[libreriscv.git] / openpower / sv / 16_bit_compressed.mdwn

# 16 bit Compressed

Similar to VLE (but without immediate-prefixing) this encoding is designed
to fit on top of OpenPOWER ISA v3.0B when a "Modeswitch" bit is set (PCR
is recommended). Note that Compressed is *mutually exclusively incompatible*
with OpenPOWER v3.1B "prefixing" due to using (requiring) both EXT000
and EXT001. Hypothetically it could be made to use anything other than
EXT001, with some inconvenience (extra gates).  The incompatibility is
"fixed" by swapping out of "Compressed" Mode and back into "Normal"
(v3.1B) Mode, at runtime, as needed.

Although initially intended to be augmented by Simple-V Prefixing (to
add Vector context, width overrides, e.g IEEE754 FP16, and predication) yet not put pressure on I-Cache power
or size, this Compressed Encoding is not critically dependent
*on* SV Prefixing, and may be used stand-alone.

See:

* <https://bugs.libre-soc.org/show_bug.cgi?id=238>
* <https://ftp.libre-soc.org/VLE_314-68105.pdf> VLE Encoding
* <http://lists.mailinglist.openpowerfoundation.org/pipermail/openpower-hdl-cores/2020-November/000210.html>

This one is a conundrum.  OpenPOWER ISA was never designed with 16
bit in mind.  VLE was added 10 years ago but only by way of marking
an entire 64k page as "VLE".  With VLE not maintained it is not
fully compatible with current PowerISA.

Here, in order to embed 16 bit into a predominantly 32 bit stream the
overhead of using an entire 16 bits just to switch into Compressed mode
is itself a significant overhead.  The situation is made worse by 6 bits
being taken up by Major Opcode space, leaving only 10 bits to allocate
to actual instructions.

Contrast this with RVC which takes 3 out of 4 
combinations of the first 2 bits for indicating 16-bit (anything with 0b00 to 0b10 in the LSBs), and uses the 4th as a Huffman-style escape-sequence, easily allowing standard 32 bit and 16 bit to intermingle cleanly.  To achieve the same thing on OpenPOWER would require a whopping 24 6-bit Major Opcodes which is clearly impractical: other schemes need to be devised.

In addition we would like to add SV-C32 which is a Vectorised version
of 16 bit Compressed, and ideally have a variant that adds the 27-bit
prefix format from SV-P64, as well.

Potential ways to reduce pressure on the 16 bit space are:

* To use more than one v3.0B Major Opcode, preferably an odd-even
  contiguous pair
* To provide "paging".  This involves bank-switching to alternative optimised encodings for specific workloads
* To enter "16 bit mode" for durations specified at the start
* To reserve one bit of every 16 bit instruction to indicate that the 16 bit mode is to continue to be sustained

This latter would be useful in the Vector context to have an alternative
meaning: as the bit which determines whether the instruction is 11-bit
prefixed or 27-bit prefixed:

    0 1 2 3 4 5 6 7 8 9 a b c d e f |
    |major op | 11 bit vector prefix|
    |16 bit opcode  alt vec. mode ^ |
    | extra vector prefix if alt set|

Using a major opcode to enter 16 bit mode, leaves 11 bits to find
something to use them for:

    0 1 2 3 4 5 6 7 8 9 a b c d e f |
    |major op | what to do here   1 |
    |16 bit    stay in 16bit mode 1 |
    |16 bit    stay in 16bit mode 1 |
    |16 bit       exit 16bit mode 0 |

One possibility is that the 11 bits are used for bank selection, with
some room for additional context such as altering the registers used
for the 16 bit operations (bank selection of which scalar regs).
However the downside is that short sequences of Compressed instructions
become penalised by the fixed overhead.  Even a single 16 bit instruction requires a 16 bit overhead to "gain access" to 16 bit "mode", making the exercise pointless.

An alternative is to use the first 11 bits for only the utmost commonly used
instructions.  That being the case then one of those 11 bits could
be dedicated to saying if 16 bit mode is to be continued, at which
point *all* 16 bits can be used for Compressed.
10 bits remain for actual opcodes, which is ridiculously tight,
however the opportunity to subsequently use all 16 bits is worth it.

