Test case for adcx and adox

问题

I'm testing Intel ADX add with carry and add with overflow to pipeline adds on large integers. I'd like to see what expected code generation should look like. From _addcarry_u64 and _addcarryx_u64 with MSVC and ICC, I thought this would be a suitable test case:

#include <stdint.h>
#include <x86intrin.h>
#include "immintrin.h"

int main(int argc, char* argv[])
{
    #define MAX_ARRAY 100
    uint8_t c1 = 0, c2 = 0;
    uint64_t a[MAX_ARRAY]={0}, b[MAX_ARRAY]={0}, res[MAX_ARRAY];
    for(unsigned int i=0; i< MAX_ARRAY; i++){ 
        c1 = _addcarryx_u64(c1, res[i], a[i], (unsigned long long int*)&res[i]);
        c2 = _addcarryx_u64(c2, res[i], b[i], (unsigned long long int*)&res[i]);
    }
    return 0;
}

When I examine the generated code from GCC 6.1 using -O3 and -madx, it reveals serialized addc. -O1 and -O2 produces similar results:

main:
        subq    $688, %rsp
        xorl    %edi, %edi
        xorl    %esi, %esi
        leaq    -120(%rsp), %rdx
        xorl    %ecx, %ecx
        leaq    680(%rsp), %r8
.L2:
        movq    (%rdx), %rax
        addb    $-1, %sil
        adcq    %rcx, %rax
        setc    %sil
        addb    $-1, %dil
        adcq    %rcx, %rax
        setc    %dil
        movq    %rax, (%rdx)
        addq    $8, %rdx
        cmpq    %r8, %rdx
        jne     .L2
        xorl    %eax, %eax
        addq    $688, %rsp
        ret

So I'm guessing the test case is not quite hitting the mark, or I am doing something wrong, or I am using something incorrectly, ...

If I am parsing Intel's docs on _addcarryx_u64 correctly, I believe the C code should generate the pipeline. So I'm guessing I am doing something wrong:

Description

Add unsigned 64-bit integers a and b with unsigned 8-bit carry-in c_in (carry or overflow flag), and store the unsigned 64-bit result in out, and the carry-out in dst (carry or overflow flag).

How can I generate the pipeline'd add with carry/add with overflow (adcx/adox)?

I've actually got a 5th generation Core i7 ready for testing (notice the adx cpu flag):

$ cat /proc/cpuinfo | grep adx
flags : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush
dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc
arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc aperfmperf eagerfpu pni
pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 fma cx16 xtpr pdcm pcid sse4_1
sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm
3dnowprefetch ida arat epb pln pts dtherm tpr_shadow vnmi flexpriority ept vpid fsgsbase
tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm rdseed adx smap xsaveopt
...

回答1:

This does look like a good test-case. It assembles to correct working code, right? It's useful for a compiler to support the intrinsic in that sense, even if it doesn't yet support making optimal code. It lets people start using the intrinsic. This is necessary for compatibility.

Next year or whenever the compiler's backend support for adcx/adox is done, the same code will compile to faster binaries with no source modification.

I assume that's what's going on for gcc.

clang 3.8.1's implementation is more literal, but it ends up doing a terrible job: flag-saving with sahf and push/pop of eax. See it on Godbolt.

I think there's even a bug in the asm source output, since mov eax, ch won't assemble. (Unlike gcc, clang/LLVM uses a built-in assembler and doesn't actually go through a text representation of asm on the way from LLVM IR to machine code). The disassembly of the machine code shows mov eax,ebp there. I think that's also a bug, because bpl (or the rest of the register) doesn't have a useful value at that point. Probably it wanted mov al, ch or movzx eax, ch.

回答2:

When GCC will be fixed to generate much better inlined code for add_carryx_... , be careful with your code, because the loop variant contains a comparison (modifies the C and O flags similarly to sub instruction) and an increment (modifies the C and O flags like an add instruction).

  for(unsigned int i=0; i< MAX_ARRAY; i++){ 
        c1 = _addcarryx_u64(c1, res[i], a[i], (unsigned long long int*)&res[i]);
        c2 = _addcarryx_u64(c2, res[i], b[i], (unsigned long long int*)&res[i]);
    }

For that reason, c1 and c2 in your code will always be pitifuly handled (saved and restored in temp registers at each loop iteration). And the resulting code generated by gcc will still look like the assembly you provided, for good reasons.

From a run-time point of view, res[i] is an immediate dependency between the 2 add_carryx instructions, the 2 instructions are not really independent and will not benefit from a possible architectural parallelism in the processor.

I understand the code is only an example, but maybe it will not be the best example to use when gcc will be modified.

The addition of 3 numbers in large integer arithmetic is a tough problem; vectorization helps, and then you better use addcarryx to handle the loop variants in parallel (increment and comparison+branch on the same variable, yet another tough problem).

来源：https://stackoverflow.com/questions/39320944/test-case-for-adcx-and-adox

标签

gcc

intrinsics

adx