-fuaddsub-overflow-match-all, -fif-conversion-gimple

Description

Arm-related instruction combination and optimization: Low-bit multiplication algorithms are identified and output the result using efficient high-bit multiplication instructions.

Usage

Use the -fuaddsub-overflow-match-all and -fif-conversion-gimple options to enable optimization.

Note:

This optimization requires the -O2 or higher optimization level and must be used together with the -ftree-fold-phiopt option.
This optimization is similar to -fmerge-mull, but it can be used in more comprehensive scenarios, such as 8-digit by 8-digit multiplication and 16-digit by 16-digit multiplication.

Result

The test case is as follows:

#include <stdint.h>

typedef unsigned __int128 uint128_t;

uint128_t mul128_perm (uint64_t a, uint64_t b)
{
    uint64_t a_lo = a & 0xFFFFFFFF;
    uint64_t b_lo = b & 0xFFFFFFFF;
    uint64_t a_hi = a >> 32;
    uint64_t b_hi = b >> 32;
    uint64_t lolo = a_lo * b_lo;
    uint64_t lohi = a_lo * b_hi;
    uint64_t hilo = a_hi * b_lo;
    uint64_t hihi = a_hi * b_hi;
    uint64_t middle = hilo + lohi;
    uint64_t middle_hi = middle >> 32;
    uint64_t middle_lo = middle << 32;
    uint64_t res_lo = lolo + middle_lo;
    uint64_t res_hi = hihi + middle_hi;
    res_hi = res_lo < middle_lo ? res_hi + 1 : res_hi;
    res_hi = middle < hilo ? res_hi + 0x100000000 : res_hi;
    uint128_t res = ((uint128_t) res_hi) << 64;
    res += res_lo;
    return res;
}

Test command:

gcc -O2 -ftree-fold-phiopt -fif-conversion-gimple -fuaddsub-overflow-match-all -S test.c -o test.s

Figure 1 Option disabled

Figure 2 Option enabled

After this option is enabled, the generated assembly code uses the umulh and mul instructions to optimize the 64-bit multiplication calculation.

Parent topic: Static Compilation Optimization