* Revert "egraphs: disable GVN of effectful idempotent ops (temporarily). (#5808)"
This reverts commit c7e2571866.
* egraphs: fix handling of effectful-but-idempotent ops and GVN.
This PR addresses #5796: currently, ops that are effectful, i.e., remain
in the side-effecting skeleton (which we keep in the `Layout` while the
egraph exists), but are idempotent and thus mergeable by a GVN pass, are
not handled properly.
GVN is still possible on effectful but idempotent ops precisely because
our GVN does not create partial redundancies: it removes an instruction
only when it is dominated by an identical instruction. An isntruction
will not be "hoisted" to a point where it could execute in the optimized
code but not in the original.
However, there are really two parts to the egraph implementation that
produce this effect: the deduplication on insertion into the egraph, and
the elaboration with a scoped hashmap. The deduplication lets us give a
single name (value ID) to all copies of an identical instruction, and
then elaboration will re-create duplicates if GVN should not hoist or
merge some of them.
Because deduplication need not worry about dominance or scopes, we use a
simple (non-scoped) hashmap to dedup/intern ops as "egraph nodes".
When we added support for GVN'ing effectful but idempotent ops (#5594),
we kept the use of this simple dedup'ing hashmap, but these ops do not
get elaborated; instead they stay in the side-effecting skeleton. Thus,
we inadvertently created potential for weird code-motion effects.
The proposal in #5796 would solve this in a clean way by treating these
ops as pure again, and keeping them out of the skeleton, instead putting
"force" pseudo-ops in the skeleton. However, this is a little more
complex than I would like, and I've realized that @jameysharp's earlier
suggestion is much simpler: we can keep an actual scoped hashmap
separately just for the effectful-but-idempotent ops, and use it to GVN
while we build the egraph. In effect, we're fusing a separate GVN pass
with the egraph pass (but letting it interact corecursively with
egraph rewrites. This is in principle similar to how we keep a separate
map for loads and fuse this pass with the egraph rewrite pass as well.
Note that we can use a `ScopedHashMap` here without the "context" (as
needed by `CtxHashMap`) because, as noted by @jameysharp, in practice
the ops we want to GVN have all their args inline. Equality on the
`InstructinoData` itself is conservative: two insts whose struct
contents compare shallowly equal are definitely identical, but identical
insts in a deep-equality sense may not compare shallowly equal, due to
list indirection. This is fine for GVN, because it is still sound to
skip any given GVN opportunity (and keep the original instructions).
Fixes#5796.
* Add comments from review.
This is a short-term fix to the same bug that #5800 is addressing
(#5796), but with less risk: it simply turns off GVN'ing of effectful
but idempotent ops. Because we have an upcoming release, and this is a
miscompile (albeit to do with trapping behavior), we would like to make
the simplest possible fix that avoids the bug, and backport it. I will
then rebase #5800 on top of a revert of this followed by the more
complete fix.
Rework br_table to use BlockCall, allowing us to avoid adding new nodes during ssa construction to hold block arguments. Additionally, many places where we previously matched on InstructionData to extract branch destinations can be replaced with a use of branch_destination or branch_destination_mut.
Move the storage for jump tables off of FunctionStencil and onto DataFlowGraph. This change is in service of #5731, making it easier to access the jump table data in the context of helpers like inst_values.
We don't have overlap in behavior for branch instructions anymore, so we can remove analyze_branch and instead match on the InstructionData directly.
Co-authored-by: Jamey Sharp <jamey@minilop.net>
Add a conditional branch instruction with two targets: brif. This instruction will eventually replace brz and brnz, as it encompasses the behavior of both.
This PR also changes the InstructionData layout for instruction formats that hold BlockCall values, taking the same approach we use for Value arguments. This allows branch_destination to return a slice to the BlockCall values held in the instruction, rather than requiring that we pattern match on InstructionData to fetch the then/else blocks.
Function generation for fuzzing has been updated to generate uses of brif, and I've run the cranelift-fuzzgen target locally for hours without triggering any new failures.
