We need to use vector instructions for these operations. Previously
we handled this with isel patterns that used extra instructions
and copies to handle the the conversions.
Now we use custom lowering to emit the conversions. This allows
them to be pattern matched and optimized on their own. For
example we can now emit vpextrw to store the result if its going
directly to memory.
I've forced the upper elements to VCVTPHS2PS to zero to keep some
code similar. Zeroes will be needed for strictfp. I've added a
DAG combine for (fp16_to_fp (fp_to_fp16 X)) to avoid extra
instructions in between to be closer to the previous codegen.
This is a step towards strictfp support for f16 conversions.
Using sign extend forces the adjacent element to either all zeros
or all ones. But all ones is a NAN. So that doesn't seem like a
great idea.
Trying to work on supporting this with strict FP where NAN would
definitely be bad.
This is based on this llvm-dev thread http://lists.llvm.org/pipermail/llvm-dev/2019-December/137521.html
The current strategy for f16 is to promote type to float every except where the specific width is required like loads, stores, and bitcasts. This results in rounding occurring in odd places instead of immediately after arithmetic operations. This interacts in weird ways with the __fp16 type in clang which is a storage only type where arithmetic is always promoted to float. InstCombine can remove some fpext/fptruncs around such arithmetic and turn it into arithmetic on half. This wouldn't be so bad if SelectionDAG was able to put those fpext/fpround back in when it promotes.
It is also not obvious how to handle to make the existing strategy work with STRICT fp. We need to use STRICT versions of the conversions which require chain operands. But if the conversions are created for a bitcast, there is no place to get an appropriate chain from.
This patch implements a different strategy where conversions are emitted directly around arithmetic operations. And otherwise its passed around as an i16 including in arguments and return values. This can result in more conversions between arithmetic operations, but is closer to matching the IR the frontend generates for __fp16. And it will allow us to use the chain from constrained arithmetic nodes to link the STRICT_FP_TO_FP16/STRICT_FP16_TO_FP that will need to be added. I've set it up so that each target can opt into the new behavior. Converting all the targets myself was more than I was able to handle.
Differential Revision: https://reviews.llvm.org/D73749
Summary:
After bugfix the undef value case here, we used more operations to implement inserting vxi1 sub vector into vXi1 vector, I optimize it by use less operations.
The history information at https://reviews.llvm.org/D68311
Reviewers: craig.topper, LuoYuanke, yubing, annita.zhang, pengfei, LiuChen3, RKSimon
Reviewed By: craig.topper
Subscribers: hiraditya, llvm-commits
Patch by Xiang Zhang (xiangzhangllvm)
Differential Revision: https://reviews.llvm.org/D71917
The previous code tried to do a trick where we would extract the subvector from the location we were inserting. Then xor that with the new value. Take the xored value and clear out the bits above the subvector size. Then shift that xored subvector to the insert location. And finally xor that with the original vector. Since the old subvector was used in both xors, this would leave just the new subvector at the inserted location. Since the surrounding bits had been zeroed no other bits of the original vector would be modified.
Unfortunately, if the old subvector came from undef we might aggressively propagate the undef. Then we end up with the XORs not cancelling because they aren't using the same value for the two uses of the old subvector. @bkramer gave me a case that demonstrated this, but we haven't reduced it enough to make it easily readable to see what's happening.
This patch uses a safer, but more costly approach. It isolate the bits above the insertion and bits below the insert point and ORs those together leaving 0 for the insertion location. Then widens the subvector with 0s in the upper bits, shifts it into position with 0s in the lower bits. Then we do another OR.
Differential Revision: https://reviews.llvm.org/D68311
llvm-svn: 373495
We have isel patterns that can put an IMPLICIT_DEF on one of
the sources for these instructions. So we should make sure
we break any dependencies there. This should be done by
just using one of the other sources.
llvm-svn: 373025
gcc and icc pass these types in zmm registers in zmm registers.
This patch implements a quick hack to override the register
type before calling convention handling to one that is legal.
Longer term we might want to do something similar to 256-bit
integer registers on AVX1 where we just split all the operations.
