Summary:
There are few oddities that occur due to v1i1, v8i1, v16i1 being legal without v2i1 and v4i1 being legal when we don't have VLX. Particularly during legalization of v2i32/v4i32/v2i64/v4i64 masked gather/scatter/load/store. We end up promoting the mask argument to these during type legalization and then have to widen the promoted type to v8iX/v16iX and truncate it to get the element size back down to v8i1/v16i1 to use a 512-bit operation. Since need to fill the upper bits of the mask we have to fill with 0s at the promoted type.
It would be better if we could just have the v2i1/v4i1 types as legal so they don't undergo any promotion. Then we can just widen with 0s directly in a k register. There are no real v4i1/v2i1 instructions anyway. Everything is done on a larger register anyway.
This also fixes an issue that we couldn't implement a masked vextractf32x4 from zmm to xmm properly.
We now have to support widening more compares to 512-bit to get a mask result out so new tablegen patterns got added.
I had to hack the legalizer for widening the operand of a setcc a bit so it didn't try create a setcc returning v4i32, extract from it, then try to promote it using a sign extend to v2i1. Now we create the setcc with v4i1 if the original setcc's result type is v2i1. Then extract that and don't sign extend it at all.
There's definitely room for improvement with some follow up patches.
Reviewers: RKSimon, zvi, guyblank
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D41560
llvm-svn: 321967
This seems to improve X86's ability to match this into an address computation. Otherwise the other operand gets assigned to the base register and the stack pointer + frame index ends up in the index register. But index registers can't encode ESP/RSP so we end up having to move it into another register to meet the constraint.
I could try to improve the address matcher in X86, but swapping the producer seemed easier. Several other places already have the operands in this order so this is at least consistent.
llvm-svn: 321370
Despite what the comment said there isn't better codegen for 512-bit vectors. The 128/256/512 bit implementation jus stores to memory and loads an element. There's no advantage to doing that with a larger size. In fact in many cases it causes a stack realignment and generates worse code.
llvm-svn: 321369
Summary:
Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1.
We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires.
For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that.
For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them.
Reviewers: RKSimon, delena, spatel, zvi
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D40942
llvm-svn: 320849
This doesn't match the semantics of the extract_vector_elt operation. Nothing downstream knows the bits were zeroed so they still get masked or sign extended after the extrat anyway.
llvm-svn: 320723
We previously only supported inserting to the LSB or MSB where it was easy to zero to perform an OR to insert.
This change effectively extracts the old value and the new value, xors them together and then xors that single bit with the correct location in the original vector. This will cancel out the old value in the first xor leaving the new value in the position.
The way I've implemented this uses 3 shifts and two xors and uses an additional register. We can avoid the additional register at the cost of another shift.
llvm-svn: 320120
There's no v2i1 or v4i1 kshift, and v8i1 is only supported with AVXDQ. Isel has fake patterns to extend these types to native shifts, but makes no guarantees about the value of any bits shifted in when shifting right.
This patch promotes the vector to a type that supports a native shift first and only allows inserting into the msb of a native sized shift.
I've constructed this in a way that doesn't do the promotion if we're going to fallback to using a xmm/ymm/zmm shuffle. I think I have a plan to remove the shuffle fall back entirely. In which case we this can be simplified, but I wanted to fix the correctness issue first.
llvm-svn: 320081
As part of the unification of the debug format and the MIR format, print
MBB references as '%bb.5'.
The MIR printer prints the IR name of a MBB only for block definitions.
* find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g'
* find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g'
* find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g'
* grep -nr 'BB#' and fix
Differential Revision: https://reviews.llvm.org/D40422
llvm-svn: 319665
As part of the unification of the debug format and the MIR format,
always print registers as lowercase.
* Only debug printing is affected. It now follows MIR.
Differential Revision: https://reviews.llvm.org/D40417
llvm-svn: 319187
The pass scans the function to find instruction chains that define
registers in the same domain (closures).
It then calculates the cost of converting the closure to another domain.
If found profitable, the instructions are converted to instructions in
the other domain and the register classes are changed accordingly.
This commit adds the pass infrastructure and a simple conversion from
the GPR domain to the Mask domain.
Differential Revision:
https://reviews.llvm.org/D37251
Change-Id: Ic2cf1d76598110401168326d411128ae2580a604
llvm-svn: 316288
Summary:
This suppresses the generation of .Lcfi labels in our textual assembler.
It was annoying that this generated cascading .Lcfi labels:
llc foo.ll -o - | llvm-mc | llvm-mc
After three trips through MCAsmStreamer, we'd have three labels in the
output when none are necessary. We should only bother creating the
labels and frame data when making a real object file.
