PR #66334 tried to renumber slot indexes before register allocation, but
the numbering was still affected by list entries for instructions which
had been erased. Fix this to make the register allocator's live range
length heuristics even less dependent on the history of how instructions
have been added to and removed from SlotIndexes's maps.
This change adjusts the cost modeling used when the target does not have a schedule model with individual instruction latencies. After this change, we use the default latency information available from TargetSchedule. The default latency information essentially ends up treating most instructions as latency 1, with a few "expensive" ones getting a higher cost.
Previously, we unconditionally applied the first legal pattern - without any consideration of profitability. As a result, this change both prevents some patterns being applied, and changes which patterns are exercised. (i.e. previously the first pattern was applied, afterwards, maybe the second one is because the first wasn't profitable.)
The motivation here is two fold.
First, this brings the default behavior in line with the behavior when -mcpu or -mtune is specified. This improves test coverage, and generally makes it less likely we will have bad surprises when providing more information to the compiler.
Second, this enables some reassociation for ILP by default. Despite being unconditionally enabled, the prior code tended to "reassociate" repeatedly through an entire chain and simply moving the first operand to the end. The result was still a serial chain, just a different one. With this change, one of the intermediate transforms is unprofitable and we end up with a partially flattened tree.
Note that the resulting code diffs show significant room for improvement in the basic algorithm. I am intentionally excluding those from this patch.
For the test diffs, I don't seen any concerning regressions. I took a fairly close look at the RISCV ones, but only skimmed the x86 (particularly vector x86) changes.
Differential Revision: https://reviews.llvm.org/D141017
This stops reporting CostPerUse 1 for `R8`-`R15` and `XMM8`-`XMM31`.
This was previously done because instruction encoding require a REX
prefix when using them resulting in longer instruction encodings. I
found that this regresses the quality of the register allocation as the
costs impose an ordering on eviction candidates. I also feel that there
is a bit of an impedance mismatch as the actual costs occure when
encoding instructions using those registers, but the order of VReg
assignments is not primarily ordered by number of Defs+Uses.
I did extensive measurements with the llvm-test-suite wiht SPEC2006 +
SPEC2017 included, internal services showed similar patterns. Generally
there are a log of improvements but also a lot of regression. But on
average the allocation quality seems to improve at a small code size
regression.
Results for measuring static and dynamic instruction counts:
Dynamic Counts (scaled by execution frequency) / Optimization Remarks:
Spills+FoldedSpills -5.6%
Reloads+FoldedReloads -4.2%
Copies -0.1%
Static / LLVM Statistics:
regalloc.NumSpills mean -1.6%, geomean -2.8%
regalloc.NumReloads mean -1.7%, geomean -3.1%
size..text mean +0.4%, geomean +0.4%
Static / LLVM Statistics:
mean -2.2%, geomean -3.1%) regalloc.NumSpills
mean -2.6%, geomean -3.9%) regalloc.NumReloads
mean +0.6%, geomean +0.6%) size..text
Static / LLVM Statistics:
regalloc.NumSpills mean -3.0%
regalloc.NumReloads mean -3.3%
size..text mean +0.3%, geomean +0.3%
Differential Revision: https://reviews.llvm.org/D133902
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
Current implementation only recognizes absolute operation implemented by
select instruction. This patch adds support for abs intrinsic.
Differential Revision: https://reviews.llvm.org/D122777
The motivation is that the update script has at least two deviations
(`<...>@GOT`/`<...>@PLT`/ and not hiding pointer arithmetics) from
what pretty much all the checklines were generated with,
and most of the tests are still not updated, so each time one of the
non-up-to-date tests is updated to see the effect of the code change,
there is a lot of noise. Instead of having to deal with that each
time, let's just deal with everything at once.
This has been done via:
```
cd llvm-project/llvm/test/CodeGen/X86
grep -rl "; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py" | xargs -L1 <...>/llvm-project/llvm/utils/update_llc_test_checks.py --llc-binary <...>/llvm-project/build/bin/llc
```
Not all tests were regenerated, however.
TargetMachine::shouldAssumeDSOLocal currently implies dso_local for such definitions.
