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.
* Allow emitting u64 constants into constant pool.
* Use constant pool for constants on x64 that do not fit in a simm32 and are needed as a RegMem or RegMemImm.
* Fix rip-relative addressing bug in pinsrd emission.
This PR refactors the x64 backend address-mode lowering to use an
incremental-build approach, where it considers each node in a tree of
`iadd`s that feed into a load/store address and, at each step, builds
the best possible `Amode`. It will combine an arbitrary number of
constant offsets (an extension beyond the current rules), and can
capture a left-shifted (scaled) index in any position of the tree
(another extension).
This doesn't have any measurable performance improvement on our Wasm
benchmarks in Sightglass, unfortunately, because the IR lowered from
wasm32 will do address computation in 32 bits and then `uextend` it to
add to the 64-bit heap base. We can't quite lift the 32-bit adds to 64
bits because this loses the wraparound semantics.
(We could label adds as "expected not to overflow", and allow *those* to
be lifted to 64 bit operations; wasm32 heap address computation should
fit this. This is `add nuw` (no unsigned wrap) in LLVM IR terms. That's
likely my next step.)
Nevertheless, (i) this generalizes the cases we can handle, which should
be a good thing, all other things being equal (and in this case, no
compile time impact was measured); and (ii) might benefit non-Wasm
frontends.
Also fix and extend the current implementation:
- AtomicRMWOp::Clr != AtomicRmwOp::And, as the input needs to be
inverted first.
- Inputs to the cmp for the RMWLoop case are sign-extended when
needed.
- Lower Xchg to Swp.
- Lower Sub to Add with a negated input.
- Added more runtests.
Copyright (c) 2022, Arm Limited.
* Cranelift: fix#3953: rework single/multiple-use logic in lowering.
This PR addresses the longstanding issue with loads trying to merge
into compares on x86-64, and more generally, with the lowering
framework falsely recognizing "single uses" of one op by
another (which would normally allow merging of side-effecting ops like
loads) when there is *indirect* duplication.
To fix this, we replace the direct `value_uses` count with a
transitive notion of uniqueness (not unlike Rust's `&`/`&mut` and how
a `&mut` downgrades to `&` when accessed through another `&`!). A
value is used multiple times transitively if it has multiple direct
uses, or is used by another op that is used multiple times
transitively.
The canonical example of badness is:
```
v1 := load
v2 := ifcmp v1, ...
v3 := selectif v2, ...
v4 := selectif v2, ...
```
both `v3` and `v4` effectively merge the `ifcmp` (`v2`), so even
though the use of `v1` is "unique", it is codegenned twice. This is
why we ~~can't have nice things~~ can't merge loads into
compares (#3953).
There is quite a subtle and interesting design space around this
problem and how we might solve it. See the long doc-comment on
`ValueUseState` in this PR for more justification for the particular
design here. In particular, this design deliberately simplifies a bit
relative to an "optimal" solution: some uses can *become* unique
depending on merging, but we don't design our data structures for such
updates because that would require significant extra costly
tracking (some sort of transitive refcounting). For example, in the
above, if `selectif` somehow did not merge `ifcmp`, then we would only
codegen the `ifcmp` once into its result register (and use that
register twice); then the load *is* uniquely used, and could be
merged. But that requires transitioning from "multiple use" back to
"unique use" with careful tracking as we do pattern-matching, which
I've chosen to make out-of-scope here for now. In practice, I don't
think it will matter too much (and we can always improve later).
With this PR, we can now re-enable load-op merging for compares. A
subsequent commit does this.
* Update x64 backend to allow load-op merging for `cmp`.
* Update filetests.
* Add test for cmp-mem merging on x64.
* Comment fixes.
* Rework ValueUseState analysis for better performance.
* Update s390x filetest: iadd_ifcout cannot merge loads anymore because it has multiple outputs (ValueUseState limitation)
* Address review comments.
