Update splitVectorIntUnary/splitVectorIntBinary to use this internally, after some operand type sanity checks.
Avoid code duplication and makes it easier to split other vector instruction forms in the future.
The first try at this patch ( bf5748a1af ) was reverted ( 5be64d4164 )
because it could crash. The cause of that problem was failing to account
for the optional peek-through-bitcast in the enclosing function.
This version of the patch adds a clause to avoid the fold in case of
bitcasts because it is unlikely to be profitable in that scenario.
A test case based on https://llvm.org/PR52504 was added to make sure
we don't have that problem again.
Original commit message:
and (pcmpgt X, -1), Y --> pandn (vsrai X, BitWidth-1), Y
This avoids the -1 constant vector in favor of an arithmetic shift
instruction if it exists (the ISA is still not complete after all
these years...).
We catch this pattern late in combining by matching PCMPGT, so it
should not interfere with more general folds.
Differential Revision: https://reviews.llvm.org/D113603
If an operand is a bitcasted or widended constant, try to more aggressively create broadcastable constants for folding, which in particular helps non-VLX modes.
I've refactored getAVX512Node so that VLX targets can make better use of this as well.
NOTE: In the future, I think we should consider removing the broadcast of constant data from DAG entirely and move this to either X86InstrInfo::foldMemoryOperand or a new pass - AVX1/2 targets has similar problems with missed (whole vector) folds that need to be improved as well.
Differential Revision: https://reviews.llvm.org/D113845
This casued assertion failures:
llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp:9446:
void llvm::SelectionDAG::ReplaceAllUsesWith(llvm::SDNode *, llvm::SDNode *):
Assertion `(!From->hasAnyUseOfValue(i) || From->getValueType(i) == To->getValueType(i))
&& "Cannot use this version of ReplaceAllUsesWith!"' failed.
See comment on the code review.
(Had to update some expectations in test/CodeGen/X86/vselect-zero.ll
manually due to other changes having landed after the reverted one.)
> and (pcmpgt X, -1), Y --> pandn (vsrai X, BitWidth-1), Y
>
> This avoids the -1 constant vector in favor of an arithmetic shift
> instruction if it exists (the ISA is still not complete after all
> these years...).
>
> We catch this pattern late in combining by matching PCMPGT, so it
> should not interfere with more general folds.
>
> Differential Revision: https://reviews.llvm.org/D113603
This reverts commit bf5748a1af.
This helper provides a more complete approach for lowering to X86ISD::PACKSS/PACKUS nodes - testing for existing suitable sign/zero extension before recreating it.
It also optionally packs the upper half instead of the lower half.
Similar to what we've done for other ops, this patch widens VPTERNLOG to a 512-bit op for non-VLX targets.
Fixes regressions in D113192
Differential Revision: https://reviews.llvm.org/D113827
For AVX512 targets without VLX, we have to widen 128/256-bit vectors to 512-bits to use some specific AVX512 instructions (or some other instructions with predicates etc.).
I've pulled out the widening code from LowerFunnelShift into the helper function, so we can convert some other widening patterns in the future.
Add support for v32i8/v64i8 converting shift-by-constant to multiply-by-constant. This helps us avoid the generic vXi8 shift lowering, and a lot of VPBLENDVB ops which can be particularly slow.
We also needed to reorder a few shift lowering patterns to prevent regressions, particularly for XOP+AVX2 (Excavator) targets (which can split to fast v16i8 shifts) and AVX512-BWI targets (which prefers to extend to fast v32i16 shifts).
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
and (pcmpgt X, -1), Y --> pandn (vsrai X, BitWidth-1), Y
This avoids the -1 constant vector in favor of an arithmetic shift
instruction if it exists (the ISA is still not complete after all
these years...).
We catch this pattern late in combining by matching PCMPGT, so it
should not interfere with more general folds.
Differential Revision: https://reviews.llvm.org/D113603
This fixes the crash due to lacking VZEXT_MOVL support with i16.
Reviewed By: LuoYuanke, RKSimon
Differential Revision: https://reviews.llvm.org/D113661
combineMulToPMADDWD is currently limited to legal types, but there's no reason why we can't handle any larger type that the existing SplitOpsAndApply code can use to split to legal X86ISD::VPMADDWD ops.
This also exposed a missed opportunity for pre-SSE41 targets to handle SEXT ops from types smaller than vXi16 - without PMOVSX instructions these will always be expanded to unpack+shifts, so we can cheat and convert this into a ZEXT(SEXT()) sequence to make it a valid PMADDWD op.
Differential Revision: https://reviews.llvm.org/D110995
Use ComputeMinSignedBits() to ensure the mul source operands at least sign-extend down from the bottom 16 bits.
This will make it easier if/when we try to support handling of source types larger than 32-bits.
As suggested on D113371, this adds a wrapper to SelectionDAG::ComputeNumSignBits, similar to the llvm::ComputeMinSignedBits wrapper.
I've included some usage, its not exhaustive, just the more obvious cases where the intention is obvious.
Differential Revision: https://reviews.llvm.org/D113396
We have several places where we need to extract the raw bits data from a BUILD_VECTOR node, so consolidate this to a single helper function that handles Undefs and Integer/FP constants, including implicit truncation.
This should make it easier to extend D113202 to handle more constant folding of bitcasted constant data.
Differential Revision: https://reviews.llvm.org/D113351
For v8i16 shuffle patterns that are lowered with AND+PACKUS, check to see if the sources are from a 256-bit vector and perform the masking using BLENDW at the 256-bit level.
With the test changes we can see more examples of duplicate XMM/YMM zero vectors (PR26018) :(
The check for whether a zero extension was needed was subtly wrong and
saw a value that was already 64 bits, so did not extend.
Fixes PR52357.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D112860
The changes in D80163 defered the assignment of MachineMemOperand (MMO)
until the X86ExpandPseudo pass. This will result in crash due to prolog
insert point been sunk across the pseudo instruction VASTART_SAVE_XMM_REGS.
Moving the assignment to the creation of the node can avoid the problem.
Reviewed By: rnk
Differential Revision: https://reviews.llvm.org/D112859
This is part of rG1cfecf4fc427 that was reverted to fix PR51226 - concating the broadcasts is OK, its the splatted loads that crash (we're not detecting extloads). I'm still creating a reduced test case so haven't added the load handling again yet.
The VINSERTF128 instruction is often much quicker, and never slower, than the more general VPERM2F128 instruction, so we should try to use that in more circumstances.
This requires a fallback to a commuted VPERM2F128 for the case where we need to fold the 256-bit vector source instead of the 128-bit subvector source.
There is one interesting side effect - DAGCombine's narrowExtractedVectorLoad combine gets called in a number of locations, this often creates an extracted subvector load without regard to other uses of the original wider load. I'm expecting AVX cpus to be capable of merging such aliased loads, but I do wonder whether narrowExtractedVectorLoad's call to X86TargetLowering::shouldReduceLoadWidth needs to be altered to check for more partial uses?
Noticed while investigating the quality of interleaved load/store codegen.
Differential Revision: https://reviews.llvm.org/D111960
As mentioned on D108539, when the gather indices are smaller than the pointer size, they are sign-extended BEFORE scale is applied, making the general fold unsafe.
If the index have sufficient sign-bits then folding the scale could be safe - I'll investigate this.
If the index operand for a gather/scatter intrinsic is being scaled (self-addition or a shl-by-immediate) then we may be able to fold that scaling into the intrinsic scale immediate value instead.
Fixes PR13310.
Differential Revision: https://reviews.llvm.org/D108539
In the instruction encoding, the passthru register is always
tied to the destination register. The CPU scheduler has to wait
for the last writer of this register to finish executing before
the gather can start. This is true even if the initial mask is
all ones so that the passthru will never be used.
By explicitly zeroing the register we can break the false
dependency. The zero idiom is executed completing by the
register renamer and so is immedately considered ready.
Authored by Craig.
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D112505
As noted in D112464, a pre-AVX target may not be able to fold an
under-aligned vector load into another op, so we shouldn't report
that as a load folding candidate. I only found one caller where
this would make a difference -- combineCommutableSHUFP() -- so
that's where I added a test to show the (minor) regression.
Differential Revision: https://reviews.llvm.org/D112545
This can avoid a vector add and a constant pool load. Or an explicit broadcast in case of non-constant.
Also reverse the transform any time we encounter a constant index addend that can't be moved to base. In that case pull the constant from base into the index. This reduces code size needed for the displacement since we needed the index add anyway. Limit this to scale of 1 to avoid divisibility and wrap issues.
Authored by Craig.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D111595
This is a small extension of D112095 to avoid another regression
seen with D112085.
In this case, we allow the same conversion from usubsat to ALU
ops if the target supports vpternlog.
That pattern will get converted later in X86DAGToDAGISel::tryVPTERNLOG().
This seems better than putting a magic immediate constant directly in
this code to create the exact vpternlog that we need. It's possible that
there are other special-cases along these lines, so we should try to
keep all of the vpternlog magic in one place.
Differential Revision: https://reviews.llvm.org/D112138
usubsat X, SMIN --> (X ^ SMIN) & (X s>> BW-1)
This would be a regression with D112085 where we combine to
usubsat more aggressively, so avoid that by matching the
special-case where we are subtracting SMIN (signmask):
https://alive2.llvm.org/ce/z/4_3gBD
Differential Revision: https://reviews.llvm.org/D112095
Inspired by D111968, provide a isNegatedPowerOf2() wrapper instead of obfuscating code with (-Value).isPowerOf2() patterns, which I'm sure are likely avenues for typos.....
