This patch updates the load insert point of the merged load in AggressiveInstCombine().
This is done to handle the reported test breaks by handling Alias Analysis correctly.
Differential Revision: https://reviews.llvm.org/D137201
Compare a relative speed of misaligned accesses before and
after vectorization, not just check the new instruction is
not going to be slower.
Since no target now returns anything but 0 or 1 for Fast
argument of the allowsMisalignedMemoryAccesses this is still NFCI.
The subsequent patch will tune actual vaues of Fast on AMDGPU.
Differential Revision: https://reviews.llvm.org/D124218
Use collectOffset to collect scaled indices and constant offset for GEP
instead of custom code. This simplifies the logic in decomposeGEP and
allows to handle all cases supported by the generic helper.
This patch splits off the logic to transform the canonical IV to a
a value for an induction with a different start and step. This
transformation only needs to be done once (independent of VF/UF) and
enables sinking of VPScalarIVStepsRecipe as follow-up.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D133758
If left unchecked, the SLPVecrtorizer can move loads/stores below a stackrestore. The move can cause issues if the loads/stores have pointer operands from `alloca`s that are reset by the stackrestores. This patch adds the dependency check.
The check is conservative, in that it does not check if the pointer operands of the loads/stores are actually from `alloca`s that may be reset. We did not observe any SPECCPU2017 performance degradation so this simple fix seems sufficient.
The test could have been added to `llvm/test/Transforms/SLPVectorizer/X86/stacksave-dependence.ll`, but that test has not been updated to use opaque pointers. I am not inclined to add tests that still use typed pointers, or to refactor `llvm/test/Transforms/SLPVectorizer/X86/stacksave-dependence.ll` to use opaque pointers in this patch. If desired, I will open a different patch to refactor and consolidate the tests.
Reviewed By: ABataev
Differential Revision: https://reviews.llvm.org/D138585
The Assignment Tracking debug-info feature is outlined in this RFC:
https://discourse.llvm.org/t/
rfc-assignment-tracking-a-better-way-of-specifying-variable-locations-in-ir
Split dbg.assign intrinsics into fragments similarly to what SROA already does
for dbg.declares, except that there's many more intrinsics to split. The
function migrateDebugInfo generates new dbg.assigns intrinsic for each part of
a split store.
Reviewed By: jmorse
Differential Revision: https://reviews.llvm.org/D133296
This is a recommit of cf624b23bc,
which was reverted in 5cfc22cafe,
because the cut-off on the number of vector elements was not low enough,
and it triggered both SDAG SDNode operand number assertions,
and caused compile time explosions in some cases.
Let's try with something really *REALLY* conservative first,
just to get somewhere, and try to bump it (to 64/128) later.
FIXME: should this respect TTI reg width * num vec regs?
Original commit message:
Now, there's a big caveat here - these bytes
are abstract bytes, not the i8 we have in LLVM,
so strictly speaking this is not exactly legal,
see e.g. https://github.com/AliveToolkit/alive2/issues/860
^ the "bytes" "could" have been a pointer,
and loading it as an integer inserts an implicit ptrtoint.
But at the same time,
InstCombine's `InstCombinerImpl::SimplifyAnyMemTransfer()`
would expand a memtransfer of 1/2/4/8 bytes
into integer-typed load+store,
so this isn't exactly a new problem.
Note that in memory, poison is byte-wise,
so we really can't widen elements,
but SROA seems to be inconsistent here.
Fixes#59116.
StoredValues only has entries for members of the interleave group. If
there are gaps, then using the index i here will either access a wrong
entry or be out-of-bounds.
Instead use a dedicated index that only gets incremented for members of
the interleave group.
Fixes#59090.
Summary:
Refactor loop peeling code by moving code for calculating phi invariance
into a separate class that does the calculation. Redescribe and rework
the algorithm in preparation for adding increased functionality. Add
test case that does not exhibit peeling that will be subsequently supported.
Author: Jamie Schmeiser <schmeise@ca.ibm.com>
Reviewed By: mkazantsev (Max Kazantsev)
Differential Revision: https://reviews.llvm.org/D138232
We were checking for a desirable integer type even when there
is no shift in the transform. This is unnecessary since we
are truncating directly to the destination type.
This removes an extractelt in more cases and seems to make the
canonicalization more uniform overall. There's still a potential
difference between patterns that need a shift vs. trunc-only.
I'm not sure if that is worth keeping at this point, but it can
be adjusted in another step (assuming this change does not cause
trouble).
In the most basic case where I noticed this, we missed a fold
that would have completely removed vector ops from a pattern
like:
https://alive2.llvm.org/ce/z/y4Qdte
We don't strip llvm.mir.debugify metadata in `llvm::stripDebugifyMetadata`. This
may lead to incorrect number of lines and variables in the metadata when we run
debugify twice, e.g. -run-pass=mir-debugify,...,mir-strip-debug,...,mir-debugify.