With these fixes, all this PR has to do is instantiate and run the
checker on the `regalloc2::Output`. This is off by default, and is
enabled by setting the `regalloc_checker` Cranelift option.
This restores the old functionality provided by e.g. the
`backtracking_checked` regalloc algorithm setting rather than
`backtracking` when we were still on regalloc.rs.
regalloc2 is a bit pickier about critical edges than regalloc.rs was,
because of how it inserts moves. In particular, if a branch has any
arguments (e.g., a conditional branch or br_table), its successors must
all have only one predecessor, so we can do edge moves at the top of
successor blocks rather than at the end of this block. Otherwise, moves
that semantically must come after the block's last uses (the branch's
args) would be placed before it.
This is almost always the case, because crit-edge splitting ensures that
if we have more than one succ, all our succs will have only one pred.
This is because branch kinds that take arguments (fixed args, not the
blockparam args) tend to have more than one successor: conditionals and
br_tables.
However, a fuzzbug recently illuminated one corner case I had missed: a
br_table can have *one* successor only, if it has a default target and
an empty table. In this case, crit-edge splitting will happily skip a
split and assume that we can insert edge moves at the end of the block
with the br_table. But this will fail.
regalloc2 explicitly checks this and bails with a panic, rather than
continue, so no miscompilation is possible; but without this fix, we
will get these panics on br_tables with empty tables.
Previously, the block successor accumulation and the blockparam branch
arg setup were decoupled. The lowering backend implicitly specified
the order of successor edges via its `MachTerminator` enum on the last
instruction in the block, while the `Lower` toplevel
machine-independent driver set up blockparam branch args in the edge
order seen in CLIF.
In some cases, these orders did not match -- for example, when the
conditional branch depended on an FP condition that was implemented by
swapping taken/not-taken edges and inverting the condition code.
This PR refactors the successor handling to be centralized in `Lower`
rather than flow through the terminator `MachInst`, and adds a
successor block and its blockparam args at the same time, ensuring the
orders match.
Following the merge of regalloc2 support, this became slower because we
are stricter about the critical-edge invariant, generating a separate
edge block for every out-edge even if two or more out-edges go to the
same successor (this is significant in cases of `br_table` with many
entries having the same target block, for example).
Many of those edge blocks are empty and end up collapsed by the
MachBuffer, which leads to a large set of aliased labels.
The invariant validation will dutifully iterate over all the data
structures at every step, validating all of our conditions. But this
gets way slower in the new context, to the point that we'll probably
have some fuzz timeouts.
This was pointed out in [1] but I missed removing this in #3989. Given
that `MachBuffer` has been around for nearly two years now, has been
fuzzed continuously with the invariant validation for that time, and
also has a correctness proof in the comments, it's probably reasonable
to remove this high (recently increased) cost from the fuzzing-specific
compilation configuration.
[1]
https://github.com/bytecodealliance/wasmtime/pull/3989#discussion_r847712263
This PR switches Cranelift over to the new register allocator, regalloc2.
See [this document](https://gist.github.com/cfallin/08553421a91f150254fe878f67301801)
for a summary of the design changes. This switchover has implications for
core VCode/MachInst types and the lowering pass.
Overall, this change brings improvements to both compile time and speed of
generated code (runtime), as reported in #3942:
```
Benchmark Compilation (wallclock) Execution (wallclock)
blake3-scalar 25% faster 28% faster
blake3-simd no diff no diff
meshoptimizer 19% faster 17% faster
pulldown-cmark 17% faster no diff
bz2 15% faster no diff
SpiderMonkey, 21% faster 2% faster
fib(30)
clang.wasm 42% faster N/A
```
This change moves the majority of the lowerings for CLIF's `load`
instruction over to ISLE. To do so, it also migrates the previous
mechanism for creating an `Amode` (`lower_to_amode`) to several ISLE
rules (see `to_amode`).
* Run the GC smoketest with epoch support enabled as well.
* Handle safepoints in cold blocks properly.