Differential Revision: https://reviews.llvm.org/D111998
If we're using an ashr to sign-extend the entire upper 16 bits of the i32 element, then we can replace with a lshr. The sign bit will be correctly shifted for PMADDWD's implicit sign-extension and the upper 16 bits are zero so the upper i16 sext-multiply is guaranteed to be zero.
The lshr also has a better chance of folding with shuffles etc.
This patch detects the absolute value pattern on the RHS of a
subtract. If we find it we swap the CMOV true/false values and
replace the subtract with an ADD.
There may be a more generic way to do this, but I'm not sure.
Targets that don't have legal or custom ISD::ABS use a generic
expand in DAG combiner already when it sees (neg (abs(x))). I
haven't checked what happens if the neg is a more general subtract.
Fixes PR50991 for X86.
Reviewed By: RKSimon, spatel
Differential Revision: https://reviews.llvm.org/D111858
CMOVGE reads SF and OF. CMOVNS only reads SF. This matches with
other recent changes to use a single flag where possible. It also
matches gcc codegen.
I believe this technically changes whether the conditioanl move happens
on INT_MIN, but for INT_MIN both registers are the same so it doesn't
matter.
Differential Revision: https://reviews.llvm.org/D111826
As noted in https://github.com/halide/Halide/pull/6302,
we hilariously fail to match PAVG if we even as much
as look at it the wrong way.
In this particular case, the problem stems from the fact that
`PAVG` root (def) is a `trunc`, and leafs (uses) are `zext`'s,
and InstCombine really loves to get rid of both of these,
for example replace them with a bit mask. So we may not have
said `zext`.
Instead of checking for that + type match,
i think we should rely on the actual active type,
as per the knownbits.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D111571
The select(pshufb,pshufb) -> or(pshufb,pshufb) fold uses getConstVector to create the refreshed pshufb masks, which treats all negative indices as undef.
PR52122 shows that if we were selecting an element that the PSHUFB has set to zero we must set it back to 0x80 when we recreate the PSHUFB mask and not just leave it as SM_SentinelZero
Without popcnt we had a special case for using the parity flag from a single test i8 test instruction if only bits 7:0 could be non-zero. That special case is still useful when we have popcnt.
To reach this special case, we enable custom lowering of parity for i16/i32/i64 even when popcnt is enabled. The check for POPCNT being enabled is now after the special case in LowerPARITY.
Fixes PR52093
Differential Revision: https://reviews.llvm.org/D111249
The X86 backend only needs to know whether structure return is via an
sret pointer. This removes the categorization enumeration and
adjusts, templatizes and renames the related functions.
Differential Revision: https://reviews.llvm.org/D109966
Noticed while investigating the regressions in D110995 - if the RHS element is already zero, then we don't need the corresponding LHS element.
Technically we could also recheck RHS once we have LHS's known zeros, but I haven't seen any missed opportunities from that yet.
Stop using APInt constructors and methods that were soft-deprecated in
D109483. This fixes all the uses I found in llvm, except for the APInt
unit tests which should still test the deprecated methods.
Differential Revision: https://reviews.llvm.org/D110807
PR52040 identified several issues with the HOP(HOP'(X,X),HOP'(Y,Y)) -> HOP(PERMUTE(HOP'(X,Y)),PERMUTE(HOP'(X,Y)) slow-HOP fold.
Not only was there a copy+paste typo when accessing the inner HOP operands, but the (unnecessary) ReplaceAllUsesOfValueWith call was missing one use checks.
Now that we have better shuffle combines of HOPs we can just return a new HOP() sequence and not use ReplaceAllUsesOfValueWith at all - this actually improved pair_sum_v8i32_v4i32 codegen as it kicks off further shuffle combines.
X86's decomposeMulByConstant never permits mul decomposition to shift+add/sub if the vector multiply is legal.
Unfortunately this isn't great for SSE41+ targets which have PMULLD for vXi32 multiplies, but is often quite slow. This patch proposes to allow decomposition if the target has the SlowPMULLD flag (i.e. Silvermont). We also always decompose legal vXi64 multiplies - even latest IceLake has really poor latencies for PMULLQ.
Differential Revision: https://reviews.llvm.org/D110588
On Windows, i128 arguments are passed as indirect arguments, and
they are returned in xmm0.
This is mostly fixed up by `WinX86_64ABIInfo::classify` in Clang, making
the IR functions return v2i64 instead of i128, and making the arguments
indirect. However for cases where libcalls are generated in the target
lowering, the lowering uses the default x86_64 calling convention for
i128, where they are passed/returned as a register pair.
Add custom lowering logic, similar to the existing logic for i128
div/mod (added in 4a406d32e9),
manually making the libcall (while overriding the return type to
v2i64 or passing the arguments as pointers to arguments on the stack).
X86CallingConv.td doesn't seem to handle i128 at all, otherwise
the windows specific behaviours would ideally be implemented as
overrides there, in generic code, handling these cases automatically.
This fixes https://bugs.llvm.org/show_bug.cgi?id=48940.
Differential Revision: https://reviews.llvm.org/D110413
This intention of this code turns out to be superfluous as we can handle this with shuffle combining, and it has a critical flaw in that it doesn't check for dependencies.
Fixes PR51974
When AVX512FP16 is enabled, FROUND(f16) cannot be dealt with
TypeLegalize, and no libcall in libm is ready for fround(f16) now.
FROUNDEVEN(f16) has related instruction in AVX512FP16.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D110312
The plan is to allow combineMulToPMADDWD to match illegal vector types (as long as they're still pow2), which should allow us to start removing the 128-bit limit on more of the PMADDWD combines.
Merge addition of VPMADDWD nodes if each element pair doesn't use the upper element in each pair (i.e. its zero) - we can generalize this to either element in the pair if we one day create VPMADDWD with zero lower elements.
There are still a number of issues with extending/shuffling with 256/512-bit VPMADDWD nodes so this initially only works for v2i32/v4i32 cases - I'm working on removing all these limitations but there's still a bit of yak shaving to go.....
If we are multiplying by a sign-extended vXi32 constant, then we can mask off the upper 16 bits to allow folding to PMADDWD and make use of its implicit sign-extension from i16
We already do this on SSE41 targets where we have sext/zext instructions, now that combineShiftToPMULH handles SSE2 targets, we can enable this here as well.
With improved shuffle combines (in particular canonicalizeShuffleWithBinOps), we can now usefully perform this on any SSE2+ target.
We should be able to remove this entirely and just use DAGCombiner's combineShiftToMULH if we can someday get it to support illegal (pre-widened) types.
As suggested on D108522, if we're sign extending a v4i16 source before multiplying as a v4i32, then we can replace that with a zero extension and rely on the implicit sign-extension of PMADDWD.
This is motivated by the examples and discussion in:
https://llvm.org/PR51245
...and related bugs.
By using vector compares and vector logic, we can convert 2 'set'
instructions into 1 'movd' or 'movmsk' and generally improve
throughput/reduce instructions.
Unfortunately, we don't have a complete vector compare ISA before
AVX, so I left SSE-only out of this patch. Ie, we'd need extra logic
ops to simulate the missing predicates for SSE 'cmpp*', so it's not
as clearly a win.
Differential Revision: https://reviews.llvm.org/D110342
This function can be adapted to solve bugs like PR51245,
but it could require differentiating the combiner timing
between the existing and new transforms.
This patch is to support transform something like
_mm512_add_ph(acc, _mm512_fmadd_pch(a, b, _mm512_setzero_ph()))
to _mm512_fmadd_pch(a, b, acc).
Differential Revision: https://reviews.llvm.org/D109953
This change adds the ASan intrinsic to the list whihc are setting hasCopyImplyingStackAdjustment.
Reviewed By: eugenis
Differential Revision: https://reviews.llvm.org/D110012
For x86 Darwin, we have a stack checking feature which re-uses some of this
machinery around stack probing on Windows. Renaming this to be more appropriate
for a generic feature.
Differential Revision: https://reviews.llvm.org/D109993
We can combine unary shuffles into either of SHUFPS's inputs and adjust the shuffle mask accordingly.
Unlike general shuffle combining, we can be more aggressive and handle multiuse cases as we're not going to accidentally create additional shuffles.
Apparently this has no test coverage before D108382,
but D108382 itself shows a few regressions that this fixes.
It doesn't seem worthwhile breaking apart broadcasts,
assuming we want the broadcasted value to be preset in several elements,
not just the 0'th one.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D108411
Split off from D108253.
Broadcast is simpler than any other shuffle we might produce
to do what we want to do here, so prefer it.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D108382
We can use `OR` instead of `BLEND` if either the element we are not picking is zero (or masked away);
or the element we are picking overwhelms (e.g. it's all-ones) whatever the element we are not picking:
https://alive2.llvm.org/ce/z/RKejao
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D109726
When searching for hidden identity shuffles (added at rG41146bfe82aecc79961c3de898cda02998172e4b), only peek through bitcasts to the source operand if it is a vector type as well.
Soft deprecrate isNullValue/isAllOnesValue and update in tree
callers. This matches the changes to the APInt interface from
D109483.
Reviewed By: lattner
Differential Revision: https://reviews.llvm.org/D109535
This library function only exists in compiler-rt not libgcc. So
this would fail to link unless we were linking with compiler-rt.
