[llvm-dev] [cfe-dev] Portable multiplication 64 x 64 -> 128 for int128 reimplementation (original) (raw)

Paweł Bylica via llvm-dev llvm-dev at lists.llvm.org
Sun Dec 30 13:45:49 PST 2018

Hi Arthur, Craig,

Thanks for you comments about GCC/Clang intrinsics. I never considered using them, but they might be better alternative to inline assembly. Is there a one for regular MUL?

Anyway, I want to go the opposite direction. If I can I relay on compiler's optimizations. If I want to use MULX in Clang I do it like that:

unsigned long mulx(unsigned long x, unsigned long y, unsigned long* hi) { auto p = (unsigned __int128){x} * y; *hi = static_cast(p >> 64); return static_cast(p); }

https://godbolt.org/z/PbgFb9

If compiled with -mbmi2 -mtune=generic it just uses MULX instruction.

mulx(unsigned long, unsigned long, unsigned long*): mov rcx, rdx mov rdx, rsi mulx rdx, rax, rdi mov qword ptr [rcx], rdx ret

What I want to do it move it further - rewrite the above mulx() helper without using __int128 type in a way that a compiler would recognize that it should use MUL/MULX instruction.

A possible implementation looks like

uint64_t mul_full_64_generic(uint64_t x, uint64_t y, uint64_t* hi) { uint64_t xl = x & 0xffffffff; uint64_t xh = x >> 32; uint64_t yl = y & 0xffffffff; uint64_t yh = y >> 32;

uint64_t t = xl * yl; uint64_t l = t & 0xffffffff; uint64_t h = t >> 32;

t = xh * yl; t += h; h = t >> 32;

t = xl * yh + (t & 0xffffffff); l |= t << 32; *hi = xh * yh + h + (t >> 32); return l; }

As expected, Clang is not able to match this pattern currently.

If we want to implement this optimization in Clang, there are some questions I have:

  1. Can we prove this pattern is equivalent of MUL 64x64 -> 128?
  2. What pass this optimization should be added to?
  3. Can this pattern be split into smaller ones? E.g. UMULH.

Paweł

On Sun, Dec 30, 2018 at 2:34 AM Craig Topper <craig.topper at gmail.com> wrote:

mulxu64 only exists when the target is x8664. That's still not very portable. I'm not opposed to removing the bmi2 check, but gcc also has the same check so it doesn't improve portability much.

~Craig

On Sat, Dec 29, 2018 at 4:44 PM Arthur O'Dwyer via llvm-dev <_ _llvm-dev at lists.llvm.org> wrote: Hi Pawel,

There is the mulxu64 intrinsic, but it currently requires the hardware flag "-mbmi2". https://github.com/Quuxplusone/WideIntProofOfConcept/blob/master/wider.h#L89-L99 On Clang 3.8.1 and earlier, the addcarryu64 and subborrowu64 intrinsics required the hardware flag -madx, even though they didn't use the hardware ADX/ADOX instructions. Modern GCC and Clang permit the use of these intrinsics (to generate ADC) even in the absence of -madx. I think it would be a very good idea for Clang to support mulxu64 (to generate MUL) even in the absence of -mbmi2. –Arthur

On Sat, Dec 29, 2018 at 6:03 PM Paweł Bylica via cfe-dev <_ _cfe-dev at lists.llvm.org> wrote: Hi, For some maybe dumb reasons I try to write a portable version of int128. What is very valuable for this implementation is access to MUL instruction on x86 which provides full 64 x 64 -> 128 bit multiplication. An equally useful on ARM would be UMULH instruction. _Well, the way you can access this on clang / GCC is to use int128 type or use inline assembly. MSVC provides an intrinsic for this instruction. This defeats the idea of portable int128 reimplementation and makes constexpr implementation of multiplication at least inconvenient. Maybe there is a hope for me in LLVM. Is there any pattern matcher that is producing MUL instruction of bigger type? If not, would it be good idea to teach LLVM about it? Bests, Paweł

cfe-dev mailing list cfe-dev at lists.llvm.org http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev

LLVM Developers mailing list llvm-dev at lists.llvm.org http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20181230/ab95ee15/attachment.html>

More information about the llvm-dev mailing list