Summary:
In revision rL278321, ExecutionDepsFix learned how to pick a better
register for undef register reads, e.g. for instructions such as
`vcvtsi2sdq`. While this revision improved performance on a good number
of our benchmarks, it unfortunately also caused significant regressions
(up to 3x) on others. This regression turned out to be caused by loops
such as:
PH -> A -> B (xmm<Undef> -> xmm<Def>) -> C -> D -> EXIT
^ |
+----------------------------------+
In the previous version of the clearance calculation, we would visit
the blocks in order, remembering for each whether there were any
incoming backedges from blocks that we hadn't processed yet and if
so queuing up the block to be re-processed. However, for loop structures
such as the above, this is clearly insufficient, since the block B
does not have any unknown backedges, so we do not see the false
dependency from the previous interation's Def of xmm registers in B.
To fix this, we need to consider all blocks that are part of the loop
and reprocess them one the correct clearance values are known. As
an optimization, we also want to avoid reprocessing any later blocks
that are not part of the loop.
In summary, the iteration order is as follows:
Before: PH A B C D A'
Corrected (Naive): PH A B C D A' B' C' D'
Corrected (w/ optimization): PH A B C A' B' C' D
To facilitate this optimization we introduce two new counters for each
basic block. The first counts how many of it's predecssors have
completed primary processing. The second counts how many of its
predecessors have completed all processing (we will call such a block
*done*. Now, the criteria to reprocess a block is as follows:
- All Predecessors have completed primary processing
- For x the number of predecessors that have completed primary
processing *at the time of primary processing of this block*,
the number of predecessors that are done has reached x.
The intuition behind this criterion is as follows:
We need to perform primary processing on all predecessors in order to
find out any direct defs in those predecessors. When predecessors are
done, we also know that we have information about indirect defs (e.g.
in block B though that were inherited through B->C->A->B). However,
we can't wait for all predecessors to be done, since that would
cause cyclic dependencies. However, it is guaranteed that all those
predecessors that are prior to us in reverse postorder will be done
before us. Since we iterate of the basic blocks in reverse postorder,
the number x above, is precisely the count of the number of predecessors
prior to us in reverse postorder.
Reviewers: myatsina
Differential Revision: https://reviews.llvm.org/D28759
llvm-svn: 293571
Previously this test case fired an assertion in getNode because we tried to create an insert_subvector with both input types the same size and the index pointing to half the vector width.
llvm-svn: 293446
Replaces an xor+movd/movq with an xorps which will be shorter in codesize, avoid an int-fpu transfer, allow modern cores to fast path the result during decode and helps other combines recognise an all-zero vector.
The only reason I can think of that we'd want to keep scalar_to_vector in this case is to help recognise the upper elts are undef but this doesn't seem to be a problem.
Differential Revision: https://reviews.llvm.org/D29097
llvm-svn: 293438
PACKUSWB converts Signed word to Unsigned byte, (the same about DW) and it can't be used for umin+truncate pattern.
AVX-512 VPMOVUS* instructions fit the pattern since they convert Unsigned to Unsigned.
See https://llvm.org/bugs/show_bug.cgi?id=31773
Differential Revision: https://reviews.llvm.org/D29196
llvm-svn: 293431
The matching code tries to canonicalize XOR to the left, but if there are two XORs and only one is a vnot, this canonicalization can prevent matching.
llvm-svn: 293402
In r292621, the recommit fixes a bug related with live interval update
after the partial redundent copy is moved.
This recommit solves an additional bug related to the lack of update of
subranges.
The original patch is to solve the performance problem described in
PR27827. Register coalescing sometimes cannot remove a copy because of
interference. But if we can find a reverse copy in one of the predecessor
block of the copy, the copy is partially redundent and we may remove the
copy partially by moving it to the predecessor block without the
reverse copy.
Differential Revision: https://reviews.llvm.org/D28585
Re-apply r292621
Revert "Revert rL292621. Caused some internal build bot failures in apple."
This reverts commit r292984.
Original patch: Wei Mi <wmi@google.com>
Subrange fix: Mostly Matthias Braun <matze@braunis.de>
llvm-svn: 293353
This commit introduces a set of experimental intrinsics intended to prevent
optimizations that make assumptions about the rounding mode and floating point
exception behavior. These intrinsics will later be extended to specify
flush-to-zero behavior. More work is also required to model instruction
dependencies in machine code and to generate these instructions from clang
(when required by pragmas and/or command line options that are not currently
supported).