This supercedes D38605, which moved the entire .seh_ implementation into
MCObjectStreamer.
This has the advantage that we do more checking when emitting textual
assembly, as a minor efficiency cost. Outputting textual assembly is not
performance critical, so this shouldn't matter.
Reviewers: majnemer, MatzeB
Subscribers: qcolombet, nemanjai, javed.absar, eraman, hiraditya, JDevlieghere, llvm-commits
Differential Revision: https://reviews.llvm.org/D38638
llvm-svn: 315259
Previously we were using one of the subvector indices twice. The included test case causes an assert without this change.
Thanks to Simon Pilgrim for catching this.
llvm-svn: 314429
NFC.
Updated 8 regression tests to use -mattr instead of -mcpu flag as follows:
-mcpu=knl --> -mattr=+avx512f
-mcpu=skx --> -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq
The updates are as part of the preparation of a large commit to add all instruction scheduling for the SKX target.
Reviewers: delena, zvi, RKSimon
Differential Revision: https://reviews.llvm.org/D38222
Change-Id: I2381c9b5bb75ecacfca017243c22d054f6eddd14
llvm-svn: 314306
This enables the use of a smaller encoding by using a VEX instruction when possible.
Differential Revision: https://reviews.llvm.org/D37092
llvm-svn: 312100
This patch completely replaces the instruction scheduling information for the Haswell architecture target by modifying the file X86SchedHaswell.td located under the X86 Target.
We used the scheduling information retrieved from the Haswell architects in order to replace and modify the existing scheduling.
The patch continues the scheduling replacement effort started with the SNB target in r307529 and r310792.
Information includes latency, number of micro-Ops and used ports by each HSW instruction.
Please expect some performance fluctuations due to code alignment effects.
Reviewers: RKSimon, zvi, aymanmus, craig.topper, m_zuckerman, igorb, dim, chandlerc, aaboud
Differential Revision: https://reviews.llvm.org/D36663
llvm-svn: 311879
There's no reason to switch instructions with and without DQI. It just creates extra isel patterns and test divergences.
There is however value in enabling the masked version of the instructions with DQI.
This required introducing some new multiclasses to enabling this splitting.
Differential Revision: https://reviews.llvm.org/D36661
llvm-svn: 311091
`llc -march` is problematic because it only switches the target
architecture, but leaves the operating system unchanged. This
occasionally leads to indeterministic tests because the OS from
LLVM_DEFAULT_TARGET_TRIPLE is used.
However we can simply always use `llc -mtriple` instead. This changes
all the tests to do this to avoid people using -march when they copy and
paste parts of tests.
See also the discussion in https://reviews.llvm.org/D35287
llvm-svn: 309774
We were already using the 32 bit element opcode if BWI isn't enabled, but there's no reason to change opcode if we have BWI. We will still use the 8/16 opcodes for masked stores though.
This allows us to use the aligned opcode when we can which makes our test output more consistent between different modes. It also reduces the number of isel patterns we need.
This is a slight inconsistency with loads which default to 64 bit element opcodes. I'll probably rectify that in a future patch.
Differential Revision: https://reviews.llvm.org/D35978
llvm-svn: 309693
These were taking priority over the aligned load instructions since there is no vmovda8/16. I don't think there is really a difference between aligned and unaligned on newer cpus so I don't think it matters which instructions we use.
But with this change we reduce the size of the isel table a little and we allow the aligned information to pass through to the evex->vec pass and produce the same output has avx/avx2 in some cases.
I also generally dislike patterns rooted in a bitcast which these were.
Differential Revision: https://reviews.llvm.org/D35977
llvm-svn: 309589
This is NFC after rerunning the "update_llc_test_checks.py" tool on the CodeGen X86 tests in order to submit a patch.
Minor differences due to added "End of Function" lines.
Reviewers: zvi
Differential Revision: https://reviews.llvm.org/D34933
llvm-svn: 306973
•static latency
•number of uOps from which the instructions consists
•all ports used by the instruction
Reviewers:
RKSimon
zvi
aymanmus
m_zuckerman
Differential Revision: https://reviews.llvm.org/D33897
llvm-svn: 306414
This patch defines the i1 type as illegal in the X86 backend for AVX512.
For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended.
This should produce better scalar code for i1 types since GPRs will be used instead of mask registers.
Differential Revision: https://reviews.llvm.org/D32273
llvm-svn: 303421
This will result in a KMOVW or KMOVD being emitted during register allocation. And in at least some cases this might allow the register coalescer to remove the copy all together.
llvm-svn: 298984
We've had several bugs(PR32256, PR32241) recently that resulted from usages of AH/BH/CH/DH either before or after a copy to/from a mask register.
