Restrict pointer loads not affected by asm memory clobber

[gooncreeper]

When a memory clobber is used in inline assembly, restrict pointers loads are incorrectly reordered after the inline assembly. GCC inline assembly documentation states

Further, the compiler does not assume that any values read from memory before an asm remain unchanged after that asm

Therefore it is invalid to assume that values pointed by restrict pointers are unchanged if they are not passed as input operands.

Example C Code

#include <stdint.h>

void ipc_copy(uint64_t p, uint64_t *restrict a, uint64_t *restrict b) {
    uint64_t s0 = a[0];
    uint64_t s1 = a[1];
    uint64_t s2 = a[2];
    uint64_t s3 = a[3];
    __asm__ __volatile__("mov %0, %%cr3" :: "r"(p) : "memory");
    b[0] = s0;
    b[1] = s1;
    b[2] = s2;
    b[3] = s3;
}

Clang

ipc_copy:                               # @ipc_copy
        mov     cr3, rdi
        movups  xmm0, xmmword ptr [rsi]
        movups  xmmword ptr [rdx], xmm0
        movups  xmm0, xmmword ptr [rsi + 16]
        movups  xmmword ptr [rdx + 16], xmm0
        ret

Expected (Clang 14.0.0 / GCC)

ipc_copy:                               # @ipc_copy
        movups  xmm0, xmmword ptr [rsi]
        movups  xmm1, xmmword ptr [rsi + 16]
        mov     cr3, rdi
        movups  xmmword ptr [rdx], xmm0
        movups  xmmword ptr [rdx + 16], xmm1
        ret

Related: #15867

[gooncreeper]

@Endilll I believe this should be labeled as a miscompilation.

[nikic]

I don't think this is a miscompilation. The asm block obviously cannot access any memory it does not have provenance to access (that would break the entire memory model), so it cannot access a restrict pointer that was not passed to the asm (or escaped).

Similarly to what would happen with a local variable. A strict reading would imply that this optimization is invalid: https://c.godbolt.org/z/v7jr5rMs3 But of course it's intended to be valid and GCC does perform the store to load forwarding across the asm block in that case. The wording in the GCC docs is just imprecise.

[gooncreeper]

I definitely agree. Thanks for clarifying.