The pmaddwd inserts a truncate, if that truncate would end up
creating additional instructions instead of making a zext
narrower, then we shouldn't do it.
I've restricted this to only sse4.1 targets since on prior
targets the zext will be done in stages. So the truncate will
probably not create additional instructions. Might need some
more investigation of mul shrinking and the other pmaddwd
transform to be sure this is the right decision.
There might be a slight regression on AVX1 targets due to add
splitting. Hard to say for sure. Maybe we need to look into
using the vector reduction flag to use 2 narrow loads and a
blend instead of extracting and inserting.
llvm-svn: 367198
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
The current pattern would trigger for scheduling changes of the
post-load computation, since those are commutable with the inline asm.
Avoid this by explicitly check the order of load vs asm block.
llvm-svn: 367180
This allows us to peek through BITCASTs, attempt to simplify the source operand, and then bitcast back.
This reapplies rL367091 which was reverted at rL367118 - we were inconsistently peeking through the bitcasts to the source value.
Fixes PR42777
llvm-svn: 367174
If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit.
This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it.
This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches.
llvm-svn: 367171
Summary:
This is an alternate approach to D57970.
Currently funclets reuse the same stack slots that are used in the
parent function for saving callee-saved xmm registers. If the parent
function modifies a callee-saved xmm register before an excpetion is
thrown, the catch handler will overwrite the original saved value.
This patch allocates space in funclets stack for saving callee-saved xmm
registers and uses RSP instead RBP to access memory.
Reviewers: andrew.w.kaylor, LuoYuanke, annita.zhang, craig.topper,
RKSimon
Subscribers: rnk, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D63396
Signed-off-by: pengfei <pengfei.wang@intel.com>
llvm-svn: 367088
This allows every serializer format to implement metaSerializer() and
return the corresponding meta serializer.
Original llvm-svn: 366946
Reverted llvm-svn: 367004
This fixes the unit tests on Windows bots.
llvm-svn: 367078
As discussed in https://bugs.llvm.org/show_bug.cgi?id=42673
there is a TTI hook hasDivRemOp() that matters here.
While -div-rem-pairs will decompose 'rem' if that hook returns false,
nothing does the opposite transform.
We can't to this in InstCombine, because it does not currently
access TTI, and i'm not sure we should change that.
We may be able to teach DivRemPairs to do this, but this really is a
per-target perf optimization, and we seem to do the opposite transform
in backend if hasDivRemOp() returned false: https://godbolt.org/z/ttt4HZ
I think it makes sense to be consistent.
https://bugs.llvm.org/show_bug.cgi?id=42673
llvm-svn: 367034
Summary:
This was originally reported in D62818.
https://rise4fun.com/Alive/oPH
InstCombine does the opposite fold, in hope that `C l>>/<< Y` expression
will be hoisted out of a loop if `Y` is invariant and `X` is not.
But as it is seen from the diffs here, if it didn't get hoisted,
the produced assembly is almost universally worse.
Much like with my recent "hoist add/sub by/from const" patches,
we should get almost universal win if we hoist constant,
there is almost always an "and/test by imm" instruction,
but "shift of imm" not so much, so we may avoid having to
materialize the immediate, and thus need one less register.
And since we now shift not by constant, but by something else,
the live-range of that something else may reduce.
Special care needs to be applied not to disturb x86 `BT` / hexagon `tstbit`
instruction pattern. And to not get into endless combine loop.
Reviewers: RKSimon, efriedma, t.p.northover, craig.topper, spatel, arsenm
Reviewed By: spatel
Subscribers: hiraditya, MaskRay, wuzish, xbolva00, nikic, nemanjai, jvesely, wdng, nhaehnle, javed.absar, tpr, kristof.beyls, jsji, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D62871
llvm-svn: 366955
This patch adds support for recognizing cases where a larger vector type is being used to reduce just the elements in the lower subvector:
e.g. <8 x i32> reduction pattern in a <16 x i32> vector:
<4,5,6,7,u,u,u,u,u,u,u,u,u,u,u,u>
<2,3,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
<1,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
matchBinOpReduction returns the lower extracted subvector in such cases, assuming isExtractSubvectorCheap accepts the extraction.
