[RISCV] missed optimization: use sh or sw when can infer the offset from alignment

[BaoshanPang]

https://godbolt.org/z/j78cddY1W

For the C code:

typedef struct {
    char a;
    char b;
    char c;
    int len;
    char e;
    int len2;
} __attribute__((packed)) str_t;

void foo(str_t *p __attribute__((align_value(4)))) {
    p->len = 0;
    p->len2 = 0;
}

The asm generated is:

foo:
        sb      zero, 3(a0)
        sb      zero, 4(a0)
        sb      zero, 5(a0)
        sb      zero, 6(a0)
        sw      zero, 8(a0)
        ret

For 'len', the middle two 'sb' can be converted to one 'sh'.

[llvmbot]

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

Author: Baoshan (BaoshanPang)

https://godbolt.org/z/YP9WP9rET

For the C code:

typedef struct {
    char a;
    char b;
    char c;
    int len;
    char e;
    int len2;
} __attribute__((packed)) str_t;

void foo(str_t *p __attribute__((aligned(4)))) {
    p->len = 0;
    p->len2 = 0;
}

The asm generated is:

foo:
        sb      zero, 3(a0)
        sb      zero, 4(a0)
        sb      zero, 5(a0)
        sb      zero, 6(a0)
        sb      zero, 8(a0)
        sb      zero, 9(a0)
        sb      zero, 10(a0)
        sb      zero, 11(a0)
        ret

For 'len', the middle two 'sb' can be converted to one 'sh'; for 'len2', the four sb can be converted to one 'sw'.

[Acthinks]

Hi @EugeneZelenko @BaoshanPang , I'd like to work on this issue.

[topperc]

Hi @EugeneZelenko @BaoshanPang , I'd like to work on this issue.

This isn't going to be easy to fix.

First I'm not sure the aligned attribute is being used correctly. gcc rejects the code.

llvm stores the alignment information in each load. But the alignment can only be a power of 2. In this case the Len load is known to start 3 bytes off a 4 byte alignment. LLVM can't represent that today.

[Acthinks]

// Aligns POINTER (useless for dereferencing)
T *__attribute__((aligned(N))) p;

// Aligns POINTEE DATA (what you need)
T *__attribute__((align_value(N))) p;

Packed struct limitations​​:
Forces unaligned accesses regardless of pointer declaration
Alignment attributes can't override fundamental type alignment

str_t *p __attribute__((align_value(4))) can solve ‘len2’ ‘sw’. I will continue to analyze the optimization of ‘len’ ‘sh’ based on this.

https://godbolt.org/z/b44sP79hd

[Acthinks]

Hi @BaoshanPang , could you update the issue to replace ‘aligned(4)’ with ‘align_value(4)’?

[BaoshanPang]

@Acthinks updated.