Bicho de sete cabeças. It's not a seven-headed beast. It's not rocket science.Cassio Neri Fast conversion from floating-point numbers CppCon 2024 - Aurora (Independent Researcher)MeMe O Neri's equationMe https://godbolt.org/z/qh4zdzqG61 2 3 Integer to string #hundreds 1 / 10 / 10 1 2 3 % 10 % 10 % 10 #units 123 #tens 121 2 3 Integer to string #hundreds 1 / 10 and thus, . F u v v − u = 10 F > 2E The larger , the shorter . F n We set to be the largest integer such that , i.e., F 10 F ≤ 2E u vTejú Jaguá - Ⅰ n × 10 F ≅ m × 2E 2E There must be some inside

match.  The Standard C getcharreturns the character read as an int, not a char.  getcharuses the integer value EOFto represent end-of-file and I/O failures.  putcharhas a parameter of type int(not char) conversions…  Calling f(int)requires no conversion at all.  It’s an exact match.  Calling f(long int)requires converting intinto “long int”.  This is an integral conversion.  Calling f(char *)requires converting converting intinto “pointer to char”.  This is a pointer conversion. 19 Exact Matches Always Win  The compiler prefers “cheap” conversions.  Exact matches are the cheapest of all.  Thus, f(0)calls:

disallowed Compiler Flag Supported Since Description -Wconversion, -Wsign-conversion GCC 2.95.3, Clang 4.0.0 Enable implicit conversion warnings -fstack-protector-strong GCC 4.9.0 Clang 6.0.0 Enable run-time s GCC 3.0.0, Clang 7.0.0 Force retention of null pointer checks -fno-strict-overflow GCC 4.2.0 Integer overflow may occur -fno-strict-aliasing GCC 2.95.3, Clang 18.0.0 Do not assume strict aliasing 0 Perform trivial auto variable initialization -Wconversion, -Wsign-conversion GCC 2.95.3, Clang 4.0.0 Enable implicit conversion warnings -fstack-clash-protection GCC 8.0.0, Clang 11.0.0 Enable run-time

generates copy constructors and copy assignment operators are deprecated. • C language style type conversion is deprecated (ie using (convert_type)) before vari- ables, and static_cast, reinterpret_cast NULL as ((void*)0), then in the following code: char *ch = NULL; C++ without the void * implicit conversion has to define NULL as 0. This still creates a new problem. Defining NULL to 0 will cause the overloading defined implements type safety. First, it cannot be implicitly converted to an integer, nor can it be compared to integer numbers, and it is even less likely to compare enumerated values of different

unanticipated conversions: For example, an unanticipated wraparound of an unsigned integer loop variable or a narrowing conversion. • Timing errors: for example, delivering a result in 1.2ms to a device supposedly Concurrent task • Message queue • Coroutine • Hash table • Graphical shape • Complex number • Infinite integer Stroustrup - C++ safety -CppCon - October 2023 18 Some are still not standardimmutability • Constants unmentioned • Often the need to inter-operate with other languages is unmentioned • Often the cost of conversion is underestimated • This is natural • Human nature • Which other language? • Must interoperate

3. Cost 4. Disabling each tool 5. Using each tool outside of LLVM ContrastRun time cost Code conversion cost Performance constraints RealtimeSanitizerRealtimeSanitizer run-time cost Comparing and thread-local integer ...fly a rocket code here... realtime_depth--; }[[nonblocking]] ● Every [[nonblocking]] function has an additional increment and decrement of an integer ● May interfere nblocking]] ● Every [[nonblocking]] function has an additional increment and decrement of an integer ● May interfere with inlining optimizations ● A few additional function calls inserted between

developers to design scalable and portable parallelization. • Target-dependent width • Visible “conversion” between scalars and simd objects (loads & stores vs. gathers & scatters) • Conditional assignment Programming Models Outlook Summary More interoperability • Conversion to/from std::array, std::span, contiguous ranges? • Mask conversion to/from std::bitset. • Initializer list constructor for simd 3 mov DWORD PTR [rdi], eax • src/dst: array of integers • throughput: 0.5/1/1 cycles (Intel) • integer multiplications: 1 1 vmovdqu32 zmm0, ZMMWORD PTR [rdi] 2 vpmulld zmm0, zmm0, zmm0 3 vmovdqu32

calling convention “The x64 ABI uses a four-register fast-call calling convention by default. … Integer arguments are passed in registers RCX, RDX, R8, and R9.” “Any argument that doesn't fit in 8 bytes calling convention “The x64 ABI uses a four-register fast-call calling convention by default. … Integer arguments are passed in registers RCX, RDX, R8, and R9.” “Any argument that doesn't fit in 8 bytes ), 13 std::integral_constant{}}; 14 } 15 }; 71RGB to Grayscale conversion 72Color spaces 1 struct RGB {}; 2 3 auto channelCount(RGB) { 4 return std::integral_constant

0 码力 | 75 页 | 1.04 MB | 6 月前

3

world (not really)Hello, world Different implementations of a simple increment function (takes an integer as argument, increments it and returns) ● C++ ● Python ● C++ with pybind11 ● C++ with boost::python be int / float / std::vector, ... ○ Most libraries support all these common types’ conversion from PyObject* and back ○ Should be optimal to let the library do the cast for you ● Can be ○ Even more optimal if your code just populates the python data structure instead of doing a conversion ● What about larger objects? ○ Large vectors, matrices? ○ Python libraries aren’t written to

preconditions: • Example: unsigned integer addition • Example: vector::size() • A function has a narrow contract if it has preconditions: • Example: signed integer addition (precondition: a + b will presentation, this was slide 1 for Timur's video. Please forgive formatting glitches caused by conversion and merging.Copyright (c) Timur Doumler | @timur_audio | https://timur.audio 18 Copyright