Author Topic: xlat is pretty fast  (Read 1420 times)

HSE

  • Member
  • *****
  • Posts: 2263
  • AMD 7-32 / i3 10-64
Re: xlat is pretty fast
« Reply #15 on: September 28, 2022, 01:28:13 AM »
Propose a better solution :thumbsup:

Tested 3 contexts for xlat operations:

Code: [Select]
function_under_glass11 macro
mov rax, rcx
xlat ; mov al,[ebx+al]
endm
function_under_glass12 macro
movzx eax, byte ptr[rbx+rcx] ; mov al,[ebx+al]
endm

function_under_glass21 macro
xlat ; mov al,[ebx+al]
endm
function_under_glass22 macro
movzx eax, byte ptr[rbx+rax] ; mov al,[ebx+al]
endm

function_under_glass31 macro
movzx eax, byte ptr [rdi]
xlat ; mov al,[ebx+al]
endm
function_under_glass32 macro
       movzx eax, byte ptr[rdi]
       movzx eax, byte ptr[rbx+rax] ; mov al,[ebx+al]
endm



measured_loop1 proc uses rbx loops:qword
    local i : qword
    mov rbx, offset somestring
    ForLp i, 0, loops, rcx
        function_under_glass12
    Next i
    ret
measured_loop1 endp

measured_loop2 proc uses rbx loops:qword
    local i : qword
    mov rbx, offset somestring
    ForLp i, 0, loops, rax
        function_under_glass22
    Next i
    ret
measured_loop2 endp

measured_loop3 proc uses rdi loops:qword
    local i : qword
    mov rdi, offset CfgTokenChars
    mov rbx, offset somestring
    ForLp i, 0, loops
        function_under_glass31
        inc rdi
    Next i
    ret
measured_loop3 endp

Always there is variations, and mean differences are less than half cycle.

Interesting, here xlat look faster in WoW64 than directly in hardware.

In WoW64:
Code: [Select]
Intel(R) Core(TM) i3-10100 CPU @ 3.60GHz (SSE4)

377     cycles for 100 * xlat
404     cycles for 100 * movzx eax, byte ptr[ebx+ecx]

377     cycles for 100 * xlat
406     cycles for 100 * movzx eax, byte ptr[ebx+ecx]

371     cycles for 100 * xlat
404     cycles for 100 * movzx eax, byte ptr[ebx+ecx]

377     cycles for 100 * xlat
405     cycles for 100 * movzx eax, byte ptr[ebx+ecx]

13      bytes for xlat
13      bytes for movzx eax, byte ptr[ebx+ecx]

72      = eax xlat
72      = eax movzx eax, byte ptr[ebx+ecx]

In UEFI:

ops           overhead             code 
11               41.46             5.3930
12               40.82             5.1589

21               34.0781          5.5607
22               41.0515          5.1583

31              -30.22              6.9508
32              -23.11              7.4392

Last test suggest a curvature, perhaps related to cache work ¿?
Equations in Assembly: SmplMath

hutch--

  • Administrator
  • Member
  • ******
  • Posts: 10031
  • Mnemonic Driven API Grinder
    • The MASM32 SDK
Re: xlat is pretty fast
« Reply #16 on: September 28, 2022, 02:08:40 AM »
xlat is one of the old instructions that still performs well, from memory direct register algos are faster but only by a small amount and xlat is convenient enough to use. Another factor is the hardware its running on where you may get variations between them.
hutch at movsd dot com
http://www.masm32.com    :biggrin:  :skrewy:

jj2007

  • Member
  • *****
  • Posts: 13336
  • Assembly is fun ;-)
    • MasmBasic
Re: xlat is pretty fast
« Reply #17 on: September 28, 2022, 02:20:37 AM »
Interesting, here xlat look faster in WoW64 than directly in hardware.

What makes you think that the 32-bit code is not running directly in hardware...?

HSE

  • Member
  • *****
  • Posts: 2263
  • AMD 7-32 / i3 10-64
Re: xlat is pretty fast
« Reply #18 on: September 28, 2022, 03:12:53 AM »
What makes you think that the 32-bit code is not running directly in hardware...?

In hardware you can't use 32 bit addresses.
Equations in Assembly: SmplMath

jj2007

  • Member
  • *****
  • Posts: 13336
  • Assembly is fun ;-)
    • MasmBasic
Re: xlat is pretty fast
« Reply #19 on: September 28, 2022, 04:24:08 AM »
What makes you think that the 32-bit code is not running directly in hardware...?

In hardware you can't use 32 bit addresses.

Sure you can, it's called "compatibility mode", and it's not an emulation. It's hardware.

HSE

  • Member
  • *****
  • Posts: 2263
  • AMD 7-32 / i3 10-64
Re: xlat is pretty fast
« Reply #20 on: September 28, 2022, 06:44:59 AM »
Sure you can, it's called "compatibility mode", and it's not an emulation. It's hardware.

