add open question trap_isans as a vector table

[libreriscv.git] / isa_conflict_resolution / isamux_isans.mdwn

# Note-form on ISAMUX (aka "ISANS")

A fixed number of additional (hidden) bits, conceptually a "namespace", that go directly and non-optionally
into the instruction decode phase, extending (in each implementation) the
opcode length to 16+N, 32+N, 48+N, where N is a hard fixed quantity on
a per-implementor basis.

Where the opcode is normally loaded from the location at the PC, the extra
bits are instead set via a CSR and mandatorially appended to every instruction: hence why they are described as "hidden" opcode bits, and as a "namespace".

The parallels with c++ "using namespace" are direct and clear.

# Hypothetical Format

Note that this is a hypothetical format, yet TBD, where particular attention
needs to be paid to the fact that there is an "immediate" version of CSRRW
(with 5 bits of immediate) that could save a lot of space in binaries.

<code>
<pre>
      3                   2                   1
   |1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0|
   |------------------------------ |-------|---------------------|-|
   |reserved reserved reserved reserved reserved   | foreignarch |1|
   |custom         | reserved    |           official|B| rvcpage |0|
</pre>
<code>

RV Mode

* when bit 0 is 0, "RV" mode is selected.
* in RV mode, bits 1 thru 5 provide up to 16 possible alternative meanings (namespaces) for 16 Bit opcodes. "pages" if you will. The top bit indicates custom meanings. When set to 0, the top bit is for official usage.
* Bits 15 thru 23 are reserved.
* Bits 24 thru 31 are for custom usage.
* bit 6 ("B") is LE/BE

16 bit page examples:

* 0b0000 STANDARD (2019) RVC
* 0b0001 RVCv2
* 0b0010 RV16
* 0b0011 RVCv3
* ...
* 0b1000 custom 16 bit opcode meanings 1
* 0b1001 custom 16 bit opcode meanings 2
* .....

Foreign Arch Mode

* when bit 0 is 1, "Foreign arch" mode is selected.
* Bits 1 thru 7 are a table of foreign arches.
* when the MSB is 1, this is for custom use.
* when the MSB is 0, bits 1 thru 6 are reserved for 64 possible official foreign archs.

Foreign archs could be (examples):

* 0b0000000 x86_32
* 0b0000001 x86_64
* 0b0000010 MIPS32
* 0b0000011 MIPS64
* ....
* 0b0010000 Java Bytecode
* 0b0010001 N.E.Other Bytecode
* ....
* 0b1000000 custom foreign arch 1
* 0b1000001 custom foreign arch 2
* ....

Note that "official" foreign archs have a binary value where the MSB is zero,
and custom foreign archs have a binary value where the MSB is 1.

# Privileged Modes / Traps <a name="privtraps"></a>

An additional WLRL CSR per priv-level named "LAST-ISANS" is required, and
another called "TRAP-ISANS"
These mirrors the ISANS CSR, and, on a trap, if the current ISANS in
that privilege level is not equal to TRAP-ISANS, its value is atomically
transferred into LAST-ISANS by the hardware, and ISANS in that trap
is set to TRAP-ISANS.  Hardware is *only then* permitted to modify the PC to
begin execution of the trap.

On exit from the trap, hardware must check to see if LAST-ISANS is equal
to TRAP-ISANS.  If it is not, LAST-ISANS is copied into the ISANS CSR,
LAST-ISANS is set to TRAP-ISANS, and *only then* is the hardware permitted
to modify the PC to begin execution where the trap left off.

Note 1: in the case of Supervisor Mode (context switches in particular),
saving and changing of LAST-ISANS (to and from the stack) must be done
atomically and under the protection of the SIE bit.  Failure to do so
could result in corruption of LAST-ISANS when multiple traps occur in
the same privilege level.

Note 2: question - should the trap due to illegal (unsupported) values
written into LAST-ISANS occur when the *software* writes to LAST-ISANS,
or when the *trap* (on exit) writes into LAST-ISANS?  this latter seems
fraught: a trap, on exit, causing another trap??