This is consistent with the recent removal of calls to mulodi4 on
32-bit targets like D108928.
I suppose maybe we could keep the libcalls for platforms like
Darwin that use compiler-rt exclusively?
Reviewed By: nickdesaulniers, MaskRay
Differential Revision: https://reviews.llvm.org/D109385
This renames the primary methods for creating a zero value to `getZero`
instead of `getNullValue` and renames predicates like `isAllOnesValue`
to simply `isAllOnes`. This achieves two things:
1) This starts standardizing predicates across the LLVM codebase,
following (in this case) ConstantInt. The word "Value" doesn't
convey anything of merit, and is missing in some of the other things.
2) Calling an integer "null" doesn't make any sense. The original sin
here is mine and I've regretted it for years. This moves us to calling
it "zero" instead, which is correct!
APInt is widely used and I don't think anyone is keen to take massive source
breakage on anything so core, at least not all in one go. As such, this
doesn't actually delete any entrypoints, it "soft deprecates" them with a
comment.
Included in this patch are changes to a bunch of the codebase, but there are
more. We should normalize SelectionDAG and other APIs as well, which would
make the API change more mechanical.
Differential Revision: https://reviews.llvm.org/D109483
integer 0/1 for the operand of bundle "clang.arc.attachedcall"
https://reviews.llvm.org/D102996 changes the operand of bundle
"clang.arc.attachedcall". This patch makes changes to llvm that are
needed to handle the new IR.
This should make it easier to understand what the IR is doing and also
simplify some of the passes as they no longer have to translate the
integer values to the runtime functions.
Differential Revision: https://reviews.llvm.org/D103000
Similar to D108842, D108844, and D108926.
__has_builtin(builtin_mul_overflow) returns true for 32b x86 targets,
but Clang is deferring to compiler RT when encountering long long types.
This breaks ARCH=i386 + CONFIG_BLK_DEV_NBD=y builds of the Linux kernel
that are using builtin_mul_overflow with these types for these targets.
If the semantics of __has_builtin mean "the compiler resolves these,
always" then we shouldn't conditionally emit a libcall.
This will still need to be worked around in the Linux kernel in order to
continue to support these builds of the Linux kernel for this
target with older releases of clang.
Link: https://bugs.llvm.org/show_bug.cgi?id=28629
Link: https://bugs.llvm.org/show_bug.cgi?id=35922
Link: https://github.com/ClangBuiltLinux/linux/issues/1438
Reviewed By: lebedev.ri, RKSimon
Differential Revision: https://reviews.llvm.org/D108928
PMADDWD(v8i16 x, v8i16 y) == (v4i32) { (int)x[0]*y[0] + (int)x[1]*y[1], ..., (int)x[6]*y[6] + (int)x[7]*y[7] }
Currently combineMulToPMADDWD only folds cases where the upper 17 bits of both vXi32 inputs are known zero (i.e. the first half is positive and the second half of the pair is zero in each 2xi16 pair), this can be relaxed to only require one zero-extended input if the other input has at least 17 sign bits.
That way the sign of the result is still preserved, and the second half is still zero.
Noticed while investigating PR47437.
Differential Revision: https://reviews.llvm.org/D108522
Please refer to
https://lists.llvm.org/pipermail/llvm-dev/2021-September/152440.html
(and that whole thread.)
TLDR: the original patch had no prior RFC, yet it had some changes that
really need a proper RFC discussion. It won't be productive to discuss
such an RFC, once it's actually posted, while said patch is already
committed, because that introduces bias towards already-committed stuff,
and the tree is potentially in broken state meanwhile.
While the end result of discussion may lead back to the current design,
it may also not lead to the current design.
Therefore i take it upon myself
to revert the tree back to last known good state.
This reverts commit 4c4093e6e3.
This reverts commit 0a2b1ba33a.
This reverts commit d9873711cb.
This reverts commit 791006fb8c.
This reverts commit c22b64ef66.
This reverts commit 72ebcd3198.
This reverts commit 5fa6039a5f.
This reverts commit 9efda541bf.
This reverts commit 94d3ff09cf.
This fixes another reproducer from https://bugs.llvm.org/show_bug.cgi?id=51615
And again, the fix lies not in the code added in D105390
In this case, we completely don't check that the "broadcast-from-mem" we create
can actually fold the load. In this case, it's operand was not a load at all:
```
Combining: t16: v8i32 = vector_shuffle<0,u,u,u,0,u,u,u> t14, undef:v8i32
Creating new node: t29: i32 = undef
RepeatLoad:
t8: i32 = truncate t7
t7: i64 = extract_vector_elt t5, Constant:i64<0>
t5: v2i64,ch = load<(load (s128) from %ir.arg)> t0, t2, undef:i64
t2: i64,ch = CopyFromReg t0, Register:i64 %0
t1: i64 = Register %0
t4: i64 = undef
t3: i64 = Constant<0>
Combining: t15: v8i32 = undef
```
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D108821
x86_fp80 format allows values that do not fit any of IEEE-754 category.
Previously they were recognized by intrinsic __builtin_isnan as NaNs.
Now this intrinsic is implemented using instruction FXAM, which
distinguish between NaNs and unsupported values. It can make some
programs behave differently.
As a solution, this fix changes lowering of the intrinsic. If floating
point exceptions are ignored, llvm.isnan is lowered into unordered
comparison, as __buildtin_isnan was implemented earlier. In strictfp
functions the intrinsic is lowered using FXAM, which does not raise
exceptions even for signaling NaN, as required by IEEE-754 and C
standards.
Differential Revision: https://reviews.llvm.org/D108037
There are no paths into LowerFormalParms that have already specified
the sret register. We always materialize a virtual and then assign it
to the physical reg at the point of the return.
Differential Revision: https://reviews.llvm.org/D108762
Even though https://bugs.llvm.org/show_bug.cgi?id=51615
appears to be introduced by D105390, the fix lies here.
We can not replace a wide volatile load with a broadcast-from-memory,
because that would narrow the load, which isn't legal for volatiles.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D108757
In-register structure returns are not special, and handled by lowering
to multiple-value tuples. We can tail-call from non-sret fns to
structure-returning functions, except on i686 where the sret pointer
is callee-pop.
Differential Revision: https://reviews.llvm.org/D105807
We don't have any vXi8 shift instructions (other than on XOP which is handled separately), so replace the shl(x,1) -> add(x,x) fold with shl(x,1) -> add(freeze(x),freeze(x)) to avoid the undef issues identified in PR50468.
Split off from D106675 as I'm still looking at whether we can fix the vXi16/i32/i64 issues with the D106679 alternative.
Differential Revision: https://reviews.llvm.org/D108139
Extend matchShuffleAsBlend to not only match against known in-place elements for BLEND shuffles, but use isElementEquivalent to determine if the shuffle mask's referenced element is the same as the in-place element.
This allows us to replace a number of insertps instructions with more general blendps instructions (better opportunities for commutation, concatenation etc.).
Before lowering shuffles, see if we can merge horizontal ops or canonicalize the shuffle mask to point to the same LHS/RHS of the HOps when an HOp's args are repeated.
This changes the lowering of saddsat and ssubsat so that instead of
using:
r,o = saddo x, y
c = setcc r < 0
s = c ? INTMAX : INTMIN
ret o ? s : r
into using asr and xor to materialize the INTMAX/INTMIN constants:
r,o = saddo x, y
s = ashr r, BW-1
x = xor s, INTMIN
ret o ? x : r
https://alive2.llvm.org/ce/z/TYufgD
This seems to reduce the instruction count in most testcases across most
architectures. X86 has some custom lowering added to compensate for
cases where it can increase instruction count.
Differential Revision: https://reviews.llvm.org/D105853
AttributeList::hasAttribute() is confusing, use clearer methods like
hasParamAttr()/hasRetAttr().
Add hasRetAttr() since it was missing from AttributeList.
The IR for pmuldq/pmuludq intrinsics uses a sext_inreg/zext_inreg
pattern on the inputs. Ideally we pattern match these away during
isel. It is possible for LICM or other middle end optimizations
to separate the extend from the mul. This prevents SelectionDAG
from removing it or depending on how the extend is lowered, we
may not be able to generate an AssertSExt/AssertZExt in the
mul basic block. This will prevent pmuldq/pmuludq from being
formed at all.
This patch teaches shouldSinkOperands to recognize this so
that CodeGenPrepare will clone the extend into the same basic
block as the mul.
Fixes PR51371.
Differential Revision: https://reviews.llvm.org/D107689
This isn't optimal, but prevents crashing when the libcall isn't
available. It just calculates the full product and makes sure the high bits
match the sign of the low half. Each of the pieces should go through their own
type legalization.
This can make D107420 unnecessary.
Needs tests, but I wanted to start discussion about D107420.
Reviewed By: FreddyYe
Differential Revision: https://reviews.llvm.org/D107581
As reported on PR51281, an internal fuzz test encountered an issue when extracting constant bits from a SUBV_BROADCAST node from a constant pool source larger than the broadcasted subvector width.
The getTargetConstantBitsFromNode was assuming that the Constant would the same size as the subvector, resulting in the incorrect packing of the per-element bits data.
This patch attempts to solve this by using the SUBV_BROADCAST node to determine the subvector width, and then ensuring we extract only the lowest bits from Constant of that subvector bitsize.
