3 ms·

If you want your (vector or scalar) integer instructions to be useful for fast strcmp/memcmp implementations, then you'll want to support unsigned big-endian in

by KMag 7y ago

If you want your (vector or scalar) integer instructions to be useful for fast strcmp/memcmp implementations, then you'll want to support unsigned big-endian integers. Also, if you want to mix integers and strings in compound b+ tree keys and use memcmp to compare keys, you'll want to store your integers big-endian.

tomxor 7y ago
Ah, but isn't that all part of the same quirk of human language, i.e that we compare strings from left most significant to right least significant?
- KMag 7y ago
  It's not some quirk. There are two common endiannesses, and the byte sort precedence of one of them matches the byte sort precedence we use for text. Since we represent characters as numbers, and for efficiency we have CPU instructions that treat aggregations of these numbers as larger numbers, we can get double-duty out of these instructions/circuits if the order they use is the same as the lexographic ordering of our text. I'm not saying big-endian is universally better. I'm just saying there are real-world advantages. Note there are more than two endiannesses, where the most common middle-endian variant is big-endian order of 16-bit words, but little-endian order within those 16-bit words. As for linguistic quirks, there's no obvious universal connection between writing order for text, the order digits are written, and the way numbers are pronounced, so I'm not sure what you're getting at. Arabic is written right-to-left, but they still write the most significant digit on the left. However, I've read that Arabic writers line-break numbers (when forced) by placing the most significant digits on the upper line and the least significant digits on the lower line, so it's not a clear-cut case of Arabic writing numbers in little-endian order right-to-left. I'm not sure about Arabic number pronunciation. I do know that German pronunciation is middle-endian: 256 is "two hundred six-and-fifty", despite the digit writing order being big-endian in German. I met a German guy who incorrectly swaps digits if you talk to him in English while he's doing mental math... forcing him to process English somehow also swaps digit orders in his head.
jart 7y ago
I don't believe you. That sounds more like bitreverse and morton interleaving. If you're serious that it's actually big endian and that it has a real use case (outside human social traditions and legacy infrastructure maintainence), then I'd love to see convincing details. As for strings, GCC/Clang are great at vectorizing normal byte-oriented code. It doesn't matter if it generates or you hand code VPCMPEQB vs. VPCMPEQQ either, since they appear to have the exact same latency. Modern compilers have also been moving in the direction of punishing folks who do things like `* (uint32_t * )p`, due to aliasing rules rather than endianness. Rob Pike had an amusing blog post on this subject: https://commandcenter.blogspot.com/2012/04/byte-order-fallacy.html https://commandcenter.blogspot.com/2012/04/byte-order-fallac...
- KMag 7y ago
  > I don't believe you. That sounds more like bitreverse and morton interleaving. If you're serious that it's actually big endian and that it has a real use case (outside human social traditions and legacy infrastructure maintainence), then I'd love to see convincing details. Back when I worked on Google's indexing system (2006-2010), one of the stages keyed documents by URL (with the host name in DNS order: com.google.www) followed by a big-endian timestamp. This put records for the same domain together, and put successive crawls for the same URL in chronological order. This improved compression ratios in our BigTable tablets, and finding the latest crawled version of <URL> was just a search for the largest key <= <URL><MAX_TIMESTAMP>. > It doesn't matter if it generates or you hand code VPCMPEQB vs. VPCMPEQQ either, since they appear to have the exact same latency. But with VPCMPEQB, you need to have 8 times as many conditional branches as CPCMPEQQ (in the naive case) in order to turn your vectorized comparison into a memcmp result. Note that VCMPEQ* don't modify any flags, so you can't just JE/JNE after your VCMPEQ*. Now, via some vector permutes, shifts, and bitwise-ors, you can cut down on the number of conditional branches. However, it's still more processing than if you can load / compare your vector registers in lexographical order.
  - jart 7y ago
    But Google doesn't use big endian CPUs in prod, so you've proven the point yourself. I agree that lexicographically arranging things can have an awesome impact on performance though. Still think that has zero to do with endianness. > But with VPCMPEQB, you need to have 8 times as many conditional branches as CPCMPEQQ in order to turn your vectorized comparison into a memcmp result Not at all. It doesn't require any branches, other than the loop itself. Here's the trick everyone uses, for the simple case of finding one character, e.g. NUL: 0: add $vectorlen,i pmovdqa mem,vector pcmpeqb query,vector pmovmskb vector,bitmask bsf bitmask,offset jz 0b add offset,i ret BSF gives you the byte offset of the match, and bam you're done.

by KMag 7y ago

tomxor 7y ago

Ah, but isn't that all part of the same quirk of human language, i.e that we compare strings from left most significant to right least significant?