Differential Revision: https://reviews.llvm.org/D138417
This reverts a change to exclude scalarizeBinopOrCmp in VectorCombine for
scalable vectors which caused poor scalable Binop codegen.
Differential Revision: https://reviews.llvm.org/D138545
When ObjCARCOpt::run() returned early, Changed and CFGChanged were never initialized. CFGChanged is read unconditionally afterwards. This came up in the course of D137942.
This relands the ODR indicator part of D138095 (reverted by 06c74b5e73):
a `__odr_asan_gen_*` symbol should use a mangled name as its associated symbol does.
Otherwise, `compiler-rt/test/asan/TestCases/pr33372.cpp` fails with an assertion:
```
clang-16: /repositories/llvm-project/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp:11988: void llvm::SelectionDAG::createOperands(llvm::SDNode *, ArrayRef<llvm::SDValue>): Assertion `SDNode::getMaxNumOperands() >= Vals.size() && "too many operands to fit into SDNode"' failed.
```
I'm not sure if this should be even more conservative,
or if we have a named constant for this in middle-end.
Now, there's a big caveat here - these bytes
are abstract bytes, not the i8 we have in LLVM,
so strictly speaking this is not exactly legal,
see e.g. https://github.com/AliveToolkit/alive2/issues/860
^ the "bytes" "could" have been a pointer,
and loading it as an integer inserts an implicit ptrtoint.
But at the same time,
InstCombine's `InstCombinerImpl::SimplifyAnyMemTransfer()`
would expand a memtransfer of 1/2/4/8 bytes
into integer-typed load+store,
so this isn't exactly a new problem.
Note that in memory, poison is byte-wise,
so we really can't widen elements,
but SROA seems to be inconsistent here.
Fixes#59116.
The KCFI sanitizer emits "kcfi" operand bundles to indirect
call instructions, which the LLVM back-end lowers into an
architecture-specific type check with a known machine instruction
sequence. Currently, KCFI operand bundle lowering is supported only
on 64-bit X86 and AArch64 architectures.
As a lightweight forward-edge CFI implementation that doesn't
require LTO is also useful for non-Linux low-level targets on
other machine architectures, add a generic KCFI operand bundle
lowering pass that's only used when back-end lowering support is not
available and allows -fsanitize=kcfi to be enabled in Clang on all
architectures.
This relands commit eb2a57ebc7 with
fixes.
Reviewed By: nickdesaulniers, MaskRay
Differential Revision: https://reviews.llvm.org/D135411
This rectifies a FIXME that dates all the way back
to 2014 about not doing so due to the backend issues.
Presumably sufficient amount of time has passes
and all the known issues have been addressed,
or at least we will find out of there are some left...
As it has been established previously by precedent,
if we see a pointer type, then that is the type we must use.
Essentially, we don't want to introduce `inttoptr`'s.
ControlHeightReduction (CHR) clones the code region to reduce the
branches in the hot code path. The number of clones is linear to the
depth of the region.
Currently it does not have control over the code size increase. We are
seeing one ~9000 BB functions get expanded to ~250000 BBs, an 25x
increase. This creates a big compile time issue for the downstream
optimizations.
This patch adds a cap for number of clones for one region.
Differential Revision: https://reviews.llvm.org/D138333
This follows 87debdadaf to further eliminate wasting time
calling helper functions only to early return to the main
run loop.
Once again, this results in significant savings based on
experimental data:
https://llvm-compile-time-tracker.com/compare.php?from=01023bfcd33f922ed8c934ce563e54abe8bfe246&to=3dce4f70b73e48ccb045decb634c185e6b4c67d5&stat=instructions:u
This is NFCI other than making the pass faster. The total
cost of VectorCombine runs in an -O3 build appears to be
well under 0.1% of compile-time now, so there's not much
left to do AFAICT.
There's a TODO about making the code cleaner, but it
probably doesn't change timing much. I didn't include those
changes here because it requires updating much more code.
An extractelt with a constant index which extracts an element from the
two vector operands of a select can be directly folded into a select.
extractelt (select %x, %vec1, %vec2), %const ->
select %x, %vec1[%const], %vec2[%const]
Note: the implementation currently only works for constant vector operands.
Reviewed By: foad, spatel
Differential Revision: https://reviews.llvm.org/D137934
The add from the IV in the inner loop was always checking for 2 uses,
the phi and the compare. The compare could be based on the phi though,
leaving one valid use of the compare. In the testcase we could be left
with the phi and a lcssa phi as the two users, invalidly allowing
flattening where we shouldn't.
Fixes 58441
Differential Revision: https://reviews.llvm.org/D138404
These limitations are too strict, and their only purpose is to avoid code
size explosion. These restrictions seem obsolete, and the size problem
is solved in other places through cheap expansion limits.
The motivation is that the old code cannot deal with comparisons against
induction variant's increment.
