This revision supports to scalarize a binary operation of two scalable splat vectors.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D122791
This was stored in LiveIntervals, but not actually used for anything
related to LiveIntervals. It was only used in one check for if a load
instruction is rematerializable. I also don't think this was entirely
correct, since it was implicitly assuming constant loads are also
dereferenceable.
Remove this and rely only on the invariant+dereferenceable flags in
the memory operand. Set the flag based on the AA query upfront. This
should have the same net benefit, but has the possible disadvantage of
making this AA query nonlazy.
Preserve the behavior of assuming pointsToConstantMemory implying
dereferenceable for now, but maybe this should be changed.
r175673 changed repairIntervalsInRange to find anchoring end points for
ranges automatically, but the calculation of Begin included the first
instruction found that already had an index. This patch changes it to
exclude that instruction:
1. For symmetry, so that the half open range [Begin,End) only includes
instructions that do not already have indexes.
2. As a possible performance improvement, since repairOldRegInRange
will scan fewer instructions.
3. Because repairOldRegInRange hits assertion failures in some cases
when it sees a def that already has a live interval.
(3) fixes about ten tests in the CodeGen lit test suite when
-early-live-intervals is forced on.
Differential Revision: https://reviews.llvm.org/D110182
The DAG Combiner unnecessarily restricts commutative CSE
to nodes with a single result value. This commit removes
that restriction.
Signed-off-by: Itay Bookstein <ibookstein@gmail.com>
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D129666
Undef tokens may appear in unreached code as result of RAUW of some optimization,
and it should not be considered as bad IR.
Patch by Dmitry Bakunevich!
Differential Revision: https://reviews.llvm.org/D128904
Reviewed By: mkazantsev
Added function to the ExpandVectorPredication pass to handle VP loads
and stores.
Reviewed By: simoll
Differential Revision: https://reviews.llvm.org/D109584
trunc (sign_ext_inreg X, iM) to iN --> sign_ext_inreg (trunc X to iN), iM
There are improvements on existing tests from this, and there are a pair
of large regressions in D127115 for Thumb2 caused by not folding this
pattern.
Differential Revision: https://reviews.llvm.org/D129890
D127595 added the ability to recurse up a (one-use) INSERT_VECTOR_ELT chain to create a BUILD_VECTOR before other combines manage to break the chain, something that is particularly bad in D127115.
The patch generalises this so it doesn't have to build the chain starting from the last element insertion, instead it can now start from any insertion and will recurse up the chain until it finds all elements or finds a UNDEF/BUILD_VECTOR/SCALAR_TO_VECTOR which represents that start of the chain.
Fixes several regressions in D127115
In https://reviews.llvm.org/D30114, support for mismatching address
spaces was introduced to CodeGenPrepare's optimizeMemoryInst, using
addrspacecast as it was argued that only no-op addrspacecasts would be
considered when constructing the address mode. However, by doing
inttoptr/ptrtoint, it's possible to get CGP to emit an addrspace
that's not actually no-op, introducing a miscompilation:
define void @kernel(i8* %julia_ptr) {
%intptr = ptrtoint i8* %julia_ptr to i64
%ptr = inttoptr i64 %intptr to i32 addrspace(3)*
br label %end
end:
store atomic i32 1, i32 addrspace(3)* %ptr unordered, align 4
ret void
}
Gets compiled to:
define void @kernel(i8* %julia_ptr) {
end:
%0 = addrspacecast i8* %julia_ptr to i32 addrspace(3)*
store atomic i32 1, i32 addrspace(3)* %0 unordered, align 4
ret void
}
In the case of NVPTX, this introduces a cvta.to.shared, whereas
leaving out the %end block and branch doesn't trigger this
optimization. This results in illegal memory accesses as seen in
https://github.com/JuliaGPU/CUDA.jl/issues/558
In this change, I introduced a check before doing the pointer cast
that verifies address spaces are the same. If not, it emits a
ptrtoint/inttoptr combination to get a no-op cast between address
spaces. I decided against disallowing ptrtoint/inttoptr with
non-default AS in matchOperationAddr, because now its still possible
to look through multiple sequences of them that ultimately do not
result in a address space mismatch (i.e. the second lit test).