* Support mergeable-but-side-effectful (idempotent) operations in general in the egraph's GVN.
This mirrors the similar change made in #5534.
* Add tests for egraph case.
Add a new type BlockCall that represents the pair of a block name with arguments to be passed to it. (The mnemonic here is that it looks a bit like a function call.) Rework the implementation of jump, brz, and brnz to use BlockCall instead of storing the block arguments as varargs in the instruction's ValueList.
To ensure that we're processing block arguments from BlockCall values in instructions, three new functions have been introduced on DataFlowGraph that both sets of arguments:
inst_values - returns an iterator that traverses values in the instruction and block arguments
map_inst_values - applies a function to each value in the instruction and block arguments
overwrite_inst_values - overwrite all values in an instruction and block arguments with values from the iterator
Co-authored-by: Jamey Sharp <jamey@minilop.net>
We have some operations defined on DataFlowGraph purely to work around borrow-checker issues with InstructionData and other data on DataFlowGraph. Part of the problem is that indexing the DFG directly hides the fact that we're only indexing the insts field of the DFG.
This PR makes the insts field of the DFG public, but wraps it in a newtype that only allows indexing. This means that the borrow checker is better able to tell when operations on memory held by the DFG won't conflict, which comes up frequently when mutating ValueLists held by InstructionData.
* egraph support: rewrite to work in terms of CLIF data structures.
This work rewrites the "egraph"-based optimization framework in
Cranelift to operate on aegraphs (acyclic egraphs) represented in the
CLIF itself rather than as a separate data structure to which and from
which we translate the CLIF.
The basic idea is to add a new kind of value, a "union", that is like an
alias but refers to two other values rather than one. This allows us to
represent an eclass of enodes (values) as a tree. The union node allows
for a value to have *multiple representations*: either constituent value
could be used, and (in well-formed CLIF produced by correct
optimization rules) they must be equivalent.
Like the old egraph infrastructure, we take advantage of acyclicity and
eager rule application to do optimization in a single pass. Like before,
we integrate GVN (during the optimization pass) and LICM (during
elaboration).
Unlike the old egraph infrastructure, everything stays in the
DataFlowGraph. "Pure" enodes are represented as instructions that have
values attached, but that are not placed into the function layout. When
entering "egraph" form, we remove them from the layout while optimizing.
When leaving "egraph" form, during elaboration, we can place an
instruction back into the layout the first time we elaborate the enode;
if we elaborate it more than once, we clone the instruction.
The implementation performs two passes overall:
- One, a forward pass in RPO (to see defs before uses), that (i) removes
"pure" instructions from the layout and (ii) optimizes as it goes. As
before, we eagerly optimize, so we form the entire union of optimized
forms of a value before we see any uses of that value. This lets us
rewrite uses to use the most "up-to-date" form of the value and
canonicalize and optimize that form.
The eager rewriting and acyclic representation make each other work
(we could not eagerly rewrite if there were cycles; and acyclicity
does not miss optimization opportunities only because the first time
we introduce a value, we immediately produce its "best" form). This
design choice is also what allows us to avoid the "parent pointers"
and fixpoint loop of traditional egraphs.
This forward optimization pass keeps a scoped hashmap to "intern"
nodes (thus performing GVN), and also interleaves on a per-instruction
level with alias analysis. The interleaving with alias analysis allows
alias analysis to see the most optimized form of each address (so it
can see equivalences), and allows the next value to see any
equivalences (reuses of loads or stored values) that alias analysis
uncovers.
- Two, a forward pass in domtree preorder, that "elaborates" pure enodes
back into the layout, possibly in multiple places if needed. This
tracks the loop nest and hoists nodes as needed, performing LICM as it
goes. Note that by doing this in forward order, we avoid the
"fixpoint" that traditional LICM needs: we hoist a def before its
uses, so when we place a node, we place it in the right place the
first time rather than moving later.
This PR replaces the old (a)egraph implementation. It removes both the
cranelift-egraph crate and the logic in cranelift-codegen that uses it.
