The current pattern for extract bits in range is typically:
Mask.lshr(BitOffset).trunc(SubSizeInBits);
Which can be particularly slow for large APInts (MaskSizeInBits > 64) as they require the allocation of memory for the temporary variable.
This is another of the compile time issues identified in PR32037 (see also D30265).
This patch adds the APInt::extractBits() helper method which avoids the temporary memory allocation.
Differential Revision: https://reviews.llvm.org/D30336
llvm-svn: 296272
The current pattern for extract bits in range is typically:
Mask.lshr(BitOffset).trunc(SubSizeInBits);
Which can be particularly slow for large APInts (MaskSizeInBits > 64) as they require the allocation of memory for the temporary variable.
This is another of the compile time issues identified in PR32037 (see also D30265).
This patch adds the APInt::extractBits() helper method which avoids the temporary memory allocation.
Differential Revision: https://reviews.llvm.org/D30336
llvm-svn: 296147
The current pattern for extract bits in range is typically:
Mask.lshr(BitOffset).trunc(SubSizeInBits);
Which can be particularly slow for large APInts (MaskSizeInBits > 64) as they require the allocation of memory for the temporary variable.
This is another of the compile time issues identified in PR32037 (see also D30265).
This patch adds the APInt::extractBits() helper method which avoids the temporary memory allocation.
Differential Revision: https://reviews.llvm.org/D30336
llvm-svn: 296141
Noticed while profiling PR32037, the target shuffle ops were being stored in SmallVector<*,8> types but the combiner could store as many as 16 ops at maximum depth (2 per depth).
llvm-svn: 296130
The current pattern for setting bits in range is typically:
Mask |= APInt::getBitsSet(MaskSizeInBits, LoPos, HiPos);
Which can be particularly slow for large APInts (MaskSizeInBits > 64) as they require the allocation memory for the temporary variable.
This is one of the key compile time issues identified in PR32037.
This patch adds the APInt::setBits() helper method which avoids the temporary memory allocation completely, this first implementation uses setBit() internally instead but already significantly reduces the regression in PR32037 (~10% drop). Additional optimization may be possible.
I investigated whether there is need for APInt::clearBits() and APInt::flipBits() equivalents but haven't seen these patterns to be particularly common, but reusing the code would be trivial.
Differential Revision: https://reviews.llvm.org/D30265
llvm-svn: 296102
The Fuchsia ABI defines slots from the thread pointer where the
stack-guard value for stack-protector, and the unsafe stack pointer
for safe-stack, are stored. This parallels the Android ABI support.
Patch by Roland McGrath
Differential Revision: https://reviews.llvm.org/D30237
llvm-svn: 296081
Minor optimization, don't create temporary mask APInts that are just going to be OR'd into the accumulate masks - insert directly instead.
llvm-svn: 295848
This patch introduces new X86ISD::FMAXS and X86ISD::FMINS opcodes. The legacy intrinsics now lower to this node. As do the AVX-512 masked intrinsics when the rounding mode is CUR_DIRECTION.
I've merged a copy of the tablegen multiclass avx512_fp_scalar into avx512_fp_scalar_sae. avx512_fp_scalar still needs to support CUR_DIRECTION appearing as a rounding mode for X86ISD::FADD_ROUND and others.
Differential revision: https://reviews.llvm.org/D30186
llvm-svn: 295810
Summary:
Rework the code that was sinking/duplicating (icmp and, 0) sequences
into blocks where they were being used by conditional branches to form
more tbz instructions on AArch64. The new code is more general in that
it just looks for 'and's that have all icmp 0's as users, with a target
hook used to select which subset of 'and' instructions to consider.
This change also enables 'and' sinking for X86, where it is more widely
beneficial than on AArch64.
The 'and' sinking/duplicating code is moved into the optimizeInst phase
of CodeGenPrepare, where it can take advantage of the fact the
OptimizeCmpExpression has already sunk/duplicated any icmps into the
blocks where they are used. One minor complication from this change is
that optimizeLoadExt needed to be updated to always mark 'and's it has
determined should be in the same block as their feeding load in the
InsertedInsts set to avoid an infinite loop of hoisting and sinking the
same 'and'.