Differential Revision: https://reviews.llvm.org/D107158
This is recommit of the patch 16ff91ebcc,
reverted in 0c28a7c990 because it had
an error in call of getFastMathFlags (base type should be FPMathOperator
but not Instruction). The original commit message is duplicated below:
Clang has builtin function '__builtin_isnan', which implements C
library function 'isnan'. This function now is implemented entirely in
clang codegen, which expands the function into set of IR operations.
There are three mechanisms by which the expansion can be made.
* The most common mechanism is using an unordered comparison made by
instruction 'fcmp uno'. This simple solution is target-independent
and works well in most cases. It however is not suitable if floating
point exceptions are tracked. Corresponding IEEE 754 operation and C
function must never raise FP exception, even if the argument is a
signaling NaN. Compare instructions usually does not have such
property, they raise 'invalid' exception in such case. So this
mechanism is unsuitable when exception behavior is strict. In
particular it could result in unexpected trapping if argument is SNaN.
* Another solution was implemented in https://reviews.llvm.org/D95948.
It is used in the cases when raising FP exceptions by 'isnan' is not
allowed. This solution implements 'isnan' using integer operations.
It solves the problem of exceptions, but offers one solution for all
targets, however some can do the check in more efficient way.
* Solution implemented by https://reviews.llvm.org/D96568 introduced a
hook 'clang::TargetCodeGenInfo::testFPKind', which injects target
specific code into IR. Now only SystemZ implements this hook and it
generates a call to target specific intrinsic function.
Although these mechanisms allow to implement 'isnan' with enough
efficiency, expanding 'isnan' in clang has drawbacks:
* The operation 'isnan' is hidden behind generic integer operations or
target-specific intrinsics. It complicates analysis and can prevent
some optimizations.
* IR can be created by tools other than clang, in this case treatment
of 'isnan' has to be duplicated in that tool.
Another issue with the current implementation of 'isnan' comes from the
use of options '-ffast-math' or '-fno-honor-nans'. If such option is
specified, 'fcmp uno' may be optimized to 'false'. It is valid
optimization in general, but it results in 'isnan' always returning
'false'. For example, in some libc++ implementations the following code
returns 'false':
std::isnan(std::numeric_limits<float>::quiet_NaN())
The options '-ffast-math' and '-fno-honor-nans' imply that FP operation
operands are never NaNs. This assumption however should not be applied
to the functions that check FP number properties, including 'isnan'. If
such function returns expected result instead of actually making
checks, it becomes useless in many cases. The option '-ffast-math' is
often used for performance critical code, as it can speed up execution
by the expense of manual treatment of corner cases. If 'isnan' returns
assumed result, a user cannot use it in the manual treatment of NaNs
and has to invent replacements, like making the check using integer
operations. There is a discussion in https://reviews.llvm.org/D18513#387418,
which also expresses the opinion, that limitations imposed by
'-ffast-math' should be applied only to 'math' functions but not to
'tests'.
To overcome these drawbacks, this change introduces a new IR intrinsic
function 'llvm.isnan', which realizes the check as specified by IEEE-754
and C standards in target-agnostic way. During IR transformations it
does not undergo undesirable optimizations. It reaches instruction
selection, where is lowered in target-dependent way. The lowering can
vary depending on options like '-ffast-math' or '-ffp-model' so the
resulting code satisfies requested semantics.
Differential Revision: https://reviews.llvm.org/D104854
Having `NewMask` outside of an if and rebinding `BaseMask` `ArrayRef`
to it is confusing. Instead, just move the `Mask` vector higher up,
and change the code that earlier had no access to it but now does
to use `Mask` instead of `BaseMask`.
This has no other intentional changes.
This is a recommit of 35c0848b57,
that was reverted to simplify reversion of an earlier change.
IR typically creates INSERT_SUBVECTOR patterns as a widening of the subvector with undefs to pad to the destination size, followed by a shuffle for the actual insertion - SelectionDAGBuilder has to do something similar for shuffles when source/destination vectors are different sizes.
This combine attempts to recognize these patterns by looking for a shuffle of a subvector (from a CONCAT_VECTORS) that starts at a modulo of its size into an otherwise identity shuffle of the base vector.
This uncovered a couple of target-specific issues as we haven't often created INSERT_SUBVECTOR nodes in generic code - aarch64 could only handle insertions into the bottom of undefs (i.e. a vector widening), and x86-avx512 vXi1 insertion wasn't keeping track of undef elements in the base vector.
Fixes PR50053
Differential Revision: https://reviews.llvm.org/D107068
Having `NewMask` outside of an if and rebinding `BaseMask` `ArrayRef`
to it is confusing. Instead, just move the `Mask` vector higher up,
and change the code that earlier had no access to it but now does
to use `Mask` instead of `BaseMask`.
This has no other intentional changes.
I want to hoist `Mask` variable higher up,
but then it would clash with this one.
So let's rename this one first.
There are no other intentional changes here other than said rename.
Given a shuffle mask, if it is picking from an input that is splat
given the current granularity of the shuffle, then adjust the mask
to pick from the same lane of the input as the mask element is in.
This may result in a shuffle being simplified into a blend.
I believe this is correct given that the splat detection matches the one
just above the new code,
My basic thought is that we might be able to get less regressions
by handling multiple insertions of the same value into a vector
if we form broadcasts+blend here, as opposed to D105390,
but i have not really thought this through,
and did not try implementing it yet.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D107009
Clang has builtin function '__builtin_isnan', which implements C
library function 'isnan'. This function now is implemented entirely in
clang codegen, which expands the function into set of IR operations.
There are three mechanisms by which the expansion can be made.
* The most common mechanism is using an unordered comparison made by
instruction 'fcmp uno'. This simple solution is target-independent
and works well in most cases. It however is not suitable if floating
point exceptions are tracked. Corresponding IEEE 754 operation and C
function must never raise FP exception, even if the argument is a
signaling NaN. Compare instructions usually does not have such
property, they raise 'invalid' exception in such case. So this
mechanism is unsuitable when exception behavior is strict. In
particular it could result in unexpected trapping if argument is SNaN.
* Another solution was implemented in https://reviews.llvm.org/D95948.
It is used in the cases when raising FP exceptions by 'isnan' is not
allowed. This solution implements 'isnan' using integer operations.
It solves the problem of exceptions, but offers one solution for all
targets, however some can do the check in more efficient way.
* Solution implemented by https://reviews.llvm.org/D96568 introduced a
hook 'clang::TargetCodeGenInfo::testFPKind', which injects target
specific code into IR. Now only SystemZ implements this hook and it
generates a call to target specific intrinsic function.
Although these mechanisms allow to implement 'isnan' with enough
efficiency, expanding 'isnan' in clang has drawbacks:
* The operation 'isnan' is hidden behind generic integer operations or
target-specific intrinsics. It complicates analysis and can prevent
some optimizations.
* IR can be created by tools other than clang, in this case treatment
of 'isnan' has to be duplicated in that tool.
Another issue with the current implementation of 'isnan' comes from the
use of options '-ffast-math' or '-fno-honor-nans'. If such option is
specified, 'fcmp uno' may be optimized to 'false'. It is valid
optimization in general, but it results in 'isnan' always returning
'false'. For example, in some libc++ implementations the following code
returns 'false':
std::isnan(std::numeric_limits<float>::quiet_NaN())
The options '-ffast-math' and '-fno-honor-nans' imply that FP operation
operands are never NaNs. This assumption however should not be applied
to the functions that check FP number properties, including 'isnan'. If
such function returns expected result instead of actually making
checks, it becomes useless in many cases. The option '-ffast-math' is
often used for performance critical code, as it can speed up execution
by the expense of manual treatment of corner cases. If 'isnan' returns
assumed result, a user cannot use it in the manual treatment of NaNs
and has to invent replacements, like making the check using integer
operations. There is a discussion in https://reviews.llvm.org/D18513#387418,
which also expresses the opinion, that limitations imposed by
'-ffast-math' should be applied only to 'math' functions but not to
'tests'.
To overcome these drawbacks, this change introduces a new IR intrinsic
function 'llvm.isnan', which realizes the check as specified by IEEE-754
and C standards in target-agnostic way. During IR transformations it
does not undergo undesirable optimizations. It reaches instruction
selection, where is lowered in target-dependent way. The lowering can
vary depending on options like '-ffast-math' or '-ffp-model' so the
resulting code satisfies requested semantics.
Differential Revision: https://reviews.llvm.org/D104854
In this episode, we are trying to avoid an x86 micro-arch quirk where complex
(3 operand) LEA potentially costs significantly more than simple LEA. So we
simultaneously push and pull the math around the CMOV to balance the operations.
I looked at the debug spew during instruction selection and decided against
trying a later DAGToDAG transform -- it seems very difficult to match if the
trailing memops are already selected and managing the creation of extra
instructions at that level is always tricky.
Differential Revision: https://reviews.llvm.org/D106918
This reverts commit 1cfecf4fc4.
This commit broke LLVM code generated through XLA by removing a
conditional on Ld->getExtensionType() == ISD::NON_EXTLOAD
This is not a perfect revert. The new function is left as other uses of
it exist now.
This reverts commit 1cfecf4fc4.
This commit broke LLVM code generated through XLA by removing a
conditional on Ld->getExtensionType() == ISD::NON_EXTLOAD
Begin replacing individual getMemIntrinsicNode calls and setup (for X86ISD::VBROADCAST_LOAD + X86ISD::SUBV_BROADCAST_LOAD opcodes) with this getBROADCAST_LOAD helper.