Per-privilege-level pseudocode (there exists UISANS, UTRAPISANS, ULASTISANS,
MISANS, MTRAPISANS, MLASTISANS and so on):

<code>
<pre>
trap_entry()
{
    if (ISANS != TRAP_ISANS) // musn't change if already there
    {
        LAST-ISANS = ISANS // record the old NS
        ISANS = TRAP_ISANS // traps are executed in "trap" NS
    }
}

and trap_exit:

trap_exit():
{
    if (LAST-ISANS != TRAP_ISANS)
    {
        ISANS = LAST-ISANS
        LAST-ISANS = TRAP_ISANS
    }
}
</pre>
</code>

# Why not have TRAP-ISANS as a vector table, matching mtvec? <a name="trap-isans-vec"></a>

Use case to be determined.  Rather than be a global per-priv-level value,
TRAP-ISANS is a table of length exactly equal to the mtvec/utvec/stvec table,
with corresponding entries that specify the assembly-code namespace in which
the trap handler routine is written.

Open question: see <https://groups.google.com/a/groups.riscv.org/d/msg/isa-dev/IAhyOqEZoWA/BM0G3J2zBgAJ>

<code>
<pre>
trap_entry(x_cause)
{
    if (ISANS != TRAP_ISANS[xcause]) // musn't change if already there
    {
        LAST-ISANS = ISANS // record the old NS
        ISANS = TRAP_ISANS[xcause] // traps are executed in "trap" NS
    }
}

and trap_exit:

trap_exit(x_cause):
{
    if (LAST-ISANS != TRAP_ISANS[x_cause])
    {
        ISANS = LAST-ISANS
        LAST-ISANS = TRAP_ISANS[x_cause]
    }
}
</pre>
</code>

# What happens if this scheme is not adopted? Why is it better than leaving things well alone? <a name="lassezfaire"></a>

At the first sign of an emergency non-backwards compatible and unavoidable
change to the *frozen* RISCV *official* Standards, the entire RISCV
community is fragmented and divided into two:

* Those vendors that are hardware compatible with the legacy standard.
* Those that are compatible with the new standard.

*These two communities would be mutually exclusively incompatible*. If
a second emergency occurs, RISCV becomes even less tenable.

Hardware that wished to be "compatible" with either flavour would require
JIT or offline static binary recompilation. No vendor would willingly
accept this as a condition of the standards divergence in the first place,
locking up decision making to the detriment of RISCV as a whole.

By providing a "safety valve" in the form of a hidden namespace, at least
newer hardware has the option to implement both (or more) variations,
*and still apply for Certification*.

However to also allow "legacy" hardware to at least be JIT soft
compatible, some very strict rules *must* be adhered to, that appear at
first sight not to make any sense.

It's complicated in other words!

# Surely it's okay to just tell people to use 48-bit encodings? <a name="use48bit"></a>

Short answer: it doesn't help resolve conflicts, and costs hardware and
redesigns to do so.  Softcores in cost-sensitive embedded applications may
even not actually be able to fit the required 48 bit instruction decode engine
into a (small, ICE40) FPGA.  48-bit instruction decoding is much more complex
than straight 32-bit decoding, requiring a queue.

Second answer: conflicts can still occur in the (unregulated, custom) 48-bit
space, which *could* be resolved by ISAMUX/ISANS as applied to the *48* bit
space in exactly the same way.  And the 64-bit space.

# Why not leave this to individual custom vendors to solve on a case by case basis?

The suggestion was raised that a custom extension vendor could create
their own CSR that selects between conflicting namespaces that resolve
the meaning of the exact same opcode.  This to be done by all and any
vendors, as they see fit, with little to no collaboration or coordination
towards standardisation in any form.

The problems with this approach are numerous, when presented to a
worldwide context that the UNIX Platform, in particular, has to face
(where the embedded platform does not)

First: lack of coordination, in the proliferation of arbitrary solutions,
has to primarily be borne by gcc, binutils, LLVM and other compilers.

Secondly: CSR space is precious. With each vendor likely needing only one
or two bits to express the namespace collision avoidance, if they make
even a token effort to use worldwide unique CSRs (an effort that would
benefit compiler writers), the CSR register space is quickly exhausted.

Thirdly: JIT Emulation of such an unregulated space becomes just as
much hell as it is for compiler writers. In addition, if two vendors
use conflicting CSR addresses, the only sane way to tell the emulator
what to do is to give the emulator a runtime commandline argument.