On `spidermonkey.wasm` running a simple recursive Fibonacci
microbenchmark, this work shows 5.5% compile-time reduction and 7.7%
runtime improvement (speedup).
Most of this implementation was done in (very productive) pair
programming sessions with Jamey Sharp, thus:
Co-authored-by: Jamey Sharp <jsharp@fastly.com>
* Review feedback.
* Review feedback.
* Review feedback.
* Bugfix: cprop rule: `(x + k1) - k2` becomes `x - (k2 - k1)`, not `x - (k1 - k2)`.
Co-authored-by: Jamey Sharp <jsharp@fastly.com>
Remove the boolean types from cranelift, and the associated instructions breduce, bextend, bconst, and bint. Standardize on using 1/0 for the return value from instructions that produce scalar boolean results, and -1/0 for boolean vector elements.
Fixes#3205
Co-authored-by: Afonso Bordado <afonso360@users.noreply.github.com>
Co-authored-by: Ulrich Weigand <ulrich.weigand@de.ibm.com>
Co-authored-by: Chris Fallin <chris@cfallin.org>
* egraph-based midend: draw the rest of the owl.
* Rename `egg` submodule of cranelift-codegen to `egraph`.
* Apply some feedback from @jsharp during code walkthrough.
* Remove recursion from find_best_node by doing a single pass.
Rather than recursively computing the lowest-cost node for a given
eclass and memoizing the answer at each eclass node, we can do a single
forward pass; because every eclass node refers only to earlier nodes,
this is sufficient. The behavior may slightly differ from the earlier
behavior because we cannot short-circuit costs to zero once a node is
elaborated; but in practice this should not matter.
* Make elaboration non-recursive.
Use an explicit stack instead (with `ElabStackEntry` entries,
alongside a result stack).
* Make elaboration traversal of the domtree non-recursive/stack-safe.
* Work analysis logic in Cranelift-side egraph glue into a general analysis framework in cranelift-egraph.
* Apply static recursion limit to rule application.
* Fix aarch64 wrt dynamic-vector support -- broken rebase.
* Topo-sort cranelift-egraph before cranelift-codegen in publish script, like the comment instructs me to!
* Fix multi-result call testcase.
* Include `cranelift-egraph` in `PUBLISHED_CRATES`.
* Fix atomic_rmw: not really a load.
* Remove now-unnecessary PartialOrd/Ord derivations.
* Address some code-review comments.
* Review feedback.
* Review feedback.
* No overlap in mid-end rules, because we are defining a multi-constructor.
* rustfmt
* Review feedback.
* Review feedback.
* Review feedback.
* Review feedback.
* Remove redundant `mut`.
* Add comment noting what rules can do.
* Review feedback.
* Clarify comment wording.
* Update `has_memory_fence_semantics`.
* Apply @jameysharp's improved loop-level computation.
Co-authored-by: Jamey Sharp <jamey@minilop.net>
* Fix suggestion commit.
* Fix off-by-one in new loop-nest analysis.
* Review feedback.
* Review feedback.
* Review feedback.
* Use `Default`, not `std::default::Default`, as per @fitzgen
Co-authored-by: Nick Fitzgerald <fitzgen@gmail.com>
* Apply @fitzgen's comment elaboration to a doc-comment.
Co-authored-by: Nick Fitzgerald <fitzgen@gmail.com>
* Add stat for hitting the rewrite-depth limit.
* Some code motion in split prelude to make the diff a little clearer wrt `main`.
* Take @jameysharp's suggested `try_into()` usage for blockparam indices.
Co-authored-by: Jamey Sharp <jamey@minilop.net>
* Take @jameysharp's suggestion to avoid double-match on load op.
Co-authored-by: Jamey Sharp <jamey@minilop.net>
* Fix suggestion (add import).
* Review feedback.
* Fix stack_load handling.
* Remove redundant can_store case.
* Take @jameysharp's suggested improvement to FuncEGraph::build() logic
Co-authored-by: Jamey Sharp <jamey@minilop.net>
* Tweaks to FuncEGraph::build() on top of suggestion.