If the rotation amount is a known splat, perform the modulo on the splat source, and then perform the splat. That way the amount-extension performed later by LowerScalarVariableShift can fold the splats away without any multiple-use issues.
Fixes one of the concerns raised on D104156
This is a minimum extension of D106607 to allow folding for
2 non-zero constantsi that can be materialized as immediates..
In the reduced test examples, we save 1 instruction by rolling
the constants into LEA/ADD. In the motivating test from the bullet
benchmark, we absorb both of the constant moves into add ops via
LEA magic, so we reduce by 2 instructions.
Differential Revision: https://reviews.llvm.org/D106684
As noticed on D105390 - we were hardwiring the depth limit for combining to VPERMI2W/VPERMI2B instructions. Not only had we made the limit too low, we hadn't accounted for slow/fast shuffles via the VariableCrossLaneShuffleDepth control
This is not the transform direction we want in general,
but by the time we have a CMOV, we've already tried
everything else that could be better.
The transform increases the uses of the other add operand,
but that is safe according to Alive2:
https://alive2.llvm.org/ce/z/Yn6p-A
We could probably extend this to other binops (not just add).
This is the motivating pattern discussed in:
https://llvm.org/PR51069
The test with i8 shows a missed fold because there's a trunc
sitting in front of the add. That can be handled with a small
follow-up.
Differential Revision: https://reviews.llvm.org/D106607
This reverts commit 2a419a0b99.
The result of a shufflevector must not propagate poison from any element
other than the one noted in the shuffle mask.
The regressions outside of fptoui-may-overflow.ll can probably be
recovered some other way; for example, using isGuaranteedNotToBePoison.
See discussion on https://reviews.llvm.org/D106053 for more background.
Differential Revision: https://reviews.llvm.org/D106222
D106053 exposed that we've not been taking into account that by bitcasting smaller elements together and then performing a ComputeKnownBits on the result we'd be allowing a poison element to influence other neighbouring elements being used in the pack. Instead we now peek through any existing bitcast to ensure that the source type already matches the width source of the pack node we're trying to match.
This has also been a chance to stop matchShuffleWithPACK creating unused nodes on the fly which could affect oneuse tests during shuffle lowering/combining.
The only regression we're seeing is due to being unable to peek through a bitcast as its on the other side of a extract_subvector - which should go away once we finally allow shuffle combining across different vector widths (by making matchShuffleWithPACK using const SelectionDAG& we've gotten closer to this - see PR45974).
Fixes issue reported on D105827 where a single shuffle of a constant (with multiple uses) was caught in an infinite loop where one shuffle (UNPCKL) used an undef arg but then that got recombined to SHUFPS as the constant value had its own undef that confused matching.....
We know that "CVTTPS2SI" returns 0x80000000 for out of range inputs (and for FP_TO_UINT, negative float values are undefined). We can use this to make unsigned conversions from vXf32 to vXi32 more efficient, particularly on targets without blend using the following logic:
small := CVTTPS2SI(x);
fp_to_ui(x) := small | (CVTTPS2SI(x - 2^31) & ARITHMETIC_RIGHT_SHIFT(small, 31))
Even on targets where "PBLENDVPS"/"PBLENDVB" exists, it is often a latency 2, low throughput instruction so this logic is applied there too (in particular for AVX2 also). It furthermore gets rid of one high latency floating point comparison in the previous lowering.
@TomHender checked the correctness of this for all possible floats between -1 and 2^32 (both ends excluded).
Original Patch by @TomHender (Tom Hender)
Differential Revision: https://reviews.llvm.org/D89697
Annoyingly, i686 cmpsd handling still fails to remove the unnecessary neg(and(x,1))
Reapplied rGe4aa6ad13216 with fix for intrinsic variants of the opcode which uses a vector return type
Its proving tricky to move this to the generic legalizer code, so manually insert the v2i32 subvector into v4i32, insert the AssertSext/AssertZext node, then extract the subvector again.
This avoids masks in the truncation/pack code, which means we avoid a PSHUFB in the fp_to_sint/uint code for sub-128 bit types (specific targets can still combine the packs to a pshufb if they have fast variable per-lane shuffles).
This was noticed when I was trying to improve fp_to_sint/uint costs with D103695 (and some targets had very high fp_to_sint costs due to the PSHUFB), so we can then update the fp_to_uint codegen from D89697.
This patch fixes PR50823.
The shuffle mask should be twisted twice before gotten the correct one due to the difference between inner HOP and outer.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D104903
This demonstrates a possible fix for PR48760 - for compares with constants, canonicalize the SGT/UGT condition code to use SGE/UGE which should reduce the number of EFLAGs bits we need to read.
As discussed on PR48760, some EFLAG bits are treated independently which can require additional uops to merge together for certain CMOVcc/SETcc/etc. modes.
I've limited this to cases where the constant increment doesn't result in a larger encoding or additional i64 constant materializations.
Differential Revision: https://reviews.llvm.org/D101074
Don't allow vectors to split into GPRs for 'r' and other scalar
constraints. Prevents assertion in getCopyToPartsVector.
Makes PR50907 give a better error instead of crashing.
This is a mechanical change. This actually also renames the
similarly named methods in the SmallString class, however these
methods don't seem to be used outside of the llvm subproject, so
this doesn't break building of the rest of the monorepo.
It looks like the fold introduced in 63f3383ece can cause crashes
if the type of the bitcasted value is not a valid vector element type,
like x86_mmx.
To resolve the crash, reject invalid vector element types. The way it is
done in the patch is a bit clunky. Perhaps there's a better way to
check?
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D104792
select (cmpeq Cond0, Cond1), LHS, (select (cmpugt Cond0, Cond1), LHS, Y) --> (select (cmpuge Cond0, Cond1), LHS, Y)
etc,
We already perform this fold in DAGCombiner for MVT::i1 comparison results, but these can still appear after legalization (in x86 case with MVT::i8 results), where we need to be more careful about generating new comparison codes.
Pulled out of D101074 to help address the remaining regressions.
Differential Revision: https://reviews.llvm.org/D104707
Since this method can apply to cmpxchg operations, make sure it's clear
what value we're actually retrieving. This will help ensure we don't
accidentally ignore the failure ordering of cmpxchg in the future.
We could potentially introduce a getOrdering() method on AtomicSDNode
that asserts the operation isn't cmpxchg, but not sure that's
worthwhile.
Differential Revision: https://reviews.llvm.org/D103338
As per the discussion in D103818, so far, this does not appear to be worthwhile.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D103818
Fixes:
- PR36507 Floating point varargs are not handled correctly with
-mno-implicit-float
- PR48528 __builtin_va_start assumes it can pass SSE registers
when using -Xclang -msoft-float -Xclang -no-implicit-float
On x86_64, floating-point parameters are normally passed in XMM
registers. For va_start, we spill those to memory so va_arg can
find them. There is an interaction here with -msoft-float and
-no-implicit-float:
When -msoft-float is in effect, instead of passing floating-point
parameters in XMM registers, they are passed in general-purpose
registers.
When -no-implicit-float is in effect, it "disables implicit
floating-point instructions" (per the LangRef). The intended
effect is to not have the compiler generate floating-point code
unless explicit floating-point operations are present in the
source code, but what exactly counts as an explicit floating-point
operation is not specified. The existing behavior of LLVM here has
led to some surprises and PRs.
This change modifies the behavior as follows:
| soft | no-implicit | old behavior | new behavior |
| no | no | spill XMM regs | spill XMM regs |
| yes | no | don't spill XMM | don't spill XMM |
| no | yes | don't spill XMM | spill XMM regs |
| yes | yes | assert | don't spill XMM |
In particular, this avoids the assert that happens when
-msoft-float and -no-implicit-float are both in effect. This
seems like a perfectly reasonable combination: If we don't want
to rely on hardware floating-point support, we want to both
avoid using float registers to pass parameters and avoid having
the compiler generate floating-point code that wasn't in the
original program. Instead of crashing the compiler, the new
behavior is to not synthesize floating-point code in this
case. This fixes PR48528.
The other interesting case is when -no-implicit-float is in
effect, but -msoft-float is not. In that case, any floating-point
parameters that are present will be in XMM registers, and so we
have to spill them to correctly handle those. This fixes
PR36507. The spill is conditional on %al indicating that
parameters are present in XMM registers, so no floating-point
code will be executed unless the function is called with
floating-point parameters.
Reviewed By: rnk
Differential Revision: https://reviews.llvm.org/D104001
Changing vector element type doesn't work for v6i32->v6i16 now
that v6i32 is an MVT and v6i16 is not.
I would like to fix this in changeVectorElementType, but you
need a LLVMContext to call getVectorVT which we can't get from
an MVT.
Fixes PR50709.
Fixes crash reported here https://reviews.llvm.org/D73607
Using a store to keep the trunc intact. Returning v16i24 would
cause the trunc to be optimized away in SelectionDAGBuilder.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D103940
Don't require a specific kind of IRBuilder for TargetLowering hooks.
This allows us to drop the IRBuilder.h include from TargetLowering.h.
Differential Revision: https://reviews.llvm.org/D103759
We were hitting an issue when the scalar_to_vector source was being implicitly truncated (in this case to i8 to vXi1) but we were also using the i8 source in a broadcast to a vXi8 value.
