Further develops the vsextload support added in #79815 / b5d35feacb7246573c6a4ab2bddc4919a4228ed5 - reduces the size of the vector constant by storing it in the constant pool in a truncated form, and zero-extend it as part of the load.
Reduce the size of the vector constant by storing it in the constant pool in a truncated form, and sign-extend it as part of the load.
I've extended the existing FixupConstant functionality to support these sext constant rebuilds - we still select the smallest stored constant entry and prefer vzload/broadcast/vextload for same bitwidth to avoid domain flips.
I intend to add the matching load+zero-extend handling in a future PR, but that requires some alterations to the existing MC shuffle comments handling first.
Extend coverage for lowering wide vector types during type legalization to allow us to use PACKSS/PACKUS patterns instead of dropping down to shuffle lowering.
First step towards avoiding premature folds of TRUNCATE to PACKSS/PACKUS nodes as described on Issue #63710 - which causes a large number of regressions on D152928 - we will next need to tweak the TRUNCATE widening in ReplaceNodeResults
Differential Revision: https://reviews.llvm.org/D154592
lowerBuildVectorAsBroadcast will not broadcast splat constants in all cases, resulting in a lot of situations where a full width vector load that has failed to fold but is loading splat constant values could use a broadcast load instruction just as cheaply, and save constant pool space.
This is an updated commit of ab4b924832ce26c21b88d7f82fcf4992ea8906bb after being reverted at 78de45fd4a902066617fcc9bb88efee11f743bc6
lowerBuildVectorAsBroadcast will not broadcast splat constants in all cases, resulting in a lot of situations where a full width vector load that has failed to fold but is loading splat constant values could use a broadcast load instruction just as cheaply, and save constant pool space.
NOTE: SSE3 targets can use MOVDDUP but not all SSE era CPUs can perform this as cheaply as a vector load, we will need to add scheduler model checks if we want to pursue this.
This is an updated commit of 98061013e01207444cfd3980cde17b5e75764fbe after being reverted at a279a09ab9524d1d74ef29b34618102d4b202e2f
lowerBuildVectorAsBroadcast will not broadcast splat constants in all cases, resulting in a lot of situations where a full width vector load that has failed to fold but is loading splat constant values could use a broadcast load instruction just as cheaply, and save constant pool space.
lowerBuildVectorAsBroadcast will not broadcast splat constants in all cases, resulting in a lot of situations where a full width vector load that has failed to fold but is loading splat constant values could use a broadcast load instruction just as cheaply, and save constant pool space.
NOTE: SSE3 targets can use MOVDDUP but not all SSE era CPUs can perform this as cheaply as a vector load, we will need to add scheduler model checks if we want to pursue this.
Noticed in D150143/D150526 - we currently create scalar Constant values using the broadcast instruction width, which might be wider than the original build vector width, making it tricky to recognise the original constant bits data.
If we have widened the broadcast value, its much more useful for asm comments if we create a ConstantVector with the original element data, add that to the constant-pool and load that with the same (wider) broadcast instruction.
This change adjusts the cost modeling used when the target does not have a schedule model with individual instruction latencies. After this change, we use the default latency information available from TargetSchedule. The default latency information essentially ends up treating most instructions as latency 1, with a few "expensive" ones getting a higher cost.
Previously, we unconditionally applied the first legal pattern - without any consideration of profitability. As a result, this change both prevents some patterns being applied, and changes which patterns are exercised. (i.e. previously the first pattern was applied, afterwards, maybe the second one is because the first wasn't profitable.)
The motivation here is two fold.
First, this brings the default behavior in line with the behavior when -mcpu or -mtune is specified. This improves test coverage, and generally makes it less likely we will have bad surprises when providing more information to the compiler.
Second, this enables some reassociation for ILP by default. Despite being unconditionally enabled, the prior code tended to "reassociate" repeatedly through an entire chain and simply moving the first operand to the end. The result was still a serial chain, just a different one. With this change, one of the intermediate transforms is unprofitable and we end up with a partially flattened tree.
Note that the resulting code diffs show significant room for improvement in the basic algorithm. I am intentionally excluding those from this patch.
For the test diffs, I don't seen any concerning regressions. I took a fairly close look at the RISCV ones, but only skimmed the x86 (particularly vector x86) changes.
Differential Revision: https://reviews.llvm.org/D141017
This stops reporting CostPerUse 1 for `R8`-`R15` and `XMM8`-`XMM31`.
