If a srl doesn't introduce any sign bits into the truncated result, then replace with a sra to let us use a PACKSS truncation - fixes a regression noticed in D56387 on pre-SSE41 targets that don't have PACKUSDW.
This caused a miscompile in Chromium, see comments on the codereview for
discussion and pointer to a reproducer.
> InstCombine already performs a fold where X == Y ? f(X) : Z is
> transformed to X == Y ? f(Y) : Z if f(Y) simplifies. However,
> if f(X) only has one use, then we can always directly replace the
> use inside the instruction. To actually be profitable, limit it to
> the case where Y is a non-expr constant.
>
> This could be further extended to replace uses further up a one-use
> instruction chain, but for now this only looks one level up.
>
> Among other things, this also subsumes D94860.
>
> Differential Revision: https://reviews.llvm.org/D94862
This also reverts the follow-up
a003f26539cf4db744655e76c41f4c4a8913f116:
> [llvm] Prevent infinite loop in InstCombine of select statements
>
> This fixes an issue where the RHS and LHS the comparison operation
> creating the predicate were swapped back and forth forever.
>
> Differential Revision: https://reviews.llvm.org/D94934
D84108 exposed a bad interaction between inlining and loop-rotation
during regular LTO, which is causing notable regressions in at least
CINT2006/473.astar.
The problem boils down to: we now rotate a loop just before the vectorizer
which requires duplicating a function call in the preheader when compiling
the individual files ('prepare for LTO'). But this then prevents further
inlining of the function during LTO.
This patch tries to resolve this issue by making LoopRotate more
conservative with respect to rotating loops that have inline-able calls
during the 'prepare for LTO' stage.
I think this change intuitively improves the current situation in
general. Loop-rotate tries hard to avoid creating headers that are 'too
big'. At the moment, it assumes all inlining already happened and the
cost of duplicating a call is equal to just doing the call. But with LTO,
inlining also happens during full LTO and it is possible that a previously
duplicated call is actually a huge function which gets inlined
during LTO.
From the perspective of LV, not much should change overall. Most loops
calling user-provided functions won't get vectorized to start with
(unless we can infer that the function does not touch memory, has no
other side effects). If we do not inline the 'inline-able' call during
the LTO stage, we merely delayed loop-rotation & vectorization. If we
inline during LTO, chances should be very high that the inlined code is
itself vectorizable or the user call was not vectorizable to start with.
There could of course be scenarios where we inline a sufficiently large
function with code not profitable to vectorize, which would have be
vectorized earlier (by scalarzing the call). But even in that case,
there probably is no big performance impact, because it should be mostly
down to the cost-model to reject vectorization in that case. And then
the version with scalarized calls should also not be beneficial. In a way,
LV should have strictly more information after inlining and make more
accurate decisions (barring cost-model issues).
There is of course plenty of room for things to go wrong unexpectedly,
so we need to keep a close look at actual performance and address any
follow-up issues.
I took a look at the impact on statistics for
MultiSource/SPEC2000/SPEC2006. There are a few benchmarks with fewer
loops rotated, but no change to the number of loops vectorized.
Reviewed By: sanwou01
Differential Revision: https://reviews.llvm.org/D94232
This patch adds a new test case which depends on AArch64 SVE support and
dynamic resize capability enabled. It created two seperate threads which
have different values of sve registers and SVE vector granule at various
points during execution.
We test that LLDB is doing the size and offset updates properly for all
of the threads including the main thread and when we VG is updated using
prctl call or by 'register write vg' command the appropriate changes are
also update in register infos.
Reviewed By: labath
Differential Revision: https://reviews.llvm.org/D82866
This patch builds on previously submitted SVE patches regarding expedited
register set and per thread register infos. (D82853 D82855 and D82857)
We need to resize SVE register based on value received in expedited list.
Also we need to resize SVE registers when we write vg register using
register write vg command. The resize will result in a updated offset
for all of fpr and sve register set. This offset will be configured
in native register context by RegisterInfoInterface and will also be
be updated on client side in GDBRemoteRegisterContext.
A follow up patch will provide a API test to verify this change.
Reviewed By: labath
Differential Revision: https://reviews.llvm.org/D82863
The test couldn't find lldb-server as it's path was being overridden by
LLDB_DEBUGSERVER_PATH environment variable (pointing to debugserver).