This change fixes a regression on X86 in the tsan runtime caused by
moving GVNHoist to a later place in the optimization pipeline (see
PR31382).
Reviewers: t.p.northover, qcolombet, MatzeB
Subscribers: aemerson, mcrosier, sebpop, llvm-commits
Differential Revision: https://reviews.llvm.org/D28813
llvm-svn: 295746
This matches what is already done during shuffle lowering and helps prevent the need for a zero-vector in cases where shuffles match both patterns.
llvm-svn: 295723
Pull out repeated code for extraction index operand and source vector value type.
Use isNullConstant helper to check for zero extraction index.
llvm-svn: 295670
Its more profitable to go through memory (1 cycles throughput)
than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) to implement EXTRACT_VECTOR_ELT with variable index.
IACA tool was used to get performace estimation (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer)
For example for var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8 test from vector-shuffle-variable-128.ll I get 26 cycles vs 79 cycles.
Removing the VINSERT node, we don't need it any more.
Differential Revision: https://reviews.llvm.org/D29690
llvm-svn: 295660
Replaces existing approach that could only search BUILD_VECTOR nodes.
Requires getTargetConstantBitsFromNode to discriminate cases with all/partial UNDEF bits in each element - this should also be useful when we get around to supporting getTargetShuffleMaskIndices with UNDEF elements.
llvm-svn: 295613
As discussed on D27692, this permits another domain to be used to combine a shuffle at high depths.
We currently set the required depth at 4 or more combined shuffles, this is probably too high for most targets but is a good starting point and already helps avoid a number of costly variable shuffles.
llvm-svn: 295608
Add the infrastructure to flag whether float and/or int domains are permitable.
A future patch will enable domain crossing based off shuffle depth and the value types of the source vectors.
llvm-svn: 295604
Minor performance speedup - if any call to getShuffleScalarElt fails to get a result, don't both calling for the remaining elements as EltsFromConsecutiveLoads will fail anyhow.
llvm-svn: 295235
Summary:
We don't seem to have great rules on what a valid VBROADCAST node looks like. And as a consequence we end up with a lot of patterns to try to catch everything. We have patterns with scalar inputs, 128-bit vector inputs, 256-bit vector inputs, and 512-bit vector inputs.
As you can see from the things improved here we are currently missing patterns for 128-bit loads being extended to 256-bit before the vbroadcast.
I'd like to propose that VBROADCAST should always take a 128-bit vector type as input. As a first step towards that this patch adds an EXTRACT_SUBVECTOR in front of VBROADCAST when the input is 256 or 512-bits. In the future I would like to add scalar_to_vector around all the scalar operations. And maybe we should consider adding a VBROADCAST+load node to avoid separating loads from the broadcasting operation when the load itself isn't foldable.
This requires an additional change in target shuffle combining to look for the extract subvector and look through it to find the original operand. I'm sure this change isn't perfect but was enough to fix a few test failures that were being caused.
Another interesting thing I noticed is that the changes in masked_gather_scatter.ll show cases were we don't remove a useless insert into element 1 before broadcasting element 0.
Reviewers: delena, RKSimon, zvi
Reviewed By: zvi
Subscribers: igorb, llvm-commits
Differential Revision: https://reviews.llvm.org/D28747
llvm-svn: 295155
We now detect that both the extract and insert indices are non-zero and convert to a shuffle. This will be lowered as a blend for 256-bit vectors or as a vshuf operations for 512-bit vectors.
llvm-svn: 294931
This results in the simplifications inside of getNode running while we're legalizing nodes popped off the worklist during the final DAG combine. This basically makes a DAG combine like operation occur during this legalize step, but we don't handle something quite the same way. I think we don't recursively added the removed nodes to the DAG combiner worklist.
llvm-svn: 294929
The target shuffle match function arguments were using the term 'Ops' but the function names referred to them as 'Inputs' - use 'Inputs' consistently.
llvm-svn: 294900
Initial 256-bit vector support - 512-bit support requires extra checks for AVX512BW support (PMOVZXBW) that will be handled in a future patch.
llvm-svn: 294896
Removes duplicate constant extraction code in getTargetShuffleMaskIndices.
getTargetConstantBitsFromNode - adds support for VZEXT_MOVL(SCALAR_TO_VECTOR) and fail if the caller doesn't support undef bits.
llvm-svn: 294856
Since r274013, we've been looking through bitcasts on broadcast inputs.
