This fixes two problems: minimum symbol alignment for the LARL
instruction, and alignment requirements for LRL/LGRL etc.
The first problem is that the LARL instruction used to load a
symbol address (PC relative) requires that the target symbol
is at least 2-byte aligned. This is always guaranteed for code
symbols (all instructions must be 2-aligned anyway), but not
necessarily for data symbols.
Other s390x compilers fix this problem by ensuring that all
global symbols are always emitted with a minimum 2-byte
alignment. This patch introduces an equivalent mechanism
for cranelift:
- Add a symbol_alignment routine to TargetIsa, similar to the
existing code_section_alignment routine.
- Respect symbol_alignment as minimum alignment for all symbols
emitted in the object backend (code and data).
The second problem is that PC-relative instructions that
directly *access* data (like LRL/LGRL, STRL/STGRL etc.)
not only have the 2-byte requirement like LARL, but actually
require that their memory operand is *naturally* aligned
(i.e. alignment is at least the size of the access).
This property (natural alignment for memory accesses) is
supposed to be provided by the "aligned" flag in MemFlags;
however, this is not implemented correctly at the moment.
To fix this, this patch:
- Only emits PC-relative memory access instructions if the
"aligned" flag is set in the associated MemFlags.
- Fixes a bug in emit_small_memory_copy and emit_small_memset
which currently set the aligned flag unconditionally, ignoring
the actual alignment info passed by their caller.
Tested with wasmtime and cg_clif.
* Cranelift: Remove `ABICallee` trait
It has only one implementation: the `ABICalleeImpl` struct. By using that
directly we can avoid unnecessary layers of generics and abstractions as well as
a couple `Box`es that were previously putting the single implementation into a
`Box<dyn>`.
* Cranelift: Rename `ABICalleeImpl` to `AbiCallee`
* Fix comments as per review
* Rename `AbiCallee` to `Callee`
This is the implementation of https://github.com/bytecodealliance/wasmtime/issues/4155, using the "inverted API" approach suggested by @cfallin (thanks!) in Cranelift, and trait object to provide a backend for an all-included experience in Wasmtime.
After the suggestion of Chris, `Function` has been split into mostly two parts:
- on the one hand, `FunctionStencil` contains all the fields required during compilation, and that act as a compilation cache key: if two function stencils are the same, then the result of their compilation (`CompiledCodeBase<Stencil>`) will be the same. This makes caching trivial, as the only thing to cache is the `FunctionStencil`.
- on the other hand, `FunctionParameters` contain the... function parameters that are required to finalize the result of compilation into a `CompiledCode` (aka `CompiledCodeBase<Final>`) with proper final relocations etc., by applying fixups and so on.
Most changes are here to accomodate those requirements, in particular that `FunctionStencil` should be `Hash`able to be used as a key in the cache:
- most source locations are now relative to a base source location in the function, and as such they're encoded as `RelSourceLoc` in the `FunctionStencil`. This required changes so that there's no need to explicitly mark a `SourceLoc` as the base source location, it's automatically detected instead the first time a non-default `SourceLoc` is set.
- user-defined external names in the `FunctionStencil` (aka before this patch `ExternalName::User { namespace, index }`) are now references into an external table of `UserExternalNameRef -> UserExternalName`, present in the `FunctionParameters`, and must be explicitly declared using `Function::declare_imported_user_function`.
- some refactorings have been made for function names:
- `ExternalName` was used as the type for a `Function`'s name; while it thus allowed `ExternalName::Libcall` in this place, this would have been quite confusing to use it there. Instead, a new enum `UserFuncName` is introduced for this name, that's either a user-defined function name (the above `UserExternalName`) or a test case name.
- The future of `ExternalName` is likely to become a full reference into the `FunctionParameters`'s mapping, instead of being "either a handle for user-defined external names, or the thing itself for other variants". I'm running out of time to do this, and this is not trivial as it implies touching ISLE which I'm less familiar with.
The cache computes a sha256 hash of the `FunctionStencil`, and uses this as the cache key. No equality check (using `PartialEq`) is performed in addition to the hash being the same, as we hope that this is sufficient data to avoid collisions.
