1STYLE REQUIREMENTS 2================== 3 41. Most code in this sub-directory is expected to be upstreamed into glibc so 5 the GNU Coding Standard and glibc specific conventions should be followed 6 to ease upstreaming. 7 82. ABI and symbols: the code should be written so it is suitable for inclusion 9 into a libc with minimal changes. This e.g. means that internal symbols 10 should be hidden and in the implementation reserved namespace according to 11 ISO C and POSIX rules. If possible the built shared libraries and static 12 library archives should be usable to override libc symbols at link time (or 13 at runtime via LD_PRELOAD). This requires the symbols to follow the glibc ABI 14 (other than symbol versioning), this cannot be done reliably for static 15 linking so this is a best effort requirement. 16 173. API: include headers should be suitable for benchmarking and testing code 18 and should not conflict with libc headers. 19 20 21CONTRIBUTION GUIDELINES FOR math SUB-DIRECTORY 22============================================== 23 241. Math functions have quality and performance requirements. 25 262. Quality: 27 - Worst-case ULP error should be small in the entire input domain (for most 28 common double precision scalar functions the target is < 0.66 ULP error, 29 and < 1 ULP for single precision, even performance optimized function 30 variant should not have > 5 ULP error if the goal is to be a drop in 31 replacement for a standard math function), this should be tested 32 statistically (or on all inputs if possible in reasonable amount of time). 33 The ulp tool is for this and runulp.sh should be updated for new functions. 34 35 - All standard rounding modes need to be supported but in non-default rounding 36 modes the quality requirement can be relaxed. (Non-nearest rounded 37 computation can be slow and inaccurate but has to be correct for conformance 38 reasons.) 39 40 - Special cases and error handling need to follow ISO C Annex F requirements, 41 POSIX requirements, IEEE 754-2008 requirements and Glibc requiremnts: 42 https://www.gnu.org/software/libc/manual/html_mono/libc.html#Errors-in-Math-Functions 43 this should be tested by direct tests (glibc test system may be used for it). 44 45 - Error handling code should be decoupled from the approximation code as much 46 as possible. (There are helper functions, these take care of errno as well 47 as exception raising.) 48 49 - Vector math code does not need to work in non-nearest rounding mode and error 50 handling side effects need not happen (fenv exceptions and errno), but the 51 result should be correct (within quality requirements, which are lower for 52 vector code than for scalar code). 53 54 - Error bounds of the approximation should be clearly documented. 55 56 - The code should build and pass tests on arm, aarch64 and x86_64 GNU linux 57 systems. (Routines and features can be disabled on specific targets, but 58 the build must complete). On aarch64, both little- and big-endian targets 59 are supported as well as valid combinations of architecture extensions. 60 The configurations that should be tested depend on the contribution. 61 623. Performance: 63 - Common math code should be benchmarked on modern aarch64 microarchitectures 64 over typical inputs. 65 66 - Performance improvements should be documented (relative numbers can be 67 published; it is enough to use the mathbench microbenchmark tool which should 68 be updated for new functions). 69 70 - Attention should be paid to the compilation flags: for aarch64 fma 71 contraction should be on and math errno turned off so some builtins can be 72 inlined. 73 74 - The code should be reasonably performant on x86_64 too, e.g. some rounding 75 instructions and fma may not be available on x86_64, such builtins turn into 76 libc calls with slow code. Such slowdown is not acceptable, a faster fallback 77 should be present: glibc and bionic use the same code on all targets. (This 78 does not apply to vector math code). 79