As mentioned on D127115, this patch that attempts to recognise shuffle masks that could be simplified to a AND mask - we already have a similar transform that will fold AND -> 'clear mask' shuffle, but this patch handles cases where the referenced elements are not from the same lane indices but are known to be zero.
Differential Revision: https://reviews.llvm.org/D129150
combineShiftAnd1ToBitTest already matches "and (not (srl X, C)), 1 --> (and X, 1<<C) == 0" patterns, but we can end up with situations where the not is before the shift.
Part of some yak shaving for D127115 to generalise the "xor (X >> ShiftC), XorC --> (not X) >> ShiftC" fold.
SimplifyDemandedBits is called slightly later which allows the not(sext(x)) -> sext(not(x)) fold to occur via foldLogicOfShifts
As mentioned on D127115, we should be able to further generalise this based off the demanded bits.
This is similar to D125680, but for llvm.experimental.patchpoint
(instead of llvm.experimental.stackmap).
Differential review: https://reviews.llvm.org/D129268
Following some recent discussions, this changes the representation
of callbrs in IR. The current blockaddress arguments are replaced
with `!` label constraints that refer directly to callbr indirect
destinations:
; Before:
%res = callbr i8* asm "", "=r,r,i"(i8* %x, i8* blockaddress(@test8, %foo))
to label %asm.fallthrough [label %foo]
; After:
%res = callbr i8* asm "", "=r,r,!i"(i8* %x)
to label %asm.fallthrough [label %foo]
The benefit of this is that we can easily update the successors of
a callbr, without having to worry about also updating blockaddress
references. This should allow us to remove some limitations:
* Allow unrolling/peeling/rotation of callbr, or any other
clone-based optimizations
(https://github.com/llvm/llvm-project/issues/41834)
* Allow duplicate successors
(https://github.com/llvm/llvm-project/issues/45248)
This is just the IR representation change though, I will follow up
with patches to remove limtations in various transformation passes
that are no longer needed.
Differential Revision: https://reviews.llvm.org/D129288
If we have a variable shift amount and the demanded mask has leading
zeros, we can propagate those leading zeros to not demand those bits
from operand 0. This can allow zero_extend/sign_extend to become
any_extend. This pattern can occur due to C integer promotion rules.
This transform is already done by InstCombineSimplifyDemanded.cpp where
sign_extend can be turned into zero_extend for example.
Reviewed By: spatel, foad
Differential Revision: https://reviews.llvm.org/D121833
Widening a G_FCONSTANT by extending and then generating G_FPTRUNC doesn't produce
the same result all the time. Instead, we can just transform it to a G_CONSTANT
of the same bit pattern and truncate using a plain G_TRUNC instead.
Fixes https://github.com/llvm/llvm-project/issues/56454
Differential Revision: https://reviews.llvm.org/D129743
If an MI will not generate a target instruction, we should not compute its
latency. Then we can compute more precise instruction sequence cost, and get
better result.
Differential Revision: https://reviews.llvm.org/D129615
isSafeToExpand() for addrecs depends on whether the SCEVExpander
will be used in CanonicalMode. At least one caller currently gets
this wrong, resulting in PR50506.
Fix this by a) making the CanonicalMode argument on the freestanding
functions required and b) adding member functions on SCEVExpander
that automatically take the SCEVExpander mode into account. We can
use the latter variant nearly everywhere, and thus make sure that
there is no chance of CanonicalMode mismatch.
Fixes https://github.com/llvm/llvm-project/issues/50506.
Differential Revision: https://reviews.llvm.org/D129630
The SI machine scheduler inherits from ScheduleDAGMI.
This patch adds support for a few features that are implemented
in ScheduleDAGMI (or its base classes) that were missing so far
because their support is implemented in overridden functions.
* Support cl::opt -view-misched-dags
This option allows to open a graphical window of the scheduling DAG.
* Support cl::opt -misched-print-dags
This option allows to print the scheduling DAG in text form.
* After constructing the scheduling DAG, call postprocessDAG()
to apply any registered DAG mutations.
Note that currently there are no mutations defined in AMDGPUTargetMachine.cpp
in case SIScheduler is used.
Still add this to avoid surprises in the future in case mutations are added.
Differential Revision: https://reviews.llvm.org/D128808
We shouldn't use getOpcodeDef() if we need to guarantee the def has only one
user since under the hood it may look through copies and optimization hints,
which themselves may have multiple users.