Adding explicit dso_local makes these tests align with the clang -fno-pic behavior
and allow the removal of the TargetMachine::shouldAssumeDSOLocal special case.
If the mask input to getV4X86ShuffleImm8 only refers to a single source element (+ undefs) then canonicalize to a full broadcast.
getV4X86ShuffleImm8 defaults to inline values for undefs, which can be useful for shuffle widening/narrowing but does leave SimplifyDemanded* calls thinking the shuffle depends on unnecessary elements.
I'm still investigating what we should do more generally to avoid these undemanded elements, but broadcast cases was a simpler win.
If we lower a v2i64 shuffle to PSHUFD, we currently clamp undef elements to 0, (elements 0,1 of the v4i32) which can result in the shuffle referencing more elements of the source vector than expected, affecting later shuffle combines and KnownBits/SimplifyDemanded calls.
By ensuring we widen the undef mask element we allow getV4X86ShuffleImm8 to use inline elements as the default, which are more likely to fold.
Previously we walked the users of any vector binop looking for
more binops with the same opcode or phis that eventually ended up
in a reduction. While this is simple it also means visiting the
same nodes many times since we'll do a forward walk for each
BinaryOperator in the chain. It was also far more general than what
we have tests for or expect to see.
This patch replaces the algorithm with a new method that starts at
extract elements looking for a horizontal reduction. Once we find
a reduction we walk through backwards through phis and adds to
collect leaves that we can consider for rewriting.
We only consider single use adds and phis. Except for a special
case if the Add is used by a phi that forms a loop back to the
Add. Including other single use Adds to support unrolled loops.
Ultimately, I want to narrow the Adds, Phis, and final reduction
based on the partial reduction we're doing. I still haven't
figured out exactly what that looks like yet. But restricting
the types of graphs we expect to handle seemed like a good first
step. As does having all the leaves and the reduction at once.
Differential Revision: https://reviews.llvm.org/D79971
This replaces the build_vector lowering code that was just added in
D80013
and matches the pattern later from the x86-specific "vzext_movl".
That seems to result in the same or better improvements and gets rid
of the 'TODO' items from that patch.
AFAICT, we always shrink wider constant vectors to 128-bit on these
patterns, so we still get the implicit zero-extension to ymm/zmm
without wasting space on larger vector constants. There's a trade-off
there because that means we miss potential load-folding.
Similarly, we could load scalar constants here with implicit
zero-extension even to 128-bit. That saves constant space, but it
means we forego load-folding, and so it increases register pressure.
This seems like a good middle-ground between those 2 options.
Differential Revision: https://reviews.llvm.org/D80131
This build vector lowering pattern came up in D79886.
I've tried to limit the improvement to cases where it looks
clearly better to load, but we could remove the 'TODO'
predicates already if we are willing to overlook some
corner cases.
Differential Revision: https://reviews.llvm.org/D80013
I think some copy/pasting was used to create loops of different
VFs. But the increment of the induction variable wasn't updated
to match the VF.
This has no effect on the pattern matching we're testing, it just
helps the test make sense to the reader.
InstCombine should ensure these don't exist.
I'm looking at making some changes to how we detect these
patterns and not having to worry about these phis will help.
These transforms rely on a vector reduction flag on the SDNode
set by SelectionDAGBuilder. This flag exists because SelectionDAG
can't see across basic blocks so SelectionDAGBuilder is looking
across and saving the info. X86 is the only target that uses this
flag currently. By removing the X86 code we can remove the flag
and the SelectionDAGBuilder code.
This pass adds a dedicated IR pass for X86 that looks across the
blocks and transforms the IR into a form that the X86 SelectionDAG
can finish.
An advantage of this new approach is that we can enhance it to
shrink the phi nodes and final reduction tree based on the zeroes
that we need to concatenate to bring the partially reduced
reduction back up to the original width.
Differential Revision: https://reviews.llvm.org/D76649
createPSADBW uses SplitsOpsAndApply so should be able to handle
any size.
Restrict the extract result type to i32 or i64 since that's what
we have coverage for today and probably matches what the
isSimple() check gave us before.
