Prep work for PR38243 - mainly adding comments on where we need to add modulo support (doing so at the moment causes massive codegen regressions).
I've also consistently added support for modulo folding for uniform constants (although at the moment we have no way to trigger this) and removed the old assertions.
llvm-svn: 348366
This is an initial patch to add a minimum level of support for funnel shifts to the SelectionDAG and to begin wiring it up to the X86 SHLD/SHRD instructions.
Some partial legalization code has been added to handle the case for 'SlowSHLD' where we want to expand instead and I've added a few DAG combines so we don't get regressions from the existing DAG builder expansion code.
Differential Revision: https://reviews.llvm.org/D54698
llvm-svn: 348353
PR17686 demonstrates that for some targets FP exceptions can fire in cases where the FP_TO_UINT is expanded using a FP_TO_SINT instruction.
The existing code converts both the inrange and outofrange cases using FP_TO_SINT and then selects the result, this patch changes this for 'strict' cases to pre-select the FP_TO_SINT input and the offset adjustment.
The X87 cases don't need the strict flag but generates much nicer code with it....
Differential Revision: https://reviews.llvm.org/D53794
llvm-svn: 348251
We only needed this because it provided really aggressive constant folding even through constant pool entries created from build_vectors. The main case was for vXi8 MULH legalization which was happening as part of legalize DAG instead of as part of legalize vector ops. Now its part of vector op legalization and we've added special handling for build vectors of all constants there. This has removed the need for this code on the list tests we have.
llvm-svn: 348237
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
Summary:
We need to unpackl and unpackh the operands to use two vXi16 multiplies. Previously it looks like the low unpack would get constant folded at least in the 128-bit case after shuffle lowering turned the unpackl into ZERO_EXTEND_VECTOR_INREG and X86 custom DAG combined it. The same doesn't happen for the high half. So we'd load a constant and then shuffle it. But the low half would just be loaded and used by the multiply directly.
After this patch we now end up with a constant pool entry for the low and high unpacks separately with no shuffle operations.
This is a step towards removing custom constant folding for ZERO_EXTEND_VECTOR_INREG/SIGN_EXTEND_VECTOR_INREG in the X86 backend.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D55165
llvm-svn: 348159
Summary:
Under -x86-experimental-vector-widening-legalization, fp_to_uint/fp_to_sint with a smaller than 128 bit vector type results are custom type legalized by promoting the result to a 128 bit vector by promoting the elements, inserting an assertzext/assertsext, then truncating back to original type. The truncate will be further legalizdd to a pack shuffle. In the case of a v8i8 result type, we'll end up with a v8i16 fp_to_sint. This will need to be further legalized during vector op legalization by promoting to v8i32 and then truncating again. Under avx2 this produces good code with two pack instructions, but Under avx512 this will result in a truncate instruction and a packuswb instruction. But we should be able to get away with a single truncate instruction.
The other option is to promote all the way to vXi32 result type during the first type legalization. But in some experimentation that seemed to require more work to produce good code for other configurations.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D54836
llvm-svn: 348158
Previously this code generated its own extracts and build_vector. But we can use a simpler concat_vectors or scalar_to_vector operation and let type legalization do additional legalization of those operations.
llvm-svn: 348087
The generic legalizer will fall back to a stack spill that uses a truncating store. That store will get expanded into a shuffle and non-truncating store on pre-avx512 targets. Once that happens the stack store/load pair will be combined away leaving behind the shuffle and bitcasts. On avx512 targets the truncating store is legal so doesn't get folded away.
By custom legalizing it we can avoid this churn and maybe produce better code.
llvm-svn: 348085
Summary: With sse4.1 we use two zero_extend_vector_inreg and a pshufd to expand the v16i8 input into two v8i16 vectors for the multiply. That's 3 shuffles to extend one operand. The other operand is usually constant as this is mostly used by division by constant optimization. Pre sse4.1 we use a punpckhbw and a punpcklbw with a zero vector. That's two shuffles and an xor and a copy due to tied register constraints. That seems maybe better than the 3 shuffles. With AVX we avoid the copy so that's obviously better.