When doing scalable vectorization, the loop vectorizer uses a urem in the computation of the vector trip count. The RHS of that urem is a (possibly shifted) call to @llvm.vscale.
vscale is effectively the number of "blocks" in the vector register. (That is, types such as <vscale x 8 x i8> and <vscale x 1 x i8> both fill one 64 bit block, and vscale is essentially how many of those blocks there are in a single vector register at runtime.)
We know from the RISCV V extension specification that VLEN must be a power of two between ELEN and 2^16. Since our block size is 64 bits, the must be a power of two numbers of blocks. (For everything other than VLEN<=32, but that's already broken.)
It is worth noting that AArch64 SVE specification explicitly allows non-power-of-two sizes for the vector registers and thus can't claim that vscale is a power of two by this logic.
Differential Revision: https://reviews.llvm.org/D129609
We have the same fold in InstCombine - though implemented via OrZero flag on isKnownToBePowerOfTwo. The reasoning here is that either a) the result of the lshr is a power-of-two, or b) we have a div-by-zero triggering UB which we can ignore.
Differential Revision: https://reviews.llvm.org/D129606
Some rework of getStackGuard() based on comments in
https://reviews.llvm.org/D129505.
- getStackGuard() now creates and returns the destination
register, simplifying calls
- the pointer type is passed to getStackGuard() to avoid
recomputation
- removed PtrMemTy in emitSPDescriptorParent(), because
this type is only used here when loading the value but
not when storing the value
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D129576
To convert CTLZ to popcount we do
x = x | (x >> 1);
x = x | (x >> 2);
...
x = x | (x >>16);
x = x | (x >>32); // for 64-bit input
return popcount(~x);
This smears the most significant set bit across all of the bits
below it then inverts the remaining 0s and does a population count.
To support non-power of 2 types, the last shift amount must be
more than half of the size of the type. For i15, the last shift
was previously a shift by 4, with this patch we add another shift
of 8.
Fixes PR56457.
Differential Revision: https://reviews.llvm.org/D129431
When lowering llvm::stackprotect intrinsic, the SDAG implementation
checks useLoadStackGuardNode() to either create a LOAD_STACK_GUARD or use
the first argument of the intrinsic. This check is not present in the
IRTranslator, which results in always generating a LOAD_STACK_GUARD even
if the target does not support it.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D129505
Noticed while investigating the SystemZ regressions in D77804, prefer handling the knownbits analysis/simplification in the bitop nodes directly before falling back to SimplifyMultipleUseDemandedBits
Verify the LiveStacks analysis after a pass that claims to preserve it,
even if there are no further passes (apart from the verifier itself)
that would use the analysis.
Differential Revision: https://reviews.llvm.org/D129200
visitInlineAsm() in SDAGBuilder was duplicating a lot of the code
in ParseConstraints(), in particular all the logic to determine the
operand value and constraint VT.
Rely on the data computed by ParseConstraints() instead, and update
its ConstraintVT implementation to match getCallOperandValEVT()
more precisely.
As far as I can tell what was happening in the original code is
that the getNode call receives the same operands as the original
node with different SDNodeFlags. The logic inside getNode detects
that the node already exists and intersects the flags into the
existing node and returns it. This results in Op and NewOp for the
TLO.CombineTo call always being the same node.
We may have already called CombineTo as part of the recursive handling.
A second call to CombineTo as we unwind the recursion overwrites
the previous CombineTo. I think this means any time we updated the
poison flags that was the only change that ends up getting made
and we relied on DAGCombiner to revisit and call SimplifyDemandedBits
again. The second time the poison flags wouldn't need to be dropped
and we would keep the CombineTo call from further down the recursion.
We can instead call setFlags to drop the poison flags and remove the
call to TLO.CombineTo. This way we keep the CombineTo from deeper in
the recursion which should be more efficient.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D129511
As suggested in the post-commit feedback for D128123,
we can ease the mask constraint to ignore the MSB
(and make the code easier to read by adjusting the check).
https://alive2.llvm.org/ce/z/bbvqWv
Pointed out in Issue #56432: the current reference models may not be
quite friendly to open source projects. Their purpose is only
illustrative - the expectation is that projects would train their own.
To avoid unintentionally pulling such a model, made the URL cmake
setting require explicit user setting.
Differential Revision: https://reviews.llvm.org/D129342
