This extends the uniform base transform used with scatter/gather to support one-use vector adds-of-splats with a non-zero base. This has the effect of essentially reassociating an add from vector to scalar domain.
The motivation is to improve the lowering of scatter/gather operations fed by complex geps.
Differential Revision: https://reviews.llvm.org/D134472
RISCV doesn't actually support a scaled form of indexed load and store. We previously handled this by forming the scaled SDNode, and then doing custom legalization during lowering. This patch instead adds a callback via TLI to prevent formation entirely.
This has two effects:
* First, the GEP gets expanded (and used). Instead of the shift being created with an SDLoc of the memory operation, it has the SDLoc of the GEP instruction. This avoids the scheduler perturbing IR order when there's no reason to.
* Second, we fix what appears to be a bug in index calculation with RV32. The rules for GEPs require index calculation be done in particular bitwidth, and it appears the custom legalization code got this wrong for the case where index type exceeds pointer width. (Or at least, I trust the generic GEP lowering to be correct a lot more.)
The DAGCombiner change to handle VPScatter/VPGather is technically separate, but is required to prevent a regression on those intrinsics.
Differential Revision: https://reviews.llvm.org/D134382
The transform to fold an add into the base of a scatter/gather was only checking to see if the LHS was a splat. Included test change indicates that splats are not canonicalized to LHS, and that we need to check both sides.
The accessibility level of a typedef or using declaration in a
struct or class was being lost when producing debug information.
Reviewed By: dblaikie
Differential Revision: https://reviews.llvm.org/D134339
This pass allows a user to dump a MIR function to a dot file
and view it as a graph. It is targeted to provide a similar
functionality as -dot-cfg pass on LLVM-IR. As of now the pass
also support below flags:
-dot-mcfg-only [optional][won't print instructions in the
graph just block name]
-mcfg-dot-filename-prefix [optional][prefix to add to output dot file]
-mcfg-func-name [optional] [specify function name or it's
substring, handy if mir file contains multiple functions and
you need to see graph of just one]
More flags and details can be introduced as per the requirements
in future. This pass is inspired from -dot-cfg IR pass and APIs
are written in almost identical format.
Patch by Yashwant Singh <Yashwant.Singh@amd.com> (yassingh)
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D133709
I don't know what was going on originally with these tests. It seems reasonable
to have the immediate be the same byte alignment unit as the IR, in which case
we need to take the log2 in order to set the right number of low bits.
This fixes a miscompile in chromium.
Differential Revision: https://reviews.llvm.org/D134380
This patch uses the API provided by LiveRangeEdit to handle rematerialization.
It will make future maintenance and improvement more easier.
No functional change.
Differential Revision: https://reviews.llvm.org/D133610
This patch changes a FADD / FMUL => FMA ISel pattern implemented
in D80801 so that it peeks through more than one FMA.
Reviewed By: foad
Differential Revision: https://reviews.llvm.org/D132837
I'm not sure why the SEXT_INREG was gated on a bitwidth check of the mask
vs element size.
This fixes a miscompile in chromium's skia library.
Differential Revision: https://reviews.llvm.org/D134236
The following changes are necessasy to get the generated tree
matcher to compile:
- In CodeExpansions::declare(), the assert() prevents connecting
two instructions. E.g. the match code
(match (MUL $t, $s1, $s2),
(SUB $d, $t, $s3)),
results in two declarations of $t, one for the def and one for
the use. Removing the assertion allows this construct.
If $t is later used, it is one of the operands, which should be
perfectly fine.
- The code emitted in GIMatchTreeVRegDefPartitioner::generatePartitionSelectorCode()
is not compilable:
- The value of NewInstrID should be emitted, not the name
- Both calls involving getOperand() end with one parenthesis too many
- Swaps generated condition for the partition code in the latter function
It also changes the rules i2p_to_p2i, fabs_fabs_fold, and fneg_fneg_fold
to use the tree matcher for a linear match. These rules are tested by:
CodeGen/AArch64/GlobalISel/combine-fabs.mir
CodeGen/AArch64/GlobalISel/combine-fneg.mir
CodeGen/AArch64/GlobalISel/combine-ptrtoint.mir
CodeGen/AMDGPU/GlobalISel/combine-add-nullptr.mir
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D133257
This patch adds in instruction based features to the regalloc advisor
gated behind a flag so a user can decide at runtime whether or not they
want to enable the feature. The features are only enabled when LLVM is
compiled in MLGO develpment mode (LLVM_HAVE_TF_API) is set to true.
To extract the instruction features, I'm taking a list of segments from
each LiveInterval and noting the start and end SlotIndices. This list is then
sorted based on the start SlotIndex and I iterate through each SlotIndex
to grab instructions, making sure to check for overlaps. This results in
a vector of opcodes and binary mapping matrix that maps live ranges to the
opcodes of the instructions within that LR.
Reviewed By: mtrofin
Differential Revision: https://reviews.llvm.org/D131930
Use salvageDebugInfo for instructions erased as trivially dead in
GlobalISel.