* Take @jameysharp's suggested clarified condition
Co-authored-by: Jamey Sharp <jamey@minilop.net>
* Clean up after suggestion (unused variable).
* Fix loop analysis.
* loop level asserts
* Revert constant-space loop analysis -- edge cases were incorrect, so let's go with the simple thing for now.
* Take @jameysharp's suggestion re: result_tys
Co-authored-by: Jamey Sharp <jamey@minilop.net>
* Fix up after suggestion
* Take @jameysharp's suggestion to use fold rather than reduce
Co-authored-by: Jamey Sharp <jamey@minilop.net>
* Fixup after suggestion
* Take @jameysharp's suggestion to remove elaborate_eclass_use's return value.
* Clarifying comment in terminator insts.
Co-authored-by: Jamey Sharp <jamey@minilop.net>
Co-authored-by: Nick Fitzgerald <fitzgen@gmail.com>
* Initial forward-edge CFI implementation
Give the user the option to start all basic blocks that are targets
of indirect branches with the BTI instruction introduced by the
Branch Target Identification extension to the Arm instruction set
architecture.
Copyright (c) 2022, Arm Limited.
* Refactor `from_artifacts` to avoid second `make_executable` (#1)
This involves "parsing" twice but this is parsing just the header of an
ELF file so it's not a very intensive operation and should be ok to do
twice.
* Address the code review feedback
Copyright (c) 2022, Arm Limited.
Co-authored-by: Alex Crichton <alex@alexcrichton.com>
This PR adds a basic *alias analysis*, and optimizations that use it.
This is a "mid-end optimization": it operates on CLIF, the
machine-independent IR, before lowering occurs.
The alias analysis (or maybe more properly, a sort of memory-value
analysis) determines when it can prove a particular memory
location is equal to a given SSA value, and when it can, it replaces any
loads of that location.
This subsumes two common optimizations:
* Redundant load elimination: when the same memory address is loaded two
times, and it can be proven that no intervening operations will write
to that memory, then the second load is *redundant* and its result
must be the same as the first. We can use the first load's result and
remove the second load.
* Store-to-load forwarding: when a load can be proven to access exactly
the memory written by a preceding store, we can replace the load's
result with the store's data operand, and remove the load.
Both of these optimizations rely on a "last store" analysis that is a
sort of coloring mechanism, split across disjoint categories of abstract
state. The basic idea is that every memory-accessing operation is put
into one of N disjoint categories; it is disallowed for memory to ever
be accessed by an op in one category and later accessed by an op in
another category. (The frontend must ensure this.)
Then, given this, we scan the code and determine, for each
memory-accessing op, when a single prior instruction is a store to the
same category. This "colors" the instruction: it is, in a sense, a
static name for that version of memory.
This analysis provides an important invariant: if two operations access
memory with the same last-store, then *no other store can alias* in the
time between that last store and these operations. This must-not-alias
property, together with a check that the accessed address is *exactly
the same* (same SSA value and offset), and other attributes of the
access (type, extension mode) are the same, let us prove that the
results are the same.
Given last-store info, we scan the instructions and build a table from
"memory location" key (last store, address, offset, type, extension) to
known SSA value stored in that location. A store inserts a new mapping.
A load may also insert a new mapping, if we didn't already have one.
Then when a load occurs and an entry already exists for its "location",
we can reuse the value. This will be either RLE or St-to-Ld depending on
where the value came from.
Note that this *does* work across basic blocks: the last-store analysis
is a full iterative dataflow pass, and we are careful to check dominance
of a previously-defined value before aliasing to it at a potentially
redundant load. So we will do the right thing if we only have a
"partially redundant" load (loaded already but only in one predecessor
block), but we will also correctly reuse a value if there is a store or
load above a loop and a redundant load of that value within the loop, as
long as no potentially-aliasing stores happen within the loop.
In particular, introduce initial support for the MOVI and MVNI
instructions, with 8-bit elements. Also, treat vector constants
as 32- or 64-bit floating-point numbers, if their value allows
it, by relying on the architectural zero extension. Finally,
stop generating literal loads for 32-bit constants.