This was previously done because instruction encoding require a REX
prefix when using them resulting in longer instruction encodings. I
found that this regresses the quality of the register allocation as the
costs impose an ordering on eviction candidates. I also feel that there
is a bit of an impedance mismatch as the actual costs occure when
encoding instructions using those registers, but the order of VReg
assignments is not primarily ordered by number of Defs+Uses.
I did extensive measurements with the llvm-test-suite wiht SPEC2006 +
SPEC2017 included, internal services showed similar patterns. Generally
there are a log of improvements but also a lot of regression. But on
average the allocation quality seems to improve at a small code size
regression.
Results for measuring static and dynamic instruction counts:
Dynamic Counts (scaled by execution frequency) / Optimization Remarks:
Spills+FoldedSpills -5.6%
Reloads+FoldedReloads -4.2%
Copies -0.1%
Static / LLVM Statistics:
regalloc.NumSpills mean -1.6%, geomean -2.8%
regalloc.NumReloads mean -1.7%, geomean -3.1%
size..text mean +0.4%, geomean +0.4%
Static / LLVM Statistics:
mean -2.2%, geomean -3.1%) regalloc.NumSpills
mean -2.6%, geomean -3.9%) regalloc.NumReloads
mean +0.6%, geomean +0.6%) size..text
Static / LLVM Statistics:
regalloc.NumSpills mean -3.0%
regalloc.NumReloads mean -3.3%
size..text mean +0.3%, geomean +0.3%
Differential Revision: https://reviews.llvm.org/D133902
Without SSE41, ANY_EXTEND_VECTOR_INREG nodes are likely to be prematurely combined to a target shuffle preventing generic sign extension folds.
Fixes a number of sign-extend regressions in D127115.
Limit this to SSE41 - AVX1 targets to avoid UNPCKL(PSHUFB,PSHUFB), pre-SSE41 we don't have PACKUSDW/BLENDW and with AVX2 we can perform this as PERMQ(PSHUFB()).
Currently we only constant fold target shuffles if any of the sources has one use, or it would remove a variable shuffle mask - the aim being to avoid constant pool bloat.
This patch proposes we should constant fold by default and only limit this if optsize is enabled - I've added a basic test for this in vector-mul.ll (the pmuludq case is by far the most common), I can add other specific test cases if people need them.
This should permit further constant folding, break some instruction dependencies and help reduce shuffle port pressure.
Differential Revision: https://reviews.llvm.org/D113748
This patch contains following enhancements to SrcRegMap and DstRegMap:
1 In findOnlyInterestingUse not only check if the Reg is two address usage,
but also check after commutation can it be two address usage.
2 If a physical register is clobbered, remove SrcRegMap entries that are
mapped to it.
3 In processTiedPairs, when create a new COPY instruction, add a SrcRegMap
entry only when the COPY instruction is coalescable. (The COPY src is
killed)
With these enhancements isProfitableToCommute can do better commute decision,
and finally more register copies are removed.
Differential Revision: https://reviews.llvm.org/D108731
Currently, X86 backend only has a global one-size-fits-all `FeatureFastVariableShuffle` feature,
which controls profitability of both the cross-lane and per-lane variable shuffles.
I guess, this has been fine so far.
But at least on AMD Zen 3, while per-line variable shuffles (e.g. `VPSHUFB`)
are as fast as as shuffles with fixed/immediate mask,
while lane-crossing shuffles, e.g. `VPERMPS` is performing worse.
So to get the benefits of variable-mask shuffles, but not the drawbacks of lane-crossing shuffles,
as suggested by @RKSimon, split the feature flag into two.
Differential Revision: https://reviews.llvm.org/D103274
Until AVX512 we don't have any vector truncation instructions, and always lower using shuffles instead.
combineVectorTruncation performs this earlier than lowering as it makes it easier to use any sign/zero-extended bits in the truncated bits with PACKSS/PACKUS to perform the shuffle.
We currently don't attempt to use combineVectorTruncation on AVX2 targets as in the past 256-bit PACKSS/PACKUS tended to cause 128-bit lane shuffle regressions - but these should now be all resolved with combineHorizOpWithShuffle and in all cases we now reduce the amount of cross-lane shuffling and variable shuffle mask usage.
Differential Revision: https://reviews.llvm.org/D96609
If we're multiplying all elements of a vector by '0' or '1' then we can more efficiently perform this as a clearing mask (that is likely to further simplify to a shuffle blend).
This was noticed when reviewing D87502 but seems to help idiv/irem by constant cases even more as '0'/'1' values are often used for 'passthrough' cases.