A basic fuzz target has been introduced that tries to do the bare minimum:
- check that a function successfully compiled and cached will be also successfully reloaded from the cache, and returns the exact same function.
- check that a trivial modification in the external mapping of `UserExternalNameRef -> UserExternalName` hits the cache, and that other modifications don't hit the cache.
- This last check is less efficient and less likely to happen, so probably should be rethought a bit.
Thanks to both @alexcrichton and @cfallin for your very useful feedback on Zulip.
Some numbers show that for a large wasm module we're using internally, this is a 20% compile-time speedup, because so many `FunctionStencil`s are the same, even within a single module. For a group of modules that have a lot of code in common, we get hit rates up to 70% when they're used together. When a single function changes in a wasm module, every other function is reloaded; that's still slower than I expect (between 10% and 50% of the overall compile time), so there's likely room for improvement.
Fixes#4155.
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!)
This implements the s390x back-end portion of the solution for
https://github.com/bytecodealliance/wasmtime/issues/4566
We now support both big- and little-endian vector lane order
in code generation. The order used for a function is determined
by the function's ABI: if it uses a Wasmtime ABI, it will use
little-endian lane order, and big-endian lane order otherwise.
(This ensures that all raw_bitcast instructions generated by
both wasmtime and other cranelift frontends can always be
implemented as a no-op.)
Lane order affects the implementation of a number of operations:
- Vector immediates
- Vector memory load / store (in big- and little-endian variants)
- Operations explicitly using lane numbers
(insertlane, extractlane, shuffle, swizzle)
- Operations implicitly using lane numbers
(iadd_pairwise, narrow/widen, promote/demote, fcvt_low, vhigh_bits)
In addition, when calling a function using a different lane order,
we need to lane-swap all vector values passed or returned in registers.
A small number of changes to common code were also needed:
- Ensure we always select a Wasmtime calling convention on s390x
in crates/cranelift (func_signature).
- Fix vector immediates for filetests/runtests. In PR #4427,
I attempted to fix this by byte-swapping the V128 value, but
with the new scheme, we'd instead need to perform a per-lane
byte swap. Since we do not know the actual type in write_to_slice
and read_from_slice, this isn't easily possible.
Revert this part of PR #4427 again, and instead just mark the
memory buffer as little-endian when emitting the trampoline;
the back-end will then emit correct code to load the constant.
- Change a runtest in simd-bitselect-to-vselect.clif to no longer
make little-endian lane order assumptions.
- Remove runtests in simd-swizzle.clif that make little-endian
lane order assumptions by relying on implicit type conversion
when using a non-i16x8 swizzle result type (this feature should
probably be removed anyway).
Tested with both wasmtime and cg_clif.
This enables the object backend for s390x, in particular the
processing of all required relocations.
This uncovered a bug: we need to use PLT relocations for the
target of calls, which we currently do not. Fixed by adding
a new S390xPLTRel32Dbl reloc type and using it where needed.
Implement the tls_value for s390 in the ELF general-dynamic mode.
Notable differences to the x86_64 implementation are:
- We use a __tls_get_offset libcall instead of __tls_get_addr.
- The current thread pointer (stored in a pair of access registers)
needs to be added to the result of __tls_get_offset.
- __tls_get_offset has a variant ABI that requires the address of
the GOT (global offset table) is passed in %r12.
This means we need a new libcall entries for __tls_get_offset.
In addition, we also need a way to access _GLOBAL_OFFSET_TABLE_.
The latter is a "magic" symbol with a well-known name defined
by the ABI and recognized by the linker. This patch introduces
a new ExternalName::KnownSymbol variant to support such names
(originally due to @afonso360).
We also need to emit a relocation on a symbol placed in a
constant pool, as well as an extra relocation on the call
to __tls_get_offset required for TLS linker optimization.
Needed by the cg_clif frontend.
This adds full i128 support to the s390x target, including new filetests
and enabling the existing i128 runtest on s390x.
The ABI requires that i128 is passed and returned via implicit pointer,
but the front end still generates direct i128 types in call. This means
we have to implement ABI support to implicitly convert i128 types to
pointers when passing arguments.
