* ABI: implement register arguments with constraints.
Currently, Cranelift's ABI code emits a sequence of moves from physical
registers into vregs at the top of the function body, one for every
register-carried argument.
For a number of reasons, we want to move to operand constraints instead,
and remove the use of explicitly-named "pinned vregs"; this allows for
better regalloc in theory, as it removes the need to "reverse-engineer"
the sequence of moves.
This PR alters the ABI code so that it generates a single "args"
pseudo-instruction as the first instruction in the function body. This
pseudo-inst defs all register arguments, and constrains them to the
appropriate registers at the def-point. Subsequently the regalloc can
move them wherever it needs to.
Some care was taken not to have this pseudo-inst show up in
post-regalloc disassemblies, but the change did cause a general regalloc
"shift" in many tests, so the precise-output updates are a bit noisy.
Sorry about that!
A subsequent PR will handle the other half of the ABI code, namely, the
callsite case, with a similar preg-to-constraint conversion.
* Update based on review feedback.
* Review feedback.
Using fallible extractors that produce no values for flag checks means
that it's not possible to pattern match cases where those flags are
false. This change reworks the existing flag-checking extractors to be
infallible, returning the flag's boolean value from the context instead.
This is a cherry-pick of a long-ago commit, 2d46637. The original
message reads:
> Now that `SyntheticAmode` can refer to constants, there is no longer a
> need for a separate instruction format--standard load instructions will
> work.
Since then, the transition to ISLE and the use of `XmmLoadConst` in many
more places makes this change a larger diff than the original. The basic
idea is the same, though: the extra indirection of `Inst::XMmLoadConst`
is removed and replaced by a direct use of `VCodeConstant` as a
`SyntheticAmode`. This has no effect on codegen, but the CLIF output is
now clearer in that the actual instruction is displayed (e.g., `movdqu`)
instead of a made-up instruction (`load_const`).
* Cranelift: Deduplicate ABI signatures during lowering
This commit creates the `SigSet` type which interns and deduplicates the ABI
signatures that we create from `ir::Signature`s. The ABI signatures are now
referred to indirectly via a `Sig` (which is a `cranelift_entity` ID), and we
pass around a `SigSet` to anything that needs to access the actual underlying
`SigData` (which is what `ABISig` used to be).
I had to change a couple methods to return a `SmallInstVec` instead of emitting
directly to work around what would otherwise be shared and exclusive borrows of
the lowering context overlapping. I don't expect any of these to heap allocate
in practice.
This does not remove the often-unnecessary allocations caused by
`ensure_struct_return_ptr_is_returned`. That is left for follow up work.
This also opens the door for further shuffling of signature data into more
efficient representations in the future, now that we have `SigSet` to store it
all in one place and it is threaded through all the code. We could potentially
move each signature's parameter and return vectors into one big vector shared
between all signatures, for example, which could cut down on allocations and
shrink the size of `SigData` since those `SmallVec`s have pretty large inline
capacity.
Overall, this refactoring gives a 1-7% speedup for compilation on
`pulldown-cmark`:
```
compilation :: cycles :: benchmarks/pulldown-cmark/benchmark.wasm
Δ = 8754213.66 ± 7526266.23 (confidence = 99%)
dedupe.so is 1.01x to 1.07x faster than main.so!
[191003295 234620642.20 280597986] dedupe.so
[197626699 243374855.86 321816763] main.so
compilation :: cycles :: benchmarks/bz2/benchmark.wasm
No difference in performance.
[170406200 194299792.68 253001201] dedupe.so
[172071888 193230743.11 223608329] main.so
compilation :: cycles :: benchmarks/spidermonkey/benchmark.wasm
No difference in performance.
[3870997347 4437735062.59 5216007266] dedupe.so
[4019924063 4424595349.24 4965088931] main.so
```
* Use full path instead of import to avoid warnings in some build configurations
Warnings will then cause CI to fail.
* Move `SigSet` into `VCode`
* x64: clean up regalloc-related semantics on several instructions.
This PR removes all uses of "modify" operands on instructions in the x64
backend, and also removes all uses of "pinned vregs", or vregs that are
explicitly tied to particular physical registers. In place of both of
these mechanisms, which are legacies of the old regalloc design and
supported via compatibility code, the backend now uses operand
constraints. This is more flexible as it allows the regalloc to see the
liveranges and constraints without "reverse-engineering" move instructions.
Eventually, after removing all such uses (including in other backends
and by the ABI code), we can remove the compatibility code in regalloc2,
significantly simplifying its liverange-construction frontend and
thus allowing for higher confidence in correctness as well as possibly a
bit more compilation speed.
Curiously, there are a few extra move instructions now; they are likely
poor splitting decisions and I can try to chase these down later.