Fixes PR42957
Differential Revision: https://reviews.llvm.org/D66708
llvm-svn: 370495
These patterns are the same as the MOVLPDmr and MOVHPDmr patterns,
but with a bitcast at the end. We can just select the PD instruction
and let execution domain fixing switch to PS.
llvm-svn: 365267
As part of the fix for rL364264 + rL364272 - limit the *_EXTEND conversion to !TLO.LegalOperations || isOperationLegal cases.
We'll improve X86 legality in future commits.
llvm-svn: 364290
We effectively had a second set of isel patterns that tried to use a
regular store instruction and an extract_subreg instruction. Or a masked move
and an extract_subreg. These patterns were intended to override the
matching of VEXTRACT instructions by taking advantage of the priority
of the explicit immediate 0 for the index.
This patch instaed just disables the immediate 0 matchin the VEXTRACT
patterns. This each of the component pieces of the larger patterns will
match by themselves.
This found a bug of sorts were we didn't use 128-bit store for 512->128
extract on KNL. Its unclear what the right thing here should be.
Using the vextract avoids constraining the register allocator to use
xmm0-15. But it always results in a longer encoding if the register
allocator ends up choosing xmm0-15 anyway.
llvm-svn: 361431
First half of PR40800, this patch adds DAG undef handling to icmp instructions to match the behaviour in llvm::ConstantFoldCompareInstruction and SimplifyICmpInst, this permits constant folding of vector comparisons where some elements had been reduced to UNDEF (by SimplifyDemandedVectorElts etc.).
This involved a lot of tweaking to reduced tests as bugpoint loves to reduce icmp arguments to undef........
Differential Revision: https://reviews.llvm.org/D59363
llvm-svn: 356938
We were lowering the last step extract_vector_elt to a bitcast+truncate. Change it to use an extract_vector_elt of index 0 instead. Add isel patterns to do the equivalent of what the bitcast would have done. Plus an isel pattern for an any_extend+extract to prevent some regressions.
Finally add a DAG combine to turn v1i1 scalar_to_vector+extract_vector_elt of 0 into an extract_subvector.
This fixes some of the regressions from D350800.
llvm-svn: 350918
This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more
opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs.
Apart from 2-3 strange cases, these are all wins.
I've structured this to be no-functional-change-intended for any target except for x86
because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those
targets have existing regression tests (4, 4, 10 files respectively) that would be
affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show
any regression test diffs. The trade-off is deciding if an extra vector load is better
than a single wide load + extract_subvector.
For x86, this is almost always better (on paper at least) because we often can fold
loads into subsequent ops and not increase the official instruction count. There's also
some unknown -- but potentially large -- benefit from using narrower vector ops if wide
ops are implemented with multiple uops and/or frequency throttling is avoided.
Differential Revision: https://reviews.llvm.org/D54073
llvm-svn: 346595
This makes X86ISD::VSEXT more similar to ISD::SIGN_EXTEND and ISD::ZERO_EXTEND.
I'm hoping to replace X86ISD::VSEXT/VZEXT with target independent nodes. Making the target specific nodes similar to the target independent nodes helps minimize test diffs in that patch.
llvm-svn: 346539
Narrowing vector binops came up in the demanded bits discussion in D52912.
I don't think we're going to be able to do this transform in IR as a canonicalization
because of the risk of creating unsupported widths for vector ops, but we already have
a DAG TLI hook to allow what I was hoping for: isExtractSubvectorCheap(). This is
currently enabled for x86, ARM, and AArch64 (although only x86 has existing regression
test diffs).
This is artificially limited to not look through bitcasts because there are so many
test diffs already, but that's marked with a TODO and is a small follow-up.
Differential Revision: https://reviews.llvm.org/D53784
llvm-svn: 345602
Summary:
I've noticed that the bitcasts we introduce for these make computeKnownBits and computeNumSignBits not work well in LegalizeVectorOps. LegalizeVectorOps legalizes bottom up while LegalizeDAG legalizes top down. The bottom up strategy for LegalizeVectorOps means operands are legalized before their uses. So we promote and/or/xor before we legalize the operands that use them making computeKnownBits/computeNumSignBits in places like LowerTruncate suboptimal. I looked at changing LegalizeVectorOps to be top down as well, but that was more disruptive and caused some regressions. I also looked at just moving promotion of binops to LegalizeDAG, but that had a few issues one around matching AND,ANDN,OR into VSELECT because I had to create ANDN as vXi64, but the other nodes hadn't legalized yet, I didn't look too hard at fixing that.