Fixes PR50374
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
SwiftTailCC has a different set of requirements than the C calling convention
for a tail call. The exact argument sequence doesn't have to match, but fewer
ABI-affecting attributes are allowed.
Also make sure the musttail diagnostic triggers if a musttail call isn't
actually a tail call.
We could previously do this by accident through the later
call to getTargetConstantBitsFromNode I think, but that only worked
if N0 had a single use. This patch makes it explicit for undef and
doesn't have a use count check.
I think this is needed to move the (shl X, 1)->(add X, X)
fold to isel for PR50468. We need to be sure X won't be IMPLICIT_DEF
which might prevent the same vreg from being used for both operands.
Differential Revision: https://reviews.llvm.org/D103192
D88631 added initial support for:
- -mstack-protector-guard=
- -mstack-protector-guard-reg=
- -mstack-protector-guard-offset=
flags, and D100919 extended these to AArch64. Unfortunately, these flags
aren't retained for LTO. Make them module attributes rather than
TargetOptions.
Link: https://github.com/ClangBuiltLinux/linux/issues/1378
Reviewed By: tejohnson
Differential Revision: https://reviews.llvm.org/D102742
This patch adds support for lowering function calls with the
`clang.arc.attachedcall` bundle. The goal is to expand such calls to the
following sequence of instructions:
callq @fn
movq %rax, %rdi
callq _objc_retainAutoreleasedReturnValue / _objc_unsafeClaimAutoreleasedReturnValue
This sequence of instructions triggers Objective-C runtime optimizations,
hence we want to ensure no instructions get moved in between them.
This patch achieves that by adding a new CALL_RVMARKER ISD node,
which gets turned into the CALL64_RVMARKER pseudo, which eventually gets
expanded into the sequence mentioned above.
The ObjC runtime function to call is determined by the
argument in the bundle, which is passed through as a
target constant to the pseudo.
@ahatanak is working on using this attribute in the front- & middle-end.
Together with the front- & middle-end changes, this should address
PR31925 for X86.
This is the X86 version of 46bc40e502,
which added similar support for AArch64.
Reviewed By: ab
Differential Revision: https://reviews.llvm.org/D94597
This is another FMF gap exposed by D90901, but I don't see a way
to show the difference in a regression test as with:
f66ba4c6025663
We will see an asm difference if we add a test as part of D90901.
Generalize the fix from rGd0902a8665b1 by ensuring we widen/narrow the indices subvector first and then perform the ZERO_EXTEND_VECTOR_INREG (if necessary), which should allow us to perform the variable permutes with source/destination/indices vectors of any widths.
D101838 incorrectly handled indices vectors of the same size but with higher element counts to just bitcast to the target indices type instead of performing a ZERO_EXTEND_VECTOR_INREG
This adds support to the X86 backend for the newly committed swiftasync
function parameter. If such a (pointer) parameter is present it gets stored
into an augmented frame record (populated in IR, but generally containing
enhanced backtrace for coroutines using lots of tail calls back and forth).
The context frame is identical to AArch64 (primarily so that unwinders etc
don't get extra complexity). Specfically, the new frame record is [AsyncCtx,
%rbp, ReturnAddr], and its presence is signalled by bit 60 of the stored %rbp
being set to 1. %rbp still points to the frame pointer in memory for backwards
compatibility (only partial on x86, but OTOH the weird AsyncCtx before the rest
of the record is because of x86).
Recommited with a fix for unwind info when i386 pc-rel thunks are
adjacent to a prologue.
This adds support to the X86 backend for the newly committed swiftasync
function parameter. If such a (pointer) parameter is present it gets stored
into an augmented frame record (populated in IR, but generally containing
enhanced backtrace for coroutines using lots of tail calls back and forth).
The context frame is identical to AArch64 (primarily so that unwinders etc
don't get extra complexity). Specfically, the new frame record is [AsyncCtx,
%rbp, ReturnAddr], and its presence is signalled by bit 60 of the stored %rbp
being set to 1. %rbp still points to the frame pointer in memory for backwards
compatibility (only partial on x86, but OTOH the weird AsyncCtx before the rest
of the record is because of x86).
Swift's new concurrency features are going to require guaranteed tail calls so
that they don't consume excessive amounts of stack space. This would normally
mean "tailcc", but there are also Swift-specific ABI desires that don't
naturally go along with "tailcc" so this adds another calling convention that's
the combination of "swiftcc" and "tailcc".
Support is added for AArch64 and X86 for now.
The intention is to be able to run this from additional locations (such as shuffle combining) in the future.
Reapplies rGb95a103808ac (after reversion at rGc012a388a15b), with SSE3/SSSE3 typo fix, test added at rG0afb10de1449.
On AVX1 targets we can handle v4i64 logical shifts by 32 bits as a pair of v8f32 shuffles with zero.
I was hoping to put this in LowerScalarImmediateShift, but performing that early causes regressions where other instructions were respliting the subvectors.
getVectorNumElements() returns a value for scalable vectors
without any warning so it is effectively getVectorMinNumElements().
By renaming it and making getVectorNumElements() forward to
it, we can insert a check for scalable vectors into getVectorNumElements()
similar to EVT. I didn't do that in this patch because there are still more
fixes needed, but I was able to temporarily do it and passed the RISCV
lit tests with these changes.
The changes to isPow2VectorType and getPow2VectorType are copied from EVT.
The change to TypeInfer::EnforceSameNumElts reduces the size of AArch64's isel table.
We're now considering SameNumElts to require the scalable property to match which
removes some unneeded type checks.
This was motivated by the bug I fixed yesterday in 80b9510806
Reviewed By: frasercrmck, sdesmalen
Differential Revision: https://reviews.llvm.org/D102262
Extend the HOP(HOP(X,Y),HOP(Z,W)) and SHUFFLE(HOP(X,Y),HOP(Z,W)) folds to handle repeating 256/512-bit vector cases.
This allows us to drop the UNPACK(HOP(),HOP()) custom fold in combineTargetShuffle.
This required isRepeatedTargetShuffleMask to be tweaked to support target shuffle masks taking more than 2 inputs.
foldShuffleOfHorizOp only handled basic shufps(hop(x,y),hop(z,w)) folds - by moving this to canonicalizeShuffleMaskWithHorizOp we can work with more general/combined v4x32 shuffles masks, float/integer domains and support shuffle-of-packs as well.
The next step will be to support 256/512-bit vector cases.
Ensure we don't try to fold when one might be an opaque constant - the constant fold will fail and then the reverse fold will happen in DAGCombine.....
Based off a discussion on D89281 - where the AARCH64 implementations were being replaced to use funnel shifts.
Any target that has efficient funnel shift lowering can handle the shift parts expansion using the same expansion, avoiding a lot of duplication.
I've generalized the X86 implementation and moved it to TargetLowering - so far I've found that AARCH64 and AMDGPU benefit, but many other targets (ARM, PowerPC + RISCV in particular) could easily use this with a few minor improvements to their funnel shift lowering (or the folding of their target ops that funnel shifts lower to).
NOTE: I'm trying to avoid adding full SHIFT_PARTS legalizer handling as I think it might actually be possible to remove these opcodes in the medium-term and use funnel shift / libcall expansion directly.
Differential Revision: https://reviews.llvm.org/D101987
Need to perfortm a bitcast on IndicesVec rather than subvector extract
if the original size of the IndicesVec is the same as the size of the
destination type.
Differential Revision: https://reviews.llvm.org/D101838
Related to PR50172.
Protects us against regressions after we will start doing cttz(zext(x)) -> zext(cttz(x)) transformation in the middle-end.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D101662
Previously we used an i32 constant to store the saturation width, but i32 isn't
legal on RISCV64. This wasn't a big deal to fix, but it is extra work for the
type legalizer.
This patch uses a VTSDNode to store the type similar to SEXT_INREG. This makes
it opaque to the type legalizer.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D101262
GCC supports negative values for -mstack-protector-guard-offset=, this
should be a signed value. Pre-req to D100919.
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D101325
PR50049 demonstrated an infinite loop between OR(SHUFFLE,SHUFFLE) <-> BLEND(SHUFFLE,SHUFFLE) patterns.
The UNDEF elements were allowing a combined shuffle mask to be widened which lost the undef element, resulting us needing to use the BLEND pattern (as the undef element would need to be zero for the OR pattern). But then bitcast folds would re-expose the undef element allowing us to use OR again.....
If we are truncating from a i32 source before comparing the result against zero, then see if we can directly compare the source value against zero.
If the upper (truncated) bits are known to be zero then we can compare against that, hopefully increasing the chances of us folding the compare into a EFLAG result of the source's operation.
Fixes PR49028.
Differential Revision: https://reviews.llvm.org/D100491
In the fold SHUFFLE(BINOP(X,Y),BINOP(Z,W)) -> BINOP(SHUFFLE(X,Z),SHUFFLE(Y,W)), check if both X/Z AND Y/W have at least one merge-able shuffle in which case the total number of shuffle should still fall.
Helps with instruction count regressions we saw while fixing PR48823
Extension to rG74f98391a7a4, we can also include any of the upper (known zero) bits in the comparison in the shuffle removal fold, just as long as we demand all the elements of the movmsk source vector.
We already allow the comparison of the upper bits of 'IsAllOf' (allbits) patterns, but we can safely compare the known zero bits for 'IsAnyOf' (zerobits) patterns as well.
This fixes an issues where we are comparing a type wide than the number of vector elements, which avoids a regression mentioned in rGbaadbe04bf75.