* Fix cranelift-codegen tests.
* Review feedback.
Ensure that constants generated for the memory case of XmmMem values are always 16 bytes, ensuring that we don't accidantally perform an unaligned load.
Fixes#4761
Lower extractlane, scalar_to_vector and splat in ISLE.
This PR also makes some changes to the SinkableLoad api
* change the return type of sink_load to RegMem as there are more functions available for dealing with RegMem
* add reg_mem_to_reg_mem_imm and register it as an automatic conversion
Lower `shuffle` and `swizzle` in ISLE.
This PR surfaced a bug with the lowering of `shuffle` when avx512vl and avx512vbmi are enabled: we use `vpermi2b` as the implementation, but panic if the immediate shuffle mask contains any out-of-bounds values. The behavior when the avx512 extensions are not present is that out-of-bounds values are turned into `0` in the result.
I've resolved this by detecting when the shuffle immediate has out-of-bounds indices in the avx512-enabled lowering, and generating an additional mask to zero out the lanes where those indices occur. This brings the avx512 case into line with the semantics of the `shuffle` op: 94bcbe8446/cranelift/codegen/meta/src/shared/instructions.rs (L1495-L1498)
* x64: Mask shift amounts for small types
* cranelift: Disable i128 shifts in fuzzer again
They are fixed. But we had a bunch of fuzzgen issues come in, and we don't want to accidentaly mark them as fixed
* cranelift: Avoid masking shifts for 32 and 64 bit cases
* cranelift: Add const shift tests and fix them
* cranelift: Remove const `rotl` cases
Now that `put_masked_in_imm8_gpr` works properly we can simplify rotl/rotr
Lower stack_addr, udiv, sdiv, urem, srem, umulhi, and smulhi in ISLE.
For udiv, sdiv, urem, and srem I opted to move the original lowering into an extern constructor, as the interactions with rax and rdx for the div instruction didn't seem meaningful to implement in ISLE. However, I'm happy to revisit this choice and move more of the embedding into ISLE.
* Add a test for iadd_pairwise with swiden input
* Implement iadd_pairwise for swiden_{low,high} input
* Add a test case for iadd_pairwise with uwiden input
* Implement iadd_pairwise with uwiden
* Cranelift: Remove the `ABICaller` trait
It has only one implementation: the `ABICallerImpl` struct. We can just use that
directly rather than having extra, unnecessary layers of generics and abstractions.
* Cranelift: Rename `ABICallerImpl` to `Caller`
The trait had only one implementation: the `Lower` struct. It is easier to just
use that directly, and not introduce unnecessary layers of generics and
abstractions.
Once upon a time, there was hope that we would have other implementations of the
`LowerCtx` trait, that did things like lower CLIF to SMTLIB for
verification. However, this is not practical these days given the way that the
trait has evolved over time, and our verification efforts are focused on ISLE
now anyways, and we're actually making some progress on that front (much more
than anyone ever did on a second `LowerCtx` trait implementation!)
* Add a test for the existing behavior of fcvt_from_unit
* Migrate the I8, I16, I32 cases of fcvt_from_uint
* Implement the I64 case of fcvt_from_uint
* Add a test for the existing behavior of fcvt_from_uint.f64x2
* Migrate fcvt_from_uint.f64x2 to ISLE
* Lower the last case of `fcvt_from_uint`
* Add a test for `fcvt_from_uint`
* Finish lowering fcmp_from_uint
* Format
This adds support for StructArgument on s390x. The ABI for this
platform requires that the address of the buffer holding the copy
of the struct argument is passed from caller to callee as hidden
pointer, using a register or overflow stack slot.
To implement this, I've added an optional "pointer" filed to
ABIArg::StructArg, and code to handle the pointer both in common
abi_impl code and the s390x back-end.
One notable change necessary to make this work involved the
"copy_to_arg_order" mechanism. Currently, for struct args
we only need to copy the data (and that need to happen before
setting up any other args), while for non-struct args we only
need to set up the appropriate registers or stack slots.
This order is ensured by sorting the arguments appropriately
into a "copy_to_arg_order" list.
However, for struct args with explicit pointers we need to *both*
copy the data (again, before everything else), *and* set up a
register or stack slot. Since we now need to touch the argument
twice, we cannot solve the ordering problem by a simple sort.
Instead, the abi_impl common code now provided *two* callbacks,
emit_copy_regs_to_buffer and emit_copy_regs_to_arg, and expects
the back end to first call copy..to_buffer for all args, and
then call copy.._to_arg for all args. This required updates
to all back ends.