Differential Revision: https://reviews.llvm.org/D76560
SplitsOpsAndApply will take care of any needed splitting correctly.
All that we need to check is that the vector element count is a
power of 2.
Differential Revision: https://reviews.llvm.org/D76558
Summary:
The 2 source operands commutable instructions are encoded in the
VEX.VVVV field and the r/m field of the MODRM byte plus the VEX.B
field.
The VEX.B field is missing from the 2-byte VEX encoding. If the
VEX.VVVV source is 0-7 and the other register is 8-15 we can
swap them to avoid needing the VEX.B field. This works as long as
the VEX.W, VEX.mmmmm, and VEX.X fields are also not needed.
Fixes PR36706.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D68550
For v2i32 we only feed 2 i8 elements into the psadbw instructions
with 0s in the other 14 bytes. The resulting psadbw instruction
will produce zeros in bits [127:16] of the output. We need to take
the result and feed it to a v2i32 add where the first element
includes bits [15:0] of the sad result. The other element should
be zero.
Prior to this patch we were using a truncate to take 0 from
bits 95:64 of the psadbw. This results in a pshufd to move those
bits to 63:32. But since we also have zeroes in bits 63:32 of
the psadbw output, we should just take those bits.
The previous code probably worked better with promoting legalization,
but now we use widening legalization. I've preserved the old
behavior if -x86-experimental-vector-widening-legalization=false
until we get that option removed.
llvm-svn: 369733
The assert that caused this to be reverted should be fixed now.
Original commit message:
This patch changes our defualt legalization behavior for 16, 32, and
64 bit vectors with i8/i16/i32/i64 scalar types from promotion to
widening. For example, v8i8 will now be widened to v16i8 instead of
promoted to v8i16. This keeps the elements widths the same and pads
with undef elements. We believe this is a better legalization strategy.
But it carries some issues due to the fragmented vector ISA. For
example, i8 shifts and multiplies get widened and then later have
to be promoted/split into vXi16 vectors.
This has the potential to cause regressions so we wanted to get
it in early in the 10.0 cycle so we have plenty of time to
address them.
Next steps will be to merge tests that explicitly test the command
line option. And then we can remove the option and its associated
code.
llvm-svn: 368183
This reverts commit 3de33245d2c992c9e0af60372043540b60f3a810.
This commit broke the MSan buildbots. See
https://reviews.llvm.org/rL367901 for more information.
llvm-svn: 368107
This patch changes our defualt legalization behavior for 16, 32, and
64 bit vectors with i8/i16/i32/i64 scalar types from promotion to
widening. For example, v8i8 will now be widened to v16i8 instead of
promoted to v8i16. This keeps the elements widths the same and pads
with undef elements. We believe this is a better legalization strategy.
But it carries some issues due to the fragmented vector ISA. For
example, i8 shifts and multiplies get widened and then later have
to be promoted/split into vXi16 vectors.
This has the potential to cause regressions so we wanted to get
it in early in the 10.0 cycle so we have plenty of time to
address them.
Next steps will be to merge tests that explicitly test the command
line option. And then we can remove the option and its associated
code.
llvm-svn: 367901
This reverts r340478 and r340631 and replaces them with a simpler
method of just letting DAG combine revisit the nodes to handle
the other operand.
llvm-svn: 367195
We catch most of these patterns (on x86 at least) by matching
a concat vectors opcode early in combining, but the pattern may
emerge later using insert subvector instead.
The AVX1 diffs for add/sub overflow show another missed narrowing
pattern. That one may be falling though the cracks because of
combine ordering and multiple uses.
llvm-svn: 360585
I noticed that we were failing to narrow an x86 ymm math op in a case similar
to the 'madd' test diff. That is because a bitcast is sitting between the math
and the extract subvector and thwarting our pattern matching for narrowing:
t56: v8i32 = add t59, t58
t68: v4i64 = bitcast t56
t73: v2i64 = extract_subvector t68, Constant:i64<2>
t96: v4i32 = bitcast t73
There are a few wins and neutral diffs in the other tests.