In the scalar-folding case (from a load, build_vector, or sc2vec),
the input type didn't matter, as we'd simply bitcast the resulting
scalar back.
However, when broadcasting a 128-bit-lane-aligned element, we create an
EXTRACT_SUBVECTOR. Use proper types, by creating an extract_subvector
of the original input type.
llvm-svn: 294774
In some cases we call getTargetConstantBitsFromNode for nodes that haven't been lowered from BUILD_VECTOR yet
Note: We're getting very close to being able to move most of the constant extraction code from getTargetShuffleMaskIndices into getTargetConstantBitsFromNode
llvm-svn: 294746
In combineOrCmpEqZeroToCtlzSrl, replace "getConstantOperand == 0" by "isNullConstant" to account for floating point constants.
Differential Revision: https://reviews.llvm.org/D29756
llvm-svn: 294588
LowerBuildVectorv16i8/LowerBuildVectorv8i16 insert values into a UNDEF vector if the build vector doesn't contain any zero elements, resulting in register dependencies with a previous use of the register.
This patch attempts to break the register dependency by either always zeroing the vector before hand or (if we're inserting to the 0'th element) by using VZEXT_MOVL(SCALAR_TO_VECTOR(i32 AEXT(Elt))) which lowers to (V)MOVD and performs a similar function. Additionally (V)MOVD is a shorter instruction than PINSRB/PINSRW. We already do something similar for SSE41 PINSRD.
On pre-SSE41 LowerBuildVectorv16i8 we go a little further and use VZEXT_MOVL(SCALAR_TO_VECTOR(i32 ZEXT(Elt))) if the build vector contains zeros to avoid the vector zeroing at the cost of a scalar zero extension, which can probably be brought over to the other cases in a future patch in some cases (load folding etc.)
Differential Revision: https://reviews.llvm.org/D29720
llvm-svn: 294581
This patch does the following.
1. Adds an Intrinsic int_x86_clzero which works with __builtin_ia32_clzero
2. Identifies clzero feature using cpuid info. (Function:8000_0008, Checks if EBX[0]=1)
3. Adds the clzero feature under znver1 architecture.
4. The custom inserter is added in Lowering.
5. A testcase is added to check the intrinsic.
6. The clzero instruction is added to assembler test.
Patch by Ganesh Gopalasubramanian with a couple formatting tweaks, a disassembler test, and using update_llc_test.py from me.
Differential revision: https://reviews.llvm.org/D29385
llvm-svn: 294558
vXi8/vXi64 vector shifts are often shifted as vYi16/vYi32 types but we weren't always remembering to bitcast the input.
Tested with a new assert as we don't currently manipulate these shifts enough for test cases to catch them.
llvm-svn: 294308
Currently we only combine shuffle nodes if they have a single user to prevent us from causing code bloat by splitting the shuffles into several different combines.
We don't take into account that in some cases we will already have combined all the users during recursively calling up the shuffle tree.
This patch keeps a list of all the shuffle nodes that have been combined so far and permits combining of further shuffle nodes if all its users are in that list.
Differential Revision: https://reviews.llvm.org/D29399
llvm-svn: 294183
Similar to what we already do for zero elt insertion, we can quickly rematerialize 'allbits' vectors so to avoid a unnecessary gpr value and insertion into a vector
llvm-svn: 294162
Similar was already done for several other shuffles in this function.
The test changes are because the old code used explicity zeroing for elements that could have been undef.