Reviewers: spatel, RKSimon
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D55138
llvm-svn: 348079
This reduces the number of shuffle operations that need to be done. The splitting strategy requires the shuffle unit for the extraction and the extension. With the unpack strategy the unpacks accomplish a splitting and extending in one operation.
llvm-svn: 348019
This does require a constant pool load instead of loading an immediate into a gpr, moving to a k register and masking. But its less instructions and more consistent with previous ISAs. It probably opens up more combine opportunities as one of the test cases demonstrates.
llvm-svn: 348018
Previously we emitted a punpcklbw/punpckhbw to move the byte elements into the upper half of 16 bit elements then shifted right by 8 to zero the upper bits. After DAG combine we end up with punpcklbw/punpckhbw into the lower bits with zeros in the uppers bits and no shifts. So just emit that directly.
llvm-svn: 347966
We had a EVT variable capturing the result of getSimpleValueType which returns an MVT. Another place using EVT that could have been MVT. And an 'int' that should be 'unsigned'.
llvm-svn: 347959
I believe we should be legalizing these with the rest of vector binary operations. If any custom lowering is required for these nodes, this will give the DAG combine between LegalizeVectorOps and LegalizeDAG to run on the custom code before constant build_vectors are lowered in LegalizeDAG.
I've moved MULHU/MULHS handling in AArch64 from Lowering to isel. Moving the lowering earlier caused build_vector+extract_subvector simplifications to kick in which made the generated code worse.
Differential Revision: https://reviews.llvm.org/D54276
llvm-svn: 347902
This is another patch for -x86-experimental-vector-widening. This pre widens narrow division by constants so that we can get pass the legal type check in the generic DAG combiner. Otherwise we end up scalarizing.
I've restricted this to splats for now because it was easy to just call DAG.getConstant. Not sure what we should do for non-splat? Increase the element size?Widen the constant vector by padding with 1?
Differential Revision: https://reviews.llvm.org/D54919
llvm-svn: 347898
This failed to select (which might be a separate bug) in
X86ISelDAGToDAG because we try to create a select node
that can be simplified away after rL347227.
This change avoids the problem by simplifying the SHRUNKBLEND
node sooner. In the test case, we manage to realize that the
true/false values of the select (SHRUNKBLEND) are the same thing,
so it simplifies away completely.
llvm-svn: 347818
Expansion of SIGN_EXTEND_INREG can create a VSRAI instruction. If there is already a VSRAI after it, we should combine them into a larger VSRAI
Differential Revision: https://reviews.llvm.org/D54959
llvm-svn: 347784
We're already mixing this APInt with other 'unsigned' variables. This allows us to use regular comparison operators instead of needing to use APInt::ult or APInt::uge. And it removes a later conversion from APInt to unsigned.
I might be adding another combine to this function and this will probably simplify the logic required for that.
llvm-svn: 347684
If we fold the bitcast into the store we'll end up creating a truncating store to vXi1 that will get scalarized. Instead allow the bitcast to be turned into a movmsk.
We probably need to do something if the store itself is a vXi1 type, but I'll leave that til a testcase appears.
llvm-svn: 347632
SplitVecOp_TruncateHelper tries to promote the result type while splitting FP_TO_SINT/UINT. It then concatenates the result and introduces a truncate to the original result type. But it does this without inserting the AssertZExt/AssertSExt that the regular result type promotion would insert. Nor does it turn FP_TO_UINT into FP_TO_SINT the way normal result type promotion for these operations does. This is bad on X86 which doesn't support FP_TO_SINT until AVX512.
This patch disables the use of SplitVecOp_TruncateHelper for these operations and just lets normal promotion handle it. I've tweaked a couple things in X86ISelLowering to avoid a few obvious regressions there. I believe all the changes on X86 are improvements. The other targets look neutral.
Differential Revision: https://reviews.llvm.org/D54906
llvm-svn: 347593
We have these 2 "isDesirable" promotion hooks (I'm not sure why we need both of them, but that's
independent of this patch), and we can adjust them to promote "mul i8 X, C" to i32. Then, all of
our existing LEA and other multiply expansion magic happens as it would for i32 ops.