This ultimately occurs because we create COPY_TO_REGCLASS with VK1 and GR8. Then in CopyToFromAsymmetricReg in X86InstrInfo we find a 32-bit super register for the GR8 to emit the KMOV with. But as these tests are demonstrating, its possible for the GR8 register to be a high register and we end up doing an accidental extra or insert from bits 15:8.
I think the best way forward is to stop making copies directly between mask registers and GR8/GR16. Instead I think we should restrict to only copies between mask registers and GR32/GR64 and use EXTRACT_SUBREG/INSERT_SUBREG to handle the conversion from GR32 to GR16/8 or vice versa.
Unfortunately, this complicates fastisel a bit more now to create the subreg extracts where we used to create GR8 copies. We can probably make a helper function to bring down the repitition.
This does result in KMOVD being used for copies when BWI is available because we don't know the original mask register size. This caused a lot of deltas on tests because we have to split the checks for KMOVD vs KMOVW based on BWI.
Differential Revision: https://reviews.llvm.org/D30968
llvm-svn: 298928
I noticed unnecessary 'sbb' instructions in D30472 and while looking at 'ptest' codegen recently.
This happens because we were transforming any 'setb' - even when we only wanted a single-bit result.
This patch moves those transforms under visitAdd/visitSub, so we we're only creating sbb/adc when it
is a win. I don't know why we need a SETCC_CARRY node type, but I'm not proposing to change that
existing behavior in this patch.
Also, I'm skeptical that sbb/adc are a win for all micro-arches, so I added comments to the test files
where this transform still fires.
The test changes here are all cases where we no longer produce sbb/adc. Avoiding partial register
stalls (generating an xor to clear a register) is not handled in some cases, but that's a separate
issue.
Differential Revision: https://reviews.llvm.org/D30611
llvm-svn: 297586
This helps in cases involving bitfields where an AND is exposed by
legalization.
Differential Revision: https://reviews.llvm.org/D30472
llvm-svn: 297249
The larger goal is to move the ADC/SBB transforms currently in
combineX86SetCC() to combineAddOrSubToADCOrSBB() because we're
creating ADC/SBB in lots of places where we shouldn't.
This was intended to be an NFC change, but avx-512 has something
strange going on. It doesn't seem like any of the affected tests
should really be using SET+TEST or ADC; a simple ADD could replace
several instructions. But that's another bug...
llvm-svn: 296978
select Cond, C +/- 1, C --> add(ext Cond), C -- with a target hook.
This is part of the ongoing process to obsolete D24480. The motivation is to
canonicalize to select IR in InstCombine whenever possible, so we need to have a way to
undo that easily in codegen.
PowerPC is an obvious winner for this kind of transform because it has fast and complete
bit-twiddling abilities but generally lousy conditional execution perf (although this might
have changed in recent implementations).
x86 also sees some wins, but the effect is limited because these transforms already mostly
exist in its target-specific combineSelectOfTwoConstants(). The fact that we see any x86
changes just shows that that code is a mess of special-case holes. We may be able to remove
some of that logic now.
My guess is that other targets will want to enable this hook for most cases. The likely
follow-ups would be to add value type and/or the constants themselves as parameters for the
hook. As the tests in select_const.ll show, we can transform any select-of-constants to
math/logic, but the general transform for any 2 constants needs one more instruction
(multiply or 'and').
ARM is one target that I think may not want this for most cases. I see infinite loops there
because it wants to use selects to enable conditionally executed instructions.
Differential Revision: https://reviews.llvm.org/D30537
llvm-svn: 296977
VZEROUPPER should not be issued on Knights Landing (KNL), but on Skylake-avx512 it should be.
Differential Revision: https://reviews.llvm.org/D29874
llvm-svn: 296859
This is part of the ongoing attempt to improve select codegen for all targets and select
canonicalization in IR (see D24480 for more background). The transform is a subset of what
is done in InstCombine's FoldOpIntoSelect().
I first noticed a regression in the x86 avx512-insert-extract.ll tests with a patch that
hopes to convert more selects to basic math ops. This appears to be a general missing DAG
transform though, so I added tests for all standard binops in rL296621
(PowerPC was chosen semi-randomly; it has scripted FileCheck support, but so do ARM and x86).
The poor output for "sel_constants_shl_constant" is tracked with:
https://bugs.llvm.org/show_bug.cgi?id=32105
Differential Revision: https://reviews.llvm.org/D30502
llvm-svn: 296699
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
Craig Topper