This test should always use lldb-server, as it tests its platform
capabilities.
There's no need for the environment override, as lldb-server tests
should test the executable they just built, so I just remote the
override capability.
This patch handles cases where we have to save/restore the link register
into the stack and and load/store instruction which use the stack are
part of the outlined region. It checks that there will be no overflow
introduced by the new offset and fixup these instructions accordingly.
Differential Revision: https://reviews.llvm.org/D92934
When a command option does not have a short version
(e.g. -f for --file), we use an arbitrary value in the
short_option field to mark it as invalid.
(though this value is unqiue to be used later for other
things)
We check that this short option is valid to print using
llvm::isPrint. This implicitly casts our int to char,
meaning we check the last char of any short_option value.
Since the arbitrary value we chose for these options is
some shortened hex version of the name, this returned true
even for invalid values.
Since llvm::isPrint returns true we later call std::islower
and/or std::isupper on the short_option value. (the int)
Calling these functions with something that cannot be validly
converted to unsigned char is undefined. Somehow we got/get
away with this but for me compiling with g++-9 I got a crash
for "help memory read".
The other command that uses this is "target variable" but that
didn't crash for unknown reasons.
Checking that short_option can fit into an unsigned char before
we call llvm::isPrint means we will not attempt to call islower/upper
on these options since we have no reason to print them.
This also fixes bogus short options being shown for "memory read"
and target variable.
For "target variable", before:
-e <filename> ( --file <filename> )
-b <filename> ( --shlib <filename> )
After:
--file <filename>
--shlib <filename>
(note that the bogus short options are just the bottom byte of our
arbitrary short_option value)
Reviewed By: labath
Differential Revision: https://reviews.llvm.org/D94917
This fixes an issue where the RHS and LHS the comparison operation
creating the predicate were swapped back and forth forever.
Differential Revision: https://reviews.llvm.org/D94934
In addition to consistency, we'll hit a wall when 11.1.0 gets released, because
we cannot represent it with lit versioning scheme.
Differential Revision: https://reviews.llvm.org/D94157
Previously uniqueCallSite could have race conditions between different
threads. Now it is accessed with an atomic RMW and will be unique
between different threads.
Differential Revision: https://reviews.llvm.org/D94784
A previous patch has already changed getInstructionCost to return
an InstructionCost type. This patch changes the other various
getXXXCost functions to return an InstructionCost too. This is a
non-functional change - I've added a few asserts that the costs
are valid in places where we're selecting between vector call
and intrinsic costs. However, since we don't yet return invalid
costs from any of the TTI implementations these asserts should
not fire.
See this patch for the introduction of the type: https://reviews.llvm.org/D91174
See this thread for context: http://lists.llvm.org/pipermail/llvm-dev/2020-November/146408.html
Differential Revision: https://reviews.llvm.org/D94065
This patch moves the parsing of `{Lang,CodeGen}Options` from `parseSimpleArgs` to the original `Parse{Lang,CodeGen}Args` functions.
This ensures all marshalled `LangOptions` are being parsed **after** the call `setLangDefaults`, which in turn enables us to marshall `LangOptions` that somehow depend on the defaults. (In a future patch.)
Now, `CodeGenOptions` need to be parsed **after** `LangOptions`, because `-cl-mad-enable` (a `CodeGenOpt`) depends on the value of `-cl-fast-relaxed-math` and `-cl-unsafe-math-optimizations` (`LangOpts`).
Unfortunately, this removes the nice property that marshalled options get parsed in the exact order they appear in the `.td` file. Now we cannot be sure that a TableGen record referenced in `ImpliedByAnyOf` has already been parsed. This might cause an ordering issues (i.e. reading value of uninitialized variable). I plan to mitigate this by moving each `XxxOpt` group from `parseSimpleArgs` back to their original parsing function. With this setup, if an option from group `A` references option from group `B` in TableGen, the compiler will require us to make the `CompilerInvocation` member for `B` visible in the parsing function for `A`. That's where we notice that `B` didn't get parsed yet.
Reviewed By: Bigcheese
Differential Revision: https://reviews.llvm.org/D94682
Element sections will also need flags, so we shouldn't squat the
WASM_SEGMENT namespace.
Depends on D90948.