I've only enabled it for X86 reduction sums so far. I intend to enable it for the bitop/minmax cases in future patches, and eventually I think its worth turning it on all the time. This is mainly just a case of ensuring calls to matchBinOpReduction don't make assumptions on the vector width based on the original vector extraction.
Fixes the x86 partial reduction sum cases in PR33758 and PR42023.
Differential Revision: https://reviews.llvm.org/D65047
llvm-svn: 366933
If we are already using the same chain for the old/new memory ops then just return.
Fixes PR42727 which had getLoad() reusing an existing node.
llvm-svn: 366922
This exposes better support to use a string table with a format through
an actual new remark::Format, called yaml-strtab.
This can now be used with -fsave-optimization-record=yaml-strtab.
llvm-svn: 366849
LegalizeDAG tries to legal the DAG by legalizing nodes before
their operands.
If we create a new node, we end up legalizing it after its operands.
This prevents some of the optimizations that can be done when the
operand is a build_vector since the build_vector will have been
legalized to something else.
Differential Revision: https://reviews.llvm.org/D65132
llvm-svn: 366835
If all the demanded elts are from one operand and are inline, then we can use the operand directly.
The changes are mainly from SSE41 targets which has blendvpd but not cmpgtq, allowing the v2i64 comparison to be simplified as we only need the signbit from alternate v4i32 elements.
llvm-svn: 366817
This patch introduces the DAG version of SimplifyMultipleUseDemandedBits, which attempts to peek through ops (mainly and/or/xor so far) that don't contribute to the demandedbits/elts of a node - which means we can do this even in cases where we have multiple uses of an op, which normally requires us to demanded all bits/elts. The intention is to remove a similar instruction - SelectionDAG::GetDemandedBits - once SimplifyMultipleUseDemandedBits has matured.
The InstCombine version of SimplifyMultipleUseDemandedBits can constant fold which I haven't added here yet, and so far I've only wired this up to some basic binops (and/or/xor/add/sub/mul) to demonstrate its use.
We do see a couple of regressions that need to be addressed:
AMDGPU unsigned dot product codegen retains an AND mask (for ZERO_EXTEND) that it previously removed (but otherwise the dotproduct codegen is a lot better).
X86/AVX2 has poor handling of vector ANY_EXTEND/ANY_EXTEND_VECTOR_INREG - it prematurely gets converted to ZERO_EXTEND_VECTOR_INREG.
The code owners have confirmed its ok for these cases to fixed up in future patches.
Differential Revision: https://reviews.llvm.org/D63281
llvm-svn: 366799
We were silently using the ABI alignment for all of the stores generated for deopt and gc values. We'd gotten the alignment of the stack slot itself properly reduced (via MachineFrameInfo's clamping), but having the MMO on the store incorrect was enough for us to generate an aligned store to a unaligned location.
The simplest fix would have been to just pass the alignment to the helper function, but once we do that, the helper function doesn't really help. So, inline it and directly call the MMO version of DAG.getStore with a properly constructed MMO.
Note that there's a separate performance possibility here. Even if we *can* realign stacks, we probably don't *want to* if all of the stores are in slowpaths. But that's a later patch, if at all. :)
llvm-svn: 366765
The build_vector will become a constant pool load. By using the
desired type initially, it ensures we don't generate a bitcast
of the constant pool load which will need to be folded with
the load.
While experimenting with another patch, I noticed that when the
load type and the constant pool type don't match, then
SimplifyDemandedBits can't handle it. While we should probably
fix that, this was a simple way to fix the issue I saw.
llvm-svn: 366732
This patch enables us to find the source loads for each element, splitting them into a Load and ByteOffset, and attempts to recognise consecutive loads that are in fact from the same source load.
