ARM: More intrinsics

[pcpa at mandriva.com.br]

> Xerxes noticed an inconsistency in my patch.  I've removed the unused
> masks, and redefined the condition in the atomic operations to use ETT
> (else, true, true) rather than the confused encoding I used before.
>
> So, it's now
>
> +      it(codebuf, COND_EQ, IT_MASK_ETT);
>
> instead of
>
> -      it(codebuf, COND_NE, IT_MASK_TEE);

  I feel there is something bogus here, because the first condition
should be 'T' always, and the mask is based on bit 0 of the test
condition. Talking this because I wrote support for arm thumb jit
recently.

> I had to do this because it's the only way the encodings make any
> sense.  There isn't really any encoding for TEE that works with
> COND_NE: it has to be expressed in the binary as COND_EQ:ETT.
> (If cond is true, don't do X, but do Y and Z.)

  For easier viewing instead of pointing to some link, the code I use
to encode it* is:

static void
_tcit(jit_state_t *_jit, unsigned int tc, int it)
{
    int		c;
    int		m;
    c = (tc >> 28) & 1;
    assert(!(tc & 0xfffffff) && tc != ARM_CC_NV);
    switch (it) {
	case THUMB2_IT:		m =   1<<3; 			break;
	case THUMB2_ITT:	m =  (c<<3)| (1<<2);		break;
	case THUMB2_ITE:	m = (!c<<3)| (1<<2);		break;
	case THUMB2_ITTT:	m =  (c<<3)| (c<<2)| (1<<1);	break;
	case THUMB2_ITET:	m = (!c<<3)| (c<<2)| (1<<1);	break;
	case THUMB2_ITTE:	m =  (c<<3)|(!c<<2)| (1<<1);	break;
	case THUMB2_ITEE:	m = (!c<<3)|(!c<<2)| (1<<1);	break;
	case THUMB2_ITTTT:	m =  (c<<3)| (c<<2)| (c<<1)|1;	break;
	case THUMB2_ITETT:	m = (!c<<3)| (c<<2)| (c<<1)|1;	break;
	case THUMB2_ITTET:	m =  (c<<3)|(!c<<2)| (c<<1)|1;	break;
	case THUMB2_ITEET:	m = (!c<<3)|(!c<<2)| (c<<1)|1;	break;
	case THUMB2_ITTTE:	m =  (c<<3)| (c<<2)|(!c<<1)|1;	break;
	case THUMB2_ITETE:	m = (!c<<3)| (c<<2)|(!c<<1)|1;	break;
	case THUMB2_ITTEE:	m =  (c<<3)|(!c<<2)|(!c<<1)|1;	break;
	case THUMB2_ITEEE:	m = (!c<<3)|(!c<<2)|(!c<<1)|1;	break;
	default:		abort();
    }
    assert(m && (tc != ARM_CC_AL || !(m & (m - 1))));
    is(0xbf00 | (tc >> 24) | m);
}

NE:TEE should be 0xbf02, but my code actually uses only IT and ITE,
so there may be bugs as I did not test all possible code paths so far.
Two IT when conditions are not inverse, e.g. in vfp code once to
ccmov after equality test and again to ccmov if unordered.

Link to most of my thumb jit code:
http://code.google.com/p/exl/source/browse/trunk/lib/ejit_arm-cpu.c

> If you want the whole patch resubmitting with these changes I will do
> so.
>
> Andrew.
>
>
>
> diff -r f2b27b309c43 arm_port/hotspot/src/cpu/zero/vm/thumb2.cpp
> --- a/arm_port/hotspot/src/cpu/zero/vm/thumb2.cpp       Wed Mar 14 05:47:19
> 2012 -0400
> +++ b/arm_port/hotspot/src/cpu/zero/vm/thumb2.cpp       Wed Mar 14 13:20:41
> 2012 -0400
> @@ -3039,13 +3039,10 @@
>  #define T_CBZ(r, uoff) (0xb100 | (((uoff) & 0x1f) << 3) | (((uoff) & 0x20) <<
> (8-5)) | ((r) & 7))
>  #define T_CBNZ(r, uoff)        (0xb900 | (((uoff) & 0x1f) << 3) | (((uoff) &
> 0x20) << (8-5)) | ((r) & 7))
>
> -#define T_IT(cond, mask) (0xbf00 | (conds[cond] << 4) | (mask))
> +#define T_IT(cond, mask) (0xbf00 | (conds[cond] << 4) ^ (mask))
>
>  #define IT_MASK_T      8
> -#define IT_MASK_TE     0x14
> -#define IT_MASK_TT     0x1e
> -#define IT_MASK_TTT    0x1e
> -#define IT_MASK_TEE    0x12
> +#define IT_MASK_ETT    0x12
>
>  #define PATCH(loc)     do {                                            \
>           unsigned oldidx = codebuf->idx;                               \
> @@ -5077,7 +5074,8 @@
>        // FIXME: The JNI StrictMath routines don't use the JNIEnv *env
>        // parameter, so it's arguably pointless to pass it here.
>        add_imm(jinfo->codebuf, ARM_R0, Rthread, THREAD_JNI_ENVIRONMENT);
> -      blx(jinfo->codebuf, (unsigned)entry_point);
> +      mov_imm(jinfo->codebuf, ARM_IP, (unsigned)entry_point);
> +      blx_reg(jinfo->codebuf, ARM_IP);
>        bcc_patch(jinfo->codebuf, COND_EQ, loc);
>        vfp_to_jstack(jinfo, VFP_D0);
>
> @@ -5125,7 +5123,7 @@
>        branch(codebuf, COND_NE, retry);
>        bcc_patch(jinfo->codebuf, COND_NE, loc_failed);
>
> -      it(codebuf, COND_NE, IT_MASK_TEE);
> +      it(codebuf, COND_EQ, IT_MASK_ETT);
>        mov_imm(codebuf, result, 0);
>        mov_imm(codebuf, result, 1);
>        fullBarrier(codebuf);
> @@ -5172,7 +5170,7 @@
>
>        unsigned result = JSTACK_REG(jinfo->jstack);
>
> -      it(codebuf, COND_NE, IT_MASK_TEE);
> +      it(codebuf, COND_EQ, IT_MASK_ETT);
>        mov_imm(codebuf, result, 0);
>        mov_imm(codebuf, result, 1);
>        fullBarrier(codebuf);

Paulo