[RISCV][LV] Subtract in reduction forcing fixed vector codegen

[preames]

In the following example, we have a summation loop. One of the terms being contributed to the accumulator involves a subtraction. This "-" appears to force us to pick a fixed vectorization factor instead of a scalable one. Simply changing the "-" to a "+" results in a scalable vector factor being chosen.

typedef struct {
    int x;
} s_t;

int example(s_t *a, s_t *b, int len) {
    int accum = 0;
    for (int i = 0; i < len; ++i) {
        accum += a[i].x - b[i].x;
        ;
    }
    return accum;
}

https://godbolt.org/z/bPebGEaj9 for the codegen.

My guess is that this is some quirk of our cost model. I haven't tried to chase it down just yet.

[preames]

@lukel97 @Mel-Chen FYI, probably an easy fix?

[llvmbot]

@llvm/issue-subscribers-backend-risc-v

Author: Philip Reames (preames)

In the following example, we have a summation loop. One of the terms being contributed to the accumulator involves a subtraction. This "-" appears to force us to pick a fixed vectorization factor instead of a scalable one. Simply changing the "-" to a "+" results in a scalable vector factor being chosen.

typedef struct {
    int x;
} s_t;

int example(s_t *a, s_t *b, int len) {
    int accum = 0;
    for (int i = 0; i < len; ++i) {
        accum += a[i].x - b[i].x;
        ;
    }
    return accum;
}

https://godbolt.org/z/bPebGEaj9 for the codegen.

My guess is that this is some quirk of our cost model. I haven't tried to chase it down just yet.

[lukel97]

This seems to be a regression after #147026, since sub reductions were split out into a new RecurKind and the RISCV TTI hook wasn't updated to handle it. Hopefully it's just a matter of adding in that enum?