki's solution.
>
> Sorry I'm a DMD asm newbie: DO i need to worry about which registers I
> clobber here? I couldn't find any info about this on the Inline Assembly
> doco page on dlang.org.

http://dlang.org/spec/abi.html#register_conventions

On Wednesday, 31 August 2016 at 07:36:16 UTC, rikki cattermole wrote:

> http://dlang.org/spec/abi.html#register_conventions

That link talks about for functions defined extern(C) and extern(D), and gives specific info for win32.

I'm using linux x86_64, does that mean I can assume standard x86_64 ABI? ie even if the function is not extern?

On 31/08/2016 7:49 PM, kookman wrote:
> On Wednesday, 31 August 2016 at 07:36:16 UTC, rikki cattermole wrote:
>
>> http://dlang.org/spec/abi.html#register_conventions
>
> That link talks about for functions defined extern(C) and extern(D), and
> gives specific info for win32.
>
> I'm using linux x86_64, does that mean I can assume standard x86_64 ABI?
> ie even if the function is not extern?

That link describes extern(D) for x86(_64) for all platforms. Keep in mind if a function body does not have an extern(X) supplied, the default is extern(D).

C is listed there as we have first class C interaction via extern(C).

The listing of DLL's on that page really should be changed to dynamic libraries. Since that covers linuxes.

On Wednesday, 31 August 2016 at 07:34:14 UTC, kookman wrote:
> On Tuesday, 30 August 2016 at 09:04:41 UTC, Basile B. wrote:
> Indeed, I want to use rdtscp to get access to the core ID it makes available (to at least know when tsc is from different cores/packages). I guess in the meantime I can use Rikki's solution.
>
> Sorry [...]

Never mind, I know Rikki's solution is better but since I've read a bit of doc and found this alternative I've just posted the alternative.

By the way maybe someone could post an ER in bugzilla to get RDTSCP available in iasm w/o using the byte code trick.

On Wednesday, 31 August 2016 at 08:23:57 UTC, Basile B. wrote:

> By the way maybe someone could post an ER in bugzilla to get RDTSCP available in iasm w/o using the byte code trick.

Someone beat me to it, but see here:
https://issues.dlang.org/show_bug.cgi?id=16449