consider the code fragment: uint64_t slow(uint64_t x) { return x / 1220703125u; } This can be replaced by: uint64_t fast(uint64_t x) { uint32_t a = ((x >> 32) * 1270091284u) >> 32; uint32_t b = ((x & 0xffffffffu) * 3777893186u) >> 32; return ((x >> 32) * 3777893186ull + a + b) >> 30; } The 'fast' code runs 50% faster than 'slow'. However, removing the redundant multiplies (see my earlier bug - fixed in 4.4 trunk) and tidying up storage, I can use the following assembler to run nearly 100% faster than 'slow': .p2align 4,,15 .globl _fast .def _fast; .scl 2; .type 32; .endef _fast: pushl %ebx movl $1270091284, %eax mull 12(%esp) movl $-517074110, %eax movl %edx, %ebx mull 8(%esp) movl $-517074110, %eax movl %edx, %ecx mull 12(%esp) addl %ebx, %ecx popl %ebx adcl $0, %edx addl %ecx, %eax adcl $0, %edx shrdl $30, %edx, %eax shrl $30, %edx ret NOTE: the 2 multipliers are derived using 96-bit arithmetic: d = 1220703125u -517074110 = 0xffe12e1342u = (((1u << 94) + d - 1) / d) >> 32 1270091284 = (((1u << 94) + d - 1) / d) & 0xffffffffu
I have tested my initial routine with more values and it turns out that the short-cut of omitting the least significant multiplication is wrong. I added the 4th multiply to generate the full 128-bit precision. I also took the multiplier used by gcc 4.2.3 on x86_64-linux-gnu (ubuntu) (4056481920730334085) and (28+64) bit shift to achieve the division by 1220703125u (largest 32-bit power of 5): uint64_t fast(uint64_t x) { static const uint64_t fact = 4056481920730334085ull; uint32_t c = ((x & 0xffffffffu) * (fact & 0xffffffffu)) >> 32; uint64_t a = ((x & 0xffffffffu) * (fact >> 32) + c); /* * Split the addition of the two middle 64-bit intermediate results * into 2 stages to avoid possible overflow. * (not actually an issue with this value of 'fact') */ uint32_t b = ((x >> 32) * (fact & 0xffffffffu) + (a & 0xffffffffu)) >> 32; return ((x >> 32) * (fact >> 32) + (a >> 32) + b) >> 28; } After hand coding the assembler on 32-bit x86, the extra multiply increased the run time from 15s to 17s - still much faster than 'slow' (64-bit divide API call).
I changed the summary as it is equally applicable to all 64-bit constant divides on 32-bit x86 that are already compiled as multiplies on 64-bit x64. It might be worth implementing as another API call, replacing the 64-bit divisor with a 64-bit multiplier and a shift count.
Created attachment 16369 [details] proposed 32-bit API calls for 64-bit constant divison I have attached a sample assembler code for a API calls for the 2 types of 64-bit division through multiplication: mul0shift - for multiplier with 64 or fewer significant bits mul1shift - for multiplier with 65 significant bits (implicit high bit)
Created attachment 16370 [details] proposed 32-bit API calls for 64-bit constant divison The original attachment did not handle shift greater than 31.
Must be a cost model issue because I can get /10u working but not /1220703125u (in GCC 11+).