Currently, the way that we find safepoint slots for a given instruction
relies on the instruction index order in the safepoint list matching the
order of instruction emission.
Previous to the introduction of cold-block support, this was trivially
satisfied by sorting the safepoint list: we emit instructions 0, 1, 2,
3, 4, ..., and so if we have safepoints at instructions 1 and 4, we will
encounter them in that order.
However, cold blocks are supported by swizzling the emission order at
the last moment (to avoid having to renumber instructions partway
through the compilation pipeline), so we actually emit instructions out
of index order when cold blocks are present.
Reference-type support in Wasm in particular uses cold blocks for
slowpaths, and has live refs and safepoints in these slowpaths, so we
can reliably "skip" a safepoint (not emit any metadata for it) in the
presence of reftype usage.
This PR fixes the emission code by building a map from instruction index
to safepoint index first, then doing lookups through this map, rather
than following along in-order as it emits instructions.
This change removes all variants of `load*_complex` and `store*_complex`
from Cranelift; this is a breaking change to the instructions exposed by
CLIF. The complete list of instructions removed is: `load_complex`,
`store_complex`, `uload8_complex`, `sload8_complex`, `istore8_complex`,
`sload8_complex`, `uload16_complex`, `sload16_complex`,
`istore16_complex`, `uload32_complex`, `sload32_complex`,
`istore32_complex`, `uload8x8_complex`, `sload8x8_complex`,
`sload16x4_complex`, `uload16x4_complex`, `uload32x2_complex`,
`sload32x2_complex`.
The rationale for this removal is that the Cranelift backend now has the
ability to pattern-match multiple upstream additions in order to
calculate the address to access. Previously, this was not possible so
the `*_complex` instructions were needed. Over time, these instructions
have fallen out of use in this repository, making the additional
overhead of maintaining them a chore.
* x64: port scalar `fcmp` to ISLE
Implement the CLIF lowering for the `fcmp` to ISLE. This adds a new
type-matcher, `ty_scalar_float`, for detecting uses of `F32` and `F64`.
* isle: rename `vec128` to `ty_vec12`
This refactoring changes the name of the `vec128` matcher function to
follow the `ty_*` convention of the other type matchers. It also makes
the helper an inline function call.
* x64: port vector `fcmp` to ISLE
Previously (as in an hour ago) #3905 landed a new ability for fuzzing to
arbitrarily insert padding between functions. Running some fuzzers
locally though this instantly hit a lot of problems on AArch64 because
the arbitrary padding isn't aligned to 4 bytes like all other functions
are. To fix this issue appending functions now correctly aligns the
output as appropriate for the platform. The alignment argument for
appending was switched to `None` where `None` means "use the platform
default" and otherwise and explicit alignment can be specified for
inserting other data (like arbitrary padding or Windows unwind tables).
The current definition of `ValueSlice` is not usable, since any call to
a constructor returning a `ValueSlice` will extend the mutable borrow
on the context taken by the constructor call, with the result that it
cannot be passed to any other constructor ever.
Re-implement `ValueSlice` as a pair of a `ValueList` identifer plus an
offset into the list. This type can simply be copied without requiring
a borrow on the context.
This changes the output of the `lower` constructor from a
`ValueRegs` to a new `InstOutput` type, which is a vector
of `ValueRegs`.
Code in `lower_common` is updated to use this new type to
handle instructions with multiple outputs. All back-ends
are updated to use the new type.
In #3721, we have been discussing what to do about the ARM32 backend in
Cranelift. Currently, this backend supports only 32-bit types, which is
insufficient for full Wasm-MVP; it's missing other critical bits, like
floating-point support; and it has only ever been exercised, AFAIK, via
the filetests for the individual CLIF instructions that are implemented.
We were very very thankful for the original contribution of this
backend, even in its partial state, and we had hoped at the time that we
could eventually mature it in-tree until it supported e.g. Wasm and
other use-cases. But that hasn't yet happened -- to the blame of no-one,
to be clear, we just haven't had a contributor with sufficient time.