Differential Revision: https://reviews.llvm.org/D27028
llvm-svn: 293226
* Simplify Consecutive Merge Store Candidate Search
Now that address aliasing is much less conservative, push through
simplified store merging search and chain alias analysis which only
checks for parallel stores through the chain subgraph. This is cleaner
as the separation of non-interfering loads/stores from the
store-merging logic.
When merging stores search up the chain through a single load, and
finds all possible stores by looking down from through a load and a
TokenFactor to all stores visited.
This improves the quality of the output SelectionDAG and the output
Codegen (save perhaps for some ARM cases where we correctly constructs
wider loads, but then promotes them to float operations which appear
but requires more expensive constant generation).
Some minor peephole optimizations to deal with improved SubDAG shapes (listed below)
Additional Minor Changes:
1. Finishes removing unused AliasLoad code
2. Unifies the chain aggregation in the merged stores across code
paths
3. Re-add the Store node to the worklist after calling
SimplifyDemandedBits.
4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is
arbitrary, but seems sufficient to not cause regressions in
tests.
5. Remove Chain dependencies of Memory operations on CopyfromReg
nodes as these are captured by data dependence
6. Forward loads-store values through tokenfactors containing
{CopyToReg,CopyFromReg} Values.
7. Peephole to convert buildvector of extract_vector_elt to
extract_subvector if possible (see
CodeGen/AArch64/store-merge.ll)
8. Store merging for the ARM target is restricted to 32-bit as
some in some contexts invalid 64-bit operations are being
generated. This can be removed once appropriate checks are
added.
This finishes the change Matt Arsenault started in r246307 and
jyknight's original patch.
Many tests required some changes as memory operations are now
reorderable, improving load-store forwarding. One test in
particular is worth noting:
CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store
forwarding converts a load-store pair into a parallel store and
a memory-realized bitcast of the same value. However, because we
lose the sharing of the explicit and implicit store values we
must create another local store. A similar transformation
happens before SelectionDAG as well.
Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle
llvm-svn: 293184
The previous patch (https://reviews.llvm.org/rL289538) got reverted because of a bug. Chandler also requested some changes to the algorithm.
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20161212/413479.html
This is an updated patch. The key difference is that collectBitProviders (renamed to calculateByteProvider) now collects the origin of one byte, not the whole value. It simplifies the implementation and allows to stop the traversal earlier if we know that the result won't be used.
From the original commit:
Match a pattern where a wide type scalar value is loaded by several narrow loads and combined by shifts and ors. Fold it into a single load or a load and a bswap if the targets supports it.
Assuming little endian target:
i8 *a = ...
i32 val = a[0] | (a[1] << 8) | (a[2] << 16) | (a[3] << 24)
=>
i32 val = *((i32)a)
i8 *a = ...
i32 val = (a[0] << 24) | (a[1] << 16) | (a[2] << 8) | a[3]
=>
i32 val = BSWAP(*((i32)a))
This optimization was discussed on llvm-dev some time ago in "Load combine pass" thread. We came to the conclusion that we want to do this transformation late in the pipeline because in presence of atomic loads load widening is irreversible transformation and it might hinder other optimizations.
Eventually we'd like to support folding patterns like this where the offset has a variable and a constant part:
i32 val = a[i] | (a[i + 1] << 8) | (a[i + 2] << 16) | (a[i + 3] << 24)
Matching the pattern above is easier at SelectionDAG level since address reassociation has already happened and the fact that the loads are adjacent is clear. Understanding that these loads are adjacent at IR level would have involved looking through geps/zexts/adds while looking at the addresses.
The general scheme is to match OR expressions by recursively calculating the origin of individual bytes which constitute the resulting OR value. If all the OR bytes come from memory verify that they are adjacent and match with little or big endian encoding of a wider value. If so and the load of the wider type (and bswap if needed) is allowed by the target generate a load and a bswap if needed.
Reviewed By: RKSimon, filcab, chandlerc
Differential Revision: https://reviews.llvm.org/D27861
llvm-svn: 293036
clang already emits this with -cl-no-signed-zeros, but codegen
doesn't do anything with it. Treat it like the other fast math
attributes, and change one place to use it.
llvm-svn: 293024
The comment talked about replacing vpmovzxwd+vpslld+vpsrad with vpmovsxwd - which isn't valid as we're sign extending a <8 x i1> bool vector not an all/nobits <8 x i16>
llvm-svn: 292948
Summary:
When conditional branches with complex conditions are split into
multiple branches in SelectionDAGBuilder::FindMergedConditions, also
handle inverted conditions. These may sometimes appear without having
been optimized by InstCombine when CodeGenPrepare decides to sink and
duplicate cmp instructions, causing them to have only one use. This
problem can be increased by e.g. GVNHoist hiding more cmps from
InstCombine by combining equivalent cmps from different blocks.