Copyright (c) 2020, Arm Limited.
The pinned register hack didn't work because the GetPinnedReg is marked
as having side-effects, so that GVN wouldn't try to common it out.
This commit tweaks the function used during lowering to vcode, so that
the GetPinnedReg opcode is specially handled. It's a bit lame, but it
makes the hack work again.
Also, use_input needs to be a no-op for real registers.
This commit adds support for generating stackmaps at safepoints to the
new backend framework and to the AArch64 backend in particular. It has
been tested to work with SpiderMonkey.
This patch includes:
- A complete rework of the way that CLIF blocks and edge blocks are
lowered into VCode blocks. The new mechanism in `BlockLoweringOrder`
computes RPO over the CFG, but with a twist: it merges edge blocks intto
heads or tails of original CLIF blocks wherever possible, and it does
this without ever actually materializing the full nodes-plus-edges
graph first. The backend driver lowers blocks in final order so
there's no need to reshuffle later.
- A new `MachBuffer` that replaces the `MachSection`. This is a special
version of a code-sink that is far more than a humble `Vec<u8>`. In
particular, it keeps a record of label definitions and label uses,
with a machine-pluggable `LabelUse` trait that defines various types
of fixups (basically internal relocations).
Importantly, it implements some simple peephole-style branch rewrites
*inline in the emission pass*, without any separate traversals over
the code to use fallthroughs, swap taken/not-taken arms, etc. It
tracks branches at the tail of the buffer and can (i) remove blocks
that are just unconditional branches (by redirecting the label), (ii)
understand a conditional/unconditional pair and swap the conditional
polarity when it's helpful; and (iii) remove branches that branch to
the fallthrough PC.
The `MachBuffer` also implements branch-island support. On
architectures like AArch64, this is needed to allow conditional
branches within plausibly-attainable ranges (+/- 1MB on AArch64
specifically). It also does this inline while streaming through the
emission, without any sort of fixpoint algorithm or later moving of
code, by simply tracking outstanding references and "deadlines" and
emitting an island just-in-time when we're in danger of going out of
range.
- A rework of the instruction selector driver. This is largely following
the same algorithm as before, but is cleaned up significantly, in
particular in the API: the machine backend can ask for an input arg
and get any of three forms (constant, register, producing
instruction), indicating it needs the register or can merge the
constant or producing instruction as appropriate. This new driver
takes special care to emit constants right at use-sites (and at phi
inputs), minimizing their live-ranges, and also special-cases the
"pinned register" to avoid superfluous moves.
Overall, on `bz2.wasm`, the results are:
wasmtime full run (compile + runtime) of bz2:
baseline: 9774M insns, 9742M cycles, 3.918s
w/ changes: 7012M insns, 6888M cycles, 2.958s (24.5% faster, 28.3% fewer insns)
clif-util wasm compile bz2:
baseline: 2633M insns, 3278M cycles, 1.034s
w/ changes: 2366M insns, 2920M cycles, 0.923s (10.7% faster, 10.1% fewer insns)
All numbers are averages of two runs on an Ampere eMAG.
- Undo temporary changes to default features (`all-arch`) and a
signal-handler test.
- Remove `SIGTRAP` handler: no longer needed now that we've found an
"undefined opcode" option on ARM64.
- Rename pp.rs to pretty_print.rs in machinst/.
- Only use empty stack-probe on non-x86. As per a comment in
rust-lang/compiler-builtins [1], LLVM only supports stack probes on
x86 and x86-64. Thus, on any other CPU architecture, we cannot refer
to `__rust_probestack`, because it does not exist.
- Rename arm64 to aarch64.
- Use `target` directive in vcode filetests.
- Run the flags verifier, but without encinfo, when using new backends.
- Clean up warning overrides.
- Fix up use of casts: use u32::from(x) and siblings when possible,
u32::try_from(x).unwrap() when not, to avoid silent truncation.
- Take immutable `Function` borrows as input; we don't actually
mutate the input IR.
- Lots of other miscellaneous cleanups.
[1] cae3e6ea23/src/probestack.rs (L39)