To do so, we add a new variant ABIArg::ImplicitArg. This acts like
StructArg, except that the value type is the actual target type,
not a pointer type. The required conversions have to be inserted
in the prologue and at function call sites.
Note that when dereferencing the implicit pointer in the prologue,
we may require a temp register: the pointer may be passed on the
stack so it needs to be loaded first, but the value register may
be in the wrong class for pointer values. In this case, we use
the "stack limit" register, which should be available at this
point in the prologue.
For return values, we use a mechanism similar to the one used for
supporting multiple return values in the Wasmtime ABI. The only
difference is that the hidden pointer to the return buffer must
be the *first*, not last, argument in this case.
(This implements the second half of issue #4565.)
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.)
Give the user the option to sign and to authenticate function
return addresses with the operations introduced by the Pointer
Authentication extension to the Arm instruction set architecture.
Copyright (c) 2021, Arm Limited.
* 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`
* Cranellift: remove Baldrdash support and related features.
As noted in Mozilla's bugzilla bug 1781425 [1], the SpiderMonkey team
has recently determined that their current form of integration with
Cranelift is too hard to maintain, and they have chosen to remove it
from their codebase. If and when they decide to build updated support
for Cranelift, they will adopt different approaches to several details
of the integration.
In the meantime, after discussion with the SpiderMonkey folks, they
agree that it makes sense to remove the bits of Cranelift that exist
to support the integration ("Baldrdash"), as they will not need
them. Many of these bits are difficult-to-maintain special cases that
are not actually tested in Cranelift proper: for example, the
Baldrdash integration required Cranelift to emit function bodies
without prologues/epilogues, and instead communicate very precise
information about the expected frame size and layout, then stitched
together something post-facto. This was brittle and caused a lot of
incidental complexity ("fallthrough returns", the resulting special
logic in block-ordering); this is just one example. As another
example, one particular Baldrdash ABI variant processed stack args in
reverse order, so our ABI code had to support both traversal
orders. We had a number of other Baldrdash-specific settings as well
that did various special things.
This PR removes Baldrdash ABI support, the `fallthrough_return`
instruction, and pulls some threads to remove now-unused bits as a
result of those two, with the understanding that the SpiderMonkey folks
will build new functionality as needed in the future and we can perhaps
find cleaner abstractions to make it all work.
[1] https://bugzilla.mozilla.org/show_bug.cgi?id=1781425
* Review feedback.
* Fix (?) DWARF debug tests: add `--disable-cache` to wasmtime invocations.
The debugger tests invoke `wasmtime` from within each test case under
the control of a debugger (gdb or lldb). Some of these tests started to
inexplicably fail in CI with unrelated changes, and the failures were
only inconsistently reproducible locally. It seems to be cache related:
if we disable cached compilation on the nested `wasmtime` invocations,
the tests consistently pass.
* Review feedback.
* Move `emit_to_memory` to `MachCompileResult`
This small refactoring makes it clearer to me that emitting to memory
doesn't require anything else from the compilation `Context`. While it's
a trivial change, it's a small public API change that shouldn't cause
too much trouble, and doesn't seem RFC-worthy. Happy to hear different
opinions about this, though!
* hide the MachCompileResult behind a method
* Add a `CompileError` wrapper type that references a `Function`
* Rename MachCompileResult to CompiledCode
* Additionally remove the last unsafe API in cranelift-codegen
* Support shadowing in isle
* Re-run the isle build.rs if the examples change
* Print error messages when isle tests fail
* Move run tests
* Refactor `let` uses that don't need to introduce unique names
This enables more runtests to be executed on s390x. Doing so
uncovered a two back-end bugs, which are fixed as well:
- The result of cls was always off by one.
- The result of popcnt.i16 has uninitialized high bits.
In addition, I found a bug in the load-op-store.clif test case:
v3 = heap_addr.i64 heap0, v1, 4
v4 = iconst.i64 42
store.i32 v4, v3
This was clearly intended to perform a 32-bit store, but
actually performs a 64-bit store (it seems the type annotation
of the store opcode is ignored, and the type of the operand
is used instead). That bug did not show any noticable symptoms
on little-endian architectures, but broke on big-endian.
