2025.02 Toolchain Codegen Regression: Increased Code Size

[bigmagic123]

When optimizing volatile memory operations, GCC 13.1 and 14.1 generate redundant instruction sequences instead of optimizing them. Below is my test case:

-march=rv64gc -mabi=lp64d  -Os

<C>
#include <stdint.h>
uint16_t test_rd(uint32_t addr)
{
    return (uint16_t)(*(volatile uint32_t *)((uintptr_t)(0xF8000000 + (addr << 2))));
}
void test(uint16_t *buf)
{
    buf[10] = test_rd(0x300ae);
    buf[11] = test_rd(0x300ad);
    buf[12] = test_rd(0x300ac);
}
Generated assembly (GCC 13.1/14.1):

<ASM>
test:
        li      a5,65024000
        slli    a5,a5,6
        addi    a5,a5,696
        lw      a5,0(a5)
        sh      a5,20(a0)
        li      a5,65024000
        slli    a5,a5,6
        addi    a5,a5,692
        lw      a5,0(a5)
        sh      a5,22(a0)
        li      a5,65024000
        slli    a5,a5,6
        addi    a5,a5,688
        lw      a5,0(a5)
        sh      a5,24(a0)
        ret

The sequence li a5,65024000 + slli a5,a5,6 is needlessly repeated for each test_rd() call. Full example: https://godbolt.org/z/1ooj71f6f

Expected behavior (GCC 12.1):

<ASM>
test:
        li      a5,65024000
        slli    a5,a5,6
        addi    a4,a5,696
        lw      a4,0(a4)
        sh      a4,20(a0)
        addi    a4,a5,692
        addi    a5,a5,688
        lw      a4,0(a4)
        lw      a5,0(a5)
        sh      a4,22(a0)
        sh      a5,24(a0)
        ret

Upstream Reference: GCC Bug #119847