We used to accept (and even test, and generate) 16-byte alignment
for 32-byte nontemporal stores, but they require 32-byte alignment,
per SDM. Found by inspection.
Instead of hardcoding 16 in the patfrag, check for natural alignment.
Also fix the autoupgrade and the various tests.
Also, use explicit -mattr instead of -mcpu: I stared at the output
several minutes wondering why I get 2x movntps for the unaligned
case (which is the ideal output, but needs some work: see FIXME),
until I remembered corei7-avx implies +slow-unaligned-mem-32.
llvm-svn: 246733
We can chain other fragments to avoid repeating conditions.
This also fixes a potential bug (that realistically can't happen),
where we would match indexed nontemporal stores for i32/i64.
llvm-svn: 246719
This commit broke the build. Numerous build bots broken, and it was
blocking my progress so reverting.
It should be trivial to reproduce -- enable the BPF backend and it
should fail when running llvm-tblgen.
llvm-svn: 242992
In this patch I have only encoding. Intrinsics and DAG lowering will be in the next patch.
I temporary removed the old intrinsics test (just to split this patch).
Half types are not covered here.
Differential Revision: http://reviews.llvm.org/D11134
llvm-svn: 242023
This patch adds support for v8i16 and v16i8 shuffle lowering using the immediate versions of the SSE4A EXTRQ and INSERTQ instructions. Although rather limited (they can only act on the lower 64-bits of the source vectors, leave the upper 64-bits of the result vector undefined and don't have VEX encoded variants), the instructions are still useful for the zero extension of any lane (EXTRQ) or inserting a lane into another vector (INSERTQ). Testing demonstrated that it wasn't typically worth it to use these instructions for v2i64 or v4i32 vector shuffles although they are capable of it.
As well as adding specific pattern matching for the shuffles, the patch uses EXTRQ for zero extension cases where SSE41 isn't available and its more efficient than the SSE2 'unpack' default approach. It also adds shuffle decode support for the EXTRQ / INSERTQ cases when the instructions are handling full byte-sized extractions / insertions.
From this foundation, future patches will be able to make use of the instructions for situations that use their ability to extract/insert at the bit level.
Differential Revision: http://reviews.llvm.org/D10146
llvm-svn: 241508
With the completion of D9746 there is now a common implementation of integer signed/unsigned min/max nodes, removing the need for the equivalent X86 specific implementations.
This patch removes the old X86ISD nodes, legalizes the relevant SSE2/SSE41/AVX2/AVX512 instructions for the ISD versions and converts the small amount of existing X86 code.
Differential Revision: http://reviews.llvm.org/D10947
llvm-svn: 241506
This patch enables support for the conversion of v2i32 to v2f64 to use the CVTDQ2PD xmm instruction and stay on the SSE unit instead of scalarizing, sign extending to i64 and using CVTSI2SDQ scalar conversions.
Differential Revision: http://reviews.llvm.org/D10433
llvm-svn: 239855
AVX-512: Implemented GETEXP instruction for KNL and SKX
Added rounding mode modifier for SQRTPS/PD
Added tests for encoding and intrinsics.
CR:
http://reviews.llvm.org/D9991
llvm-svn: 238923
This patch removes the old X86ISD::FSRL op - which allowed float vectors to use the byte right shift operations (causing a domain switch....).
Since the refactoring of the shuffle lowering code this no longer has any use.
Differential Revision: http://reviews.llvm.org/D10169
llvm-svn: 238906
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
Added intrinsics for the instructions. CC parameter of the intrinsics was changed from i8 to i32 according to the spec.
By Igor Breger (igor.breger@intel.com)
llvm-svn: 236714
[DebugInfo] Add debug locations to constant SD nodes
This adds debug location to constant nodes of Selection DAG and updates
all places that create constants to pass debug locations
(see PR13269).
Can't guarantee that all locations are correct, but in a lot of cases choice
is obvious, so most of them should be. At least all tests pass.
Tests for these changes do not cover everything, instead just check it for
SDNodes, ARM and AArch64 where it's easy to get incorrect locations on
constants.
This is not complete fix as FastISel contains workaround for wrong debug
locations, which drops locations from instructions on processing constants,
but there isn't currently a way to use debug locations from constants there
as llvm::Constant doesn't cache it (yet). Although this is a bit different
issue, not directly related to these changes.
