While we can do that, it doesn't increase instruction count,
if the old `sub` sticks around then the transform is not only
not a unlikely win, but a likely regression, since we likely
now extended live range and use count of both of the `sub` operands,
as opposed to just the result of `sub`.
As Kostya Serebryany notes in post-commit review in
https://reviews.llvm.org/D68408#1998112
this indeed can degrade final assembly,
increase register pressure, and spilling.
This isn't what we want here,
so at least for now let's guard it with an use check.
it to fusing different kinds of linalg operations on tensors.
The implementation of fusion on tensor was initially planned for just
GenericOps (and maybe IndexedGenericOps). With addition of
linalg.tensor_reshape, and potentially other such non-structured ops,
refactor the existing implementation to allow easier specification of
fusion between different linalg operations on tensors.
Differential Revision: https://reviews.llvm.org/D78463
This was backwards from intended and missing a test. We perhaps should
just ignored the FP mode here, since it shouldn't be legal to mix code
with different default modes in the absence of strictfp.
If f32 denormals were enabled pre-gfx9, we would still try to
implement this with v_max_f32. Pre-gfx9, these instructions ignored
the denormal mode and did not flush. Switch to the multiply form for
f32 as a workaround which should always work in any case.
This fixes conformance failures when the library implementation of
fmin/fmax were accidentally not inlined, forcing the assumption of no
flushing on targets where denormals are not enabled by default. This
is a workaround, since really we should not be mixing code with
different FP mode expectations, but prefer the lowering that will work
in any mode.
Now this will always use max to implement canonicalize on gfx9+. This
is only really beneficial for f64. For f32/f16 it's a neutral choice
(and worse in terms of code size in 1 case), but possibly worse for
the compiler since it does add an extra register use operand. Leave
this change for later.
LC_LOAD_WEAK_DYLIB is analogous to LC_LOAD_DYLIB and doesn't require any special handling.
Test plan: make check-all
Differential revision: https://reviews.llvm.org/D78602
This will fix building the wrapper shared library when
BUILD_SHARED_LIBS is ON.
Reviewers: PaulkaToast
Differential Revision: https://reviews.llvm.org/D78737
367229e100 retired ValueHandle but
mistakenly removed the implementation for `negate` which was not
tested and would result in linking errors.
This revision adds the implementation back and provides a test.
Summary:
llvm::getInlineCost starts off by determining whether inlining should
happen or not because of user directives or easily determinable
unviability. This CL refactors this functionality as a reusable API.
Reviewers: davidxl, eraman
Reviewed By: davidxl, eraman
Subscribers: hiraditya, haicheng, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73825
Summary:
Piggy-back off of TypeSize's STRICT_FIXED_SIZE_VECTORS flag and:
- if it is defined, assert that the vector is not scalable
- if it is not defined, complain if the vector is scalable
Reviewers: efriedma, sdesmalen, c-rhodes
Reviewed By: sdesmalen
Subscribers: hiraditya, mgorny, tschuett, rkruppe, psnobl, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78576
The FSEvents APIs are available on iOS6+: however, the DirectoryWatcher
code isn't wired up to really use FSEvents on embedded platforms.
I've duplicated code from DirectoryWatcher-not-implemented.cpp here and
used TargetConditionals instead of adding cmakery to check try_compile;
I couldn't get that to work properly.
Static analyzer has a mechanism of clearing redundant nodes when
analysis hits a certain threshold with a number of nodes in exploded
graph (default is 1000). It is similar to GC and aims removing nodes
not useful for analysis. Unfortunately nodes corresponding to array
subscript expressions (that actively participate in data propagation)
get removed during the cleanup. This might prevent the analyzer from
generating useful notes about where it thinks the data came from.
This fix is pretty much consistent with the way analysis works
already. Lvalue "interestingness" stands for the analyzer's
possibility of tracking values through them.
Differential Revision: https://reviews.llvm.org/D78638
This reverts commit f969c2aa65.
There are some msan buildbot failures sanitzer-x86_64-linux-fast that
I need to investigate.
Differential Revision: https://reviews.llvm.org/D78422
Summary:
Move InstantiateComponent and InstantiateIntrinsicType from symbol.cpp
to type.cpp as that is where they are called.
Put both in InstantiateHelper class to better isolate them.
Add CreateDerivedTypeSpec in InstantiateHelper.
Add non-const forms for AsIntrinsic and AsDerived to avoid const_cast.
No functional changes.
Reviewers: DavidTruby, klausler, PeteSteinfeld, jdoerfert
Reviewed By: klausler
Subscribers: llvm-commits
Tags: #flang, #llvm
Differential Revision: https://reviews.llvm.org/D78678
1. Use Subtarget.isUsingPCRelativeCalls() in LowerConstantPool to
check if using PCRelative addressing.
2. Change MO_GOT_FLAG = 32 to MO_GOT_FLAG = 8 in PPC.h to use
consecutive bits.
Differential Revision: https://reviews.llvm.org/D78406
The contents of the .ARM.exidx section must be ordered by SHF_LINK_ORDER
rules. We don't need to know the precise address for this order, but we
do need to know the relative order of sections. We have been using the
sectionIndex for this purpose, this works when the OutputSection order
has a monotonically increasing virtual address, but it is possible to
write a linker script with non-monotonically increasing virtual address.
For these cases we need to evaluate the base address of the OutputSection
so that we can order the .ARM.exidx sections properly.
This change moves the finalisation of .ARM.exidx till after the first
call to AssignAddresses. This permits us to sort on virtual address which
is linker script safe. It also permits a fix for part of pr44824 where
we generate .ARM.exidx section for the vector table when that table is so
far away it is out of range of the .ARM.exidx section. This fix will come
in a follow up patch.
Differential Revision: https://reviews.llvm.org/D78422
Calling Disconnect while the read thread is running is racy because the
thread can also call Disconnect. This is a follow-up to b424b0bf, which
reorders other occurences of Disconnect/StopReadThread I can find, and also
adds an assertion to guard against new occurences being introduced.
The current Liveness analysis does not support operations with nested regions.
This causes issues when querying liveness information about blocks nested within
operations. Furthermore, the live-in and live-out sets are not computed properly
in these cases.
Differential Revision: https://reviews.llvm.org/D77714
Sanjay Patel