Differential Revision: https://reviews.llvm.org/D88225
Shuffle combining can now handle this output, and by performing this early in combineVectorTruncation we avoid a scalarization that caused a regression on D87502.
This patch uses partial DemandedElts masks to further simplify target shuffle chains and finally starts making target shuffle combining part of SimplifyDemandedBits/SimplifyDemandedVectorElts.
We already manage this for Depth == 0 cases, where combineX86ShuffleChain would early-out if the shuffle combined to the same op, but the patch generalizes this by manipulating the depth handling of combineX86ShufflesRecursively - calling with a new Depth = 0 and reducing the maximum shuffle combine depth accordingly.
Differential Revision: https://reviews.llvm.org/D66004
We already do this on AVX (+ for ZERO_EXTEND_VECTOR_INREG), but this enables it for all SSE targets - we attempted something similar back at rL357057 but hit issues with the ZERO_EXTEND_VECTOR_INREG handling (PR41249).
I'm still looking at the vector-mul.ll regression - which is due to 32-bit targets performing the load as a f64, resulting in the shuffle combiner thinking it has to create a shuffle in the float domain.
Pulled out from the ongoing work on D66004, currently we don't do a good job of simplifying variable shuffle masks that have already lowered to constant pool entries.
This patch adds SimplifyDemandedVectorEltsForTargetShuffle (a custom x86 helper) to first try SimplifyDemandedVectorElts (which we already do) and then constant pool simplification to help mark undefined elements.
To prevent lowering/combines infinite loops, we only handle basic constant pool loads instead of creating new BUILD_VECTOR nodes for lowering - e.g. we don't try to convert them to broadcast/vzext_load - there might be some benefit to this but if so I'd rather we come up with some way to reuse existing code than reimplement a lot of BUILD_VECTOR code.
Differential Revision: https://reviews.llvm.org/D81791
Default legalization will create two v8i64 truncs to v8i32, concat
them to v16i32, and then truncate the rest of the way to v16i8.
Instead we can truncate directly from v8i64 to v8i8 in the lower
half of an xmm. Then concat the two halves to use vpunpcklqdq.
This is the same number of uops, but the dependency chain through
the uops is better since the halves are merged at the end.
I had to had SimplifyDemandedBits support for VTRUNC to prevent
a regression on vector-trunc-math.ll. combineTruncatedArithmetic
no longer gets a chance to shrink vXi64 mul so we were producing
the v8i64 multiply sequence using multiple PMULUDQs. With the
demanded bits fix we are able to prune out the extra ops leaving
just two PMULUDQs, one for each v8i64 half. This is twice the
width of the 2 v8i32 PMULLDs we had before, but PMULUDQ is 1
uop and PMULLD is 2. We also save some truncates. It's probably
worth using PMULUDQ even when PMULLQ is available since the latter
is 3 uops, but that will require a different change.
Differential Revision: https://reviews.llvm.org/D79231
We can improve computeKnownBits results by avoiding excess bitcasts.
For this pattern we were doing:
(v16i8 PACKUS(v8i16 BITCAST(v16i8 AND(V1, MASK)), v8i16 BITCAST(v16i8 AND(V2, MASK))))
By performing the MASK/AND with a v8i16 type and bitcasting V1/V2 directly we can help computeKnownBits see that the mask is clearing the upper bits and allows shuffle combining to peek through later on.
This will be necessary to extend rG9d1721ce3926 to AVX2+ targets in a future patch.
Now that rG18c19441d105 has improved VPERM2X128 handling, we can perform this to improve x64->x32 truncation without poor cross-lane issues.
Someday combineX86ShufflesRecursively will handle this, but we're still really bad at dealing with different vector widths.
We can only report the knownbits for a SCALAR_TO_VECTOR node if we only demand the 0'th element - the upper elements are undefined and shouldn't be trusted.
This is causing a number of regressions that need addressing but we need to get the bugfix in first.
Summary:
The 2 source operands commutable instructions are encoded in the
VEX.VVVV field and the r/m field of the MODRM byte plus the VEX.B
field.
The VEX.B field is missing from the 2-byte VEX encoding. If the
VEX.VVVV source is 0-7 and the other register is 8-15 we can
swap them to avoid needing the VEX.B field. This works as long as
the VEX.W, VEX.mmmmm, and VEX.X fields are also not needed.
Fixes PR36706.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D68550
The build_vector will become a constant pool load. By using the
desired type initially, it ensures we don't generate a bitcast
of the constant pool load which will need to be folded with
the load.