A helper function, findEltLoadSrc, recurses to find a LoadSDNode and determines the element's byte offset within it. When attempting to match consecutive loads, byte offsetted loads then attempt to matched against a previous load that has already been confirmed to be a consecutive match.
Next step towards PR16739 - after this we just need to account for shuffling/repeated elements to create a vector load + shuffle.
Fixed out of bounds load assert identified in rL366501
Differential Revision: https://reviews.llvm.org/D64551
llvm-svn: 366681
Much like with `urem`, the same optimization (albeit with slightly
different algorithm) applies for the signed case, too.
I'm simply copying the test coverage from `urem` case for now,
i believe it should be (close to?) sufficient.
llvm-svn: 366640
Summary:
Four things here:
1. Generalize the fold to handle non-splat divisors. Reasonably trivial.
2. Unban power-of-two divisors. I don't see any reason why they should
be illegal.
* There is no ban in Hacker's Delight
* I think the ban came from the same bug that caused the miscompile
in the base patch - in `floor((2^W - 1) / D)` we were dividing by
`D0` instead of `D`, and we **were** ensuring that `D0` is not `1`,
which made sense.
3. Unban `1` divisors. I no longer believe Hacker's Delight actually says
that the fold is invalid for `D = 0`. Further considerations:
* We know that
* `(X u% 1) == 0` can be constant-folded to `1`,
* `(X u% 1) != 0` can be constant-folded to `0`,
* Also, we know that
* `X u<= -1` can be constant-folded to `1`,
* `X u> -1` can be constant-folded to `0`,
* https://godbolt.org/z/7jnZJXhttps://rise4fun.com/Alive/oF6p
* We know will end up with the following:
`(setule/setugt (rotr (mul N, P), K), Q)`
* Therefore, for given new DAG nodes and comparison predicates
(`ule`/`ugt`), we will still produce the correct answer if:
`Q` is a all-ones constant; and both `P` and `K` are *anything*
other than `undef`.
* The fold will indeed produce `Q = all-ones`.
4. Try to re-splat the `P` and `K` vectors - we don't care about
their values for the lanes where divisor was `1`.
Reviewers: RKSimon, hermord, craig.topper, spatel, xbolva00
Reviewed By: RKSimon
Subscribers: hiraditya, javed.absar, dexonsmith, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D63963
llvm-svn: 366637
As detailed on PR42674, we can reduce a vXi8 down until we have the final <8 x i8>, and then use PSADBW with zero, to sum those values. We then extract the bottom i8, discarding any overflow from the upper bits of the i16 result.
llvm-svn: 366636
Summary:
Current PRE hoists common computations into
CMBB = DT->findNearestCommonDominator(MBB, MBB1).
However, if CMBB is in a hot loop body, we might get performance
degradation.
Differential Revision: https://reviews.llvm.org/D64394
llvm-svn: 366570
Summary:
For split-stack, if the nested argument (i.e. R10) is not used, no need to save/restore it in the prologue.
Reviewers: thanm
Reviewed By: thanm
Subscribers: mstorsjo, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D64673
llvm-svn: 366569
Summary:
Inline asm doesn't use labels when compiled as an object file. Therefore, we
shouldn't create one for the (potential) callbr destination. Instead, use the
symbol for the MachineBasicBlock.
Reviewers: nickdesaulniers, craig.topper
Reviewed By: nickdesaulniers
Subscribers: xbolva00, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D64888
llvm-svn: 366523
I plan on adding memcpy optimizations in the GlobalISel pipeline, but we can't
do that unless we delay lowering to actual function calls. This patch changes
the translator to generate G_INTRINSIC_W_SIDE_EFFECTS for these functions, and
then have each target specify that using the new custom legalizer for intrinsics
hook that they want it expanded it a libcall.
Differential Revision: https://reviews.llvm.org/D64895
llvm-svn: 366516
This is a small extension of !associated, mostly useful for the implementation
convenience of instrumentation passes that RAUW globals with aliases, such
as LowerTypeTests.
