two-instruction pass actually worsens codegen

[AZero13]

https://godbolt.org/z/vrexqnWoe

There is nothing wrong with:
add edi, esi
mov eax, edi
sub edi, edx
cmovnc eax, edi
ret

But the two-address pass makes it:

    lea     eax, [rdi + rsi]
    xor     ecx, ecx
    cmp     eax, edx
    cmovae  ecx, edx
    sub     eax, ecx
    ret

which is actually less optimal.

[AZero13]

Here, the moment where it turns the add into a lea when it doesn't need to.

the reason is that you need an intermediate register for the result, so you might as well just use the register where it already is instead of doing an early move to EAX

[topperc]

There is nothing wrong with:
add edi, esi
mov eax, edi
sub edi, edx
cmovnc eax, edi
ret

This doesn't look like the input you're showing for the two address instruction pass. The MOV32r0 and SUB after the cmov are missing.

[AZero13]

There is nothing wrong with:
add edi, esi
mov eax, edi
sub edi, edx
cmovnc eax, edi
ret

This doesn't look like the input you're showing for the two-address instruction pass. The MOV32r0 and SUB after the cmov are missing.

I must have made a mistake, but I know that this is still worse codegen.