The reason for picking 2 contiguous Major v3.0B opcodes is illustrated below:

    |0 1 2 3 4 5 6 7 8 9 a b c d e f|
    |major op..0| LO Half C space   |
    |major op..1| HI Half C space   |
    |N N N N N|<--11 bits C space-->|

If NNNNN is the same value (two contiguous Major v3.0B Opcodes) this saves gates at a critical part of the decode phase.

# Opcode Allocation Ideas

* one bit from the 16-bit mode is used to indicate that standard
  (v3.0B) mode is to be dropped into for only one single instruction
  <https://bugs.libre-soc.org/show_bug.cgi?id=238#c2>

## Opcodes exploration (Attempt 1)

Switching between different encoding modes is controlled by M (alone)
in 10-bit mode, and M and N in 16-bit mode.

* M in 10-bit mode if zero indicates that following instructions are
  standard OpenPOWER ISA 32-bit encoded (including, redundantly,
  further 10/16-bit instructions)
* M in 10-bit mode if 1 indicates that following instructions are
  in 16-bit encoding mode

Once in 16-bit mode:

* 0b01 (M=1, N=0): stay in 16-bit mode
* 0b00: leave 16-bit mode permanently (return to standard OpenPOWER ISA)
* 0b10: leave 16-bit mode for one cycle (return to standard OpenPOWER ISA)
* 0b11: free to be used for something completely different.

The current "top" idea for 0b11 is to use it for a new encoding format
of predominantly "immediates-based" 16-bit instructions (branch-conditional,
addi, mulli etc.)

* The Compressed Major Opcode is in bits 5-7.
* Minor opcode in bit 8.
* In some cases bit 9 is taken as an additional sub-opcode, followed
  by bits 0-4 (for CR operations)
* M+N mode-switching is not available for C-Major.minor 0b001.1
* 10 bit mode may be expanded by 16 bit mode, adding capabilities
  that do not fit in the extreme limited space.

Mode-switching FSM showing relationship between v3.0B, C 10bit and C 16bit.
16-bit immediate mode remains in 16-bit.

    | 0 | 1234 | 567  8 | 9abcde | f | explanation
    | EXT000/1 | Cmaj.m | fields | 0 | 10bit then v3.0B
    | EXT000/1 | Cmaj.m | fields | 1 | 10bit then 16bit
    | 0 | flds | Cmaj.m | fields | 0 | 16bit then v3.0B
    | 0 | flds | Cmaj.m | fields | 1 | 16bit then 16bit
    | 1 | flds | Cmaj.m | fields | 0 | 16b then 1x v3.0B
    | 1 | flds | Cmaj.m | fields | 1 | 16b/imm then 16bit

Notes:

* Cmaj.m is the C major/minor opcode: 3 bits for major, 1 for minor
* EXT000 and EXT001 are v3.0B Major Opcodes.  The first 5 bits
  are zero, therefore the 6th bit is actually part of Cmaj.
* "10bit then 16bit" means "this instruction is encoded C 10bit
  and the following one in C 16bit"

### C Instruction Encoding types

10-bit Opcode formats (all start with v3.0B EXT000 or EXT001
Major Opcodes)

    | 01234    | 567  8 | 9  | a b | c  | d e | f | enc
    | E01      | Cmaj.m | fld1     | fld2     | M | 10b
    | E01      | Cmaj.m | offset              | M | 10b b
    | E01      | 001.1  | S1 | fd1 | S2 | fd2 | M | 10b sub
    | E01      | 111.m  | fld1     | fld2     | M | 10b LDST

16-bit Opcode formats (including 10/16/v3.0B Switching)

    | 0 | 1234 | 567  8 | 9  | a b | c  | d e | f | enc
    | N | immf | Cmaj.m | fld1     | fld2     | M | 16b
    | 1 | immf | Cmaj.m | fld1     | imm      | 1 | 16b imm
    | fd3      | 001.1  | S1 | fd1 | S2 | fd2 | M | 16b sub
    | N | fd4  | 111.m  | fld1     | fld2     | M | 16b LDST

Notes:

* fld1 and fld2 can contain reg numbers, immediates, or opcode
  fields (BO, BI, LK)
* S1 and S2 are further sub-selectors of C 001.1

