powerpc: failure to optimize manual vec_nmsub implementation

[folkertdev]

given this code

https://godbolt.org/z/3WxTM4Yao

#include <altivec.h>

vector float old(vector float a, vector float b, vector float c) {
    return vec_nmsub(a, b, c);
}

vector float new(vector float a, vector float b, vector float c) {
    return vec_neg(vec_madd(a, b, vec_neg(c)));
}

on newer powerpc cpus, these both generate the exact same assembly as expected:

        xvnmsubasp 36, 34, 35
        vmr     2, 4
        blr

however for older cpus, the non-intrinsic implementation fails to optimize

old:
        vnmsubfp 2, 2, 3, 4
        blr

new:
        vspltisb 5, -1
        vslw 5, 5, 5
        vsubfp 4, 5, 4
        vmaddfp 2, 2, 3, 4
        vsubfp 2, 5, 2
        blrasm

this came up here rust-lang/stdarch#1734

[llvmbot]

@llvm/issue-subscribers-backend-powerpc

Author: Folkert de Vries (folkertdev)

given this code

https://godbolt.org/z/3WxTM4Yao

#include <altivec.h>

vector float old(vector float a, vector float b, vector float c) {
    return vec_nmsub(a, b, c);
}

vector float new(vector float a, vector float b, vector float c) {
    return vec_neg(vec_madd(a, b, vec_neg(c)));
}

on newer powerpc cpus, these both generate the exact same assembly as expected:

        xvnmsubasp 36, 34, 35
        vmr     2, 4
        blr

however for older cpus, the non-intrinsic implementation fails to optimize

old:
        vnmsubfp 2, 2, 3, 4
        blr

new:
        vspltisb 5, -1
        vslw 5, 5, 5
        vsubfp 4, 5, 4
        vmaddfp 2, 2, 3, 4
        vsubfp 2, 5, 2
        blrasm

this came up here rust-lang/stdarch#1734