On s390x, we do not have a frame pointer that can be used to chain
stack frames for easy unwinding. Instead, our ABI defines a stack
"backchain" mechanism that can be used to the same effect.
This PR uses that backchain mechanism to implement the new
preserve_frame_pointers flags introduced here:
https://github.com/bytecodealliance/wasmtime/pull/4469
This adds full support for all Cranelift SIMD instructions
to the s390x target. Everything is matched fully via ISLE.
In addition to adding support for many new instructions,
and the lower.isle code to match all SIMD IR patterns,
this patch also adds ABI support for vector types.
In particular, we now need to handle the fact that
vector registers 8 .. 15 are partially callee-saved,
i.e. the high parts of those registers (which correspond
to the old floating-poing registers) are callee-saved,
but the low parts are not. This is the exact same situation
that we already have on AArch64, and so this patch uses the
same solution (the is_included_in_clobbers callback).
The bulk of the changes are platform-specific, but there are
a few exceptions:
- Added ISLE extractors for the Immediate and Constant types,
to enable matching the vconst and swizzle instructions.
- Added a missing accessor for call_conv to ABISig.
- Fixed endian conversion for vector types in data_value.rs
to enable their use in runtests on the big-endian platforms.
- Enabled (nearly) all SIMD runtests on s390x. [ Two test cases
remain disabled due to vector shift count semantics, see below. ]
- Enabled all Wasmtime SIMD tests on s390x.
There are three minor issues, called out via FIXMEs below,
which should be addressed in the future, but should not be
blockers to getting this patch merged. I've opened the
following issues to track them:
- Vector shift count semantics
https://github.com/bytecodealliance/wasmtime/issues/4424
- is_included_in_clobbers vs. link register
https://github.com/bytecodealliance/wasmtime/issues/4425
- gen_constant callback
https://github.com/bytecodealliance/wasmtime/issues/4426
All tests, including all newly enabled SIMD tests, pass
on both z14 and z15 architectures.
Introduce a new concept in the IR that allows a producer to create
dynamic vector types. An IR function can now contain global value(s)
that represent a dynamic scaling factor, for a given fixed-width
vector type. A dynamic type is then created by 'multiplying' the
corresponding global value with a fixed-width type. These new types
can be used just like the existing types and the type system has a
set of hard-coded dynamic types, such as I32X4XN, which the user
defined types map onto. The dynamic types are also used explicitly
to create dynamic stack slots, which have no set size like their
existing counterparts. New IR instructions are added to access these
new stack entities.
Currently, during codegen, the dynamic scaling factor has to be
lowered to a constant so the dynamic slots do eventually have a
compile-time known size, as do spill slots.
The current lowering for aarch64 just targets Neon, using a dynamic
scale of 1.
Copyright (c) 2022, Arm Limited.
This defines the full set of 32 128-bit vector registers on s390x.
(Note that the VRs overlap the existing FPRs.) In addition, this
adds support to use all 32 vector registers to implement floating-
point operations, by using vector floating-point instructions with
the 'W' bit set to operate only on the first element.
This part of the vector instruction set mostly matches the old FP
instruction set, with two exceptions:
- There is no vector version of the COPY SIGN instruction. Instead,
now use a VECTOR SELECT with an appropriate bit mask to implement
the fcopysign operation.
- There are no vector version of the float <-> int conversion
instructions where source and target differ in bit size. Use
appropriate multiple conversion steps instead. This also requires
use of explicit checking to implement correct overflow handling.
As a side effect, this version now also implements the i8 / i16
variants of all conversions, which had been missing so far.
For all operations except those two above, we continue to use the
old FP instruction if applicable (i.e. if all operands happen to
have been allocated to the original FP register set), and use the
vector instruction otherwise.
Now that lowering is fully done in ISLE, clean up some code remnants
in lower.rs. In particular, move code to lower/isle.rs where
possible, and inline lower_insn_to_regs into its caller and simplify.
This adds infrastructure to allow implementing call and return
instructions in ISLE, and migrates the s390x back-end.
