The GlobalISel TableGen backend didn't check for predicates on the
source children. This caused it to generate code for ARM patterns such
as SMLABB or similar, but without properly checking for the sext_16_node
part of the operands. This in turn meant that we would select SMLABB
instead of MLA for simple sequences such as s32 + s32 * s32, which is
wrong (we want a MLA on the full operands, not just their bottom 16
bits).
This patch forces TableGen to skip patterns with predicates on the src
children, so it doesn't generate code for SMLABB and other similar ARM
instructions at all anymore. AArch64 and X86 are not affected.
Differential Revision: https://reviews.llvm.org/D39554
llvm-svn: 317313
This will enable us to prefer VALIGND/Q during shuffle lowering in order to get the extended register encoding space when BWI isn't available. But if we end up not using the extended registers we can switch VPALIGNR for the shorter VEX encoding.
Differential Revision: https://reviews.llvm.org/D39401
llvm-svn: 317122
The importer will now accept nested instructions in the result pattern such as
(ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction
def's a single vreg and the parent instruction consumes a single vreg where a
nested instruction is specified. The importer will automatically create a vreg
to connect the two using the type information from the pattern. This vreg will
be constrained to the register classes given in the instruction definitions*.
* REG_SEQUENCE is explicitly rejected because of this. The definition doesn't
constrain to a register class and it therefore needs special handling.
llvm-svn: 317117
The next commit will add support for multi-instruction emission so we need to
start allocating instruction ID's instead of hard-coding them to 0.
llvm-svn: 317057
I need a test that only runs in a reasonable amount of time on systems
that have sparse files. The broadest class of systems that support
sparse files are linux systems. So restricting my test to linux systems
should suffice. This change adds the system-linux feature to llvm-lit so
that it can be required.
Differential Revision: https://reviews.llvm.org/D39482
llvm-svn: 317055
Multi-instruction emission needs to ensure the the instructions are generated
a depth-first fashion. For example:
(ADDWrr (SUBWrr a, b), c)
needs to emit the SUBWrr before the ADDWrr. However, our walk over
TreePatternNode's is highly context sensitive which makes it difficult to append
BuildMIActions in the order we want. To fix this, we now keep track of the
insertion point as we add actions. This will allow multi-insn emission to insert
BuildMI's in the correct place.
The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the
const_iterator forms of insert() and emplace() that were added in C++11. As a
result I've switched the const_iterators to iterators.
llvm-svn: 317049
The same bots fail but I believe I know what the issue is now. These bots are
missing the const_iterator versions of insert/emplace/etc. that were introduced
in C++11.
llvm-svn: 317042
Multi-instruction emission needs to ensure the the instructions are generated
a depth-first fashion. For example:
(ADDWrr (SUBWrr a, b), c)
needs to emit the SUBWrr before the ADDWrr. However, our walk over
TreePatternNode's is highly context sensitive which makes it difficult to append
BuildMIActions in the order we want. To fix this, we now keep track of the
insertion point as we add actions. This will allow multi-insn emission to insert
BuildMI's in the correct place.
The previous commit failed on the Ubuntu bots using GCC 4.8. These bots didn't
like a call to emplace(). I've replaced it with insert() to see if it's a quirk
of the C++11 support.
llvm-svn: 317040
Multi-instruction emission needs to ensure the the instructions are generated
a depth-first fashion. For example:
(ADDWrr (SUBWrr a, b), c)
needs to emit the SUBWrr before the ADDWrr. However, our walk over
TreePatternNode's is highly context sensitive which makes it difficult to append
BuildMIActions in the order we want. To fix this, we now keep track of the
insertion point as we add actions. This will allow multi-insn emission to insert
BuildMI's in the correct place.
llvm-svn: 317029
Multi-instruction emission will require that we have separate handling for
the defs between the implicitly created temporaries and the rule outputs.
The former require new temporary vregs while the latter should copy existing
operands. Factor out the implicit def/use renderers to minimize the code
duplication when we implement that.
llvm-svn: 317025
Prepare for multiple instruction emission by allowing BuildMIAction to
search for a suitable matcher that will support mutation.