Differential Revision: http://reviews.llvm.org/D9084
llvm-svn: 235989
This adds debug location to constant nodes of Selection DAG and updates
all places that create constants to pass debug locations
(see PR13269).
Can't guarantee that all locations are correct, but in a lot of cases choice
is obvious, so most of them should be. At least all tests pass.
Tests for these changes do not cover everything, instead just check it for
SDNodes, ARM and AArch64 where it's easy to get incorrect locations on
constants.
This is not complete fix as FastISel contains workaround for wrong debug
locations, which drops locations from instructions on processing constants,
but there isn't currently a way to use debug locations from constants there
as llvm::Constant doesn't cache it (yet). Although this is a bit different
issue, not directly related to these changes.
Differential Revision: http://reviews.llvm.org/D9084
llvm-svn: 235977
Gather and scatter instructions additionally write to one of the source operands - mask register.
In this case Gather has 2 destination values - the loaded value and the mask.
Till now we did not support code gen pattern for gather - the instruction was generated from
intrinsic only and machine node was hardcoded.
When we introduce the masked_gather node, we need to select instruction automatically,
in the standard way.
I added a flag "hasTwoExplicitDefs" that allows to handle 2 destination operands.
(Some code in the X86InstrFragmentsSIMD.td is commented out, just to split one big
patch in many small patches)
llvm-svn: 230471
Change the memory operands in sse12_fp_packed_scalar_logical_alias from scalars to vectors.
That's what the hardware packed logical FP instructions define: 128-bit memory operands.
There are no scalar versions of these instructions...because this is x86.
Generating the wrong code (folding a scalar load into a 128-bit load) is still possible
using the peephole optimization pass and the load folding tables. We won't completely
solve this bug until we either fix the lowering in fabs/fneg/fcopysign and any other
places where scalar FP logic is created or fix the load folding in foldMemoryOperandImpl()
to make sure it isn't changing the size of the load.
Differential Revision: http://reviews.llvm.org/D7474
llvm-svn: 229531
Implement a BITCAST dag combine to transform i32->mmx conversion patterns
into a X86 specific node (MMX_MOVW2D) and guarantee that moves between
i32 and x86mmx are better handled, i.e., don't use store-load to do the
conversion..
llvm-svn: 228293
r224330 introduced a bug by misinterpreting the "FeatureVectorUAMem" bit.
The commit log says that change did not affect anything, but that's not correct.
That change allowed SSE instructions to have unaligned mem operands folded into
math ops, and that's not allowed in the default specification for any SSE variant.
The bug is exposed when compiling for an AVX-capable CPU that had this feature
flag but without enabling AVX codegen. Another mistake in r224330 was not adding
the feature flag to all AVX CPUs; the AMD chips were excluded.
This is part of the fix for PR22371 ( http://llvm.org/bugs/show_bug.cgi?id=22371 ).
This feature bit is SSE-specific, so I've renamed it to "FeatureSSEUnalignedMem".
Changed the existing test case for the feature bit to reflect the new name and
renamed the test file itself to better reflect the feature.
Added runs to fold-vex.ll to check for the failing codegen.
Note that the feature bit is not set by default on any CPU because it may require a
configuration register setting to enable the enhanced unaligned behavior.
llvm-svn: 227983
By Asaf Badouh and Elena Demikhovsky
Added special nodes for rounding: FMADD_RND, FMSUB_RND..
It will prevent merge between nodes with rounding and other standard nodes.
llvm-svn: 227303
According to AVX specification:
"Most arithmetic and data processing instructions encoded using the VEX prefix and
performing memory accesses have more flexible memory alignment requirements
than instructions that are encoded without the VEX prefix. Specifically,
With the exception of explicitly aligned 16 or 32 byte SIMD load/store instructions,
most VEX-encoded, arithmetic and data processing instructions operate in
a flexible environment regarding memory address alignment, i.e. VEX-encoded
instruction with 32-byte or 16-byte load semantics will support unaligned load
operation by default. Memory arguments for most instructions with VEX prefix
operate normally without causing #GP(0) on any byte-granularity alignment
(unlike Legacy SSE instructions)."
