| Age | Commit message (Collapse) | Author |
|---|
|
hipe_range_split is a complex live range splitter, more sophisticated
thatn hipe_restore_reuse, but still targeted specifically at temporaries
forced onto stack by being live over call instructions.
hipe_range_split partitions the control flow graph at call instructions,
like hipe_regalloc_prepass. Splitting decisions are made on a per
partition and per temporary basis.
There are three different ways in which hipe_range_split may choose to
split a temporary in a program partition:
* Mode1: Spill the temp before calls, and restore it after them
* Mode2: Spill the temp after definitions, restore it after calls
* Mode3: Spill the temp after definitions, restore it before uses
To pick which of these should be used for each temp×partiton pair,
hipe_range_split uses a cost function. The cost is simply the sum of the
cost of all expected stack accesses, and the cost for an individual
stack access is based on the probability weight of the basic block that
it resides in. This biases the range splitter so that it attempts moving
stack accesses from a functions hot path to the cold path.
hipe_bb_weights is used to compute the probability weights.
mode3 is effectively the same as what hipe_restore_reuse does. Because
of this, hipe_restore_reuse reuses the analysis pass of
hipe_restore_reuse in order to compute the minimal needed set of spills
and restores. The reason mode3 was introduced to hipe_range_split rather
than simply composing it with hipe_restore_reuse (by running both) is
that such a composition resulted in poor register allocation results due
to insufficiently strong move coalescing in the register allocator.
The cost function heuristic has a couple of tuning knobs:
* {range_split_min_gain, Gain} (default: 1.1, range: [0.0, inf))
The minimum proportional improvement that the cost of all stack
accesses to a temp must display in order for that temp to be split.
* {range_split_mode1_fudge, Factor} (default: 1.1, range: [0.0, inf))
Costs for mode1 are multiplied by this factor in order to discourage
it when it provides marginal benefits. The justification is that
mode1 causes temps to be live for longest, thus leading to higher
register pressure.
* {range_split_weight_power, Factor} (default: 2, range: (0.0, inf))
Adjusts how much effect the basic block weights have on the cost of a
stack access. A stack access in a block with weight 1.0 has cost 1.0,
a stack access in a block with weight 0.01 has cost 1/Factor.
Additionally, the option range_split_weights chooses whether the basic
block weights are used at all.
In the case that the input is very big, hipe_range_split automatically
falls back to hipe_restore_reuse only in order to keep compile times
under control. Note that this is not only because of hipe_range_split
being slow, but also due to the resulting program being slow to register
allocate, and is not as partitionable by hipe_regalloc_prepass.
hipe_restore_reuse, on the other hand, does not affect the programs
partitionability.
The hipe_range_split pass is controlled by a new option ra_range_split.
ra_range_split is added to o2, and ra_restore_reuse is disabled in o2.
|
|
hipe_restore_reuse is a simplistic range splitter that splits temps that
are forced onto the stack by being live over call instructions. In
particular, it attempts to avoid cases where there are several accesses
to such stack allocated temps in straight-line code, uninterrupted by
any calls. In order to achieve this it splits temps between just before
the first access(es) and just after the last access(es) in such
straight-line code groups.
The hipe_restore_reuse pass is controlled by a new option
ra_restore_reuse.
ra_restore_reuse is added to o1.
|
|
|
|
This allows us to pass around the context data that
hipe_regalloc_prepass needs cleanly, without using process dictionary or
parameterised modules (like it was previous to this change).
|
|
This is sound because the liveness data structure only stores liveness
info at basic block boundaries, and the rewrites that happen in
TargetSpecific:check_and_rewrite/2 preserves all existing definitions
and uses, and all new liveness intervals, belonging to newly introduced
temporaries, are always local to a basic block, and thus do not show up
in the liveout or livein sets for the basic block.
|
|
If temps introduced by hipe_regalloc_prepass end up above SpillLimit,
the register allocators will not spill them. This constraint is
unnecessarily limiting the allocators and might theoretically lead to
unallocatable programs (more temps above SpillLimit alive at a time than
there are physical registers).
|
|
|
|
Now that all backends do register allocation on a CFG directly and
define the defun_to_cfg/1 callback as the identity function, it can be
removed.
|
|
|
|
hipe_regalloc_prepass speeds up register allocation by spilling any temp
that is live over a call (which clobbers all register).
In order to detect these, a new function was added to the target
interface; defines_all_alloc/1, that takes an instruction and returns a
boolean.
|
|
For x86, additionally reuse liveness from float LSRA for the GP LSRA.
|
|
|
|
|
|
|
