Different instructions may be generated for different immediates

[nemequ]

For instructions with immediate-mode parameters, the compiler may generate different instruction(s). For example,

#include <stdint.h>

typedef int32_t i32x4 __attribute__((__vector_size__(16)));

i32x4
replace_lane_0(i32x4 v, int32_t value) {
    v[0] = value;
    return v;
}

i32x4
replace_lane_1(i32x4 v, int32_t value) {
    v[1] = value;
    return v;
}

Compiler Explorer: https://godbolt.org/z/9beqbP4Wj

Generates:

replace_lane_0(int __vector(4), int):               # @replace_lane_0(int __vector(4), int)
        movd    xmm1, edi
        movss   xmm0, xmm1                      # xmm0 = xmm1[0],xmm0[1,2,3]
        ret
replace_lane_1(int __vector(4), int):               # @replace_lane_1(int __vector(4), int)
        movd    xmm1, edi
        punpcklqdq      xmm1, xmm0              # xmm1 = xmm1[0],xmm0[0]
        shufps  xmm1, xmm0, 226                 # xmm1 = xmm1[2,0],xmm0[2,3]
        movaps  xmm0, xmm1
        ret

I haven't worked out all the details yet, but for most instructions I think it will be possible to work around this by generating all possible values; with the exception of shuffle, worst case is only 16. We'll have to do some work parsing the assembly to deduplicate the results, but nothing too complex.

Shuffle is obviously the exception here, but that's nothing new. For that, I think the best option is just to link to Compiler Explorer and let people plug in their own values.

[nemequ]

I tihnk this is basically resolved by 2be4bca. We now try all possible values for the immediates, and merge implementations where the output differs only in by the value of an immediate parameter in the assembly. The code is a mess, but it seems to basically work.