Unfortunately, the existence of the backend and lack of active
maintainer now potentially pose a bit of a burden as we hope to make
continuing changes to the backend framework. For example, the ISLE
migration, and the use of regalloc2 that it will allow, would need all
of the existing lowering patterns in the hand-written ARM32 backend to
be rewritten as ISLE rules.
Given that we don't currently have the resources to do this, we think
it's probably best if we, sadly, for now remove this partial backend.
This is not in any way a statement of what we might accept in the
future, though. If, in the future, an ARM32 backend updated to our
latest codebase with an active maintainer were to appear, we'd be happy
to merge it (and likewise for any other architecture!). But for now,
this is probably the best path. Thanks again to the original contributor
@jmkrauz and we hope that this work can eventually be brought back and
reused if someone has the time to do so!
We already defined the `Gpr` newtype and used it in a few places, and we already
defined the `Xmm` newtype and used it extensively. This finishes the transition
to using the newtypes extensively in lowering by making use of `Gpr` in more
places.
Fixes#3685
Add accessors to prelude.isle to access data fields of
`func_addr` and `symbol_value` instructions.
These are based on similar versions I had added to the s390x
back-end, but are a bit more straightforward to use.
- func_ref_data: Extract SigRef, ExternalName, and RelocDistance
fields given a FuncRef.
- symbol_value_data: Extract ExternalName, RelocDistance, and
offset fields given a GlobalValue representing a Symbol.
- reloc_distance_near: Test for RelocDistance::Near.
The s390x back-end is changed to use these common versions.
Note that this exposed a bug in common isle code: This extractor:
(extractor (load_sym inst)
(and inst
(load _ (def_inst (symbol_value
(symbol_value_data _
(reloc_distance_near) offset)))
(i64_from_offset
(memarg_symbol_offset_sum <offset _)))))
would raise an assertion in sema.rs due to a supposed cycle in
extractor definitions. But there was no actual cycle, it was
simply that the extractor tree refers twice to the `insn_data`
extractor (once via the `load` and once via the `symbol_value`
extractor). Fixed by checking for pre-existing definitions only
along one path in the tree, not across the whole tree.
The debuginfo analyses are written with the assumption that the order of
instructions in the VCode is the order of instructions in the final
machine ocde. This was previously a strong invariant, until we
introduced support for cold blocks. Cold blocks are implemented by
reordering during emission, because the VCode ordering has other
requirements related to lowering (respecting def-use dependencies in the
reverse pass), so it is much simpler to reorder instructions at the last
moment. Unfortunately, this causes the breakage we now see.
This commit fixes the issue by skipping all cold instructions when
emitting value-label ranges (which are translated into debuginfo). This
means that variables defined in cold blocks will not have DWARF
metadata. But cold blocks are usually compiler-inserted slowpaths, not
user code, so this is probably OK. Debuginfo is always best-effort, so
in any case this does not violate any correctness constraints.
This primary motivation of this large commit (apologies for its size!) is to
introduce `Gpr` and `Xmm` newtypes over `Reg`. This should help catch
difficult-to-diagnose register class mixup bugs in x64 lowerings.
But having a newtype for `Gpr` and `Xmm` themselves isn't enough to catch all of
our operand-with-wrong-register-class bugs, because about 50% of operands on x64
aren't just a register, but a register or memory address or even an
immediate! So we have `{Gpr,Xmm}Mem[Imm]` newtypes as well.
Unfortunately, `GprMem` et al can't be `enum`s and are therefore a little bit
noisier to work with from ISLE. They need to maintain the invariant that their
registers really are of the claimed register class, so they need to encapsulate
the inner data. If they exposed the underlying `enum` variants, then anyone
could just change register classes or construct a `GprMem` that holds an XMM
register, defeating the whole point of these newtypes. So when working with
these newtypes from ISLE, we rely on external constructors like `(gpr_to_gpr_mem
my_gpr)` instead of `(GprMem.Gpr my_gpr)`.
