Re: VLE load/store multiple instructions

[Peter Bergner <bergner at vnet dot ibm dot com>]

On 9/29/17 3:46 PM, Alexander Fedotov wrote:
> +00000000 <prolog>:
> +   0:	18 81 11 00 	e_stmvsrrw 0\(r1\)
> +   4:	18 22 11 04 	e_stmvsprw 4\(r2\)
> +   8:	18 03 11 08 	e_stmvgprw 8\(r3\)
> +   c:	18 a4 11 0c 	e_stmvcsrrw 12\(r4\)
> +  10:	18 c5 11 10 	e_stmvdsrrw 16\(r5\)
> +00000014 <epilog>:
> +  14:	18 26 10 14 	e_ldmvsprw 20\(r6\)
> +  18:	18 07 10 18 	e_ldmvgprw 24\(r7\)
> +  1c:	18 88 10 1c 	e_ldmvsrrw 28\(r8\)
> +  20:	18 a9 10 20 	e_ldmvcsrrw 32\(r9\)
> +  24:	18 ca 10 24 	e_ldmvdsrrw 36\(r10\)
> +00000028 <epilog_alt>:
> +  28:	18 2b 10 28 	e_ldmvsprw 40\(r11\)
> +  2c:	18 0c 10 2c 	e_ldmvgprw 44\(r12\)
> +  30:	18 8d 10 30 	e_ldmvsrrw 48\(r13\)
> +  34:	18 ae 10 34 	e_ldmvcsrrw 52\(r14\)
> +  38:	18 cf 10 38 	e_ldmvdsrrw 56\(r15\)
> \ No newline at end of file

I'll note that having fixed offsets in your file will make adding or
removing lines "harder" later, in that you'll have to update the offsets
for every line after the one you add or delete.  Also, is it possible
one can use VLE in either Big Endian or Little Endian mode?  If so, your
fixed instruction hex values will only match one of them.  If you look
at what we do in power9.d, you'll see we match offsets with ".*" and
we allow matching with either BE or LE instruction encodings, so our
tests work on both BE and LE.  For example:

.*:     (7d a3 04 74|74 04 a3 7d)       cnttzd  r3,r13
.*:     (7d c4 04 75|75 04 c4 7d)       cnttzd\. r4,r14
.*:     (7d e5 04 34|34 04 e5 7d)       cnttzw  r5,r15
.*:     (7e 06 04 35|35 04 06 7e)       cnttzw\. r6,r16
.*:     (7d 54 ae 12|12 ae 54 7d)       modsd   r10,r20,r21
.*:     (7d 75 b6 16|16 b6 75 7d)       modsw   r11,r21,r22

Just a friendly suggestion, so take or ignore it. :-)


>>> In current PPC opcodes we have only e_ldm*. But it seems that a lot of
>>> developers uses e_lm* in their software and it fails with GAS. Other
>>> compilers also recognizes only e_lm*.

...

>  {"e_ldmvgprw",	OPVUPRT(6,16,0),OPVUPRT_MASK,	PPCVLE,	0,		{D8, RA0}},
> +{"e_lmvgprw",	OPVUPRT(6,16,0),OPVUPRT_MASK,	PPCVLE,	0,		{D8, RA0}},

If e_lmvgprw is the more common usage over e_ldmvgprw, which your
comment above seems to imply, then you'll want to add it above
e_ldmvgprw, rather than after it.  That way, objdump will disassemble
the instruction to e_lmvgprw rather than e_ldmvgprw.


Peter