Fourthly: with each vendor coming up with their own way of handling
conflicts, not only are the chances of mistakes higher, it is against the
very principles of collaboration and cooperation that save vendors money
on development and ongoing maintenance.  Each custom vendor will have
to maintain their own separate hard fork of the toolchain and software,
which is well known to result in security vulnerabilities.

By coordinating and managing the allocation of namespace bits (unary
or binary) the above issues are solved. CSR space is no longer wasted,
compiler and JIT software writers have an easier time, clashes are
avoided, and RISCV is stabilised and has a trustable long term future.

# Why ISAMUX / ISANS has to be WLRL and mandatory trap on illegal writes <a name="wlrlmandatorytrap"></a>

The namespaces, set by bits in the CSR, are functionally directly
equivalent to c++ namespaces, even down to the use of braces.

WARL, by allowing implementors to choose the value, prevents and prohibits
the critical and necessary raising of an exception that would begin the
JIT process in the case of ongoing standards evolution.

Without this opportunity, an implementation has no way of knowing
how to JIT emulate any given conflicting opcode. It is as if the c++
standard was given the similar opportunity to completely ignore the
"using namespace" prefix!

--

Ok so I trust it's now clear why WLRL (thanks Allen) is needed.

When Dan raised the WARL concern initially a situation was masked by
the conflict, that if gone unnoticed would jeapordise ISAMUX/ISANS
entirely. Actually, two separate errors. So thank you for raising the
question.

The situation arises when foreign archs are to be given their own NS
bit. MIPS is allocated bit 8, x86 bit 9, whilst LE/BE is given bit 0,
RVCv2 bit 1 andso on. All of this potential rather than actual, clearly.

Imagine then that software tries to write and set not just bit 8 and
bit 9, it also tries to set bit 0 and 1 as well.

This *IS* on the face of it a legitimate reason to make ISAMUX/ISANS WARL.

However it masks a fundamental flaw that has to be addressed, which
brings us back much closer to the original design of 18 months ago,
and it's highlighted thus:

x86 and simultaneous RVCv2 modes are total nonsense in the first place!

The solution instead is to have a NS bit (bit0) that SPECIFICALLY
determines if the arch is RV or not.  If 0, the rest of the ISAMUX/ISANS
is very specifically RV *only*, and if 1, the ISAMUX/ISANS is a *binary*
table of foreign architectures and foreign architectures only.

Exactly how many bits are used for the foreign arch table, is to
be determined. 7 bits, one of which is reserved for custom usage,
leaving a whopping 64 possible "official" foreign instruction sets to
be hardware-supported/JIT-emulated seems to be sufficiently gratuitous,
to me.

One of those could even be Java Bytecode!

Now, it could *hypothetically* be argued that the permutation of setting
LE/BE and MIPS for example is desirable. A simple analysis shows this
not to be the case: once in the MIPS foreign NS, it is the MIPS hardware
implementation that should have its own way of setting and managing its
LE/BE mode, because to do otherwise drastically interferes with MIPS
binary compatibility.

Thus, it is officially Not Our Problem: only flipping into one foreign
arch at a time makes sense, thus this has to be reflected in the
ISAMUX/ISANS CSR itself, completely side-stepping the (apparent) need
to make the NS CSR WARL (which would not work anyway, as previously
mentioned).

So, thank you, again, Dan, for raising this. It would have completely
jeapordised ISAMUX/NS if not spotted.

The second issue is: how does any hardware system, whether it support
ISANS or not, and whether any future hardware supports some Namespaces
and, in a transitive fashion, has to support *more* future namespaces,
through JIT emulation, if this is not planned properly in advance?

Let us take the simple case first: a current 2019 RISCV fully compliant
RV64GC UNIX capable system (with mandatory traps on all unsupported CSRs).

Fast forward 20 years, there are now 5 ISAMUX/NS unary bits, and 3
foreign arch binary table entries.

Such a system is perfectly possible of software JIT emulating ALL of these
options because the write to the (illegal, for that system) ISAMUX/NS
CSR generates the trap that is needed for that system ti begin JIT mode.

