ARM32 Not using immediate offset

[felixjones]

https://godbolt.org/z/Ka89GqGfj

This affects performance for ARM32 embedded systems.

No matter what flags I change, I can never get the resulting ldr to use an immediate offset. It seems to always be handled via an extra register move, and a register offset.

Clang 14.0.0 -Oz -marm -mcpu=arm7tdmi -mabi=aapcs

mov     r1, #16
lsl     r0, r0, #2
orr     r1, r1, #4096
ldr     r0, [r0, r1]
bx      lr

These two instructions could be combined into a single load immediate offset.

mov     r1, #16
...
ldr     r0, [r0, r1]

Re-producing C code:

int read(int idx) {
    const volatile int* const DATA = ( const volatile int* ) 0x1010;
    return DATA[idx];
}

GCC produces the desired output.

GCC 12.0.0 -Os -marm -mcpu=arm7tdmi -mabi=aapcs

lsl     r0, r0, #2
add     r0, r0, #4096
ldr     r0, [r0, #16]
bx      lr

[llvmbot]

@llvm/issue-subscribers-backend-arm