While I was here I also changed other shuffle vectors in the same function to use the same input twice instead of creating UNDEF nodes. getVectorShuffle can create the UNDEF for us.
llvm-svn: 294130
On one test this seems to have given more chance for DAG combine to do other INSERT_SUBVECTOR/EXTRACT_SUBVECTOR combines before the BLENDI was created. Looks like we can still improve more by teaching DAG combine to optimize INSERT_SUBVECTOR/EXTRACT_SUBVECTOR with BLENDI.
llvm-svn: 293944
This is a first attempt at using the MOVMSK instructions to replace all_of/any_of reduction patterns (i.e. an and/or + shuffle chain).
So far this only matches patterns where we are reducing an all/none bits source vector (i.e. a comparison result) but we should be able to expand on this in conjunction with improvements to 'bool vector' handling both in the x86 backend as well as the vectorizers etc.
Differential Revision: https://reviews.llvm.org/D28810
llvm-svn: 293880
This moves creation of SUBV_BROADCAST and merging of adjacent loads that are being inserted together.
This is a step towards removing legalizing of INSERT_SUBVECTOR except for vXi1 cases.
llvm-svn: 293872
Also add the ability to recognise PINSR(Vex, 0, Idx).
Targets shuffle combines won't replace multiple insertions with a bit mask until a depth of 3 or more, so we avoid codesize bloat.
The unnecessary vpblendw in clearupper8xi16a will be fixed in an upcoming patch.
llvm-svn: 293627
combineX86ShufflesRecursively can still only handle a maximum of 2 shuffle inputs but everything before it now supports any number of shuffle inputs.
This will be necessary for combining OR(SHUFFLE, SHUFFLE) patterns.
llvm-svn: 293560
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
Pulled out code that removed unused inputs from a target shuffle mask into a helper function to allow it to be reused in a future commit.
llvm-svn: 293175
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
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
r291670 doesn't crash on the original testcase from PR31589,
but it crashes on a slightly more complex one.
PR31589 has the new reproducer.
llvm-svn: 292444
This patch improves the mul instruction combine function (combineMul)
by adding new layer of logic.
In this patch, we are adding the ability to fold (mul x, -((1 << c) -1))
or (mul x, -((1 << c) +1)) into (neg(X << c) -x) or (neg((x << c) + x) respective.
Differential Revision: https://reviews.llvm.org/D28232
llvm-svn: 292358
Even with the fix from r291630, this still causes problems. I get
widespread assertion failures in the Swift runtime's WeakRefCount::increment()
function. I sent a reduced testcase in reply to the commit.
llvm-svn: 292242
Use v8i64 variable ASHR instructions if we don't have VLX.
This is a reduced version of D28537 that just adds support for variable shifts - I'll continue with that patch (for just constant/uniform shifts) once I've fixed the type legalization issue in avx512-cvt.ll.
Differential Revision: https://reviews.llvm.org/D28604
llvm-svn: 291901
Rename from addOperand to just add, to match the other method that has been
added to MachineInstrBuilder for adding more than just 1 operand.
See https://reviews.llvm.org/D28057 for the whole discussion.
Differential Revision: https://reviews.llvm.org/D28556
llvm-svn: 291891
Some shuffles can be lowered to blend mask instruction (VPBLENDMB/VPBLENDMW/VPBLENDMD/VPBLENDMQ) .
In this patch, I added new pattern match for this case.
Reviewers:
1. craig.topper
2. guyblank
3. RKSimon
4. igorb
Differential Revision: https://reviews.llvm.org/D28483
llvm-svn: 291888
Emit SHRQ/SHLQ instead of ANDQ with a 64 bit constant mask if the result
is unused and the mask has only higher/lower bits set. For example, with
this patch LLVM emits
shrq $41, %rdi
je
instead of
movabsq $0xFFFFFE0000000000, %rcx
testq %rcx, %rdi
je
This reduces number of instructions, code size and register pressure.
The transformation is applied only for cases where the mask cannot be
encoded as an immediate value within TESTQ instruction.