Differential Revision: https://reviews.llvm.org/D92315
This patch changes to make call_indirect explicitly refer to the
corresponding function table, residualizing TABLE_NUMBER relocs against
it.
With this change, wasm-ld now sees all references to tables, and can
link multiple tables.
Differential Revision: https://reviews.llvm.org/D90948
`ssize_t` is from POSIX and is not standard unfortunately.
Rewritting the code so it doesn't depend on it.
Differential Revision: https://reviews.llvm.org/D94760
Previously committed as 9e08e51a20, and
reverted because a dependency commit was reverted. This incorporates the
following follow-on commits that were also reverted:
7e84aa1b81 by Simon Pilgrim
ed13d8c667 by me
95c7b6cadb by Sam McCall
430d5d8429 by Dave Zarzycki
to dependent declarations.
Treat an id-expression that names a local variable in a templated
function as being instantiation-dependent.
This addresses a language defect whereby a reference to a dependent
declaration can be formed without any construct being value-dependent.
Fixing that through value-dependence turns out to be problematic, so
instead this patch takes the approach (proposed on the core reflector)
of allowing the use of pointers or references to (but not values of)
dependent declarations inside value-dependent expressions, and instead
treating template arguments as dependent if they evaluate to a constant
involving such dependent declarations.
This ends up affecting a bunch of OpenMP tests, due to OpenMP
imprecisely handling instantiation-dependent constructs, bailing out
early instead of processing dependent constructs to the extent possible
when handling the template.
Previously committed as 8c1f2d15b8, and
reverted because a dependency commit was reverted.
the nested-name-specifier when determining whether a qualified type is
instantiation-dependent.
Previously reverted in 25a02c3d1a due to
causing us to reject some code. It turns out that the rejected code was
ill-formed (no diagnostic required).
if E is merely instantiation-dependent.
Previously reverted in 34e72a146111dd986889a0f0ec8767b2ca6b2913;
re-committed with a fix to an issue that caused name mangling to assert.
Add nounwind attribute to avoid generating cfi instructions. Also make
global buffer 64 bytes align in lit test case.
Differential Revision: https://reviews.llvm.org/D94910
As of 8dacca943a, we sign extend the atomic loaded
operand for signed subword comparisons. However, the assumption that the other
operand is correctly sign extended doesn't always hold. This patch sign extends
the other operand if it needs to be sign extended.
This is a second fix for https://bugs.llvm.org/show_bug.cgi?id=30451
Differential revision: https://reviews.llvm.org/D94058
The C++ standard wording doesn't appear to properly handle the case
where a class inherits a default constructor from a base class. Various
properties of classes are defined in terms of the corresponding property
of the default constructor, and in this case, the class does not have a
default constructor despite being default-constructible, which the
wording doesn't handle properly.
This change implements a tentative fix for these problems, which has
also been proposed to the C++ committee: if a class would inherit a
default constructor, and does not explicitly declare one, then one is
implicitly declared.
This is a additional bug fix for c5be0e0cc0. The distance for
the spill instructions is wrong in previous patch.
Differential Revision: https://reviews.llvm.org/D94772
This patch teaches SimplifyCFG::SimplifyBranchOnICmpChain to understand select form of
(x == C1 || x == C2 || ...) / (x != C1 && x != C2 && ...) and optimize them into switch if possible.
D93065 has more context about the transition, including links to the list of optimizations being updated.
Differential Revision: https://reviews.llvm.org/D93943
The code here is checking to see if two sets are identical.
OtherBlocksSet should point to OtherL->getBlocksSet() instead.
Differential Revision: https://reviews.llvm.org/D94926
The implementation had a lot of boilerplate and was more complicated than
necessary. This NFC refactoring introduces a few macros to reduce code
duplication, and uses a consistent style and formatting for the whole file.
Differential Revision: https://reviews.llvm.org/D94544
After much refactoring over the last 2 weeks to the reduction
matching code, I think this change is finally ready.
We effectively broke fmax/fmin vector reduction optimization
when we started canonicalizing to intrinsics in instcombine,
so this should restore that functionality for SLP.
There are still FMF problems here as noted in the code comments,
but we should be avoiding miscompiles on those for fmax/fmin by
restricting to full 'fast' ops (negative tests are included).
Fixing FMF propagation is a planned follow-up.
Differential Revision: https://reviews.llvm.org/D94913
Alexander Belyaev