Differential Revision: https://reviews.llvm.org/D61806
llvm-svn: 360541
Summary:
Extract the logic for doing reassociations
from DAGCombiner::reassociateOps into a helper
function DAGCombiner::reassociateOpsCommutative,
and use that helper to trigger reassociation
on the original operand order, or the commuted
operand order.
Codegen is not identical since the operand order will
be different when doing the reassociations for the
commuted case. That causes some unfortunate churn in
some test cases. Apart from that this should be NFC.
Reviewers: spatel, craig.topper, tstellar
Reviewed By: spatel
Subscribers: dmgreen, dschuff, jvesely, nhaehnle, javed.absar, sbc100, jgravelle-google, hiraditya, aheejin, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D61199
llvm-svn: 359476
This is the sibling fold for insert-of-insert that was added with D56604.
Now that we have x86 shuffle narrowing (D57156), this change shows improvements for
lots of AVX512 reduction code (not sure that we would ever expect extract-of-extract otherwise).
There's a small regression in some of the partial-permute tests (extracting followed by splat).
That is tracked by PR40500:
https://bugs.llvm.org/show_bug.cgi?id=40500
Differential Revision: https://reviews.llvm.org/D57336
llvm-svn: 352528
This isn't the final fix for our reduction/horizontal codegen, but it takes care
of a lot of the problems. After we narrow the shuffle, existing combines for
insert/extract and binops kick in, and we end up with cheaper 128-bit ops.
The avg and mul reduction tests show an existing shuffle lowering hole for
AVX2/AVX512. I think in its most minimal form this is:
https://bugs.llvm.org/show_bug.cgi?id=40434
...but we might need multiple fixes to get it right.
Differential Revision: https://reviews.llvm.org/D57156
llvm-svn: 352209
Summary:
Make machine PHIs optimization to work for single value register taken from
several different copies. This is the first step to fix PR38917. This change
allows to get rid of redundant PHIs (see opt_phis2.mir test) to make
the subsequent optimizations (like CSE) possible and simpler.
For instance, before this patch the code like this:
%b = COPY %z
...
%a = PHI %bb1, %a; %bb2, %b
could be optimized to:
%a = %b
but the code like this:
%c = COPY %z
...
%b = COPY %z
...
%a = PHI %bb1, %a; %bb2, %b; %bb3, %c
would remain unchanged.
With this patch the latter case will be optimized:
%a = %z```.
Committed on behalf of: Anton Afanasyev anton.a.afanasyev@gmail.com
Reviewers: RKSimon, MatzeB
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D54839
llvm-svn: 349271
This is another step in vector narrowing - a follow-up to D53784
(and hoping to eventually squash potential regressions seen in
D51553).
The x86 test diffs are wins, but the AArch64 diff is probably not.
That problem already exists independent of this patch (see PR39722), but it
went unnoticed in the previous patch because there were no regression tests
that showed the possibility.
The x86 diff in i64-mem-copy.ll is close. Given the frequency throttling
concerns with using wider vector ops, an extra extract to reduce vector
width is the right trade-off at this level of codegen.
Differential Revision: https://reviews.llvm.org/D54392
llvm-svn: 347356
This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more
opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs.
Apart from 2-3 strange cases, these are all wins.
I've structured this to be no-functional-change-intended for any target except for x86
because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those
targets have existing regression tests (4, 4, 10 files respectively) that would be
affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show
any regression test diffs. The trade-off is deciding if an extra vector load is better
than a single wide load + extract_subvector.
For x86, this is almost always better (on paper at least) because we often can fold
loads into subsequent ops and not increase the official instruction count. There's also
some unknown -- but potentially large -- benefit from using narrower vector ops if wide
ops are implemented with multiple uops and/or frequency throttling is avoided.
Differential Revision: https://reviews.llvm.org/D54073
llvm-svn: 346595
Narrowing vector binops came up in the demanded bits discussion in D52912.
I don't think we're going to be able to do this transform in IR as a canonicalization
because of the risk of creating unsupported widths for vector ops, but we already have
a DAG TLI hook to allow what I was hoping for: isExtractSubvectorCheap(). This is
currently enabled for x86, ARM, and AArch64 (although only x86 has existing regression
test diffs).