A bit of extra lines of code are included to add support for register mapping
for all of these newtypes as well. Ultimately this is all a bit wordier than I'd
hoped it would be when I first started authoring this commit, but I think it is
all worth it nonetheless!
In the process of adding these newtypes, I didn't want to have to update both
the ISLE `extern` type definition of `MInst` and the Rust definition, so I move
the definition fully into ISLE, similar as aarch64.
Finally, this process isn't complete. I've introduced the newtypes here, and
I've made most XMM-using instructions switch from `Reg` to `Xmm`, as well as
register class-converting instructions, but I haven't moved all of the GPR-using
instructions over to the newtypes yet. I figured this commit was big enough as
it was, and I can continue the adoption of these newtypes in follow up commits.
Part of #3685.
In order to migrate branches to ISLE, we define a second entry
point `lower_branch` which gets the list of branch targets as
additional argument.
This requires a small change to `lower_common`: the `isle_lower`
callback argument is changed from a function pointer to a closure.
This allows passing the extra argument via a closure.
Traps make use of the recently added facility to emit safepoints
from ISLE, but are otherwise straightforward.
Change the implementation of emitted_insts in IsleContext from
a plain vector of instructions into a vector of tuples, where
the second element is a boolean that indicates whether this
instruction should be emitted as a safepoint.
This allows targets to emit safepoint insns via ISLE.
If a block is marked cold but has side-effect-free code that is only
used by side-effectful code in non-cold blocks, we will erroneously fail
to emit it, causing a regalloc failure.
This is due to the interaction of block ordering and lowering: we rely
on block ordering to visit uses before defs (except for backedges) so
that we can effectively do an inline liveness analysis and skip lowering
operations that are not used anywhere. This "inline DCE" is needed
because instruction lowering can pattern-match and merge one instruction
into another, removing the need to generate the source instruction.
Unfortunately, the way that I added cold-block support in #3698 was
oblivious to this -- it just changed the block sort order. For
efficiency reasons, we generate code in its final order directly, so it
would not be tenable to generate it in e.g. RPO first and then reorder
cold blocks to the bottom; we really do want to visit in the same order
as the final code.
This PR fixes the bug by moving the point at which cold blocks are sunk
to emission-time instead. This is cheaper than either trying to visit
blocks during lowering in RPO but add to VCode out-of-order, or trying
to do some expensive analysis to recover proper liveness. It's not clear
that the latter would be possible anyway -- the need to lower some
instructions depends on other instructions' isel results/merging
success, so we really do need to visit in RPO, and we can't simply lower
all instructions as side-effecting roots (some can't be toplevel nodes).
The one downside of this approach is that the VCode itself still has
cold blocks inline; so in the text format (and hence compile-tests) it's
not possible to see the sinking. This PR adds a test for cold-block
sinking that actually verifies the machine code. (The test also includes
an add-instruction in the cold path that would have been incorrectly
skipped prior to this fix.)
Fortunately this bug would not have been triggered by the one current
use of cold blocks in #3699, because there the only operation in the
cold block was an (always effectful) call instruction. The worst-case
effect of the bug in other code would be a regalloc panic; no silent
miscompilations could result.
In preparing to move the s390x back-end to ISLE, I noticed a few
missing pieces in the common prelude code. This patch:
- Defines the reference types $R32 / $R64.
- Provides a trap_code_bad_conversion_to_integer helper.
- Provides an avoid_div_traps helper. This requires passing the
generic flags in addition to the ISA-specifc flags into the
ISLE lowering context.
This PR adds a flag to each block that can be set via the frontend/builder
interface that indicates that the block will not be frequently
executed. As such, the compiler backend should place the block "out of
line" in the final machine code, so that the ordinary, more frequent
execution path that excludes the block does not have to jump around it.
This is useful for adding handlers for exceptional conditions
(slow-paths, guard violations) in a way that minimizes performance cost.
Fixes#2747.