Fixes the issues noted in PR48768, where the and/or/xor instruction had been promoted to avoid i8/i16 partial-dependencies, but the test against zero had not.
We can almost certainly relax this fold to work for any truncation, although it breaks a number of existing folds (notable movmsk folds which tend to rely on the truncate to determine the demanded bits/elts in the source vector).
There is a reverse combine in TargetLowering.SimplifySetCC so we must wait until after legalization before attempting this.
Improve AVX512 mask inversion, rG38c799bce801 exposed some missing opportunities to move scalar not() back onto the boolvector types for folding with setcc etc.
Followup to D100177, handle an similar (demorgan inverse style) case from PR47797 as well
The AVX512 test cases could be further improved if we folded not(iX bitcast(vXi1)) -> (iX bitcast(not(vXi1)))
Alive2: https://alive2.llvm.org/ce/z/AnA_-W
I've initially just enabled this for BMI which has the ANDN instruction for i32/i64 - the i16/i8 cases give an idea of what'd we get when we enable it in all cases (I'll do this as a later commit).
Additionally, the i16/i8 cases could be freely promoted to i32 (as the args are already zeroext) and we could then make use of ANDN + the free cmp0 there as well - this has come up in PR48768 and PR49028 so I'm going to look at this soon.
https://alive2.llvm.org/ce/z/QVWHP_https://alive2.llvm.org/ce/z/pLngT-
Vector cases do not appear to benefit from this as we end up with having to generate the zero vector as well - this is one of the reasons I didn't try to tie this into hasAndNot/hasAndNotCompare.
Differential Revision: https://reviews.llvm.org/D100177
Fixes PR47603
This should probably be transferable to DAGCombine - the main limitation with the existing trunc(logicop) DAG fold is we don't know if legalization has tried to promote truncated logicops already. We might be able to peek through extensions as well.
Use the getTargetShuffleInputs helper for all shuffle decoding
Reapplied (after reversion in rGfa0aff6d6960) with fix+test for subvector splitting - we weren't accounting for peeking through bitcasts changing the vector element count of the shuffle sources.
The default expansion creates a MUL and either a MULHS/MULHU. Each
of those separately expand to sequences that use one or more
PMULLW instructions as well as additional instructions to
extend the types to vXi16. The MULHS/MULHU expansion computes the
whole 16-bit product, but only keeps the high part.
We can improve the lowering of SMULO/UMULO for some cases by using the MULHS/MULHU
expansion, but keep both the high and low parts. And we can use
those parts to calculate the overflow.
For AVX512 we might have vXi1 overflow outputs. We can improve those by using
vpcmpeqw to produce a k register if AVX512BW is enabled. This is a little better
than truncating the high result to use vpcmpeqb. If we don't have avx512bw we
can extend up to v16i32 to use vpcmpeqd to produce a k register.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D97624
Currently needsStackRealignment returns false if canRealignStack returns false.
This means that the behavior of needsStackRealignment does not correspond to
it's name and description; a function might need stack realignment, but if it
is not possible then this function returns false. Furthermore,
needsStackRealignment is not virtual and therefore some backends have made use
of canRealignStack to indicate whether a function needs stack realignment.
This patch attempts to clarify the situation by separating them and introducing
new names:
- shouldRealignStack - true if there is any reason the stack should be
realigned
- canRealignStack - true if we are still able to realign the stack (e.g. we
can still reserve/have reserved a frame pointer)
- hasStackRealignment = shouldRealignStack && canRealignStack (not target
customisable)
Targets can now override shouldRealignStack to indicate that stack realignment
is required.
This change will make it easier in a future change to handle the case where we
need to realign the stack but can't do so (for example when the register
allocator creates an aligned spill after the frame pointer has been
eliminated).
Differential Revision: https://reviews.llvm.org/D98716
Change-Id: Ib9a4d21728bf9d08a545b4365418d3ffe1af4d87
This was crashing with the example from:
https://llvm.org/PR49716
...and that was avoided with a283d72583 ,
but as we can see from the SSE vs. AVX test code diff,
we can try harder to match the pattern.
This matcher code was adapted from another pmadd pattern
match in D49636, but it needs different ops to deal with
size mismatches.
Differential Revision: https://reviews.llvm.org/D99531
For these cases we need to extract the upper or lower elements,
multiply them using 16-bit multiplies and repack them.
Previously we used punpcklbw/punpckhbw+psraw or pmovsxbw+pshudfd to
extract and sign extend so we could use pmullw to compute the 16-bit
product and then shift down the high bits.
We can avoid the need to sign extend if we unpack the bytes into
the high byte of each word and fill the lower byte with 0 using
pxor. This puts the sign bit of each byte into the sign bit of
each word. Since the LHS and RHS have 8 trailing zeros, the full
32-bit product of those 16-bit values will have 16 trailing zeros.
This means the 16-bit product of the original bytes is in the upper
16 bits which we can calculate using pmulhw.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D98587
Remove VBROADCAST/MOVDDUP/splat-shuffle handling from foldShuffleOfHorizOp
This can all be handled by canonicalizeShuffleMaskWithHorizOp along as we check that the HADD/SUB are only used once (to prevent infinite loops on slow-horizop targets which will try to reuse the nodes again followed by a post-hop shuffle).
See if the combined shuffle mask is equivalent to an identity shuffle, typically this is due to repeated LHS/RHS ops in horiz-ops, but isTargetShuffleEquivalent might see other patterns as well.
This is another small step towards getting rid of foldShuffleOfHorizOp and relying on canonicalizeShuffleMaskWithHorizOp and generic shuffle combining.
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
LowerVSETCC calls splitIntVSETCC after canonicalizing certain patterns, in particular (X & CPow2 != 0) -> (X & CPow2 == CPow2).
Unfortunately if we're splitting for AVX1/non-AVX512BW cases, we lose these canonicalizations as we call the split with the original SetCC node, and when the split nodes are later lowered in LowerVSETCC the patterns are lost behind extract_subvector etc. But if we pass the canonicalized operands for splitting we retain the optimizations.
Differential Revision: https://reviews.llvm.org/D99256
If we're truncating to vXi1 from a wider type, then prefer the original wider vector as is simplifies folding the separate truncations + extensions.
AVX1 this is only worth it for v8i1 cases, not v4i1 where we're always better off truncating down to v4i32 for movmsk.
Helps with some regressions encountered in D96609
Prefer broadcast from scalar on AVX targets as this makes it easier for later folds to strip away bitcasts etc.
This helps a lot with the AVX1 poor codegen from PR49658.
There's a trivial regression in bitcast-int-to-vector-bool-*ext.ll tests due to SimplifyDemandedBits not being able to see a multi-use case, but there's bigger existing codegen issues to be addressed first in those tests (unnecessary NOTs).
Helps fix cases where we've splatted smaller types to a wider vector element type without needing the upper bits.
Avoid this on AVX512 targets as that can affect broadcast folding.
The target shuffle code handles vector sources, but X86ISD::VBROADCAST can also accept a scalar source for splatting.
Suggested by @craig.topper on PR49658
For slow-hop targets, see if any single-op hops are duplicating work already done on another (dual-op) hop, which can sometimes occur as isHorizontalBinOp tries to find potential duplicates (but can't merge them itself). If so, reuse the other hop and shuffle the result.
That review is extracted from D69372.
It fixes https://bugs.llvm.org/show_bug.cgi?id=42219 bug.
For the noimplicitfloat mode, the compiler mustn't generate
floating-point code if it was not asked directly to do so.
This rule does not work with variable function arguments currently.
Though compiler correctly guards block of code, which copies xmm vararg
parameters with a check for %al, it does not protect spills for xmm registers.
Thus, such spills are generated in non-protected areas and could break code,
which does not expect floating-point data. The problem happens in -O0
optimization mode. With this optimization level there is used
FastRegisterAllocator, which spills virtual registers at basic block boundaries.
Register Allocator does not protect spills with additional control-flow modifications.
Thus to resolve that problem, it is suggested to not copy incoming physical
registers into virtual registers. Instead, store incoming physical xmm registers
into the memory from scratch.
Differential Revision: https://reviews.llvm.org/D80163
Generalize the shuffle(not(x)) -> not(shuffle(x)) fold to handle any binop with 0/-1.
Hopefully we can further generalize to help push target unary/binary shuffles through binops similar to what we do in DAGCombiner::visitVECTOR_SHUFFLE
As discussed on D97478. The removal of the custom tag causes some changes in the add/sub-overflow expansion as it no longer expands to sat-arith codegen.
We should be able to extend this "canonicalizeShuffleWithBinOps" to handle more generic binop cases where either/both operands can be cheaply shuffled.
In conjunction with the 'vperm2x128(bitcast(x),bitcast(y),c) -> bitcast(vperm2x128(x,y,c))' fold in combineTargetShuffle, this should remove any unnecessary bitcasts around vperm2x128 lane shuffles.
Extend the existing combine that handles bitcasting for fp-logic ops to also help remove logic ops across bitcasts to/from the same integer types.
This helps improve AVX512 predicate handling for D/Q logic ops and also allows DAGCombine's scalarizeExtractedBinop to remove some annoying gpr->simd->gpr transfers.
The concat_vectors regression in pr40891.ll will be addressed in a followup commit on this patch.
Differential Revision: https://reviews.llvm.org/D96206
This patch fixes some crashes coming from
X86ISelLowering::getSetCCResultType, which would occasionally return
an EVT constructed from an invalid MVT, which has a null Type pointer.