In Windows you don't access memory directly, but through "Virtual addresses". Then everything is an emulation in some sense.

Proof is that, just by accident, I was forgotten to change your 32 bits addresses, and program jump nowhere  :biggrin: :biggrin:

I know kernels and drivers can access some 32 bit addresses (in BIOS I think).

Equations in Assembly: SmplMath

hutch--

  • Administrator
  • Member
  • ******
  • Posts: 10031
  • Mnemonic Driven API Grinder
    • The MASM32 SDK
Re: xlat is pretty fast
« Reply #21 on: September 28, 2022, 10:11:01 AM »
It has been the case since antiquity that only the OS has direct access at memory, it is via the OS that an instance handle in one app may be at the same virtual address as any other app but not at the same physical address.

In long mode, all addresses must be 64 bit (anything that does not squark /LARGEADDRESSAWARE errors). Compatibility mode gets you all of the extra complexity of 64 bit but is strangled at 2 gig so I wonder why anyone bothers. I think it had some to do with Microsoft porting some of their older 32 bit apps to 64 bit.
hutch at movsd dot com
http://www.masm32.com    :biggrin:  :skrewy:

HSE

  • Member
  • *****
  • Posts: 2263
  • AMD 7-32 / i3 10-64
Re: xlat is pretty fast
« Reply #22 on: September 28, 2022, 12:37:22 PM »
It has been the case since antiquity that only the OS has direct access at memory, it is via the OS that an instance handle in one app may be at the same virtual address as any other app but not at the same physical address.

 :thumbsup: I read something about 1958, or so.

In long mode, all addresses must be 64 bit (anything that does not squark /LARGEADDRESSAWARE errors). Compatibility mode gets you all of the extra complexity of 64 bit but is strangled at 2 gig so I wonder why anyone bothers.

Exactly. There is no compatibility mode in hardware. In 64 bits machines only there are 64 bits addresses, compatibility mode is mostly an OS' trick.

Debuggers make the inverse trick, and you can think that compatibility mode happen in hardware  :biggrin:

I think it had some to do with Microsoft porting some of their older 32 bit apps to 64 bit.

Legacy support involve all hardware and software industries. Is not just MS, apparently some opcodes are only maintained in processors design to facilitate legacy software support (and that complete the compatibility trick). 
 
Equations in Assembly: SmplMath

hutch--

  • Administrator
  • Member
  • ******
  • Posts: 10031
  • Mnemonic Driven API Grinder
    • The MASM32 SDK
Re: xlat is pretty fast
« Reply #23 on: September 28, 2022, 01:58:23 PM »
One thing I am a fan of is keeping the complete mnemonic set in later hardware. Whenever things like dropping some instructions gets done, they often whack out some of the useful ones. You cannot use PUSHAD and POPAD in 64 bit which is unfortunate as they can be useful in debugging an algorithm.
hutch at movsd dot com
http://www.masm32.com    :biggrin:  :skrewy:

NoCforMe

  • Member
  • ****
  • Posts: 866
Re: xlat is pretty fast
« Reply #24 on: September 28, 2022, 02:06:08 PM »
In Windows you don't access memory directly, but through "Virtual addresses". Then everything is an emulation in some sense.

Proof is that, just by accident, I was forgotten to change your 32 bits addresses, and program jump nowhere  :biggrin: :biggrin:

I have no idea what you mean by that second sentence. But I believe the first one is just plain wrong.

So you're saying that if I code MOV EAX, SomeMemoryAddress, that I'm not actually executing the X86 MOV instruction, moving all 32 bits of data (with all the MOD/R/M stuff)? I don't believe that.

Now of course when we do access memory like that, that memory is virtual, and it might turn out that it's in a page that's been paged out to virtual storage (disk). In that case, a page fault is generated and the OS swaps that page back into memory, so we can physically access that data. (In other words, MOV does exactly what the Intel or AMD manual tells us it does.)

Or am I wrong about this? This is my understanding of how things work.

If you're right, think how incredibly S-L-O-O-O-W the whole process would be. Can you say "emulator"?

jj2007

  • Member
  • *****
  • Posts: 13336
  • Assembly is fun ;-)
    • MasmBasic
Re: xlat is pretty fast
« Reply #25 on: September 28, 2022, 07:31:04 PM »
x86: mov eax, [ebx]
x64: mov rax, [rbx]

In both cases, ebx/rbx is a pointer to a virtual address. The MMU, a highly integrated part of the CPU, translates that address to a physical address.

Claiming that "all addresses are 64 bit" is misleading. In compatibility mode, the MMU certainly translates a 32-bit address to a physical one. Unfortunately, I have not been able to find a doc that clearly says whether that physical address is in the section below 4GB. or whether it's possible that mov eax, [ebx] talks to a physical address above that area. If that is the case, the MMU will still hand over a 32-bit address to the CPU in compatibility mode. Btw the name is misleading: "32-bit mode" would be simpler and more correct. The cpu has more than one mode, fullstop.