### Immediate Opcodes

only available in 16-bit mode, only available when M=1 and N=1
and when Cmaj.min is not 0b001.1.

instruction counts from objdump on /bin/bash:

      466 extsw r1,r1
      649 stw r1,1(r1)
      691 lwz r1,1(r1)
      705 cmpdi r1,1
      791 cmpwi r1,1
      794 addis r1,r1,1
     1474 std r1,1(r1)
     1846 li r1,1
     2031 mr r1,r1
     2473 addi r1,r1,1
     3012 nop
     3028 ld r1,1(r1)


    | 0 | 1  | 2 | 3 4 | | 567.8 | 9ab  | cde | f |
    | 1 | 0  | 0   0 0 | | 001.0 |      | 000 | 1 | TBD
    | 1 | 0  |  sh2    | | 001.0 | RA   | sh  | 1 | sradi.
    | 1 | 1  | 0   0 0 | | 001.0 |      | 000 | 1 | TBD
    | 1 | 1  | 0 | sh2 | | 001.0 | RA   | sh  | 1 | srawi.
    | 1 | 1  | 1 |     | | 001.0 |      |     | 1 | TBD
    | 1 | i2 |  RT     | | 010.0 | RA|0 | imm | 1 | addi
    | 1 | 0 | i2       | | 010.1 | RA   | imm | 1 | cmpdi
    | 1 | 1 | i2       | | 010.1 | RA   | imm | 1 | cmpwi
    | 1 | 0 | i2       | | 011.0 | RT!=1| imm | 1 | ldspi
    | 1 | 1 | i2       | | 011.0 | RT!=1| imm | 1 | lwspi
    | 1 | 0 | i2       | | 011.1 | RT!=1| imm | 1 | stwspi
    | 1 | 1 | i2       | | 011.1 | RT!=1| imm | 1 | stdspi
    | 1 |              | | 011.0 | 001  |     | 1 | TBD
    | 1 |              | | 011.1 | 001  |     | 1 | TBD
    | 1 | i2           | | 100.0 | RT   | imm | 1 | stwi
    | 1 | i2           | | 100.1 | RT   | imm | 1 | stdi
    | 1 | i2           | | 101.0 | RA   | imm | 1 | ldi
    | 1 | i2           | | 101.1 | RA   | imm | 1 | lwi
    | 1 | i2 | RA      | | 110.0 | RT   | imm | 1 | fsti
    | 1 | i2 | RA      | | 110.1 | RT   | imm | 1 | fstdi
    | 1 | i2 | RT      | | 111.0 | RA   | imm | 1 | flwi
    | 1 | i2 | RT      | | 111.1 | RA   | imm | 1 | fldi

Construction of immediate:

* LD/ST r1 (SP) variants should be offset by -256
 see <https://bugs.libre-soc.org/show_bug.cgi?id=238#c43>
  - SP variants map to e.g ld RT, imm(r1)
  - SV Prefixing can be used to map r1 to alternate regs
* [1] not the same as v3.0B addis: the shift amount is smaller and actually
  still maps to within the v3.0B addi immediate range.
* addi is EXTS(i2||imm) to give a 4-bit range -8 to +7
* addis is EXTS(i2||imm||000) to give a 11-bit range -1024 to +1023 in increments of 8 
* all others are EXTS(i2||imm) to give a 7-bit range -128 to +127
  (further for LD/ST due to word/dword-alignment)

Further Notes:

* bc also has an immediate mode, listed separately below in Branch section
* for LD/ST, offset is aligned.  8-byte: i2||imm||0b000 4-byte: 0b00
* SV Prefix over-rides help provide alternative bitwidths for LD/ST
* RA|0 if RA is zero, addi. becomes "li"
  - this only works if RT takes part of opcode
  - mv is also possible by specifying an immediate of zero

### Illegal and nop

Note that illeg is all zeros, including in the 16-bit mode.
Given that C is allocated to OpenPOWER ISA Major opcodes EXT000 and
EXT001 this ensures that in both 10-bit *and* 16-bit mode, a 16-bit
run of all zeros is considered "illegal" whilst 0b0000.0000.1000.0000
is "nop"

    | 16-bit mode | | 10-bit mode                 |
    | 0 | 1 | 234 | | 567.8  | 9  ab | c   de | f |
    | 0 | 0   000 | | 000.0  | 0  00 | 0   00 | 0 | illeg
    | 0 | 0   000 | | 000.0  | 0  00 | 0   00 | 1 | nop