While experimenting with another patch, I noticed that when the
load type and the constant pool type don't match, then
SimplifyDemandedBits can't handle it. While we should probably
fix that, this was a simple way to fix the issue I saw.
llvm-svn: 366732
combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together.
This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356).
This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit.
My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great.
Differential Revision: https://reviews.llvm.org/D60375#inline-539623
llvm-svn: 358692
Similar to D57867 - this is a small patch with lots of test diffs.
With half-vector-width narrowing potential, using an extract + 128-bit vshufps
is a win because it replaces a 256-bit shuffle with a 128-bit shufle.
This seems like it should be a win even for targets with 'fast-variable-shuffle',
but we are intentionally deferring that to an independent change to make sure
that is true.
Differential Revision: https://reviews.llvm.org/D58181
llvm-svn: 354279
This is effectively re-committing the changes from:
rL347917 (D54640)
rL348195 (D55126)
...which were effectively reverted here:
rL348604
...because the code had a bug that could induce infinite looping
or eventual out-of-memory compilation.
The bug was that this code did not guard against transforming
opaque constants. More details are in the post-commit mailing
list thread for r347917. A reduced test for that is included
in the x86 bool-math.ll file. (I wasn't able to reduce a PPC
backend test for this, but it was almost the same pattern.)
Original commit message for r347917:
The motivating case for this is shown in:
https://bugs.llvm.org/show_bug.cgi?id=32023
and the corresponding rot16.ll regression tests.
Because x86 scalar shift amounts are i8 values, we can end up with trunc-binop-trunc
sequences that don't get folded in IR.
As the TODO comments suggest, there will be regressions if we extend this (for x86,
we mostly seem to be missing LEA opportunities, but there are likely vector folds
missing too). I think those should be considered existing bugs because this is the
same transform that we do as an IR canonicalization in instcombine. We just need
more tests to make those visible independent of this patch.
llvm-svn: 348706
As discussed in the post-commit thread of r347917, this
transform is fighting with an existing transform causing
an infinite loop or out-of-memory, so this is effectively
reverting r347917 and its follow-up r348195 while we
investigate the bug.
llvm-svn: 348604
This is the smallest vector enhancement I could find to D54640.
Here, we're allowing narrowing to only legal vector ops because we'll see
regressions without that. All of the test diffs are wins from what I can tell.
With AVX/AVX512, we can shrink ymm/zmm ops to xmm.
x86 vector multiplies are the problem case that we're avoiding due to the
patchwork ISA, and it's not clear to me if we can dance around those
regressions using TLI hooks or if we need preliminary patches to plug those
holes.
Differential Revision: https://reviews.llvm.org/D55126
llvm-svn: 348195
For bitcast nodes from larger element types, add the ability for SimplifyDemandedVectorElts to call SimplifyDemandedBits by merging the elts mask to a bits mask.
I've raised https://bugs.llvm.org/show_bug.cgi?id=39689 to deal with the few places where SimplifyDemandedBits's lack of vector handling is a problem.
Differential Revision: https://reviews.llvm.org/D54679
llvm-svn: 347301
Previously we used an arithmetic shift right by 31, but that requires a copy to preserve the input. So we might as well materialize a zero and compare to it since the comparison will overwrite the register that contains the zeros. This should be one byte shorter.
llvm-svn: 347181
Summary: This is an improvement over the two pshufbs and punpcklqdq we'd get otherwise.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D54671
llvm-svn: 347171
We were using the 'normalized' shuffle mask from resolveTargetShuffleInputs, which replaces zero/undef inputs with sentinel values. For SimplifyDemandedVectorElts we need the raw mask so we can correctly demand those 'zero' inputs that got normalized away, this requires an extra bit of logic to locally normalize undef inputs.
llvm-svn: 347158
With avx512f but not avx512bw we need to extend to v16i32 then truncate that to to v16i8. Previously we emitted both nodes during lowering, but I'm trying to switch to using target independent nodes and with that switched the extend+truncate wou
This patch changes the implementation to what will be necessary with that patch which helps minimize test diffs.
llvm-svn: 346552
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
These promotions add additional bitcasts to the SelectionDAG that can pessimize computeKnownBits/computeNumSignBits. It also seems to interfere with broadcast formation.
This patch removes the promotion and adds isel patterns instead.
The increased table size is more than I would like, but hopefully we can find some canonicalizations or other tricks to start pruning out patterns going forward.
Differential Revision: https://reviews.llvm.org/D53268
llvm-svn: 345408
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
