[Bug middle-end/109057] Does GCC interpret assembly when deciding to optimize away a variable?

["hbucher at gmail dot com" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109057

--- Comment #1 from Henry <hbucher at gmail dot com> ---
Two caveats: 

1. If you add something like `xor %0,%0` inside the assembly text, LUT is not
optimized

inline void DoNotOptimize( uint8_t value) {
  asm volatile("xor %0,%0" : : "r,m"(value) : "memory");
}
void func2(uint8_t val) {
    DoNotOptimize(LUT[val]); 
}

Produces

func2(unsigned char):
        movzbl  %dil, %edi
        xor LUT(%rdi),LUT(%rdi)
        ret
LUT:
        .string "\001\005\003"
        .ascii  "\002\007\001\002"

https://godbolt.org/z/Mn5asGWe4

2. If you make value a uint32_t instead of a uint8_t, LUT is not optimized

#include <stdint.h>
static const uint8_t LUT[8] = {1,5,3,0,2,7,1,2};

inline void DoNotOptimize( uint32_t value) {
  asm volatile("" : : "r,m"(value) : "memory");
}
void func2(uint8_t val) {
    DoNotOptimize(LUT[val]); 
}

Produces 

func2(unsigned char):
        movzbl  %dil, %edi
        movzbl  LUT(%rdi), %eax
        ret
LUT:
        .string "\001\005\003"
        .ascii  "\002\007\001\002"

https://godbolt.org/z/rTfExvEbb