It would be helpful to implement support of G_PTR_ADD and G_FRAME_INDEX
in salvageDebugInfo in future in order to preserve more variable
location.
Reviewed by: arsenm
Differential Revision: https://reviews.llvm.org/D133986
This is a follow-up to D133777, which resolved a use-after-free case but
did not cover all possible memory bugs due to misplacement of loads.
In short, the overall problem was that sinked loads could be moved after
state-modifying instructions leading to memory bugs.
The solution is to restrict load sinking unless it is found to be sound.
i) Within a basic block (to-be-sinked load and select-user are in the same BB),
loads can be sinked only if there is no intervening state-modifying instruction.
This is a conservative approach to avoid resorting to alias analysis to detect
potential memory overlap.
ii) Across basic blocks, sinking of loads is avoided. This is because going over
multiple basic blocks looking for memory conflicts could be computationally
expensive and also unlikely to allow loads to sink. Further, experiments showed
that not sinking these loads has a slight positive performance effect.
Maybe for some of these loads, having some separation allows enough time
for the load to be executed in time for its user. This is not the case for
floating point operations that benefit more from sinking.
The solution in D133777 was essentially undone in this patch,
since the latter is a complete solution to the observed problem.
Overall, the performance impact of this patch is minimal.
Tested on two internal Google workloads with instrPGO.
Search application showed <0.05% perf difference,
while the database one showed a slight improvement,
but not statistically significant.
Reviewed By: davidxl
Differential Revision: https://reviews.llvm.org/D133999
This is a partial port of the code used by the SelectionDAGBuilder to
translate selects.
In particular, see matchSelectPattern in ValueTracking.cpp. This is a
GISel-equivalent of the portion which handles fminnum/fmaxnum/fminimum/fmaximum.
I tried to set it up so it'd be easy to add the non-FP cases. Those are simpler.
On the AArch64-end, it seems like the FP cases are more important for perf
right now, so I bit the bullet and went at the more complicated problem. :)
I elected to do this as a post-legalize combine rather than in the
IRTranslator because
Deciding which fmax/fmin to use can depend on legalization rules
Philosophically-speaking (TM), putting it in a combine just feels cleaner
Being able to enable/disable the combine is handy
Another option would be to use the ValueTracking code in the IRTranslator and
match what SelectionDAGBuilder::visitSelect does. I think that may be somewhat
annoying since we'd need to write lowerings back into the selects in the
legalizer. I'm not strongly opposed to the approach.
We'd also want to be careful with vector selects once that's implemented,
which explicitly check if a vector select is legal on the target. That'd
probably need a hook.
From what I can tell, doing this as a combine is probably a cleaner option
long-term.
Differential Revision: https://reviews.llvm.org/D116702
Currently, it can become extremely costly to compute MayIncreasePressure if the
size of CSUses turns out to be very large. In that case, it's no longer cost
effective to keep computing MayIncreasePressure. Therefore, to limit the amount
of time spent in isProfitableToCSE, we simply conservatively assume
MayIncreasePressure if the size of CSUses is too large. This can reduce overall
compile time by 30% for some benchmarks.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D134003
I want to default all VP operations to Expand. These 2 were blocking
because VE doesn't support them and the tests were expecting them
to fail a specific way. Using Expand caused them to fail differently.
Seemed better to emulate them using operations that are supported.
@simoll mentioned on Discord that VE has some expansion downstream. Not
sure if its done like this or in the VE target.
Reviewed By: frasercrmck, efocht
Differential Revision: https://reviews.llvm.org/D133514
used, and MUL_LOHI is available
Folding into a sra(mul) / srl(mul) into a mulh introduces an extra
multiplication to compute the high half of the multiplication,
while it is more profitable to compute the high and lower halfs with a
single mul_lohi.
Differential Revision: https://reviews.llvm.org/D133768
The IR stack protector pass should insert stack checks before the tail
calls not only the musttail calls. So that the attributes `ssqreq` and
`tail call`, which are emited by llvm-opt, could be both enabled by
llvm-llc.
Reviewed By: compnerd
Differential Revision: https://reviews.llvm.org/D133860
On AArch64, doing the vector truncate separately after the fptoui
conversion can be lowered more efficiently using tbl.4, building on
D133495.
https://alive2.llvm.org/ce/z/T538CC
Depends on D133495
Reviewed By: t.p.northover
Differential Revision: https://reviews.llvm.org/D133496
Similar to using tbl to lower vector ZExts, tbl4 can be used to lower
vector truncates.
The initial version support i32->i8 conversions.
Depends on D120571
Reviewed By: t.p.northover
Differential Revision: https://reviews.llvm.org/D133495
Callee save registers must be preserved, so -fzero-call-used-regs
should not be zeroing them. The previous implementation only did
not zero callee save registers that were saved&restored inside the
function, but we need preserve all of them.
Fixes https://github.com/llvm/llvm-project/issues/57692.
