-------------------------
*** unexpected failure for jtod_circint(opt)
New back end thoughts
-----------------------------------------------------------------------------
Codegen notes
* jumps to ImpossibleBranch should be removed.
* Profiling:
- when updating a closure with an indirection to a function,
we should make a permanent indirection.
- check that we're bumping the scc count appropriately
* check perf & binary sizes against the HEAD
-----------------------------------------------------------------------------
C backend notes
* use STGCALL macros for foreign calls (doesn't look like volatile regs
are handled properly at the mo).
-----------------------------------------------------------------------------
Cmm parser notes
* switches
* need to cater for unexported procedures/info tables?
* We should be able to get rid of entry labels, use info labels only.
- we need a %ENTRY_LBL(info_lbl) macro, so that instead of
JMP_(foo_entry) we can write jump %ENTRY_LBL(foo_info).
-----------------------------------------------------------------------------
* Move arg-descr from LFInfo to ClosureInfo?
But: only needed for functions
* Move all of CgClosure.link_caf into NewCaf, and newDynCaf
* If the case binder is dead, and the constr is nullary,
do we need to assign to Node?
-------------------------
* Relation between separate type sigs and pattern type sigs
f :: forall a. a->a
f :: b->b = e -- No: monomorphic
f :: forall a. a->a
f :: forall a. a->a -- OK
f :: forall a. [a] -> [a]
f :: forall b. b->b = e ???
-------------------------------
NB: all floats are let-binds, but some non-rec lets
may be unlifted (with RHS ok-for-speculation)
simplArg: [use strictness]
[used for non-top-lvl non-rec RHS or function arg]
if strict-type || demanded
simplStrictExpr
else
simplExpr ---> (floats,expr)
float all the floats if exposes constr app, return expr
simpl (applied lambda) ==> simplNonRecBind
simpl (Let (NonRec ...) ..) ==> simplNonRecBind
simpl (Let (Rec ...) ..) ==> simplRecBind
simplRecBind:
simplify binders (but not its IdInfo)
simplify the pairs one at a time
using simplRecPair
simplNonRecBind: [was simplBeta]
[used for non-top-lvl non-rec bindings]
- check for PreInlineUnconditionally
- simplify binder, including its IdInfo
- simplArg
- if strict-type
addCaseBind [which makes a let if ok-for-spec]
else
completeLazyBind
simplLazyBind: [binder already simplified, but not its IdInfo]
[used for both rec and top-lvl non-rec]
[must not be strict/unboxed; case not allowed]
- check for PreInlineUnconditionally
- substituteIdInfo and add result to in-scope
[so that rules are available in rec rhs]
- simplExpr --> (floats,expr)
- float: lifted floats only
if exposes constructor or pap (even if non-triv args)
or if top level
- completeLazyBind
completeLazyBind: [given a simplified RHS]
[used for both rec and non-rec bindings, top level and not]
- try discarding dead
- try PostInlineUnconditionally
- let-bind coerce arg and repeat
- try rhs tylam (float)
- try eta expand (float) [not if any float is unlifted && (non-spec || top_lvl || rec)]
- let-bind constructor args [not if any float is ..as above..]
- add unfolding [this is the only place we add an unfolding]
add arity
Right hand sides and arguments
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In many ways we want to treat
(a) the right hand side of a let(rec), and
(b) a function argument
in the same way. But not always! In particular, we would
like to leave these arguments exactly as they are, so they
will match a RULE more easily.
f (g x, h x)
g (+ x)
It's harder to make the rule match if we ANF-ise the constructor,
or eta-expand the PAP:
f (let { a = g x; b = h x } in (a,b))
g (\y. + x y)
On the other hand if we see the let-defns
p = (g x, h x)
q = + x
then we *do* want to ANF-ise and eta-expand, so that p and q
can be safely inlined.
Even floating lets out is a bit dubious. For let RHS's we float lets
out if that exposes a value, so that the value can be inlined more vigorously.
For example
r = let x = e in (x,x)
Here, if we float the let out we'll expose a nice constructor. We did experiments
that showed this to be a generally good thing. But it was a bad thing to float
lets out unconditionally, because that meant they got allocated more often.
For function arguments, there's less reason to expose a constructor (it won't
get inlined). Just possibly it might make a rule match, but I'm pretty skeptical.
So for the moment we don't float lets out of function arguments either.
Eta expansion
~~~~~~~~~~~~~~
For eta expansion, we want to catch things like
case e of (a,b) -> \x -> case a of (p,q) -> \y -> r
If the \x was on the RHS of a let, we'd eta expand to bring the two
lambdas together. And in general that's a good thing to do. Perhaps
we should eta expand wherever we find a (value) lambda? Then the eta
expansion at a let RHS can concentrate solely on the PAP case.