(This again emphasises exactly why the trap is mandatory).

Now let us take the case of a hypothetical system from say 2021 that
implements RVCv2 at the hardware level.

Fast forward 20 years: if the CSR were made WARL, that system would be
absolutely screwed. The implementor would be under the false impression
that ignoring setting of "illegal" bits was acceptable, making the
transition to JIT mode flat-out impossible to detect.

When this is considered transitively, considering all future additions to
the NS, and all permutations, it can be logically deduced that there is
a need to reserve a *full* set of bits in the ISAMUX/NS CSR *in advance*.

i.e. that *right now*, in the year 2019, the entire ISAMUX/NS CSR cannot
be added to piecemeal, the full 32 (or 64) bits *has* to be reserved,
and reserved bits set at zero.

Furthermore, if any software attempts to write to those reserved bits,
it *must* be treated just as if those bits were distinct and nonexistent
CSRs, and a trap raised.

It makes more sense to consider each NS as having its own completely
separate CSR, which, if it does not exist, clearly it should be obvious
that, as an unsupported CSR, a trap should be raised (and JIT emulation
activated).

However given that only the one bit is needed (in RV NS Mode, not
Foreign NS Mode), it would be terribly wasteful of the CSRs to do this,
despite it being technically correct and much easier to understand why
trap raising is so essential (mandatory).

This again should emphasise how to mentally get one's head round this
mind-bendingly complex problem space: think of each NS bit as its own
totally separate CSR that every implementor is free and clear to implement
(or leave to JIT Emulation) as they see fit.

Only then does the mandatory need to trap on write really start to hit
home, as does the need to preallocate a full set of reserved zero values
in the RV ISAMUX/NS.

Lastly, I *think* it's ok to only reserve say 32 bits, and, in 50 years
time if that genuinely is not enough, start the process all over again
with a new CSR.  ISAMUX2/NS2.

Subdivision of the RV NS (support for RVCv3/4/5/RV16 without wasting
precious CSR bits) best left for discussion another time, the above is
a heck of a lot to absorb, already.

# Alternative RVC 16 Bit Opcode meanings

Ok, here is appropriate to raise an idea how to cover RVC and future
variants, including RV16.

Just as with foreign archs, and you quite rightly highlight above, it
makes absolutely no sense to try to select both RVCv1, v2, v3 and so on,
all simultaneously. An unary bit vector for RVC modes, changing the 16
BIT opcode space meaning, is wasteful and again has us believe that WARL
is the "solution".

The correct thing to do is, again, just like with foreign archs, to
treat RVCs as a *binary* namespace selector. Bits 1 thru 3 would give
8 possible completely new alternative meanings, just like how the Z80
and the 286 and 386 used to do bank switching.

All zeros is clearly reserved for the present RVC. 0b001 for RVCv2. 0b010
for RV16 (look it up) and there should definitely be room reserved here
for custom reencodings of the 16 bit opcode space.

# Why WARL will not work and why WLRL is required

WARL requires a follow-up read of the CSR to ascertain what heuristic
the hardware *might* have applied, and if that procedure is followed in
this proposal, performance even on hardware would be severely compromised.

In addition when switching to foreign architectures, the switch has to
be done atomically and guaranteed to occur.

In the case of JIT emulation, the WARL "detection" code will be in an
assembly language that is alien to hardware.

Support for both assembly languages immediately after the CSR write
is clearly impossible, this leaves no other option but to have the CSR
be WLRL (on all platforms) and for traps to be mandatory (on the UNIX
Platform).

# Is it strictly necessary for foreign archs to switch back? <a name="foreignswitch"></a>

No, because LAST-ISANS handles the setting and unsetting of the ISANS CSR
in a completely transparent fashion as far as the foreign arch is concerned.

Thus, in e.g. Hypervisor Mode, the foreign guest arch has no knowledge
or need to know that the hypervisor is flipping back to RV at the time of
a trap.

Note however that this is **not** the same as the foreign arch executing
*foreign* traps!  Foreign architecture trap and interrupt handling mechanisms
are **out of scope** of this document and MUST be handled by the foreign
architecture implementation in a completely transparent fashion that in
no way interacts or interferes with this proposal.