generated code of inline assembly modifier 'P' with -fPIE is different between gcc and clang

[bysui]

GCC version: 9.3.0
Clang version: 15.0.7

The output assembly code of following code is different between gcc and clang with -c -O2 -fPIE:

#include <stdio.h>

#pragma GCC visibility push(hidden)

#define current(var) ({\
	typeof(var) pfo_ret__;\
	asm("movq %%gs:%P1, %0\n"\
	    : "=r"(pfo_ret__)\
	    : "p" (&(var))\
	);\
	pfo_ret__;\
})


struct foo {
	int field1;
	int field2;
};

extern struct foo a;

int test ()
{
	int total;

	total = current(a).field1 + current(a).field2;
	printf ("total: %d\n", total);

	return 0;
}

The expected assembly code should be like this:

movq %gs:a(%rip), %rax

Clang could generate right code but gcc generates absolute reference:

movq %gs:a, %rax

It seems that modifier 'P' is not compatible with PIE in gcc, so I try to use modifier 'a'
and it can generate right code. But modifier 'a' leads to compilation error in clang.

My question is why the generated code of modifier 'P' is different between gcc and clang?
Which one is right?

Thanks.

[nickdesaulniers]

Documentation: https://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html#x86-Operand-Modifiers

[bysui]

Documentation: https://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html#x86-Operand-Modifiers

Hi Nick,

I had looked up the documentation before. Regarding the 'P' modifier, it states that it "Prints the raw symbol name if used for a constant". If I understand correctly, it should generate an absolute reference. I have tested it by compiling with both -fPIE and -fno-PIE, but it seems to have no impact. So, GCC always generates an absolute reference, whereas CLANG always generates a RIP-relative reference, regardless of -fPIE being used or not.

I think the documentation is not clear enough for the user.

Additionally, I was unable to find an explanation for the 'a' modifier, even in the source code of gcc. Furthermore, it appears that the 'p' modifier is also not supported in CLANG.

I think that the documentation may be insufficient for users to fully understand the behavior of these modifiers. :(

[llvmbot]

@llvm/issue-subscribers-backend-x86

GCC version: 9.3.0 Clang version: 15.0.7

The output assembly code of following code is different between gcc and clang with -c -O2 -fPIE:

#include <stdio.h>

#pragma GCC visibility push(hidden)

#define current(var) ({\
	typeof(var) pfo_ret__;\
	asm("movq %%gs:%P1, %0\n"\
	    : "=r"(pfo_ret__)\
	    : "p" (&(var))\
	);\
	pfo_ret__;\
})


struct foo {
	int field1;
	int field2;
};

extern struct foo a;

int test ()
{
	int total;

	total = current(a).field1 + current(a).field2;
	printf ("total: %d\n", total);

	return 0;
}

The expected assembly code should be like this:

movq %gs:a(%rip), %rax

Clang could generate right code but gcc generates absolute reference:

movq %gs:a, %rax

It seems that modifier 'P' is not compatible with PIE in gcc, so I try to use modifier 'a'
and it can generate right code. But modifier 'a' leads to compilation error in clang.

My question is why the generated code of modifier 'P' is different between gcc and clang?
Which one is right?

Thanks.