Widened pointers not optimized back to scalar for vectorized interleaved accesses on RISC-V

[lukel97]

https://compiler-explorer.com/z/bcnocK139

#include <stddef.h>

struct Foo { int a, b; };

void f(struct Foo *x, size_t n) {
    for (size_t i = 0; i < n; i++) {
        x[i].a += 1;
        x[i].b += 2;
    }
}

void g(struct Foo *x, size_t n) {
    for (size_t i = 0; i < n; i++)
        x[i].a += x[i].b;
}

In the loop f, the loop vectorizer is able to use a scalar address throughout the segmented accesses.

In g, where not every segmented field is written back to, it's kept around as a widened vector of pointers, even though we only use the first element of it.

legalizeAndOptimizeInductions in VPlanTransforms.cpp may not be catching this case.

f:
        beqz    a1, .LBB0_8
        csrr    a2, vlenb
        srli    a7, a2, 1
        bgeu    a1, a7, .LBB0_3
        li      a6, 0
        j       .LBB0_6
.LBB0_3:
        srli    a2, a2, 3
        neg     a4, a7
        and     a6, a1, a4
        slli    a4, a2, 5
        mv      a5, a0
        mv      a2, a6
        vsetvli a3, zero, e32, m2, ta, ma
.LBB0_4:
        vlseg2e32.v     v8, (a5)
        sub     a2, a2, a7
        vadd.vi v8, v8, 1
        vadd.vi v10, v10, 2
        vsseg2e32.v     v8, (a5)
        add     a5, a5, a4
        bnez    a2, .LBB0_4
        beq     a1, a6, .LBB0_8
.LBB0_6:
        sh3add  a2, a6, a0
        sh3add  a1, a1, a0
        addi    a0, a2, 4
        addi    a1, a1, 4
.LBB0_7:
        lw      a2, -4(a0)
        lw      a3, 0(a0)
        addi    a2, a2, 1
        addi    a3, a3, 2
        sw      a2, -4(a0)
        sw      a3, 0(a0)
        addi    a0, a0, 8
        bne     a0, a1, .LBB0_7
.LBB0_8:
        ret

g:
        beqz    a1, .LBB1_8
        csrr    a3, vlenb
        srli    a2, a3, 1
        bgeu    a1, a2, .LBB1_3
        li      a6, 0
        j       .LBB1_6
.LBB1_3:
        srli    a3, a3, 3
        neg     a4, a2
        vsetvli a5, zero, e64, m4, ta, ma
        vid.v   v8
        and     a6, a1, a4
        slli    t0, a3, 5
        li      a7, 8
        mv      a4, a0
        mv      a5, a6
.LBB1_4:
        vsetvli zero, zero, e64, m4, ta, ma
        vsll.vi v12, v8, 3
        vadd.vx v12, v12, a0
        vmv.x.s a3, v12
        vsetvli zero, zero, e32, m2, ta, ma
        vlseg2e32.v     v12, (a3)
        vsetvli zero, zero, e64, m4, ta, ma
        vadd.vx v8, v8, a2
        sub     a5, a5, a2
        vsetvli zero, zero, e32, m2, ta, ma
        vadd.vv v12, v12, v14
        vsse32.v        v12, (a4), a7
        add     a4, a4, t0
        bnez    a5, .LBB1_4
        beq     a1, a6, .LBB1_8
.LBB1_6:
        sh3add  a2, a6, a0
        sh3add  a0, a1, a0
.LBB1_7:
        lw      a1, 0(a2)
        lw      a3, 4(a2)
        add     a1, a1, a3
        sw      a1, 0(a2)
        addi    a2, a2, 8
        bne     a2, a0, .LBB1_7
.LBB1_8:
        ret

[llvmbot]

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

Author: Luke Lau (lukel97)

https://compiler-explorer.com/z/bcnocK139

In the loop f, the loop vectorizer is able to use a scalar address throughout the segmented accesses.

In g, where not every segmented field is written back to, it's kept around as a widened vector of pointers, even though we only use the first element of it.

legalizeAndOptimizeInductions in VPlanTransforms.cpp may not be catching this case.

[topperc]

Would this be fixed if the loop vectorizer emitted strided.store natively instead of masked.scatter?

[lukel97]

Would this be fixed if the loop vectorizer emitted strided.store natively instead of masked.scatter?

I'm not sure. The strided store is already using a scalar pointer, it's the segmented load that carries around the vector pointer

[topperc]

Would this be fixed if the loop vectorizer emitted strided.store natively instead of masked.scatter?

I'm not sure. The strided store is already using a scalar pointer, it's the segmented load that carries around the vector pointer

The scalar pointer for the strided store was created by RISCVGatherScatterLowering wasn't it? Hopefully if the vectorizer was emitting a strided store, it wouldn't make a vector induction variable.

[lukel97]

Would this be fixed if the loop vectorizer emitted strided.store natively instead of masked.scatter?

I'm not sure. The strided store is already using a scalar pointer, it's the segmented load that carries around the vector pointer

The scalar pointer for the strided store was created by RISCVGatherScatterLowering wasn't it? Hopefully if the vectorizer was emitting a strided store, it wouldn't make a vector induction variable.

Oh I see what you mean, that makes sense.

I did some poking about and it looks like we never create a widened pointer to begin with in f because none of the widening decisions are CM_GatherScatter. The new CM_Strided decision type in #128718 would fix this for g