Some of the test diffs show that we could end up with an actual 32-bit mul instruction here
because we choose not to expand to simpler ops. That instruction could be slower depending on the
subtarget. On the plus side, this means we don't need a separate instruction to load the constant
operand and possibly an extra instruction to move the result. If we need to tune mul i32 further,
we could add a later transform that tries to shrink it back to i8 based on subtarget timing.
I did not bother to duplicate all of the 32-bit test file RUNs and target settings that exist to
test whether LEA expansion is cheap or not. The diffs here assume a default target, so that means
LEA is generally cheap.
Differential Revision: https://reviews.llvm.org/D54803
llvm-svn: 347557
This should likely be adjusted to limit this transform
further, but these diffs should be clear wins.
If we have blendv/conditional move, then we should assume
those are cheap ops. The loads become independent of the
compare, so those can be speculated before we need to use
the values in the blend/mov.
llvm-svn: 347526
We can't guarantee that demanded bits passing through the vector shuffle won't cause the AND in front of this to be removed. This would prevent the PACKUS from being matched during shuffle lowering.
Unfortunately, this adds a packuswb to one of the vector-reduce-mul.ll tests since we were removing the shuffle via SimplifyDemandedVectorElts. We appear to have similar issues with vpmovwb on the same test case on other targets.
llvm-svn: 347361
Previously we emitted to separate shuffles, one for unpcklbw and one for unpcklwd. Instead emit a single shuffle equivalent to both of the original shuffles. Shuffle lowering seems able to handle it. This avoids a bitcast between the two shuffles which seems helpful to DAG combine.
Remove the custom type legalization for v8i8->v8i32. I had put that in to avoid some almost duplicate punpcklbw instructions I was seeing, but this lowering change seems to fix that. It also fixes some duplicate shuffles seen in vector-sext.ll
llvm-svn: 347348
Pull out getPackDemandedElts demanded elts remapping helper from computeKnownBitsForTargetNode and use in computeKnownBits/ComputeNumSignBits.
llvm-svn: 347303
Previously if V2 was unused we ended up using V1 for both inputs as part of the code that follows the new code. By using lowerVectorShuffleWithUNPCK we keep the undef nature of V2 in the output.
As near as I can tell this makes v16i8 behavior consistent with every other VT now.
This does mean that we give the register allocator freedom to fill in random registers now and create false dependencies. But like I said we're already doing that for other types.
llvm-svn: 347296
getZeroVector produces a specifically canonicalized zero vector, but we can just let DAG legalization take care of it.
The test changes are because MULH lowering happens later than it should and this change gave us the opportunity to constant fold away a multiply during a DAG combine before the build_vector got legalized with a bitcast.
llvm-svn: 347290
SSE PSHUFB vector ctlz lowering works at the i4 nibble level. As detailed in PR39703, we were masking the lower nibble off but we only actually use it in the case where the upper nibble is known to be zero, making it safe to remove the mask and save an instruction.
Differential Revision: https://reviews.llvm.org/D54707
llvm-svn: 347242
Previously we split the vectors in half to allow the two halves to be any extended then concatenated the results back together.
This patch instead instead extends the v16i8 sse algorithm to extend half of each 128-bit lane using punpcklbw/punpckhbw. Multiplies all the low half lanes and high half lanes together in separate operations. Then merges the half lane results back together using packuswb.
Unfortunately, some of the cases in vector-reduce-mul.ll regress because we aren't narrowing the vector width of the multiplies as we reduce. The splitting was somewhat making up for that before by causing halves to be discarded after the split.
Differential Revision: https://reviews.llvm.org/D54668
llvm-svn: 347240
The shift requires a copy to avoid clobbering a register. Comparing with 0 uses an xor to produce 0 that will be overwritten with the compare results. So still requires 2 instructions, but should be one byte shorter since it doesn't need to encode an immediate.
llvm-svn: 347185
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
Simon Pilgrim