* Make haltreq and resumereq proper write-only.

They already mostly were, but then I clarified some stuff and broke
keeping haltreq high during reset, which is the only mechanism to catch
a hart out of reset if optional reselthaltreq isn't implemented.

Also:
* resume ack is cleared by the DM when resumereq is written 1.
* haltreq can be high together with hartreset.

This is the minimal version of #378. It has all the same functionality,
but lacks some of #378's clarifications, which may be making this harder
to review.

* Leave haltreq high to catch a hart out of all resets.

Not just unexpected ones.

* Rephrase halt resume explanation.

* DM contains->tracks bits of state.

* Minimal implementations can't access all registers

* Provide more detail on minimal implementation.

For inches, apparently the thing to do is spell them out, so I did that.

This resolves a contradiction in the spec.
Fixes #365.

* Explicitly state what's required for compliance.

* Clarify "all registers."

* Clarify W1.

The definition used to be vague. I think this is the one that makes the
most sense, and shouldn't cause any problems. (Possibly we don't want
this for ackhavereset. I sent e-mail about that.)

The bits that are defined as W1 are:
DM: ackhavreset, setresethaltreq, clrresethaltreq
JTAG: dmihardreset, dmireset

* Revert my "clarification" of W1.

* Add dcsr.cause for being halted out of reset.

* Halt out of reset may be reported as just halted.

* Punctuation.

* Clarify halt out of reset due to resethaltreq.

* Clean up language, formatting, consistency.

* Remove unused R/W0, R/W1.

* Define haltresetreq.

* Add short name to register tables.

* Little language and formatting cleanups.

* Device Tree -> configuration string

The privileged spec talks about a configuration string, which may be a
Device Tree.

Now it's consistent with dtmcs.dmistat.

This reverts 7f0f09a0. It is no longer true that a debugger can assume
dmi.data contains a valid value unless a read was explicitly executed,
so there's no sense in accommodating that case.

* Distinguish draft and release builds.

Rename the output PDFs so it's less likely to cause confusion.
The release build is not labelled as DRAFT, does not include the preface
warning, and does not include the change log.

* Add serial.tex, which got lost somewhere.

(Only used by future.tex.)

* Fix build.

Also add riscv-debug-release.tex which I forgot earlier.

* Stepping over wfi does not enter wait state.

* Halt breaks wfi out of stalled state.

* Consistently use stall instead of wait.

Explicitly state which arguments abstract commands do or do not change.

Fixes #346.

* Harts can be in exactly 1 of 4 states.

Fixes #351.

* Grammar fixes.

* While held in reset, harts are unavailable.

Fixes #353.

* Remove "impl specific [reset] indicators"

They don't supersede the unavailable status, and their mention is
confusing. (They were added in ce1b3593.)

Fixes #353.

* Unavailable harts can't execute abstract commands.

Being in reset doesn't mean you have to be unavailable. #355

* Clarify reset states.

* Abstract behavior when hart is/becomes unavailable

Just a rewording that (hopefully) makes more sense.

* Mostly match "official" style for credits.

* Fix Megan's e-mail address.

Fixes #358.

* Abstract commands might work on a hung hart.

From a discussion on today's call.

* Rephrase.

Can't change harts during operations, and the current hart becoming unavailable may terminate the abstract command with error. (#322)

* Can't change harts during operations.

See #315.

* Only restrict hartsel changes during abstract cmds

reset/halt could take a very long time if a hart is asleep, and we don't
want to prevent debugging other harts in that case.

* Set cmderr for hartsel change during abstract cmd.

* Clarify reset and hasel during abstract command.

* Incorporate review feedback.

* Make debuggers handle harts becoming unavailable.

Instead of having the DM correctly handle a hart that becomes
unavailable, make debuggers responsible for that. This seems like a good
tradeoff because this should be rare.

* Clarifications requested in review.

* Clarify per review.

Fixes #301.

* Document breakpoint exception + enter debug mode

* Multiple triggers firing with different timing.

* Be a more lenient when trigger actions conflict

This came out of the July 11 meeting, where consensus was that requiring
both actions to be taken might significantly complicate hardware to
handle what is an uncommon corner case.

* Define MXLEN.

* Introduce DXLEN.

This lets us be more precise about register widths, to deal with the
fact that MXLEN can be changed by software.

* Mention debugger's responsibility when XLEN!=DXLEN

* DXLEN is the widest supported XLEN.

This simplifies implementations and validation by explicitly preventing
the debugger from doing things that aren't very useful anyway.

Fixes #320.

Note, the e.g. must be followed by a backslash to tell LaTeX that it
isn't the end of a sentence.

$0x400$ tells LaTeX to read "0 x 400" as math, not what you want

This is an alternate to #326.