This patch deliberately neglects to add matchers aside from the root to
preserve NFC. That said, it should be noted that until we support mutations
other than just the opcode the chances of finding a non-root instruction
for which canMutate() is true, is essentially zero. Furthermore in the
presence of multi-instruction emission the chances of finding any
instruction for which canMutate() is true is also zero. Nevertheless, we
can't continue to require that all BuildMIAction's consider the root of the match
to be recyclable due to the risk of recycling it twice in the same rule.
llvm-svn: 317022
Based on similar python tool - utils/shuffle-fuzz.py - this tool extends the ability of it's previous by optionally attaching select instruction to the generated shufflevector instructions.
This was mainly developed to perform exhaustive testing of the X86 AVX512 masked shuffle instructions. But yet it can be used for various other targets.
The general design of the implementation is much modular than the original shuffle_fuzz.py tool, which makes it easier for anyone to extend it further.
Differential Revision: https://reviews.llvm.org/D38031
Change-Id: I0efc2aaa091b61a8a9552311c21cc77916a97111
llvm-svn: 316989
gtest depends on this #define to determine whether it can
use various classes like std::tuple, or whether it has to fall
back to experimental classes in the std::tr1 namespace. The
check in the current version of gtest relies on the value of
the `__cplusplus` macro, but MSVC provides a non-conformant
value of this macro, making it effectively impossible to detect
C++11. In short, LLVM compiled with MSVC has been silently
using the tr1 versions of several classes since the beginning of
time.
This would normally be pretty benign, except that in the latest
preview of MSVC they have marked all of the tr1 classes
deprecated, so it spews thousands of warnings.
llvm-svn: 316798
When multi-instruction emission is supported, it will no longer be guaranteed
that every BuildMIAction has a corresponding matched instruction. BuildMIAction
should support not having one to cover the case where a rule produces more
instructions than it matched.
llvm-svn: 316463
Ideally, we should compare 32- and 64-bit versions to see if the
ret line is the only difference and then insert the regex only
in that case. But this is a quick hack to avoid a bunch of noise
as existing tests are updated.
llvm-svn: 316443
This patch enables the import of stores. Unfortunately, doing so by itself,
loses an optimization where storing 0 to memory makes use of WZR/XZR.
To mitigate this, this patch also introduces a new feature that allows register
operands to nominate a zero register. When this is done, GlobalISel will
substitute (G_CONSTANT 0) with the nominated register automatically. This
is currently configured to only apply to the stores.
Applying it to GPR32/GPR64 register classes in general will be done after
review see (https://reviews.llvm.org/D39150).
llvm-svn: 316360
Summary: test/CodeGen/PowerPC/pr33093.ll uses both powerpc64 (big-endian) and powerpc64le while the former was unsupported.
Subscribers: nemanjai
Differential Revision: https://reviews.llvm.org/D39164
llvm-svn: 316297
This is similar to how we generate the VEX tables.
More fixes are still needed for the instructions that use EVEX.b (broadcast and embedded rounding).
llvm-svn: 316294
This introduces a new operand type to encode the whether the index register should be XMM/YMM/ZMM. And new code to fixup the results created by readSIB.
This has the nice effect of removing a bunch of code that hard coded the name of every GATHER and SCATTER instruction to map the index type.
This fixes PR32807.
llvm-svn: 316273
This is a temporary hack to support adding checks for the "registers:"
block of mir functions. This is necessary to convert a number of tests
so that there's less churn when we change the MIR printer to put the
vreg classes on defs instead of in their own block.
llvm-svn: 316134
An empty livein block doesn't make much sense (why not just omit it?)
but they're legal and some tests have them, so its best to handle it.
llvm-svn: 316089
Matching prefixes isn't good enough, because it leads to things like
calling the first constant C3 just because there were two copies
before it. Tighten up the check to match more precisely, but also be
careful about ambiguity when dealing with target opcodes that end in a
number.
llvm-svn: 316088
This adds update_mir_test_checks, which updates the check lines in mir
tests. This can only update tests that start and end with .mir
currently (ie, -run-pass) but it should be sufficient for updating at
least some of the GlobalISel tests.
llvm-svn: 316057
The new scheme should match the normalization of embedded paths in
linkrepro tar files.
Differential Revision: https://reviews.llvm.org/D39023
llvm-svn: 316044
MSVC doesn't seem to like implicitly instantiating addPredicate and then
explicitly specializing it later. It causes an internal compiler error.
llvm-svn: 315930