The same for AVX-512.
This change does not affect anything right now, because only the "memop pattern fragment"
depends on FeatureVectorUAMem and it is not used in AVX patterns.
All AVX patterns are based on the "unaligned load" anyway.
llvm-svn: 224330
including SAE mode and memory operand.
Added AVX512_maskable_scalar template, that should cover all scalar instructions in the future.
The main difference between AVX512_maskable_scalar<> and AVX512_maskable<> is using X86select instead of vselect.
I need it, because I can't create vselect node for MVT::i1 mask for scalar instruction.
http://reviews.llvm.org/D6378
llvm-svn: 222820
trick that I missed.
VPERMILPS has a non-immediate memory operand mode that allows it to do
asymetric shuffles in the two 128-bit lanes. Use this rather than two
shuffles and a blend.
However, it turns out the variable shuffle path to VPERMILPS (and
VPERMILPD, although that one offers no functional differenc from the
immediate operand other than variability) wasn't even plumbed through
codegen. Do such plumbing so that we can reasonably emit
a variable-masked VPERMILP instruction. Also plumb basic comment parsing
and printing through so that the tests are reasonable.
There are still a few tests which don't show the shuffle pattern. These
are tests with undef lanes. I'll teach the shuffle decoding and printing
to handle undef mask entries in a follow-up. I've looked at the masks
and they seem reasonable.
llvm-svn: 218300
td pattern). Currently we only model the immediate operand variation of
VPERMILPS and VPERMILPD, we should make that clear in the pseudos used.
Will be adding support for the variable mask variant in my next commit.
llvm-svn: 218282
introducing a synthetic X86 ISD node representing this generic
operation.
The relevant patterns for mapping these nodes into the concrete
instructions are also added, and a gnarly bit of C++ code in the
target-specific DAG combiner is replaced with simple code emitting this
primitive.
The next step is to generically combine blends of adds and subs into
this node so that we can drop the reliance on an SSE4.1 ISD node
(BLENDI) when matching an SSE3 feature (ADDSUB).
llvm-svn: 217819
parsing (and latent bug in the instruction definitions).
This is effectively a revert of r136287 which tried to address
a specific and narrow case of immediate operands failing to be accepted
by x86 instructions with a pretty heavy hammer: it introduced a new kind
of operand that behaved differently. All of that is removed with this
commit, but the test cases are both preserved and enhanced.
The core problem that r136287 and this commit are trying to handle is
that gas accepts both of the following instructions:
insertps $192, %xmm0, %xmm1
insertps $-64, %xmm0, %xmm1
These will encode to the same byte sequence, with the immediate
occupying an 8-bit entry. The first form was fixed by r136287 but that
broke the prior handling of the second form! =[ Ironically, we would
still emit the second form in some cases and then be unable to
re-assemble the output.
The reason why the first instruction failed to be handled is because
prior to r136287 the operands ere marked 'i32i8imm' which forces them to
be sign-extenable. Clearly, that won't work for 192 in a single byte.
However, making thim zero-extended or "unsigned" doesn't really address
the core issue either because it breaks negative immediates. The correct
fix is to make these operands 'i8imm' reflecting that they can be either
signed or unsigned but must be 8-bit immediates. This patch backs out
r136287 and then changes those places as well as some others to use
'i8imm' rather than one of the extended variants.
Naturally, this broke something else. The custom DAG nodes had to be
updated to have a much more accurate type constraint of an i8 node, and
a bunch of Pat immediates needed to be specified as i8 values.
The fallout didn't end there though. We also then ceased to be able to
match the instruction-specific intrinsics to the instructions so
modified. Digging, this is because they too used i32 rather than i8 in
their signature. So I've also switched those intrinsics to i8 arguments
in line with the instructions.
In order to make the intrinsic adjustments of course, I also had to add
auto upgrading for the intrinsics.
I suspect that the intrinsic argument types may have led everything down
this rabbit hole. Pretty happy with the result.
llvm-svn: 217310
No functional change. This will be used by the new FMA intrinsic lowering
code.