In any case, it's hardware, not an emulation. An emulator is quite a different animal. Steem, for example, translates...
Code: [Select]
               MOVE.W  MODE(A6),D7
               BTST    #0,D7
               BEQ.S   SCR_ENDE
               BSR.S   PPL
               BTST    #1,D7
               BEQ.S   SCR_ENDE
               BSR.S   PPL
               BSR.S   PPL
... into something that can be understood by an Intel or AMD CPU. Below an example of a BASIC program that firmly believes it's running on a 68000 CPU.

Yet another story is Wow64: Instead of running a complete parallel 32-bit version of Windows, all 32-bit processes that need to access the core Windows APIs go through a "gate" that translates their 32-bit handles etc to 64-bit equivalents and then call the native 64-bit DLLs. That process is so fast that you won't find any significant performance difference between 32- and 64-bit processes. In contrast, an emulator is typically a factor 10-20 slower than native code.

HSE

  • Member
  • *****
  • Posts: 2263
  • AMD 7-32 / i3 10-64
Re: xlat is pretty fast
« Reply #26 on: September 29, 2022, 12:13:30 AM »
Or am I wrong about this? This is my understanding of how things work.

I don't know. I think is close to what JJ say. OS store a base address somewhere, and perform an hardware function.


In both cases, ebx/rbx is a pointer to a virtual address. The MMU, a highly integrated part of the CPU, translates that address to a physical address.

I think that is correct if you are running a program inside an OS. I'm searching, from time to time, how to do that without an OS.

In compatibility mode,

Must not be confused CPU modes with OS modes. UEFI put CPU in 64 bits mode. Once CPU is running 64 bits, there is no comeback (is a little different with Legacy BIOS, but this machine run from UEFI because integrated graphic card requiere that :biggrin:).

I have not been able to find a doc that clearly says whether that physical address is in the section below 4GB

I think User Virtual Address Space in Windows is 2GB by default and System Virtual Address Space also 2GB, and in Linux is optional 1/3, 2/2 or 3/1. That for 32-bits. For 64 bits is 128TB ( not sure this last).

In any case, it's hardware, not an emulation.

Is mostly hardware, but not exactly your code  :thumbsup:.
Equations in Assembly: SmplMath

jj2007

  • Member
  • *****
  • Posts: 13336
  • Assembly is fun ;-)
    • MasmBasic
Re: xlat is pretty fast
« Reply #27 on: September 29, 2022, 12:40:02 AM »
I have not been able to find a doc that clearly says whether that physical address is in the section below 4GB

I think User Virtual Address Space in Windows is 2GB by default and System Virtual Address Space also 2GB, and in Linux is optional 1/3, 2/2 or 3/1. That for 32-bits. For 64 bits is 128TB ( not sure this last).

The question is actually what happens if 16GB RAM are installed: Can the MMU hand over a physical address above 4GB to a 32-bit process, obviously translated to a 32-bit pointer? My suspicion is that it can serve only the 0...4GB range, but I can't find evidence on the web.

hutch--

  • Administrator
  • Member
  • ******
  • Posts: 10031
  • Mnemonic Driven API Grinder
    • The MASM32 SDK
Re: xlat is pretty fast
« Reply #28 on: September 29, 2022, 12:59:33 AM »
Some of this stuff sounds like it comes out of Alice In Wonderland.

Look up the details of a 32 bit PE (portable executable) file and you will find the expression "Relative Virtual Address" which tells you that the ancient form of direct memory access (early MS-DOS) is not available in "Protected Mode".

Protected mode means the OS controls the entire memory space including the memory that you allocate and its why you get access violations when you try and write past the end of allocated memory.

EVERY APP (no exceptions) runs in its own memory space which is allocated and controlled by the OS and there is no leakage across applications due to this isolation. The technique to share data between apps is memory mapped files which the OS provides for that purpose.

It is both contained in the hardware and the OS to be able to run both 32 and 64 bit apps. They both can run in a 64 bit OS version because the OS supports it. You cannot run 16 bit apps natively as the OS no longer supports it.

There is an ugly hybrid where you can run some 32 bit code in a 64 bit PE file but it comes at a cost of a 2 gig memory limit so you get the worst of both worlds, 64 bit complexity and 32 bit memory limitations.

Go back to 16 bit Windows before true hardware multitasking and you had the joys of writing a perfect app, only to have it trashed by some piece of chyte that trashed something in the OS and crashed Windows. Protected mode and true memory isolation is a blessing.  :thumbsup:
hutch at movsd dot com
http://www.masm32.com    :biggrin:  :skrewy:

HSE

  • Member
  • *****
  • Posts: 2263
  • AMD 7-32 / i3 10-64
Re: xlat is pretty fast
« Reply #29 on: September 29, 2022, 01:06:52 AM »
 :thumbsup:
Equations in Assembly: SmplMath