This patch refers to D95434.
Differential Revision: https://reviews.llvm.org/D97036
Fold shuffle(bop(shuffle(x,y),shuffle(z,w)),bop(shuffle(a,b),shuffle(c,d))) -> bop(shuffle(x,y),shuffle(z,w)),bop(shuffle(a,b),shuffle(c,d))
Attempt to fold from a shuffle of a pair of binops to a binop of shuffles, as long as one/both of the binop sources are also shuffles that can be merged with the outer shuffle. This should guarantee that we remove one binop without introducing any additional shuffles.
Technically there's potential for a merged shuffle's lowering to be poorer than the original shuffle, but it could also be better, and I'm not seeing any regressions as long as we keep the 'don't merge splats' rule already present in MergeInnerShuffle.
This expands and generalizes an existing X86 combine and attempts to merge either of each binop's sources (with an on-the-fly commutation of the shuffle mask) - we couldn't do that in the x86 version as it had to stay in a form that DAGCombine's MergeInnerShuffle would still recognise.
Fixes issue raised by @saugustine in rG5aa8f4c0843a where we were failing to replace null shuffle operands from MergeInnerShuffle to UNDEFs.
Differential Revision: https://reviews.llvm.org/D96345
This reverts commit 5dfba562dd.
That commit causes an assertion failure with the following repro:
typedef long b __attribute__((__vector_size__(16)));
b *d;
b e;
b __attribute__((__always_inline__)) c(b h, b i) {
return (__attribute__((__vector_size__(8 * sizeof(short)))) short)h + i;
}
j() {
b k, l, m, n, o[6], p, q;
m = d[5];
b r = m;
b s = f(r, 8);
q = s;
l = d[1];
p = l;
t(q);
n = c(m, l);
o[1] = c(s, f(p, 8));
k = __builtin_shufflevector(n, o[1], 0, 2);
e = __builtin_ia32_psrlwi128(k, j);
}
./bin/clang -cc1 -triple x86_64-grtev4-linux-gnu -emit-obj -O1 -std=c99 test.c
Fold shuffle(bop(shuffle(x,y),shuffle(z,w)),bop(shuffle(a,b),shuffle(c,d))) -> bop(shuffle(x,y),shuffle(z,w)),bop(shuffle(a,b),shuffle(c,d))
Attempt to fold from a shuffle of a pair of binops to a binop of shuffles, as long as one/both of the binop sources are also shuffles that can be merged with the outer shuffle. This should guarantee that we remove one binop without introducing any additional shuffles.
Technically there's potential for a merged shuffle's lowering to be poorer than the original shuffle, but it could also be better, and I'm not seeing any regressions as long as we keep the 'don't merge splats' rule already present in MergeInnerShuffle.
This expands and generalizes an existing X86 combine and attempts to merge either of each binop's sources (with an on-the-fly commutation of the shuffle mask) - we couldn't do that in the x86 version as it had to stay in a form that DAGCombine's MergeInnerShuffle would still recognise.
Differential Revision: https://reviews.llvm.org/D96345
Begin transitioning the X86 vector code to recognise sub(umax(a,b) ,b) or sub(a,umin(a,b)) USUBSAT patterns to make it more generic and available to all targets.
This initial patch just moves the basic umin/umax patterns to DAG, removing some vector-only checks on the way - these are some of the patterns that the legalizer will try to expand back to so we can be reasonably relaxed about matching these pre-legalization.
We can handle the trunc(sub(..))) variants as well, which helps with patterns where we were promoting to a wider type to detect overflow/saturation.
The remaining x86 code requires some cleanup first - some of it isn't actually tested etc. I also need to resurrect D25987.
Differential Revision: https://reviews.llvm.org/D96413
Fold shufps(hop(x,y),hop(z,w)) -> permute(hop(x,z)) - this is very similar to the equivalent unpack fold.
I did start trying to convert foldShuffleOfHorizOp to handle generic shuffle masks but we're relying on a lot of special cases at the moment.
I don't think we have any reason to believe the FP_ROUND here doesn't change the value.
Found while trying to see if we still need the fp128 block in CanCombineFCOPYSIGN_EXTEND_ROUND.
Removing that check caused this FP_ROUND to fire for fp128 which introduced a libcall expansion that asserted for this being a 1.
Reviewed By: RKSimon, pengfei
Differential Revision: https://reviews.llvm.org/D96098
Now that PR48908 has been dealt with, we can handle v4f64 permute cases by extracting the low/high lane VPERMILPD masks and creating a new mask based on which lanes are referenced by the VPERM2F128 mask.
Extends D95779 to permit insertion into float/doubles vectors while avoiding a lot of aliased memory traffic.
The scalar value is already on the simd unit, so we only need to transfer and splat the index value, then perform the select.
SSE4 codegen is a little bulky due to the tied register requirements of (non-VEX) BLENDPS/PD but the extra moves are cheap so shouldn't be an actual problem.
Differential Revision: https://reviews.llvm.org/D95866
With predicate masks, AVX512 can efficiently perform variable-index vector insertion with 2 broadcasts + 1 comparison, avoiding a lot of aliased memory traffic.
Differential Revision: https://reviews.llvm.org/D95779
This is directly analogous to the existing no_caller_saved_registers, but with the opposite intention. A function or call so marked shifts the responsibility of spilling the usual CSRs to it's caller.
An indirect call site and callee which don't agree on the attribute is ill defined.
The motivation for this change is that being able to prune callee saves (without modifying other details of the calling convention) is sometimes useful when generating stubs and adapters. There's no intention to expose this as a source language feature; this is expected to be used by frontends to implement adapters where warranted.
Some specific examples of use cases:
* GC compatible compiled code wants to call an externally defined library function without needing to track pointer values through CSRs.
* debug enabled code wants to call precompiled library which doesn't provide enough information to track CSRs while preserving debug quality in caller.
* adapter stub entering hand written assembler which doesn't follow normal calling conventions.
We can only legally extract from the lowest 128-bit subvector, so extract the correct subvector to allow us to handle 256/512-bit vector element extracts.
If the shuffle mask can't be widened to match the original extracted element width, see if the upper bits are zeroable - which allows us to extract+zero-extend the smaller extraction.
Ensure we check the valuetypes of all the HOP(SHUFFLE(X,Y),SHUFFLE(X,Y)) shuffle input ops - there was a copy+paste typo (noticed by MSVC analyzer) that meant we were checking the same input from one of the shuffles twice.
I haven't been able to create a test case for this yet - I don't think its currently possible to create a target/faux binary shuffle that scales to a 2x128 shuffle mask from two different value types.
Unlike VPERMILPS, VPERMILPD can have non-repeating masks in each 128-bit subvector, we weren't accounting for this when folding vperm2f128(vpermilpd(x,c),vpermilpd(y,c)) -> vpermilpd(vperm2f128(x,y),c).
I'm intending to add support for this but wanted to get a minimal fix in first for merging into 12.xx.
Fixes PR48908
We're relying on the source inputs for shuffle combining having already been widened to the root size (otherwise the offset logic falls over) - we're going to be supporting different sized shuffle inputs soon, so we need to explicitly make the minimum widened width the original root size.
We allow insert_subvector lowering of all legal types, so don't always cast to the vXi64/vXf64 shuffle types - this is only necessary for X86ISD::SHUF128/X86ISD::VPERM2X128 patterns later.
Simplify vperm2x128(concat(X,Y),concat(Z,W)) folding.
Use collectConcatOps / ISD::INSERT_SUBVECTOR to find the source subvectors instead of hardcoded immediate matching.
combineX86ShufflesConstants/canonicalizeShuffleMaskWithHorizOp can both handle/earlyout shuffles with inputs of different widths, so delay widening as late as possible to make it easier to match constant folds etc.
The plan is to eventually move the widening inside combineX86ShuffleChain so that we don't create any new nodes unless we successfully combine the shuffles.
rGbe69e66b1cd8 added the fold, but DAGCombiner.visitVECTOR_SHUFFLE doesn't merge shuffles if the inner shuffle is a splat, so we need to bail.
The non-fast-horiz-ops paths see some minor regressions, we might be able to improve on this after lowering to target shuffles.
Fix PR48823
We already have an experimental option to tune loop alignment. Its impact
is very wide (and there is a suspicion that it's not always profitable). We want
to have something more narrow to play with. This patch adds similar option that
overrides preferred alignment for innermost loops. This is for experimental
purposes, default values do not change the existing behavior.
Differential Revision: https://reviews.llvm.org/D94895
Reviewed By: pengfei
We already handle "vperm2x128 (ins ?, X, C1), (ins ?, X, C1), 0x31" for shuffling of the upper subvectors, but we weren't dealing with the case when we were splatting the upper subvector from a single source.
As discussed on D56387, if we're shifting to extract the upper/lower half of a vXi64 vector then we're actually better off performing this at the subvector level as its very likely to fold into something.
combineConcatVectorOps can perform this in reverse if necessary.
If a srl doesn't introduce any sign bits into the truncated result, then replace with a sra to let us use a PACKSS truncation - fixes a regression noticed in D56387 on pre-SSE41 targets that don't have PACKUSDW.
Specify LHS/RHS operands in matchShuffleWithUNPCK's calls to isTargetShuffleEquivalent, and handle VBROADCAST/VBROADCAST_LOAD matching in IsElementEquivalent
Fangrui Song