16 bit mode only:

    | 1 | 0   000 | | 000.0  | 0  00 | 0   00 | 0 | nop
    | 1 | nonzero | | 000.0  | 0  00 | 0   00 | 0 | TBD

Notes:

* All-zeros being an illegal instruction is normal for ISAs.  Ensuring that 
  this remains true at all times i.e. for both 10 bit and 16 bit mode is
  common sense.
* The 10-bit nop (bit 15, M=1) is intended for circumstances
  where alignment to 32-bit before returning to v3.0B is required.
  M=1 being an indication "return to Standard v3.0B Encoding Mode". 
* The 16-bit nop (bit 0, N=1) is intended for circumstances where a
  return to Standard v3.0B Encoding is required for one cycle
  but one cycle where alignment to a 32-bit boundary is needed.
  Examples of this would be to return to "strict" (non-C) mode
  where the PC may not be on a non-word-aligned boundary.
* If for any reason multiple 16 bit nops are needed in succession
  the M=1 variant can be used, because each one returns to
  Standard v3.0B Encoding Mode, each time.

In essence the 2 nops are needed due to there being 2 different C forms: 10 and 16 bit.  

### Branch

    | 16-bit mode | | 10-bit mode                 |
    | 0 | 1 | 234 | | 567.8  | 9  ab | c   de | f |
    | N | offs2   | | 000.LK | offs!=0        | M | b, bl
    | 1 | offs2   | | 000.LK | BI    | BO1 oo | 1 | bc, bcl
    | N | BO3 BI3 | | 001.0  | LK BI | BO     | M | bclr, bclrl

16 bit mode:

* bc only available when N,M=0b11
* offs2 extends offset in MSBs
* BI3 extends BI in MSBs to allow selection of full CR
* BO3 extends BO
* bc offset constructed from oo as LSBs and offs2 as MSBs
* bc BI allows selection of all bits from CR0 or CR1
* bc CR check is always active (as if BO0=1) therefore BO1 inverts

10 bit mode:

* illegal (all zeros) covers part of branch (offs=0,M=0,LK=0)
* nop also covers part of branch (offs=0,M=0,LK=1)
* bc **not available** in 10-bit mode
* BO[0] enables CR check, BO[1] inverts check
* BI refers to CR0 only (4 bits of)
* no Branch Conditional with immediate
* no Absolute Address
* CTR mode allowed with BO[2] for b only.
* offs is to 2 byte (signed) aligned
* all branches to 2 byte aligned

### LD/ST

    | 16-bit mode      | | 10-bit mode               |
    | 0   | 1  | 2 3 4 | | 567.8 | 9 a b | c d e | f |
    | RA2 | SZ |  RB   | | 001.1 | 1  RA | 0  RT | M | st
    | RA2 | SZ |  RB   | | 001.1 | 1  RA | 1  RT | M | fst
    | N   | SZ |  RT   | | 111.0 |  RA   |  RB   | M | ld
    | N   | SZ |  RT   | | 111.1 |  RA   |  RB   | M | fld

* elwidth overrides can set different widths

16 bit mode:

* SZ=1 is 64 bit, SZ=0 is 32 bit
* RA2 extends RA to 3 bits (MSB)
* RT2 extends RT to 3 bits (MSB)

10 bit mode:

* RA and RB are only 2 bit (0-3)
* for LD, RT is implicitly RB: "ld RT=RB, RA(RB)"
* for ST, there is no offset: "st RT, RA(0)"

### Arithmetic

    | 16-bit mode | | 10-bit mode             |
    | 0 | 1 | 234 | | 567.8 | 9ab | c d e | f |
    | N | 0 | RT  | | 010.0 | RB  | RA!=0 | M | add
    | N | 0 | RT  | | 010.1 | RB  | RA|0  | M | sub.
    | N | 0 | BF  | | 011.0 | RB  | RA|0  | M | cmpl