KMag 7y ago
It's not some quirk. There are two common endiannesses, and the byte sort precedence of one of them matches the byte sort precedence we use for text. Since we represent characters as numbers, and for efficiency we have CPU instructions that treat aggregations of these numbers as larger numbers, we can get double-duty out of these instructions/circuits if the order they use is the same as the lexographic ordering of our text. I'm not saying big-endian is universally better. I'm just saying there are real-world advantages. Note there are more than two endiannesses, where the most common middle-endian variant is big-endian order of 16-bit words, but little-endian order within those 16-bit words. As for linguistic quirks, there's no obvious universal connection between writing order for text, the order digits are written, and the way numbers are pronounced, so I'm not sure what you're getting at. Arabic is written right-to-left, but they still write the most significant digit on the left. However, I've read that Arabic writers line-break numbers (when forced) by placing the most significant digits on the upper line and the least significant digits on the lower line, so it's not a clear-cut case of Arabic writing numbers in little-endian order right-to-left. I'm not sure about Arabic number pronunciation. I do know that German pronunciation is middle-endian: 256 is "two hundred six-and-fifty", despite the digit writing order being big-endian in German. I met a German guy who incorrectly swaps digits if you talk to him in English while he's doing mental math... forcing him to process English somehow also swaps digit orders in his head.

jart 7y ago

I don't believe you. That sounds more like bitreverse and morton interleaving. If you're serious that it's actually big endian and that it has a real use case (outside human social traditions and legacy infrastructure maintainence), then I'd love to see convincing details. As for strings, GCC/Clang are great at vectorizing normal byte-oriented code. It doesn't matter if it generates or you hand code VPCMPEQB vs. VPCMPEQQ either, since they appear to have the exact same latency. Modern compilers have also been moving in the direction of punishing folks who do things like `* (uint32_t * )p`, due to aliasing rules rather than endianness. Rob Pike had an amusing blog post on this subject: https://commandcenter.blogspot.com/2012/04/byte-order-fallacy.html https://commandcenter.blogspot.com/2012/04/byte-order-fallac...

KMag 7y ago
> I don't believe you. That sounds more like bitreverse and morton interleaving. If you're serious that it's actually big endian and that it has a real use case (outside human social traditions and legacy infrastructure maintainence), then I'd love to see convincing details. Back when I worked on Google's indexing system (2006-2010), one of the stages keyed documents by URL (with the host name in DNS order: com.google.www) followed by a big-endian timestamp. This put records for the same domain together, and put successive crawls for the same URL in chronological order. This improved compression ratios in our BigTable tablets, and finding the latest crawled version of <URL> was just a search for the largest key <= <URL><MAX_TIMESTAMP>. > It doesn't matter if it generates or you hand code VPCMPEQB vs. VPCMPEQQ either, since they appear to have the exact same latency. But with VPCMPEQB, you need to have 8 times as many conditional branches as CPCMPEQQ (in the naive case) in order to turn your vectorized comparison into a memcmp result. Note that VCMPEQ* don't modify any flags, so you can't just JE/JNE after your VCMPEQ*. Now, via some vector permutes, shifts, and bitwise-ors, you can cut down on the number of conditional branches. However, it's still more processing than if you can load / compare your vector registers in lexographical order.
- jart 7y ago
  But Google doesn't use big endian CPUs in prod, so you've proven the point yourself. I agree that lexicographically arranging things can have an awesome impact on performance though. Still think that has zero to do with endianness. > But with VPCMPEQB, you need to have 8 times as many conditional branches as CPCMPEQQ in order to turn your vectorized comparison into a memcmp result Not at all. It doesn't require any branches, other than the loop itself. Here's the trick everyone uses, for the simple case of finding one character, e.g. NUL: 0: add $vectorlen,i pmovdqa mem,vector pcmpeqb query,vector pmovmskb vector,bitmask bsf bitmask,offset jz 0b add offset,i ret BSF gives you the byte offset of the match, and bam you're done.