For legacy (arithmetic) instructions, the operand size override prefix (0x66)
is used to switch the operand data size from 32b to 16b (in 32/64-bit mode),
16b to 32b (in 16-bit mode). That's why we set OpSize16 for 16-bit instructions
and set OpSize32 for 32-bit instructions.
But it's not a generic rule any more after APX. APX adds 4 variants for
arithmetic instructions: promoted EVEX, NDD (new data destination), NF (no flag),
NF_NDD. All the 4 variants are in EVEX space and only legal in 64-bit
mode. EVEX.pp is set to 01 for the 16-bit instructions to encode 0x66.
For APX, we should set OpSizeFixed for 8/16/32/64-bit variants and set PD for
the 16-bit variants.
Hence, to reuse the classes ITy and its subclasses BinOp* for APX instructions,
we extract the OpSize setting from the class ITy.
It helps simplify the class definitions. Now, the only explicit usage of PS is
to check prefix 0x66/0xf2/0xf3 can not be used a prefix, e.g. wbinvd.
See 82974e0114f02ffc07557e217d87f8dc4e100a26 for more details.
Commit 1b531d54f623 (#74203) removed the usage of unique_ptrs of arrays
in favour of using vectors, but inadvertently increased peak memory
usage by removing the ability to deallocate vector memory that was no
longer needed mid-LDV.
In that same review, it was pointed out that `FuncValueTable` typedef
could be removed, since it was "just a vector".
This commit addresses both issues by making `FuncValueTable` a real data
structure, capable of mapping BBs to ValueTables and able to free
ValueTables as needed.
This reduces peak memory usage in the compiler by 10% in the benchmarks
flagged by the original review.
As a consequence, we had to remove a handful of instances of the
"declare-then-initialize" antipattern in unittests, as the
FuncValueTable class is no longer default-constructible.
The original brute force dominates algorithm is O(n) complexity so it is
very slow for very large machine basic block which is very common with
O0. This patch added InstrPosIndexes to assign index for each
instruction and use it to determine dominance. The complexity is now
O(1).
This implements assembly support for the Memory Systems Extensions
introduced as part of the Armv9.5-A architecture version.
The changes include:
* New subtarget feature for FEAT_TLBIW.
* New system registers for FEAT_HDBSS:
* HDBSSBR_EL2 and HDBSSPROD_EL2.
* New system registers for FEAT_HACDBS:
* HACDBSBR_EL2 and HACDBSCONS_EL2.
* New TLBI instructions for FEAT_TLBIW:
* VMALLWS2E1(nXS), VMALLWS2E1IS(nXS) and VMALLWS2E1OS(nXS).
* New system register for FEAT_FGWTE3:
* FGWTE3_EL3.
This adds the following instructions which are added in PAuthLR:
- PACIA171615
- PACIB171615
- AUTIA171615
- AUTIB171615
Also updates some encodings to match final published values.
Documentation can be found here:
https://developer.arm.com/documentation/ddi0602/2023-12/Base-Instructions
Co-authored-by: Lucas Prates <lucas.prates@arm.com>
Instead, we add a `BranchOnly` parameter to indicate that only
branches with its predecessors will be fused.
X86 is the only user of `createBranchMacroFusionDAGMutation`.
This can help us aovid introducing new classes T_MAP*PS|PD|XS|XD
when a new opcode map is supported.
And, T_MAP*PS|PD|XS|XD does not look better than T_MAP*, PS|PD|XS|XD.
This patch adds proper detection support for more znver2 CPUs.
Specifically, this adds in support for CPUs codenamed Renoir, Lucienne,
and Mendocino.
This was originally proposedfor Renoir in
https://reviews.llvm.org/D96220 and
got approved, but slipped through the cracks. However, there is still a
demand for this feature.
In addition to adding support for more znver2 CPUs, this patch also includes
some additional refactoring and comments related to cpu model
information for zen CPUs.
Fixes https://github.com/llvm/llvm-project/issues/74934.
`VEX_4V` does not look simpler than `VEX, VVVV`. It's kind of confusing
b/c classes like `VEX_L`, `VEX_LIG` do not imply `VEX` but it does.
For APX, we have promote EVEX, NDD, NF and NDD_NF instructions. All of
the 4 variants are in EVEX space and NDD/NDD_NF set the VVVV fields.
To extract the common fields (e.g EVEX) into a class and set VVVV
conditionally, we need VVVV to not imply other prefixes.
ISD::VP_MERGE treats the false operand as the source for elements past
VL. The vmerge instruction encodes 3 registers and treats the vd
register as the source for the tail.
This patch adds a new ISD opcode that models the tail source explicitly.
During lowering we copy the false operand to this operand.
I think we can merge RISCVISD::VSELECT_VL with this new opcode by using
an UNDEF passthru, but I'll save that for another patch.
LLVM ObjectFile currently records the start offsets of sections as the
start of the section header, whereas most other tools (WABT, emscripten,
wasm-tools) record it as the start of the section content, after the
header. This affects binutils tools such as objdump and nm, but not
compilation/assembly (since that is driven by symbols and assembler
labels which already have their values inside the section payload rather
in the header. This patch updates LLVM to match the other tools.
Renaming a member variable from "Endoding" to "Encoding".
Also replace inlined code for "isNormalized" with a call to the
function, so that if the definition of normalization ever changes, we
only need to change the one place.
By looking at whether a global is large instead of looking at the code
model.
This also fixes references to large data in the small code model.
We now always fold any 32-bit offset into the addressing mode with the
large code model since it uses 64-bit relocations.
Depositing value into the lowest byte/word is a common code pattern.
This patch improves the code generation for it to avoid redundant AND
and OR operations.
IR intrinsics were already defined, but no codegen support had been
added.
I extracted this code from our downstream. Some of it may have come from
https://repo.hca.bsc.es/gitlab/rferrer/llvm-epi/ originally.
Currently isGuaranteedNotToBeUndef() is the same as
isGuaranteedNotToBeUndefOrPoison(). This function is used in places
where we only care about undef (due to multi-use issues), not poison.
Make it more precise by only considering instructions that can create
undef (like loads or call), and ignore those that can only create
poison. In particular, we can ignore poison-generating flags.
This means that inferring more flags has less chance to pessimize other
transforms.
According to Intel SDE, ADCX reads CF and ADOX reads OF. `Uses` was
set to empty by accident, the bug was not exposed b/c compiler never
emits these instructions.