We can probably add NO_EXC here as well, I am just not too familiar with this
part of AVX512 yet. We can add that later.
llvm-svn: 215662
They have different semantics (valign is interlane while palingr is intralane)
and palingr is still needed even in the AVX512 context. According to the
latest spec AVX512BW provides these.
llvm-svn: 214887
instructions available as synthetic SDNodes PACKSS and PACKUS that will
select to the correct instruction variants based on the return type.
This allows us to use these rather important instructions when lowering
vector shuffles.
Also moves the relevant instruction definitions to be split out from
the fully generic multiclasses to allow them to match these new SDNodes
in the same way that the UNPCK instructions do.
No functionality should actually be changed here.
llvm-svn: 211332
Summary:
The INSERTPS pattern fragment was called insrtps (mising 'e'), which
would make it harder to grep for the patterns related to this instruction.
Renaming it to use the proper instruction name.
Reviewers: nadav
CC: llvm-commits
Differential Revision: http://reviews.llvm.org/D3443
llvm-svn: 206779
I believe VZEXT_MOVL means "zero all vector elements except the first" (and
should have identical input & output types) whereas VZEXT means "zero extend
each element of a vector (discarding higher elements if necessary)".
For example:
(v4i32 (vzext (v16i8 ...)))
should zero extend the low 4 bytes of the incoming vector to 32-bits,
discarding higher bytes.
However, somewhere in the past, these two concepts had become confused, even
leading to a nonsensical VSEXT_MOVL.
This re-merges the nodes where appropriate (all VSEXT_MOVL -> VSEXT, VZEXT_MOVL
-> VZEXT when it's an actual extension).
rdar://problem/15981990
llvm-svn: 200918
Added scalar compare VCMPSS, VCMPSD.
Implemented LowerSELECT for scalar FP operations.
I replaced FSETCCss, FSETCCsd with one node type FSETCCs.
Node extract_vector_elt(v16i1/v8i1, idx) returns an element of type i1.
llvm-svn: 197384
All insertf*/extractf* functions replaced with insert/extract since we have insertf and inserti forms.
Added lowering for INSERT_VECTOR_ELT / EXTRACT_VECTOR_ELT for 512-bit vectors.
Added lowering for EXTRACT/INSERT subvector for 512-bit vectors.
Added a test.
llvm-svn: 187491
We match the pattern "x >= y ? x-y : 0" into "subus x, y" and two special cases
if y is a constant. DAGCombiner canonicalizes those so we first have to undo the
canonicalization for those cases. The pattern occurs in gzip when the loop
vectorizer is enabled. Part of PR14613.
llvm-svn: 170273
- Due to the current matching vector elements constraints in
ISD::FP_ROUND, rounding from v2f64 to v4f32 (after legalization from
v2f32) is scalarized. Add a customized v2f32 widening to convert it
into a target-specific X86ISD::VFPROUND to work around this
constraints.
llvm-svn: 165631
this allows for better code generation.
Added a new DAGCombine transformation to convert FMAX and FMIN to FMANC and
FMINC, which are commutative.
For example:
movaps %xmm0, %xmm1
movsd LC(%rip), %xmm0
minsd %xmm1, %xmm0
becomes:
minsd LC(%rip), %xmm0
llvm-svn: 162187
- FP_EXTEND only support extending from vectors with matching elements.
This results in the scalarization of extending to v2f64 from v2f32,
which will be legalized to v4f32 not matching with v2f64.
- add X86-specific VFPEXT supproting extending from v4f32 to v2f64.
- add BUILD_VECTOR lowering helper to recover back the original
extending from v4f32 to v2f64.
- test case is enhanced to include different vector width.
llvm-svn: 161894
Original message:
Modify the code that lowers shuffles to blends from using blendvXX to vblendXX.
blendV uses a register for the selection while Vblend uses an immediate.
On sandybridge they still have the same latency and execute on the same execution ports.
llvm-svn: 154483
blendv uses a register for the selection while vblend uses an immediate.
On sandybridge they still have the same latency and execute on the same execution ports.
llvm-svn: 154396
Original commit message from r147481:
DAGCombine for transforming 128->256 casts into a vmovaps, rather
then a vxorps + vinsertf128 pair if the original vector came from a load.
Fix:
Unaligned loads need to generate a vmovups.
rdar://10974078
llvm-svn: 152366
Igor Breger