Differential Revision: https://reviews.llvm.org/D64951
llvm-svn: 366502
This reverts r366441 (git commit 48104ef7c9c653bbb732b66d7254957389fea337)
This causes clang to fail to compile some file in Skia. Reduction soon.
llvm-svn: 366501
This causes sections with relative pointers to be marked as read only,
which means that they won't end up sharing pages with writable data.
Differential Revision: https://reviews.llvm.org/D64948
llvm-svn: 366494
This patch enables us to find the source loads for each element, splitting them into a Load and ByteOffset, and attempts to recognise consecutive loads that are in fact from the same source load.
A helper function, findEltLoadSrc, recurses to find a LoadSDNode and determines the element's byte offset within it. When attempting to match consecutive loads, byte offsetted loads then attempt to matched against a previous load that has already been confirmed to be a consecutive match.
Next step towards PR16739 - after this we just need to account for shuffling/repeated elements to create a vector load + shuffle.
Differential Revision: https://reviews.llvm.org/D64551
llvm-svn: 366441
LEA doesn't affect flags, so use it more liberally to replace an ADD when
we know that the ADD operands affect flags.
In the motivating example from PR40483:
https://bugs.llvm.org/show_bug.cgi?id=40483
...this lets us avoid duplicating a math op just to avoid flag conflict.
As mentioned in the TODO comments, this heuristic can be extended to
fire more often if that leads to more improvements.
Differential Revision: https://reviews.llvm.org/D64707
llvm-svn: 366431
I'm not convinced the code this calls is properly vetted for
vXi1 vectors. Experimental vector widening legalization testing
for D55251 is now hitting an assertion failure inside
EltsFromConsecutiveLoads. This is occurring from a v2i1 load
having a store size different than its VT size. Hopefully
this commit will keep such issues from happening.
llvm-svn: 366405
This is part of what is requested by PR42023:
https://bugs.llvm.org/show_bug.cgi?id=42023
There's an extension needed for FP add, but exactly how we would specify
that using flags is not clear to me, so I left that as a TODO.
We're still missing patterns for partial reductions when the input vector
is 256-bit or 512-bit, but I think that's a failure of vector narrowing.
If we can reduce the widths, then this matching should work on those tests.
Differential Revision: https://reviews.llvm.org/D64760
llvm-svn: 366268
Summary:
As per title. DAGCombiner only mathes the special case where b = 0, this patches extends the pattern to match any value of b.
Depends on D57302
Reviewers: hfinkel, RKSimon, craig.topper
Subscribers: llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D59208
llvm-svn: 366214
We mostly avoid sub with immediate but there are a couple cases that can create them. One is the add 128, %rax -> sub -128, %rax trick in isel. The other is when a SUB immediate gets created for a compare where both the flags and the subtract value is used. If we are unable to linearize the SelectionDAG to satisfy the flag user and the sub result user from the same instruction, we will clone the sub immediate for the two uses. The one that produces flags will eventually become a compare. The other will have its flag output dead, and could then be considered for LEA creation.
I added additional test cases to add.ll to show the the sub -128 trick gets converted to LEA and a case where we don't need to convert it.
This showed up in the current codegen for PR42571.
Differential Revision: https://reviews.llvm.org/D64574
llvm-svn: 366151
inttofp (trunc (extelt X, 0)) --> inttofp (extelt (bitcast X), 0)
We have pseudo-vectorization of scalar int to FP casts, so this tries to
make that more likely by replacing a truncate with a bitcast. I didn't see
any test diffs starting from 'uitofp', so I left that as a TODO. We can't
only match the shorter trunc+extract pattern because there's an opposing
transform somewhere, so we infinite loop. Waiting to try this during
lowering is another possibility.
A motivating case is shown in PR39975 and included in the test diffs here:
https://bugs.llvm.org/show_bug.cgi?id=39975
Differential Revision: https://reviews.llvm.org/D64710
llvm-svn: 366098
