| Age | Commit message (Collapse) | Author |
|---|
|
Some huge functions would compile very slowly because of a bottleneck
in `beam_ssa:def_used/2`. One example is the `cuter_binlib` module in
https://github.com/cuter-testing/cuter. On my computer, this commit
reduces the compilatation time for `cuter_binlib` to 45 seconds down
from more than 4 minutes.
Noticed-by: Kostis Sagonas
|
|
Eliminate the Dialyzer warnings shown when the limits in
lib/cerl/erl_types.erl were raised as follows:
-define(TUPLE_TAG_LIMIT, 10).
-define(TUPLE_ARITY_LIMIT, 10).
-define(SET_LIMIT, 64).
|
|
The expansion of record field updates, when more than one field is
updated, but not a majority of the fields, will create a sequence of
calls to `erlang:setelement(Index, Value, Tuple)` where Tuple in the
first call is the original record tuple, and in the subsequent calls
Tuple is the result of the previous call. Furthermore, all Index
values are constant positive integers, and the first call to
`setelement` will have the greatest index. Thus all the following
calls do not actually need to test at run-time whether Tuple has type
tuple, nor that the index is within the tuple bounds.
Since OTP R7, the `sys_core_dsetel` pass, run as the very last Core
Erlang pass, has optimized this sequence of `setelement` calls to use
a special destructive version of `setelement` (called
`set_tuple_element`) for all but the very first `setelement` in the
sequence.
It turns out that the presence of the `set_tuple_element` in SSA code
is awkward and can prevent or complicate type analysis and aggressive
optimizations.
Therefore, this commit removes the `sys_core_dsetel` pass and
reimplements it for SSA code. The optimization will be done in the
`beam_ssa_pre_codegen` pass (that is, just before code generation and
after running all other SSA code optimization passes).
In most cases, the resulting BEAM code is identical to previous
code. For a few modules, the BEAM code is actually slightly better,
with smaller stack frames.
|
|
The ssa_opt_sink optimization of beam_ssa_opt could get very slow
for certain huge functions. 9a190cae9bd7 partly addressed this issue
by terminating the optimization early if there happened to be no
get_tuple_element instructions at all in the function.
This commit addresses the issue more directly by making the dominator
calculation in beam_ssa:dominators/1 more efficient. The same
algorithm as before is used, but it is implemented in a more efficient
way based on the ideas in "A Simple, Fast Dominance Algorithm"
(http://www.hipersoft.rice.edu/grads/publications/dom14.pdf).
As well as being more efficient, the new implementation also gives
an explicit representation of the dominator tree, which makes it
possible to simplify and optimize the ssa_opt_sink optimization.
|
|
beam_ssa:def_used/2 is used by beam_ssa_pre_codegen when reserving
Y registers.
Do the following optimizations:
* Use an ordset instead of a gb_set. When the only operation performed
on a set is union/2, an ordset will usually be faster, especially when
the result is an ordset.
* Use a cerl_set instead of a gb_set for the set of all possible
predecessors. cerl_sets is usually faster than gb_sets.
|
|
Avoiding calls usually reduces the size of the stack frame and reduces
register shuffling.
|
|
The `get_map_element` instruction has no side effects, and should be
removed if its value is not used.
|
|
* bjorn/compiler/misc-fixes:
beam_ssa: Remove unnecessary beam_ssa: prefixes
beam_ssa_bsm: Fix replacement of variables in a remote call
|
|
|
|
If a fun is defined locally and only used for calls, it can be replaced
with direct calls to the relevant function. This greatly speeds up "named
functions" (which rely on make_fun to recreate themselves) and macros that
wrap their body in a fun.
|
|
This has been superseded by bs_get_tail/3. Note that it is NOT
removed from the emulator or beam_disasm, as old modules are still
legal.
|
|
get_anno/3: as get_anno but with a default value
definitions/1-2: returns a map of variable definitions (#b_set{})
uses/1-2: returns a map of all uses of a given variable
mapfold_blocks_rpo/4: mapfolds over blocks
|
|
We chose to refer to variables through their var_name() because we
anticipated the need to annotate them, but it turned out we didn't
really need that, and many things become a lot cleaner if the
entire #b_var{} is used to represent variables.
|
|
Add more instructions to the list of functions that can be safely
removed if their values are not used. This is necessary for
correctness when doing more aggressive optimizations. Without this
change, the 'succeeded' instruction could be optimized away leaving
just the instruction followed by an unconditional branch, which the
beam_ssa_codegen does not know how to handle. Here is an example:
_3 = bs_start_match _1
br label 13
By adding bs_start_match to the list, the bs_start_match instruction
will be removed too. (If the result of bs_start_match is actually
used, the succeeded instruction would not be removed.)
While we are it, rename the misnamed function is_pure/1 to
no_side_effect/1 and move it to beam_ssa. is_pure/1 is a bad name
because bif:get has no side effect, but is not pure.
|
|
It is faster to use cerl_sets instead of gb_sets to keep track of
seen blocks.
|
|
Add trim_unreachable/1 to remove unreachable blocks
and adjust phi nodes.
|
|
Since beam_ssa:linearize/1 may remove blocks that are unreachable,
adjust phi nodes to make sure that they don't refer to discarded
blocks or to blocks that no longer branch to the phi node in
question.
|
|
Add normalize/1 to simplify optimizations.
|
|
v3_codegen is replaced by three new passes:
* beam_kernel_to_ssa which translates the Kernel Erlang format
to a new SSA-based intermediate format.
* beam_ssa_pre_codegen which prepares the SSA-based format
for code generation, including register allocation. Registers
are allocated using the linear scan algorithm.
* beam_ssa_codegen which generates BEAM assembly code from the
SSA-based format.
It easier and more effective to optimize the SSA-based format before X
and Y registers have been assigned. The current optimization passes
constantly have to make sure no "holes" in the X register assignments
are created (that is, that no X register becomes undefined that an
allocation instruction depends on).
This commit also introduces the following optimizations:
* Replacing of tuple matching of records with the is_tagged_tuple
instruction. (Replacing beam_record.)
* Sinking of get_tuple_element instructions to just before the first
use of the extracted values. As well as potentially avoiding
extracting tuple elements when they are not actually used on all
executions paths, this optimization could also reduce the number
values that will need to be stored in Y registers. (Similar to
beam_reorder, but more effective.)
* Live optimizations, removing the definition of a variable that is
not subsequently used (provided that the operation has no side
effects), as well strength reduction of binary matching by replacing
the extraction of value from a binary with a skip instruction. (Used
to be done by beam_block, beam_utils, and v3_codegen.)
* Removal of redundant bs_restore2 instructions. (Formerly done
by beam_bs.)
* Type-based optimizations across branches. More effective than
the old beam_type pass that only did type-based optimizations in
basic blocks.
* Optimization of floating point instructions. (Formerly done
by beam_type.)
* Optimization of receive statements to introduce recv_mark and
recv_set instructions. More effective with far fewer restrictions
on what instructions are allowed between creating the reference
and entering the receive statement.
* Common subexpression elimination. (Formerly done by beam_block.)
|