This patch seems to produce better results overall than my other attempts. We now form broadcasts of constants better in some cases. For at least some of them the AND was being introduced in LegalizeDAG, promoted to vXi64, and the BUILD_VECTOR was also legalized there. I think we got bad ordering of that. Now the promotion is out of the legalizer so we handle this better.
In the longer term I think we really should evaluate whether we should be doing this promotion at all. It's really there to reduce isel pattern count, but I'm wondering if we'd be better served just eating the pattern cost or doing C++ based isel for vector and/or/xor in X86ISelDAGToDAG. The masked and/or/xor will definitely be difficult in patterns if a bitcast gets between the vselect and the and/or/xor node. That becomes a lot of permutations to cover.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D53107
llvm-svn: 344487
Enable enableMultipleCopyHints() on X86.
Original Patch by @jonpa:
While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling.
Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates.
Differential Revision: https://reviews.llvm.org/D38128
llvm-svn: 342578
AVX512 doesn't have an immediate controlled blend instruction. But blend throughput is still better than movss/sd on SKX.
This commit changes AVX512 to use the AVX blend instructions instead of MOVSS/MOVSD. This constrains the register allocation since it won't be able to use XMM16-31, but hopefully the increased throughput and reduced port 5 pressure makes up for that.
llvm-svn: 337083
This teaches tryToFoldExtOfLoad to set the right location on a
newly-created extload. With that in place, the logic for performing a
certain ([s|z]ext (load ...)) combine becomes identical for sexts and
zexts, and we can get rid of one copy of the logic.
The test case churn is due to dependencies on IROrders inherited from
the wrong SDLoc.
Part of: llvm.org/PR37262
Differential Revision: https://reviews.llvm.org/D46158
llvm-svn: 332118
This supports things like
(setcc ugt X, 0) -> (setcc ne X, 0)
I've restricted to only make changes to vectors before legalize ops because I doubt all targets have accurate condition code legality information for vectors given how little we did before.
Differential Revision: https://reviews.llvm.org/D42948
llvm-svn: 326495
This allows us to avoid an opsize prefix. And forcing some move immediates to i32 avoids a length changing prefix on those instructions.
This mostly replaces the existing combine we had for zext/sext+cmov of constants. I left in a case for sign extending a 32 bit cmov of constants to 64 bits.
Differential Revision: https://reviews.llvm.org/D43327
llvm-svn: 325601
Summary:
Currently we only use min/max to help with ule/uge compares because it removes an invert of the result that would otherwise be needed. But we can also use it for ult/ugt compares if it will prevent the need for a sign bit flip needed to use pcmpgt at the cost of requiring an invert after the compare.
I also refactored the code so that the max/min code is self contained and does its own return instead of setting up a flag to manipulate the rest of the function's behavior.
Most of the test cases look ok with this. I did notice that we added instructions when one of the operands being sign flipped is a constant vector that we were able to constant fold the flip into.
I also noticed that sometimes the SSE min/max clobbers a register that is needed after the compare. This resulted in an extra move being inserted before the min/max to preserve the register. We could try to detect this and switch from min to max and change the compare operands to use the operand that gets reused in the compare.
Reviewers: spatel, RKSimon
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D42935
llvm-svn: 324842
Discussed here:
http://lists.llvm.org/pipermail/llvm-dev/2018-January/120320.html
In preparation for adding support for named vregs we are changing the sigil for
physical registers in MIR to '$' from '%'. This will prevent name clashes of
named physical register with named vregs.
llvm-svn: 323922
The existing code was already doing something very similar to subvector insertion so this allows us to remove the nearly duplicate code.
This patch is a little larger than it should be due to differences between the DQI handling between the two today.
llvm-svn: 323212
Summary:
For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register.
I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes.
There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined.
I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc))
Overall this removes something like 13000 CHECK lines from lit tests.
Reviewers: zvi, RKSimon, delena, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D42031
llvm-svn: 323201
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
Craig Topper