Differential Revision: https://reviews.llvm.org/D138412
Reviewed By: lebedev.ri, reames
This amends commit 00be3578e0 to demangle symbol
names in global descriptors. We keep the mangled name for the `__odr_gen_asan_*`
variables and the runtime __cxa_demangle call site change (which fixed possible
leaks for other scenarios: non-fatal diagnostics).
compiler-rt/lib/sanitizer_common/sanitizer_symbolizer_posix_libcdep.cpp uses
an undefined weak `__cxa_demangle` which does not pull in an archive definition.
A -static-libstdc++ executable link does not get demangled names.
Unfortunately this means we cannot rely on runtime demangling.
See compiler-rt/test/asan/TestCases/global-demangle.cpp
The option was added with https://reviews.llvm.org/D102496,
and currently the name is accurate, but I am hoping to add
a load transform that is not a scalarization. See issue #17113.
extractelements.
If the resulting type is going to be scalarized, no need to adjust the
cost of removed extractelement and insert/extract subvector costs.
Otherwise, the compiler can crash because of the wrong type sizes.
Minor refactoring in LoopVectorizationCostModel::calculateRegisterUsage.
Also adding some FIXME:s related to what appears to be some short
comings related to how the register usage is calculated.
Differential Revision: https://reviews.llvm.org/D138342
This patch replaces NoneType() and NoneType::None with None in
preparation for migration from llvm::Optional to std::optional.
In the std::optional world, we are not guranteed to be able to
default-construct std::nullopt_t or peek what's inside it, so neither
NoneType() nor NoneType::None has a corresponding expression in the
std::optional world.
Once we consistently use None, we should even be able to replace the
contents of llvm/include/llvm/ADT/None.h with something like:
using NoneType = std::nullopt_t;
inline constexpr std::nullopt_t None = std::nullopt;
to ease the migration from llvm::Optional to std::optional.
Differential Revision: https://reviews.llvm.org/D138376
The runtime calls `MaybeDemangleGlobalName` for error reporting and
`__cxxabiv1::__cxa_demangle` is called if available, so demanging Itanium
mangled names in global metadata is unnecessary and wastes data size.
Add `MaybeDemangleGlobalName` in ODR violation detection to support demangled
names in a suppressions file. `MaybeDemangleGlobalName` may call
`DemangleCXXABI` and leak memory. Use an internal allocation to prevent lsan
leak (in case there is no fatal asan error).
The debug feature `report_globals=2` prints information for all instrumented
global variables. `MaybeDemangleGlobalName` would be slow, so don't do that.
The output looks like `Added Global[0x56448f092d60]: beg=0x56448fa66d60 size=4/32 name=_ZL13test_global_2`
and I think the mangled name is fine.
Other mangled schemes e.g. Windows (see win-string-literal.ll) remain the
current behavior.
Reviewed By: hctim
Differential Revision: https://reviews.llvm.org/D138095
The Assignment Tracking debug-info feature is outlined in this RFC:
https://discourse.llvm.org/t/
rfc-assignment-tracking-a-better-way-of-specifying-variable-locations-in-ir
This reduces peak memory overhead by 15% when building CTMark's tramp3d-v4 with
-O2 -g with assignment tracking enabled.
Reviewed By: jmorse
Differential Revision: https://reviews.llvm.org/D133321
The Assignment Tracking debug-info feature is outlined in this RFC:
https://discourse.llvm.org/t/
rfc-assignment-tracking-a-better-way-of-specifying-variable-locations-in-ir
The inliner requires two additions:
fixupAssignments - Update inlined instructions' DIAssignID metadata so that
inlined DIAssignID attachments are unique to the inlined instance.
trackInlinedStores - Treat inlined stores to caller-local variables
(i.e. callee stores to argument pointers that point to the caller's allocas) as
assignments. Track them using trackAssignments, which is the same method as is
used by the AssignmentTrackingPass. This means that we're able to detect stale
memory locations due to DSE after inlining. Because the stores are only tracked
_after_ inlining, any DSE or movement of stores _before_ inlining will not be
accounted for. This is an accepted limitation mentioned in the RFC.
One change is also required:
Update CloneBlock to preserve debug use-before-defs. Otherwise the assignments
will be dropped due to having the intrinsic operands replaced with empty
metadata (see use-before-def.ll in this patch and this related discourse post.
Reviewed By: jmorse
Differential Revision: https://reviews.llvm.org/D133318
The Assignment Tracking debug-info feature is outlined in this RFC:
https://discourse.llvm.org/t/
rfc-assignment-tracking-a-better-way-of-specifying-variable-locations-in-ir
Update simplifycfg:
sinkLastInstruction - preserve debug use-before-defs.
SpeculativelyExecuteBB - replace the value component of dbg.assign intrinsics
when stores are hoisted and merged using a select, and don't delete them.
Reviewed By: jmorse
Differential Revision: https://reviews.llvm.org/D133310
Juan Manuel MARTINEZ CAAMAÑO