Unnecessary copying of struct field when the same field is later overwritten

[Kmeakin]

These functions should all produce the same assembly

https://godbolt.org/z/4c674fKxE

typedef unsigned long u64;

typedef struct {
    u64 x;
    u64 y;
    u64 z;
} Big;

// Not optimized

Big assign(Big b) {
    b.x = 0;
    return b;
}

Big copy_then_assign(Big b) {
    Big c = b;
    c.x = 0;
    return c;
}

Big memcpy_then_assign(Big b) {
    Big c;
    __builtin_memcpy(&c, &b, sizeof(b));
    c.x = 0;
    return c;
}

// Optimized

Big c99_initializer(Big b) { return (Big){0, b.y, b.z}; }

Big c99_initializer2(Big b) {
    Big c = {b.x, b.y, b.z};
    c.x = 0;
    return c;
}

Big copy_individually(Big b) {
    Big c;
    c.x = 0;
    c.y = b.y;
    c.z = b.z;
    return c;
}

Big copy_individually_then_assign(Big b) {
    Big c;
    c.x = b.x;
    c.y = b.y;
    c.z = b.z;
    c.x = 0;
    return c;
}

Unoptimized assembly:

assign:
        str     xzr, [x0]
        ldr     x9, [x0, #16]
        ldr     q0, [x0]
        str     x9, [x8, #16]
        str     q0, [x8]
        ret

Optimized assembly:

c99_initializer:
        ldur    q0, [x0, #8]
        str     xzr, [x8]
        stur    q0, [x8, #8]
        ret

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Karl Meakin (Kmeakin)

These functions should all produce the same assembly

https://godbolt.org/z/4c674fKxE

typedef unsigned long u64;

typedef struct {
    u64 x;
    u64 y;
    u64 z;
} Big;

// Not optimized

Big assign(Big b) {
    b.x = 0;
    return b;
}

Big copy_then_assign(Big b) {
    Big c = b;
    c.x = 0;
    return c;
}

Big memcpy_then_assign(Big b) {
    Big c;
    __builtin_memcpy(&c, &b, sizeof(b));
    c.x = 0;
    return c;
}

// Optimized

Big c99_initializer(Big b) { return (Big){0, b.y, b.z}; }

Big c99_initializer2(Big b) {
    Big c = {b.x, b.y, b.z};
    c.x = 0;
    return c;
}

Big copy_individually(Big b) {
    Big c;
    c.x = 0;
    c.y = b.y;
    c.z = b.z;
    return c;
}

Big copy_individually_then_assign(Big b) {
    Big c;
    c.x = b.x;
    c.y = b.y;
    c.z = b.z;
    c.x = 0;
    return c;
}

Unoptimized assembly:

assign:
        str     xzr, [x0]
        ldr     x9, [x0, #<!-- -->16]
        ldr     q0, [x0]
        str     x9, [x8, #<!-- -->16]
        str     q0, [x8]
        ret

Optimized assembly:

c99_initializer:
        ldur    q0, [x0, #<!-- -->8]
        str     xzr, [x8]
        stur    q0, [x8, #<!-- -->8]
        ret

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Karl Meakin (Kmeakin)

These functions should all produce the same assembly

https://godbolt.org/z/4c674fKxE

typedef unsigned long u64;

typedef struct {
    u64 x;
    u64 y;
    u64 z;
} Big;

// Not optimized

Big assign(Big b) {
    b.x = 0;
    return b;
}

Big copy_then_assign(Big b) {
    Big c = b;
    c.x = 0;
    return c;
}

Big memcpy_then_assign(Big b) {
    Big c;
    __builtin_memcpy(&c, &b, sizeof(b));
    c.x = 0;
    return c;
}

// Optimized

Big c99_initializer(Big b) { return (Big){0, b.y, b.z}; }

Big c99_initializer2(Big b) {
    Big c = {b.x, b.y, b.z};
    c.x = 0;
    return c;
}

Big copy_individually(Big b) {
    Big c;
    c.x = 0;
    c.y = b.y;
    c.z = b.z;
    return c;
}

Big copy_individually_then_assign(Big b) {
    Big c;
    c.x = b.x;
    c.y = b.y;
    c.z = b.z;
    c.x = 0;
    return c;
}

Unoptimized assembly:

assign:
        str     xzr, [x0]
        ldr     x9, [x0, #<!-- -->16]
        ldr     q0, [x0]
        str     x9, [x8, #<!-- -->16]
        str     q0, [x8]
        ret

Optimized assembly:

c99_initializer:
        ldur    q0, [x0, #<!-- -->8]
        str     xzr, [x8]
        stur    q0, [x8, #<!-- -->8]
        ret

[efriedma-quic]

For the memcpy-then-store case, there's a FIXME in DeadStoreElimination.cpp, in isShortenableAtTheBeginning.

store-then-memcpy is a little more complicated. Maybe we can do something along the lines of MemCpyOptPass::performMemCpyToMemSetOptzn.