Re: build musl for armv7m

[Rich Felker <dalias@libc.org>]

On Tue, Jun 14, 2016 at 01:49:40AM -0700, weimingz@codeaurora.org wrote:
> diff --git a/src/setjmp/arm/longjmp.s b/src/setjmp/arm/longjmp.s
> index e28d8f3..5c85cba 100644
> --- a/src/setjmp/arm/longjmp.s
> +++ b/src/setjmp/arm/longjmp.s
> @@ -7,8 +7,11 @@ _longjmp:
>  longjmp:
>  	mov ip,r0
>  	movs r0,r1
> +  it eq
>  	moveq r0,#1
> -	ldmia ip!, {v1,v2,v3,v4,v5,v6,sl,fp,sp,lr}
> +	ldmia ip!, {v1,v2,v3,v4,v5,v6,sl,fp}
> + ldr sp, [ip]!
> + ldr lr, [ip]!

Small note: your indention is inconsistent -- using a single space
instead of a tab.

In my earlier work trying to get an all-thumb build working, I had
tried doing this as:

+       ldmia ip!, {r2,v1,v2,v3,v4,v5,v6,sl,fp,lr}
+       mov sp,r2

which might be marginally faster but of course it reorders the fields
in the jmp_buf. Nothing should be depending on them anyway but I like
your version better I think. Likewise for setjmp.

> diff --git a/src/thread/arm/atomics.s b/src/thread/arm/atomics.s
> index 673fc03..05fd9b9 100644
> --- a/src/thread/arm/atomics.s
> +++ b/src/thread/arm/atomics.s
> @@ -49,6 +49,7 @@ __a_cas_dummy:
>  	mov r3,r0
>  	ldr r0,[r2]
>  	subs r0,r3,r0
> + it eq
>  	streq r1,[r2]
>  	bx lr
>  .global __a_cas_v6
> @@ -58,6 +59,7 @@ __a_cas_v6:
>  	mcr p15,0,r0,c7,c10,5
>  1:	.word 0xe1920f9f        /* ldrex r0,[r2] */
>  	subs r0,r3,r0
> + itt eq
>  	.word 0x01820f91        /* strexeq r0,r1,[r2] */
>  	teqeq r0,#1
>  	beq 1b

You're ignoring that the .word-encoded instruction is not valid thumb,
and that the two 'eq' conditions are independent. The second one
(teqeq) is using the result of the strexeq, not the result of the
subs. But there are much deeper reasons this file cannot work as thumb
without significant changes: the arithmetic on pc is not
thumb-compatible.

> @@ -70,6 +72,7 @@ __a_cas_v7:
>  	.word 0xf57ff05b        /* dmb ish */
>  1:	.word 0xe1920f9f        /* ldrex r0,[r2] */
>  	subs r0,r3,r0
> + itt eq
>  	.word 0x01820f91        /* strexeq r0,r1,[r2] */
>  	teqeq r0,#1
>  	beq 1b

Same here.

There's also the matter of the code to read the thread pointer, which
requires a special CP15 register that cortex-m's lack, I think. I
don't see any alternate way to do this functionality (and it would be
ABI-incompatible anyway) so I think the kernel (or bare-metal
undefined instruction exception trap) needs to just trap and emulate
it.

Rich