Differential Revision: https://reviews.llvm.org/D133946
The method of counting resource consumption is modified to be based on
"Cycles" value when DFA is not used.
The calculation of ResMII is modified to total "Cycles" and divide it
by the number of units for each resource. Previously, ResMII was
excessive because it was assumed that resources were consumed for
the cycles of "Latency" value.
The method of resource reservation is modified similarly. When a
value of "Cycles" is larger than 1, the resource is considered to be
consumed by 1 for cycles of its length from the scheduled cycle.
To realize this, ResourceManager maintains a resource table for all
slots. Previously, resource consumption was always 1 for 1 cycle
regardless of the value of "Cycles" or "Latency".
In addition, the number of micro operations per cycle is modified to
be constrained by "IssueWidth". To disable the constraint,
--pipeliner-force-issue-width=100 can be used.
For the case of using DFA, the scheduling results are unchanged.
Reviewed By: dpenry
Differential Revision: https://reviews.llvm.org/D133572
This patch contains changes necessary to carry physical condition register (SCC) dependencies through the SDNode scheduler. It adds the edge in the SDNodeScheduler dependency graph instead of inserting the SCC copy between each definition and use. This approach lets the scheduler place instructions in an optimal way placing the copy only when the dependency cannot be resolved.
Reviewed By: rampitec
Differential Revision: https://reviews.llvm.org/D133593
This patch extends CodeGenPrepare to lower zext v16i8 -> v16i32 in loops
using a wide shuffle creating a v64i8 vector, selecting groups of 3
zero elements and an element from the input.
This is profitable on AArch64 where such shuffles can be lowered to tbl
instructions, but only in loops, because it requires materializing 4
masks, which can be done in the loop preheader.
This is the only reason the transform is part of CGP. If there's a
better alternative I missed, please let me know. The same goes for the
shouldReplaceZExtWithShuffle hook which guards this. I am not sure if
this transform will be beneficial on other targets, but it seems like
there is no way other convenient way.
This improves the generated code for loops like the one below in
combination with D96522.
int foo(uint8_t *p, int N) {
unsigned long long sum = 0;
for (int i = 0; i < N ; i++, p++) {
unsigned int v = *p;
sum += (v < 127) ? v : 256 - v;
}
return sum;
}
https://clang.godbolt.org/z/Wco866MjY
Reviewed By: t.p.northover
Differential Revision: https://reviews.llvm.org/D120571
All in-tree targets pass pointer-sized ConstantSDNodes to the
method. This overload reduced amount of boilerplate code a bit. This
also makes getCALLSEQ_END consistent with getCALLSEQ_START, which
already takes uint64_ts.
The class priority is expected to be at most 5 bits before it starts
clobbering bits used for other fields. Also clamp the instruction
distance in case we have millions of instructions.
AMDGPU was accidentally overflowing into the global priority bit in
some cases. I think in principal we would have wanted this, but in the
cases I've looked at, it had the counter intuitive effect and
de-prioritized the large register tuple.
Avoid using weird bit hack PPC uses for global priority. The
AllocationPriority field is really 5 bits, and PPC was relying on
overflowing this to 6-bits to forcibly set the global priority
bit. Split this out as a separate flag to avoid having magic behavior
for values above 31.
MachineMemOperand::print can dereference a NULL pointer if TII
is not passed from the printMemOperand. This does not happen while
dumping the DAG/MIR from llc but crashes the debugger if a dump()
method is called from gdb.
Differential Revision: https://reviews.llvm.org/D133903
Get some load-store forwarding cases for big-endian where a larger store covers
a smaller load, and the offset would be 0 and handled on little-endian but on
big-endian the offset is adjusted to be non-zero. The idea is just to shift the
data to make it look like the offset 0 case.
Differential Revision: https://reviews.llvm.org/D130115
https://reviews.llvm.org/D133637 fixes the problem where we should hash raw content of
register mask instead of the pointer to it.
Fix the same issue in `llvm::hash_value()`.
Remove the added API `MachineOperand::getRegMaskSize()` to avoid potential confusion.
Add an assert to emphasize that we probably should hash a machine operand iff it has
associated machine function, but keep the fallback logic in the original change.
Reviewed By: MatzeB
Differential Revision: https://reviews.llvm.org/D133747
There's no real read of the register, so the copy introduced a new
live value. Make sure we introduce a replacement implicit_def instead
of just erasing the copy.
Found from llvm-reduce since it tries to set undef on everything.
When a select is converted to a branch and load instructions are sinked to the true/false blocks,
lifetime intrinsics (if present) could be made unsound if not moved.
This conservatively moves all lifetime intrinsics in a transformed BB to the end block to ensure
preserved lifetime semantics.
Reviewed By: davidxl
Differential Revision: https://reviews.llvm.org/D133777
This is the ultimate fallback code if UADDO isn't supported.
If the target uses 0/1 we used one compare, but if the target doesn't
use 0/1 we emitted two compares. Regardless of boolean constants we
should only need to check that the Result is less than one of the
original operands. So we only need one compare.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D133708
