x86-64 optimization: (x <= 0xFFFFF) in '-Os' mode can convert to ((x >> 20) == 0)

[Explorer09]

Test code

#include <stdbool.h>
#include <stdint.h>

bool func1(uint32_t x) {
    return x < 0x100000U;
}

bool func2(uint32_t x) {
    return ((x >> 20) == 0);
}

bool func3(uint32_t x) {
    return ((x / 0x100000U) == 0);
}

The three functions are equivalent. Specifically the threshold constant to be compared with is a power-of-two. If there is no shortage of temporary registers, then the right shift version (that is, func2) will likely make the smallest code.

My expected result is like this (assuming -Os optimization, not -O2)

        shrl    $20, %edi
        sete    %al
        retq

Instead I got this (x86-64 clang 20.1.0 with -Os option, tested in Compiler Explorer):

        cmpl    $1048576, %edi
        setb    %al
        retq

(Clang can recognize the three functions are equivalent. It just made suboptimal code among the possible choices.)

Note that in RISC-V targets of Clang, it does use the right shift to reduce code size

        srliw   a0, a0, 20
        seqz    a0, a0
        ret

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Kang-Che Sung (宋岡哲) (Explorer09)

Test code

#include <stdbool.h>
#include <stdint.h>

bool func1(uint32_t x) {
    return x < 0x100000U;
}

bool func2(uint32_t x) {
    return ((x >> 20) == 0);
}

bool func3(uint32_t x) {
    return ((x / 0x100000U) == 0);
}

The three functions are equivalent. Specifically the threshold constant to be compared with is a power-of-two. If there is no shortage of temporary registers, then the right shift version (that is, func2) will likely make the smallest code.

My expected result is like this (assuming -Os optimization, not -O2)

        shrl    $20, %edi
        sete    %al
        retq

Instead I got this (x86-64 clang 20.1.0 with -Os option, tested in Compiler Explorer):

        cmpl    $1048576, %edi
        setb    %al
        retq

(Clang can recognize the three functions are equivalent. It just made suboptimal code among the possible choices.)

Note that in RISC-V targets of Clang, it does use the right shift to reduce code size

        srliw   a0, a0, 20
        seqz    a0, a0
        ret

[topperc]

I think the decision is controlled by isLegalICmpImmediate which supports simm32 for X86 and simm12 for RISC-V.