This was added as an incremental step to improve AArch64 code quality in
PR #2278. At the time, we did not have a way to pattern-match the load +
splat opcode sequence that the relevant Wasm opcodes lowered to.
However, now with PR #2366, we can merge effectful instructions such as
loads into other ops, and so we can do this pattern matching directly.
The pattern-matching update will come in a subsequent commit.
It corresponds to WebAssembly's `load*_splat` operations, which
were previously represented as a combination of `Load` and `Splat`
instructions. However, there are architectures such as Armv8-A
that have a single machine instruction equivalent to the Wasm
operations. In order to generate it, it is necessary to merge the
`Load` and the `Splat` in the backend, which is not possible
because the load may have side effects. The new IR instruction
works around this limitation.
The AArch64 backend leverages the new instruction to improve code
generation.
Copyright (c) 2020, Arm Limited.
This converts an `i32x4` into an `f32x4` with some rounding either by using an AVX512VL/F instruction--VCVTUDQ2PS--or a long sequence of SSE4.1 compatible instructions.
The `convert_i64x2_imul` custom legalization checks the ISA flags for AVX512DQ or AVX512VL support and legalizes `imul.i64x2` to an `x86_pmullq` in this case; if not, it uses a lengthy SSE2-compatible instruction sequence.
Without this special instruction, legalizing to the AVX512 instruction AND the SSE instruction sequence is impossible. This extra instruction would be rendered unnecessary by the x64 backend.
The InsertLane format has an ordering (`value().imm().value()`) and immediate name (`"lane"`) that make it awkward to use for other instructions. This changes the ordering (`value().value().imm()`) and uses the default name (`"imm"`) throughout the codebase.
* Encode vselect using BLEND instructions on x86
* Legalize vselect to bitselect
* Optimize bitselect to vselect for some operands
* Add run tests for bitselect-vselect optimization
* Address review feedback
This involves some large mask tables that may hurt code size but reduce the number of instructions. See https://github.com/WebAssembly/simd/issues/117 for a more in-depth discussion on this.
This involves some large mask tables that may hurt code size but reduce the number of instructions. See https://github.com/WebAssembly/simd/issues/117 for a more in-depth discussion on this.
The Intel manual uses `CMPNLT` and `CMPNLE` to denote not-less-than and not-less-than-or-equals. These were translated previously to `FloatCC::GreaterThan` and `FloatCC::GreaterThanOrEqual` but should be correctly translated to `FloatCC::UnorderedOrGreaterThanOrEqual` and `FloatCC::UnorderedOrGreaterThan`. This change adds the necessary legalizations to make use of these new encodings.
error[E0425]: cannot find value `ones` in this scope
--> cranelift-codegen/meta/src/isa/x86/legalize.rs:564:33
|
564 | def!(c = vconst(ones)),
| ^^^^ not found in this scope
Previously `fsub` was used but this fails when negating -0.0 and +0.0 in the SIMD spec tests; using more instructions, this change uses shifts to create a constant for flipping the most significant bit of each lane with `bxor`.
In order to implement SIMD's all_true (https://github.com/WebAssembly/simd/blob/master/proposals/simd/SIMD.md#all-lanes-true), we must legalize some instruction (I chose `vall_true`) to a comparison against 0 and a similar reduction as vany_true using `PTEST` and `SETNZ`. Since `icmp` only allows integers but `vall_true` could allow more vector types, `raw_bitcast` is used to convert the lane types into integers, e.g. b32x4 to i32x4. To do so without runtime type-checking, the `raw_bitcast` instruction (which emits no instruction) can now bitcast from any vector type to the same type, e.g. i32x4 to i32x4.
Only the shifts with applicable SSE2 instructions are implemented here: PSRL* (for ushr) only has 16-64 bit instructions and PSRA* (for sshr) only has 16-32 bit instructions.
This avoids doing multiple unpacking of the InstructionData for a single
legalization, improving readability and reducing size of the generated
code. For instance, icmp had to unpack the format once per IntCC
condition code.
This change should make the code more clear (and less code) when adding encodings for instructions with specific immediates; e.g., a constant with a 0 immediate could be encoded as an XOR with something like `const.bind(...)` without explicitly creating the necessary predicates. It has several parts:
* Introduce Bindable trait to instructions
* Convert all instruction bindings to use Bindable::bind()
* Add ability to bind immediates to BoundInstruction
This is an attempt to reduce some of the issues in #955.
This commit is based on the assumption that floats are already stored in XMM registers in x86. When extracting a lane, cranelift was moving the float to a regular register and back to an XMM register; this change avoids this by shuffling the float value to the lowest bits of the XMM register. It also assumes that the upper bits can be left as is (instead of zeroing them out).
raw_bitcast matches the intent of this legalization more clearly (to simply change the CLIF type without changing any bits) and the additional null encodings added are necessary for later instructions