For example codegen X & !(Y | Z) as:
jmp_if_X TmpBB
jmp FBB
TmpBB:
jmp_if_notY Tmp2BB
jmp FBB
Tmp2BB:
jmp_if_notZ TBB
jmp FBB
Reviewers: bogner, MatzeB, qcolombet
Subscribers: llvm-commits, hiraditya, mcrosier, sebpop
Differential Revision: https://reviews.llvm.org/D28380
llvm-svn: 292944
Summary:
This teaches getNode to simplify extracting from Undef. This is similar to what is done for EXTRACT_VECTOR_ELT. It also adds support for extracting from CONCAT_VECTOR when we can reuse one of the inputs to the concat. These seem like simple non-target specific optimizations.
For X86 we currently handle undef in extractSubvector, but not all EXTRACT_SUBVECTOR creations go through there.
Ultimately, my motivation here is to simplify extractSubvector and remove custom lowering for EXTRACT_SUBVECTOR since we don't do anything but handle undef and BUILD_VECTOR optimizations, but those should be DAG combines.
Reviewers: RKSimon, delena
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D29000
llvm-svn: 292876
The recommit fixes a bug related with live interval update after the partial
redundent copy is moved.
The original patch is to solve the performance problem described in PR27827.
Register coalescing sometimes cannot remove a copy because of interference.
But if we can find a reverse copy in one of the predecessor block of the copy,
the copy is partially redundent and we may remove the copy partially by moving
it to the predecessor block without the reverse copy.
Differential Revision: https://reviews.llvm.org/D28585
llvm-svn: 292621
This patch improves the knownbits logic for unsigned integer min/max opcodes.
For UMIN we know that the result will have the maximum of the inputs' known leading zero bits in the result, similarly for UMAX the maximum of the inputs' leading one bits.
This is particularly useful for simplifying clamping patterns,. e.g. as SSE doesn't have a uitofp instruction we want to use sitofp instead where possible and for that we need to confirm that the top bit is not set.
Differential Revision: https://reviews.llvm.org/D28853
llvm-svn: 292528
As discussed on D28219 - it is profitable to combine trunc(binop (s/zext(x), s/zext(y)) to binop(trunc(s/zext(x)), trunc(s/zext(y))) assuming the trunc(ext()) will simplify further
llvm-svn: 292493
As discussed on D28219 - it is profitable to combine trunc(binop (s/zext(x), s/zext(y)) to binop(trunc(s/zext(x)), trunc(s/zext(y))) assuming the trunc(ext()) will simplify further
llvm-svn: 292487
Summary:
The SDNodeOrder is saved in the IROrder field in the SDNode, and this
field may affects scheduling. Thus, letting dbg.value/declare increase
the order numbers may in turn affect scheduling.
Because of this change we also need to update the code deciding when
dbg values should be output, in ScheduleDAGSDNodes.cpp/ProcessSDDbgValues.
Dbg values now have the same order as the SDNode they are connected to,
not the following orders.
Test cases provided by Florian Hahn.
Reviewers: bogner, aprantl, sunfish, atrick
Reviewed By: atrick
Subscribers: fhahn, probinson, andreadb, llvm-commits, MatzeB
Differential Revision: https://reviews.llvm.org/D25318
llvm-svn: 292485
If the subvector comes from a load, we convert to SUBV_BROADCAST and use a broadcast instruction. But if there is no load we keep the inserts. I think we should create the SUBV_BROADCAST even without the load and let isel use the fallback patterns that are used if the load can't be folded. This will use the SHUFF32X4 or similar instruction for the 128-bit into 512-bit case and a single insert for 128 into 256 or 256 into 512.
This should be fixed so subvector broadcast intrinsics can be replaced with native IR since some of those currently lower directly to SHUFF32X4.
llvm-svn: 292475
Summary:
Currently we expand and scalarize these operations, but I think we should be able to implement ADD/SUB with KXOR and MUL with KAND.
We already do this for scalar i1 operations so I just extended it to vectors of i1.
Reviewers: zvi, delena
Reviewed By: delena
Subscribers: guyblank, llvm-commits
Differential Revision: https://reviews.llvm.org/D28888
llvm-svn: 292474
Craig Topper