Differential Revision: https://reviews.llvm.org/D28198
llvm-svn: 291806
64-bit integer division in Intel CPUs is extremely slow, much slower
than 32-bit division. On the other hand, 8-bit and 16-bit divisions
aren't any faster. The only important exception is Atom where DIV8
is fastest. Because of that, the patch
1) Enables bypassing of 64-bit division for Atom, Silvermont and
all big cores.
2) Modifies 64-bit bypassing to use 32-bit division instead of
16-bit one. This doesn't make the shorter division slower but
increases chances of taking it. Moreover, it's much more likely
to prove at compile-time that a value fits 32 bits and doesn't
require a run-time check (e.g. zext i32 to i64).
Differential Revision: https://reviews.llvm.org/D28196
llvm-svn: 291800
r289653 added a case where `vselect <cond> <vector1> <all-zeros>`
is transformed to:
`vselect xor(cond, DAG.getConstant(1, DL, CondVT) <all-zeros> <vector1>`
This was not aimed to catch cases where Cond is not a vXi1
mask but it does. Moreover, when Cond type is VxiN (N > 1)
then xor(cond, DAG.getConstant(1, DL, CondVT) != NOT(cond).
This patch changes the above to xor with allones, and avoids
entering the case for non-mask Conds.
llvm-svn: 291745
DAG patterns optimization: truncate + unsigned saturation supported by VPMOVUS* instructions in AVX-512.
And VPACKUS* instructions on SEE* targets.
Differential Revision: https://reviews.llvm.org/D28216
llvm-svn: 291670
This was reverted because it would miscompile code where the cmp had
multiple uses. That was due to a deficiency in the existing code, which
was fixed in r291630 (see the PR for details).
This re-commit includes an extra test for the kind of code that got
miscompiled: @test_sub_1_setcc_jcc.
llvm-svn: 291640
We would miscompile the following:
void g(int);
int f(volatile long long *p) {
bool b = __atomic_fetch_add(p, 1, __ATOMIC_SEQ_CST) < 0;
g(b ? 12 : 34);
return b ? 56 : 78;
}
into
pushq %rax
lock incq (%rdi)
movl $12, %eax
movl $34, %edi
cmovlel %eax, %edi
callq g(int)
testq %rax, %rax <---- Bad.
movl $56, %ecx
movl $78, %eax
cmovsl %ecx, %eax
popq %rcx
retq
because the code failed to take into account that the cmp has multiple
uses, replaced one of them, and left the other one comparing garbage.
llvm-svn: 291630
This patch fix PR31351: https://llvm.org/bugs/show_bug.cgi?id=31351
1. This patch adds new type of shuffle lowering
2. We can use the expand instruction, When the shuffle pattern is as following:
{ 0*a[0]0*a[1]...0*a[n] , n >=0 where a[] elements in a ascending order}.
Reviewers: 1. igorb
2. guyblank
3. craig.topper
4. RKSimon
Differential Revision: https://reviews.llvm.org/D28352
llvm-svn: 291584
Use the existing AVX2 v8i16 vector shift lowering for v16i8 (extending to v16i32) on AVX512 targets and v32i8 (extending to v32i16) on AVX512BW targets.
Cost model updates to follow.
llvm-svn: 291451
Use the existing AVX2 v8i16 vector shift lowering for v16i16 on AVX512 targets (AVX512BW will have already have lowered with vpsravw).
Cost model updates to follow.
llvm-svn: 291445
I noticed this problem as part of the ongoing attempt to canonicalize min/max ops in IR.
The debug output shows nodes like this:
t4: i32 = xor t2, Constant:i32<-1>
t21: i8 = setcc t4, Constant:i32<0>, setlt:ch
t14: i32 = select t21, t4, Constant:i32<-1>
And because the select is holding onto the t4 (xor) node while EmitTest creates a new
x86-specific xor node, the lowering results in:
t4: i32 = xor t2, Constant:i32<-1>
t25: i32,i32 = X86ISD::XOR t2, Constant:i32<-1>
t28: i32,glue = X86ISD::CMOV Constant:i32<-1>, t4, Constant:i8<15>, t25:1
Differential Revision: https://reviews.llvm.org/D28374
llvm-svn: 291392
Craig Topper