Bit fields are unspecified about riscv-elf-psabi-doc HOT 8 CLOSED

I'm not convinced we're doing something sane here -- it looks like we're 32-bit aligned unless the base type is 8 bits or the bit field width is larger than 32. I think we should at least be consistent about whether we align to bit field length or to the base type length.

@aswaterman @sorear Thoughts?

palmer gamma02 ~ $ cat test.c 
#include <stdint.h>

struct bf {
	uint32_t a: 1;
	uint32_t b;
	uint32_t c: 3;
	uint32_t d;
	uint64_t e: 1;
	uint32_t f;
	uint64_t g: 32;
	uint32_t h;
	uint64_t i: 64;
	uint32_t j;
	uint8_t k: 1;
	uint8_t l;
};

void func(uint64_t);
extern struct bf bf;

uint32_t main() {
	func(bf.a);
	func(bf.b);
	func(bf.c);
	func(bf.d);
	func(bf.e);
	func(bf.f);
	func(bf.g);
	func(bf.h);
	func(bf.i);
	func(bf.j);
	func(bf.k);
	func(bf.l);
}
palmer gamma02 ~ $ riscv64-unknown-linux-gnu-gcc test.c -O3 -S -o -
	.file	"test.c"
	.option nopic
	.section	.text.startup,"ax",@progbits
	.align	2
	.globl	main
	.type	main, @function
main:
	add	sp,sp,-16
	sd	s0,0(sp)
	lui	s0,%hi(bf)
	lbu	a0,%lo(bf)(s0)
	sd	ra,8(sp)
	add	s0,s0,%lo(bf)
	and	a0,a0,1
	call	func
	lwu	a0,4(s0)
	call	func
	lbu	a0,8(s0)
	and	a0,a0,7
	call	func
	lwu	a0,12(s0)
	call	func
	lbu	a0,16(s0)
	and	a0,a0,1
	call	func
	lwu	a0,20(s0)
	call	func
	lwu	a0,24(s0)
	call	func
	lwu	a0,28(s0)
	call	func
	ld	a0,32(s0)
	call	func
	lwu	a0,40(s0)
	call	func
	lbu	a0,44(s0)
	and	a0,a0,1
	call	func
	lbu	a0,45(s0)
	call	func
	ld	ra,8(sp)
	ld	s0,0(sp)
	add	sp,sp,16
	jr	ra
	.size	main, .-main
	.ident	"GCC: (GNU) 6.1.0"

from riscv-elf-psabi-doc.

It looks like rv64 matches aarch64 and x86-64 in this example, which is sane enough for me. I propose keeping as-is and documenting the behavior.

from riscv-elf-psabi-doc.

If it matches then we should keep it -- plus, that makes it easier to document. @sorear assuming you're OK with this, can you write it up?

from riscv-elf-psabi-doc.

Does anyone have time to document what GCC actually does for bit fields?

from riscv-elf-psabi-doc.

I'll take a stab at it

from riscv-elf-psabi-doc.

Not sure if this is everything, but it is a start:

Bitfields are packed in little-endian fashion. A bitfield that would span the alignment boundary of its integer type is padded to begin at the next alignment boundary. For example, struct { int x : 10; int y : 12; } is a 32-bit type with x in bits 9-0, y in bits 21-10, and bits 31-22 undefined. By contrast, struct { short x : 10; short y : 12; } is a 32-bit type with x in bits 9-0, y in bits 27-16, and bits 31-28 and 15-10 undefined.

from riscv-elf-psabi-doc.

from riscv-elf-psabi-doc.

Thanks, Jim. I'll make a PR.

from riscv-elf-psabi-doc.