Failures to use optimal math routine

[runer112]

Absolute value

All of them synthesize the operation (poorly)

• [ ] int8_t
• [ ] int16_t
• [ ] int24_t
• [ ] int32_t
• [ ] int48_t
• [ ] int64_t

Bit reverse

• [ ] uint48_t foo(uint48_t x) { return __builtin_bitreverse64(x) >> 16; } uses __llbitrev instead of __i48bitrev (but __builtin_bitreverse32(x >> 16) | (uint48_t)__builtin_bitreverse16(x) << 32 is fine?)

Byte swap

• [ ] int24_t foo(int24_t x) { return __builtin_bswap32(x) >> 8; } uses __lbswap instead of __ibswap (__ibswap doesn't exist yet, but probably should?)
• [ ] int48_t foo(int48_t x) { return __builtin_bswap64(x) >> 16; } uses __llbswap instead of __i48bswap? (can't confirm because of https://github.com/jacobly0/llvm-project/issues/44#issuecomment-2468738761)

Division

• [ ] int8_t foo(int8_t x, int8_t y) { return x / y; } uses __sdivs instead of __bdivs
• [ ] int16_t foo(int16_t x, int16_t y) { return x / y; } uses __idivs instead of __sdivs

Negation

• [ ] int48_t foo(int48_t x) { return -x; } synthesizes the operation (poorly) instead of using __i48abs

Remainder

• [ ] int8_t foo(int8_t x, int8_t y) { return x % y; } uses __srems instead of __brems
• [ ] int16_t foo(int16_t x, int16_t y) { return x % y; } uses __irems instead of __srems

Shift left

• [ ] int8_t foo(int8_t x, int8_t y) { return x << y; } uses __ishl instead of __bshl
• [ ] int16_t foo(int16_t x, int16_t y) { return x << y; } uses __ishl instead of __sshl

Shift right

• [ ] int8_t foo(int8_t x, int8_t y) { return x >> y; } uses __ishrs instead of __bshl
• [ ] uint8_t foo(uint8_t x, uint8_t y) { return x >> y; } uses __ishru instead of __bshru
• [ ] int16_t foo(int16_t x, int16_t y) { return x >> y; } uses __ishrs instead of __sshrs
• [ ] uint16_t foo(uint16_t x, uint16_t y) { return x >> y; } uses __ishru instead of __sshru