Re: [9fans] 5l bug

[zephyr.pellerin@gmail.com]

Patching p -> scond = q -> scond in /sys/src/cmd/5l/noop.c fixes this unfortunate linker bug.

On Sunday, April 28, 2013 10:54:18 PM UTC-7, Paul Patience wrote:
> Mischief reported a crash with on arm with
>
> winwatch when closing all windows excluding those
>
> ignored through the -e flag. Cinap_lenrek
>
> narrowed the problem down to a bug in the 5l
>
> linker: it generates incorrect code when using
>
> the conditional operator and dividing.
>
> A minimal test case is the following:
>
>
>
> term% cat foo.c
>
> #include <u.h>
>
> #include <libc.h>
>
>
>
> void
>
> main(void)
>
> {
>
> 	int i, j, k;
>
>
>
> 	i = 1;
>
> 	j = 0;
>
> 	k = j <= 0 ? 1 : 2/i;
>
> 	print("%d\n", k);
>
> 	exits(nil);
>
> }
>
>
>
> The output of asm(main) in acid is:
>
>
>
> acid: asm(main)
>
> main 0x00001020	MOVW.W	R14,#-0x18(R13)
>
> main+0x4 0x00001024	MOVW	$#0x1,R1
>
> main+0x8 0x00001028	MOVW	$#0x0,R2
>
> main+0xc 0x0000102c	CMP.S	$#0x0,R2
>
> main+0x10 0x00001030	MOVW.LE	R1,R3
>
> main+0x14 0x00001034	MOVW.GT	$#0x2,R3
>
> main+0x18 0x00001038	SUB.GT	$#0x8,R13,R13
>
> main+0x1c 0x0000103c	MOVW	R1,#0x4(R13)
>
> main+0x20 0x00001040	MOVW	R3,R11
>
> main+0x24 0x00001044	BL	_div
>
> main+0x28 0x00001048	MOVW	R11,R3
>
> main+0x2c 0x0000104c	ADD	$#0x8,R13,R13
>
> main+0x30 0x00001050	MOVW	$#0x7080,R0
>
> main+0x34 0x00001054	MOVW	R3,#0x8(R13)
>
> main+0x38 0x00001058	BL	print
>
> main+0x3c 0x0000105c	MOVW	$#0x0,R0
>
> main+0x40 0x00001060	BL	exits
>
> main+0x44 0x00001064	RET.P	#0x18(R13)
>
>
>
> The problem is that the division at main+0x24
>
> happens regardless of the comparison. The output
>
> from the compiler is correct, however:
>
>
>
> term% 5c -S foo.c
>
> 	TEXT	main+0(SB),0,$20
>
> 	MOVW	$1,R1
>
> 	MOVW	$0,R2
>
> 	CMP	$0,R2,
>
> 	MOVW.LE	R1,R3
>
> 	MOVW.GT	$2,R3
>
> 	DIV.GT	R1,R3,R3
>
> 	MOVW	$.string<>+0(SB),R0
>
> 	MOVW	R3,8(R13)
>
> 	BL	,print+0(SB)
>
> 	MOVW	$0,R0
>
> 	BL	,exits+0(SB)
>
> 	RET	,
>
> 	DATA	.string<>+0(SB)/8,$"%d\n\z\z\z\z\z"
>
> 	GLOBL	.string<>+0(SB),$8
>
> 	END	,
>
>
>
> A (temporary) workaround for this is using an
>
> if-else statement, for which the linker generates
>
> correct code.
>
>
>
> --
>
> pap