Notes:

* sub. and cmpl: default CR target is CR0
* for (RA|0) when RA=0 the input is a zero immediate,
  meaning that sub. becomes neg. and cmp becomes cmpi against zero
* RT is implicitly RB: "add RT(=RB), RA, RB"
* Opcode 0b010.0 RA=0 is not missing from the above:
  it is a system-wide instruction, "cbank" (section below)

16 bit mode only:

    | 0 | 1 | 234 | | 567.8 | 9ab | cde   | f |
    | N | 1 | RA  | | 010.0 | RB  | RS    | 0 | sld.
    | N | 1 | RA  | | 010.1 | RB  | RS!=0 | 0 | srd.
    | N | 1 | RA  | | 010.1 | RB  | 000   | 0 | srad.
    | N | 1 | BF  | | 011.0 | RB  | RA|0  | 0 | cmpw

Notes:

* for srad, RS=RA: "srad. RA(=RS), RS, RB"


### Logical

    | 16-bit mode   | | 10-bit mode             |
    | 0 | 1 | 2 3 4 | | 567.8 | 9ab | c d e | f |
    | N | 0 |  RT   | | 100.0 | RB  | RA!=0 | M | and
    | N | 0 |  RT   | | 100.1 | RB  | RA!=0 | M | nand
    | N | 0 |  RT   | | 101.0 | RB  | RA!=0 | M | or
    | N | 0 |  RT   | | 101.1 | RB  | RA!=0 | M | nor
    | N | 0 |  RT   | | 100.0 | RB  | 0 0 0 | M | extsw
    | N | 0 |  RT   | | 100.1 | RB  | 0 0 0 | M | cntlz
    | N | 0 |  RT   | | 101.0 | RB  | 0 0 0 | M | popcnt
    | N | 0 |  RT   | | 101.1 | RB  | 0 0 0 | M | not

16-bit mode only:

    | 0 | 1 | 2 3 4 | | 567.8 | 9ab | c d e | f |
    | N | 1 |  RT   | | 100.0 | RB  | RA!=0 | 0 | TBD
    | N | 1 |  RT   | | 100.1 | RB  | RA!=0 | 0 | TBD
    | N | 1 |  RT   | | 101.0 | RB  | RA!=0 | 0 | xor
    | N | 1 |  RT   | | 101.1 | RB  | RA!=0 | 0 | eqv (xnor)
    | N | 1 |  RT   | | 100.0 | RB  | 0 0 0 | 0 | extsb
    | N | 1 |  RT   | | 100.1 | RB  | 0 0 0 | 0 | cnttz
    | N | 1 |  RT   | | 101.0 | RB  | 0 0 0 | 0 | TBD
    | N | 1 |  RT   | | 101.1 | RB  | 0 0 0 | 0 | extsh

10 bit mode:

* for (RA|0) when RA=0 the input is a zero immediate,
  meaning that nor becomes not
* cntlz, popcnt, exts **not available** in 10-bit mode
* RT is implicitly RB: "and RT(=RB), RA, RB"

### Floating Point

Note here that elwidth overrides (SV Prefix) can be used to select FP16/32/64

    | 16-bit mode   | | 10-bit mode             |
    | 0 | 1 | 2 3 4 | | 567.8 | 9ab | c d e | f |
    | N |   |  RT   | | 011.1 | RB  | RA!=0 | M | fsub.
    | N | 0 |  RT   | | 110.0 | RB  | RA!=0 | M | fadd
    | N | 0 |  RT   | | 110.1 | RB  | RA!=0 | M | fmul
    | N | 0 |  RT   | | 011.1 | RB  | 0 0 0 | M | fneg.
    | N | 0 |  RT   | | 110.0 | RB  | 0 0 0 | M |
    | N | 0 |  RT   | | 110.1 | RB  | 0 0 0 | M |

16-bit mode only:

    | 0 | 1 | 2 3 4 | | 567.8 | 9ab | c d e | f |
    | N | 1 |  RT   | | 011.1 | RB  | RA!=0 | 0 |
    | N | 1 |  RT   | | 110.0 | RB  | RA!=0 | 0 |
    | N | 1 |  RT   | | 110.1 | RB  | RA!=0 | 0 | fdiv
    | N | 1 |  RT   | | 011.1 | RB  | 0 0 0 | 0 | fabs.
    | N | 1 |  RT   | | 110.0 | RB  | 0 0 0 | 0 | fmr.
    | N | 1 |  RT   | | 110.1 | RB  | 0 0 0 | 0 |