In the s390x back end, in addition to the new ABI code, I'm now
adding code to actually copy the struct data, using the MVC
instruction (for small buffers) or a memcpy libcall (for larger
buffers). This also requires a bit of new infrastructure:
- MVC is the first memory-to-memory instruction we use, which
needed a bit of memory argument tweaking
- We also need to set up the infrastructure to emit libcalls.
(This implements the first half of issue #4565.)
* Cranelift: Add instructions for getting the current stack/frame pointers and return address
This is the initial part of https://github.com/bytecodealliance/wasmtime/issues/4535
* x64: Remove `Amode::RbpOffset` and use `Amode::ImmReg` instead
We just special case getting operands from `Amode`s now.
* Fix s390x `get_return_address`; require `preserve_frame_pointers=true`
* Assert that `Amode::ImmRegRegShift` doesn't use rbp/rsp
* Handle non-allocatable registers in Amode::with_allocs
* Use "stack" instead of "r15" on s390x
* r14 is an allocatable register on s390x, so it shouldn't be used with `MovPReg`
* cranelift: Reorganize test suite
Group some SIMD operations by instruction.
* cranelift: Deduplicate some shift tests
Also, new tests with the mod behaviour
* aarch64: Lower shifts with mod behaviour
* x64: Lower shifts with mod behaviour
* wasmtime: Don't mask SIMD shifts
* x64: port `atomic_rmw` to ISLE
This change ports `atomic_rmw` to ISLE for the x64 backend. It does not
change the lowering in any way, though it seems possible that the fixed
regs need not be as fixed and that there are opportunities for single
instruction lowerings. It does rename `inst_common::AtomicRmwOp` to
`MachAtomicRmwOp` to disambiguate with the IR enum with the same name.
* x64: remove remaining hardcoded register constraints for `atomic_rmw`
* x64: use `SyntheticAmode` in `AtomicRmwSeq`
* review: add missing reg collector for amode
* review: collect memory registers in the 'late' phase
Move from passing and returning u8 and u16 values to u32 in many of
the functions. This removes a number of type conversions and gives
a small compilation time speedup, around ~0.7% on my aarch64 machine.
Copyright (c) 2022, Arm Limited.
- Handle call instructions' clobbers with the clobbers API, using RA2's
clobbers bitmask (bytecodealliance/regalloc2#58) rather than clobbers
list;
- Pull in changes from bytecodealliance/regalloc2#59 for much more sane
edge-case behavior w.r.t. liverange splitting.
The current lowering helper for `cmpxchg` returns the literal RealReg
`rax` as its result. However, this breaks a number of invariants, and
eventually causes a regalloc panic if used as a blockparam arg (pinned
vregs cannot be used in this way).
In general we have to return regular vregs, not a RealReg, as results of
instructions during lowering. However #4223 added a helper for
`x64_cmpxchg` that returns a literal `rax`.
Fortunately we can do the right thing here by just giving a fresh vreg
to the instruction; the regalloc constraints mean that this vreg is
constrained to `rax` at the instruction (at its def/late point), so the
generator of the instruction need not worry about `rax` here.
* 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.
This PR removes "argument polarity": the feature of ISLE extractors that lets them take
inputs aside from the value to be matched.
Cases that need this expressivity have been subsumed by #4072 with if-let clauses;
we can now finally remove this misfeature of the language, which has caused significant
confusion and has always felt like a bit of a hack.
This PR (i) removes the feature from the ISLE compiler; (ii) removes it from the reference
documentation; and (iii) refactors away all uses of the feature in our three existing
backends written in ISLE.
* 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.
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`).
* x64: port GPR-held `icmp` to ISLE
* x64: port equality `icmp` for i128 type
* x64: port `icmp` for vector types
* x64: rename from_intcc to intcc_to_cc
* x64: port `select` using an FP comparison to ISLE
This change includes quite a few interlocking parts, required mainly by
the current x64 conventions in ISLE:
- it adds a way to emit a `cmove` with multiple OR-ing conditions;
because x64 ISLE cannot currently safely emit a comparison followed
by several jumps, this adds `MachInst::CmoveOr` and
`MachInst::XmmCmoveOr` macro instructions. Unfortunately, these macro
instructions hide the multi-instruction sequence in `lower.isle`
- to properly keep track of what instructions consume and produce
flags, @cfallin added a way to pass around variants of
`ConsumesFlags` and `ProducesFlags`--these changes affect all
backends
- then, to lower the `fcmp + select` CLIF, this change adds several
`cmove*_from_values` helpers that perform all of the awkward
conversions between `Value`, `ValueReg`, `Reg`, and `Gpr/Xmm`; one
upside is that now these lowerings have much-improved documentation
explaining why the various `FloatCC` and `CC` choices are made the
the way they are.
Co-authored-by: Chris Fallin <chris@cfallin.org>