This is artificially limited to not look through bitcasts because there are so many
test diffs already, but that's marked with a TODO and is a small follow-up.
Differential Revision: https://reviews.llvm.org/D53784
llvm-svn: 345602
This is the planned follow-up to D52997. Here we are reducing horizontal vector math codegen
by default. AMD Jaguar (btver2) should have no difference with this patch because it has
fast-hops. (If we want to set that bit for other CPUs, let me know.)
The code changes are small, but there are many test diffs. For files that are specifically
testing for hops, I added RUNs to distinguish fast/slow, so we can see the consequences
side-by-side. For files that are primarily concerned with codegen other than hops, I just
updated the CHECK lines to reflect the new default codegen.
To recap the recent horizontal op story:
1. Before rL343727, we were producing hops for all subtargets for a variety of patterns.
Hops were likely not optimal for all targets though.
2. The IR improvement in r343727 exposed a hole in the backend hop pattern matching, so
we reduced hop codegen for all subtargets. That was bad for Jaguar (PR39195).
3. We restored the hop codegen for all targets with rL344141. Good for Jaguar, but
probably bad for other CPUs.
4. This patch allows us to distinguish when we want to produce hops, so everyone can be
happy. I'm not sure if we have the best predicate here, but the intent is to undo the
extra hop-iness that was enabled by r344141.
Differential Revision: https://reviews.llvm.org/D53095
llvm-svn: 344361
Previously we asumed a vector reduction add is part of a loop and one of the input is a phi. But the code in SelectionDAGBuilder that sets vector reduction flag handles more cases than that. It just requires that the use chain ends in a horizontal reduction. And there are no other uses. This means it can handle unrolled reduction loops.
If the initial value of the reduction was 0, an unrolled loop would begin with a vector reduction add that has two sad inputs. Previously we would only transform one side of the add, but for this case we need to transform both sides.
I've created a lambda to reuse some of the code for both sides. And fixed the variables names to remove reference to "phi".
Differential Revision: https://reviews.llvm.org/D50817
llvm-svn: 340478
When nodes are reassociated the vector-reduction flag gets lost.
The test case is here is what would happen if you had a sum of absolute differences loop that started with a non-zero but contant sum and that loop was unrolled. The vectorizer will generate a constant vector for the initial value. And DAGCombiner reassociate tries to move it down the addition tree erasing the vector-reduction flag. Interestingly this moves constants the opposite direction of the reassociate IR pass.
I've chosen to just punt on the reassociate, but I suppose we could maybe preserve the flag if both nodes have it set.
Differential Revision: https://reviews.llvm.org/D50827
llvm-svn: 339946
Re-enable commit r323991 now that r325931 has been committed to make
MachineOperand::isRenamable() check more conservative w.r.t. code
changes and opt-in on a per-target basis.
llvm-svn: 326208
This reverts commit r323991.
This commit breaks target that don't model all the register constraints
in TableGen. So far the workaround was to set the
hasExtraXXXRegAllocReq, but it proves that it doesn't cover all the
cases.
For instance, when mutating an instruction (like in the lowering of
COPYs) the isRenamable flag is not properly updated. The same problem
will happen when attaching machine operand from one instruction to
another.
Geoff Berry is working on a fix in https://reviews.llvm.org/D43042.
llvm-svn: 325421
Summary:
This change extends MachineCopyPropagation to do COPY source forwarding
and adds an additional run of the pass to the default pass pipeline just
after register allocation.
This version of this patch uses the newly added
MachineOperand::isRenamable bit to avoid forwarding registers is such a
way as to violate constraints that aren't captured in the
Machine IR (e.g. ABI or ISA constraints).
This change is a continuation of the work started in D30751.
Reviewers: qcolombet, javed.absar, MatzeB, jonpa, tstellar
Subscribers: tpr, mgorny, mcrosier, nhaehnle, nemanjai, jyknight, hfinkel, arsenm, inouehrs, eraman, sdardis, guyblank, fedor.sergeev, aheejin, dschuff, jfb, myatsina, llvm-commits
Differential Revision: https://reviews.llvm.org/D41835
llvm-svn: 323991
