Pnop extended mnemonic

Public Discussion
#1

Hi,

ISA 3.1 added pnop (prefixed nop) instruction, with the second word allowed to contain any value that isn’t a branch instruction.

I was wondering if it would be possible to add an extended mnemonic that functions a bit like a tag?

The reason why I’m asking for this is because when generating bytecode in libxsmm it would greatly help with debugging to have a tag that I can insert into the instruction stream to know where I’ve gotten to. I’m currently using pnop with an index, but objdump doesn’t do anything with the second word.

Will

#2

Hi!

WillTrojak September 3 Hi, ISA 3. 1 added pnop (prefixed nop) instruction, with the second word allowed to contain any value that isn’t a branch instruction. I was wondering if it would be possible to add an extended mnemonic that functions a

WillTrojak
September 3

Hi,

ISA 3.1 added pnop (prefixed nop) instruction, with the second word allowed to contain any value that isn’t a branch instruction.

Almost: it cannot be an rfebb, an attn, or any CSI either. It’s pretty hard to avoid all of those.

I was wondering if it would be possible to add an extended mnemonic that functions a bit like a tag?

You can code this easily already?

.long 0x07000000, 0x12345678

Segher
Unless otherwise stated above:

IBM Nederland B.V.
Gevestigd te Amsterdam
Inschrijving Handelsregister Amsterdam Nr. 33054214

#3

Thanks Segher. Yeah that’s what I was doing. But I was thinking that maybe if something was added as an extended mnemonic to the ISA, then it would mean tools like objdump could interpret the pnop with a non-zero value as a tag and even the assembler could generate these pnops?

I think if you want to avoid those 3 cases, you can do that by enforcing that bit 0-5 of the second word are zero? (Although I think bits 0-5 on the effected instructions are only 17, 19, or 31, so that wouldn’t be too hard to filter).

For example 0x0700000000000001 currently resolves as

      00:       00 00 00 07     pnop                                         
      04:       01 00 00 00

but maybe it could be something like

      00:       00 00 00 07     pnop     1                                         
      04:       01 00 00 00

or even

      00:       00 00 00 07     tag      1                                         
      04:       01 00 00 00

Just a shower thought, but these might even be useful if you want to sneak a 32-bit constant to be used but want to avoid the potential penalty of cache miss from reading from a data region.

#4

Hi!

WillTrojak September 4 Thanks Segher. Yeah that’s what I was doing. But I was thinking that maybe if something was added as an extended mnemonic to the ISA, then it would mean tools like objdump could interpret the pnop with a non-zero value

WillTrojak
September 4

Thanks Segher. Yeah that’s what I was doing. But I was thinking that maybe if something was added as an extended mnemonic to the ISA, then it would mean tools like objdump could interpret the pnop with a non-zero value as a tag and even the assembler could generate these pnops?

Hrm. The insn with payload should be the base instruction, and the one with everything zero (so “pnop”) an extended mnemonic, then.

I think if you want to avoid those 3 cases, you can do that by enforcing that bit 0-5 of the second word are zero? (Although I think bits 0-5 on the effected instructions are only 17, 19, or 31, so that wouldn’t be too hard to filter).

You can simply outlaw all of primary ops 16…19, as well as attn.

That will unnecessarily disallow crlogicals and addpcis, but that’s about it. Allowing the first half of primary 19 again will make this better (but less regular).

You cannot just decide to always use small numbers, because attn is one (it is hex 100).

For example 0x0700000000000001 currently resolves as

      00:       00 00 00 07     pnop                                         
      04:       01 00 00 00

but maybe it could be something like

      00:       00 00 00 07     pnop     1                                         
      04:       01 00 00 00

or even

      00:       00 00 00 07     tag      1                                         
      04:       01 00 00 00

Just a shower thought, but these might even be useful if you want to sneak a 32-bit constant to be used but want to avoid the potential penalty of cache miss from reading from a data region.

Ah, have pnop both as an extended mnemonic and as a base one, if it has no args. Works for me!

Segher

Unless otherwise stated above:

IBM Nederland B.V.
Gevestigd te Amsterdam
Inschrijving Handelsregister Amsterdam Nr. 33054214