16 bit only, FP to INT convert (using C 0b001.1 subencoding)

    | 0123 | 4 | | 567.8 | 9 ab | cde  | f |
    | 0010 | X | | 001.1 | 0 RA | Y RT | M | fp2int
    | 0011 | X | | 001.1 | 0 RA | Y RT | M | int2fp

* X: signed=1, unsigned=0
* Y: FP32=0, FP64=1

10 bit mode:

* fsub. fneg. and fmr. default target is CR1
* fmr. is **not available** in 10-bit mode
* fdiv is **not available** in 10-bit mode

16 bit mode:

* fmr. copies RB to RT (and sets CR1)

### Condition Register

    | 16-bit mode   | | 10-bit mode            |
    | 0 1 2 3 | 4   | | 567.8 | 9 ab | cde | f |
    | 0 0 0 0 | BF2 | | 001.1 | 0 BF | BFA | M | mcrf
    | 0 0 0 1 | BA2 | | 001.1 | 0 BA | BB  | M | crnor
    | 0 1 0 0 | BA2 | | 001.1 | 0 BA | BB  | M | crandc
    | 0 1 1 0 | BA2 | | 001.1 | 0 BA | BB  | M | crxor
    | 0 1 1 1 | BA2 | | 001.1 | 0 BA | BB  | M | crnand
    | 1 0 0 0 | BA2 | | 001.1 | 0 BA | BB  | M | crand
    | 1 0 0 1 | BA2 | | 001.1 | 0 BA | BB  | M | creqv
    | 1 1 0 1 | BA2 | | 001.1 | 0 BA | BB  | M | crorc
    | 1 1 1 0 | BA2 | | 001.1 | 0 BA | BB  | M | cror

10 bit mode:

* mcrf BF is only 2 bits which means the destination is only CR0-CR3
* CR operations: **not available** in 10-bit mode (but mcrf is)

16 bit mode:

* mcrf BF2 extends BF (in MSB) to 3 bits
* CR operations: destination register is same as BA.
* CR operations: only possible on CR0 and CR1

SV (Vector Mode):

* CR operations: greatly extended reach/range (useful for predicates)

### System

cbank: Selection of Compressed-encoding "Bank".  Different "banks"
give different meanings to opcodes.  Example: CBank=0b001 is heavily
optimised to A/Video Encode/Decode.  cbank borrows from add's encoding
space (when RA==0)

    | 16-bit mode | | 10-bit mode             |
    | 0 | 1 2 3 4 | | 567.8 | 9ab   | cde | f |
    | N | 0 Bank2 | | 010.0 | CBank | 000 | M | cbank

**not available** in 10-bit mode:

    | 0 1 2 3 | 4  | | 567.8 | 9 ab | cde  | f |
    | 1 1 1 1 | 0  | | 001.1 | 0 00 |  RT  | M | mtlr
    | 1 1 1 1 | 0  | | 001.1 | 0 01 |  RT  | M | mtctr
    | 1 1 1 1 | 0  | | 001.1 | 0 11 |  RT  | M | mtcr
    | 1 1 1 1 | 1  | | 001.1 | 0 00 |  RA  | M | mflr
    | 1 1 1 1 | 1  | | 001.1 | 0 01 |  RA  | M | mfctr
    | 1 1 1 1 | 1  | | 001.1 | 0 11 |  RA  | M | mfcr

### Unallocated

    | 0 1 2 3 | 4  | | 567.8 | 9 ab | cde  | f |
    | 0 1 0 1 |    | | 001.1 | 0    |      | M |
    | 1 0 1 0 |    | | 001.1 | 0    |      | M |
    | 1 0 1 1 |    | | 001.1 | 0    |      | M |
    | 1 1 0 0 |    | | 001.1 | 0    |      | M |
    | 1 1 1 1 |    | | 001.1 | 0 10 |      | M |