use of %gs by an app on linux not handled properly

During the dev of the test case for #1823 I found a problem when a program use the gs segment.
Just run the program following program with drrun.

``` C
#include

#ifdef __i386__
# include
# include
# include
# include
# include
# include
#else
# include
# include
#endif

int arch_prctl(int code, unsigned long addr);

int test_func()
{
return (42);
}

int main()
{
#ifdef __i386__
void *seg = mmap(NULL, getpagesize(), PROT_WRITE | PROT_READ,
MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
struct user_desc u_info;
int val;

((void **)seg)[0x10 / sizeof(void *)] = &test_func;
u_info.entry_number = 7;
u_info.base_addr = (unsigned long)seg;
u_info.limit = getpagesize();
u_info.seg_32bit = 1;
u_info.contents = MODIFY_LDT_CONTENTS_DATA;
u_info.read_exec_only = 0;
u_info.limit_in_pages = 0;
u_info.seg_not_present = 0;
u_info.useable = 1;
val = (7 << 3 | 0 << 2 | 3);

if (syscall(SYS_set_thread_area, &u_info) < 0 && errno == EINVAL) {
u_info.entry_number = 13;
val = (13 << 3 | 0 << 2 | 3);
syscall(SYS_set_thread_area, &u_info);
}

__asm__ volatile ("push %%gs; \n\
mov %0, %%gs; \n\
call *%%gs:0x10; \n\
mov $0x10, %%eax; \n\
call *%%gs:(%%eax); \n\
pop %%gs" \
: : "m" (val) : "eax");
#else
void (*funcs[10])();
void *old_fs;

funcs[0x10 / sizeof(void *)] = (void *)&test_func;

arch_prctl(ARCH_SET_GS, (unsigned long)funcs);
__asm__ volatile ("call *%gs:0x10; \n\
mov $0x10, %rax; \n\
call *%gs:(%rax);");

#endif
return 0;
}

```

On 32bit dynamoRIO segfault and on 64bit it's look like an infinite loop.

This program run correctly without dynamoRIO in both 32bit and 64bit.

Start by compiling the C reproducer in the issue and running it with drrun on both 32-bit and 64-bit Linux, then compare its behavior with a direct run. Trace the %gs accesses and segment setup involved in the reproducer; done means both variants execute test_func and return normally under DynamoRIO, as they do without it.

Written by the indexing model from the issue text.