To implement ABI details, this patch creates public accessors
for `ABISig` and makes them accessible in ISLE. All actual
code generation is then done in ISLE rules, following the
information provided by that signature.
[ Note that the s390x back end never requires multiple slots for
a single argument - the infrastructure to handle this should
already be present, however. ]
To implement loops in ISLE rules, this patch uses regular tail
recursion, employing a `Range` data structure holding a range
of integers to be looped over.
- 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.
RA2 recently removed the need for a dedicated scratch register for
cyclic moves (bytecodealliance/regalloc2#51). This has moderate positive
performance impact on function bodies that were register-constrained, as
it means that one more register is available. In Sightglass, I measured
+5-8% on `blake3-scalar`, at least among current benchmarks.
* Remove unused srcloc in MachReloc
* Remove unused srcloc in MachTrap
* Use `into_iter` on array in bench code to suppress a warning
* Remove unused srcloc in MachCallSite
Previously, the pinned register (enabled by the `enable_pinned_reg`
Cranelift setting and used via the `get_pinned_reg` and `set_pinned_reg`
CLIF ops) was only used when Cranelift was embedded in SpiderMonkey, in
order to support a pinned heap register. SpiderMonkey has its own
calling convention in Cranelift (named after the integration layer,
"Baldrdash").
However, the feature is more general, and should be usable with the
default system calling convention too, e.g. SysV or Windows Fastcall.
This PR fixes the ABI code to properly treat the pinned register as a
globally allocated register -- and hence an implicit input and output to
every function, not saved/restored in the prologue/epilogue -- for SysV
on x86-64 and aarch64, and Fastcall on x86-64.
Fixes#4170.
Current codegen had a number of logic errors confusing
NAND with AND WITH COMPLEMENT, and NOR with OR WITH COMPLEMENT.
Add support for the missing z15 instructions and fix logic.
This PR fixes#4066: it modifies the Cranelift `build.rs` workflow to
invoke the ISLE DSL compiler on every compilation, rather than only
when the user specifies a special "rebuild ISLE" feature.
The main benefit of this change is that it vastly simplifies the mental
model required of developers, and removes a bunch of failure modes
we have tried to work around in other ways. There is now just one
"source of truth", the ISLE source itself, in the repository, and so there
is no need to understand a special "rebuild" step and how to handle
merge errors. There is no special process needed to develop the compiler
when modifying the DSL. And there is no "noise" in the git history produced
by constantly-regenerated files.
The two main downsides we discussed in #4066 are:
- Compile time could increase, by adding more to the "meta" step before the main build;
- It becomes less obvious where the source definitions are (everything becomes
more "magic"), which makes exploration and debugging harder.
This PR addresses each of these concerns:
1. To maintain reasonable compile time, it includes work to cut down the
dependencies of the `cranelift-isle` crate to *nothing* (only the Rust stdlib),
in the default build. It does this by putting the error-reporting bits
(`miette` crate) under an optional feature, and the logging (`log` crate) under
a feature-controlled macro, and manually writing an `Error` impl rather than
using `thiserror`. This completely avoids proc macros and the `syn` build slowness.
The user can still get nice errors out of `miette`: this is enabled by specifying
a Cargo feature `--features isle-errors`.
2. To allow the user to optionally inspect the generated source, which nominally
lives in a hard-to-find path inside `target/` now, this PR adds a feature `isle-in-source-tree`
that, as implied by the name, moves the target for ISLE generated source into
the source tree, at `cranelift/codegen/isle_generated_source/`. It seems reasonable
to do this when an explicit feature (opt-in) is specified because this is how ISLE regeneration
currently works as well. To prevent surprises, if the feature is *not* specified, the
build fails if this directory exists.
* 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.
* Narrow `allow(dead_code)` declarations
Having module wide `allow(dead_code)` may hide some code that's really
dead. In this commit I just narrowed the declarations to the specific
enum variants that were not used (as it seems reasonable to keep them
and their handling in all the matches, for future use). And the compiler
found more dead code that I think we can remove safely in the short
term.
With this, the only files annotated with a module-wide
`allow(dead_code)` are isle-generated files.
* resurrect some functions as test helpers
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.
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.
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.
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`).
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.
