This pattern turned a vector sqrt/rcp/rsqrt operation of sse_load_f32/f64 into the the scalar instruction for the operation and put undef into the upper bits. For correctness, the resulting code should still perform the sqrt/rcp/rsqrt on the upper bits after the load is extended since that's what the operation asked for. Particularly in the case where the upper bits are 0, in that case we need calculate the sqrt/rcp/rsqrt of the zeroes and keep the result in the upper-bits. This implies we should be using the packed instruction still.
The only test case for this pattern is one I just added so there was no coverage of this.
llvm-svn: 288784
This occurs due to a pattern that uses sse_load_f32/f64 with vector sqrt/rcp/rsqrt operations and turns them into scalar instructions. Perhaps for the case were the upper bits come from undef this is ok. I believe a (vzmovl load64) would do the same thing but those seems to become vzload instead and selectScalarSSELoad doesn't handle that today. In that case we should be performing the vector operation on the zeros in the upper bits which is not equivalent to using a scalar instruction.
I will remove this pattern in a follow up patch. There appears to be no other test content for it.
llvm-svn: 288783
The intrinsics are supposed to pass the upper bits straight through to their output register. This means we need to make sure we still perform the 128-bit load to get those upper bits to pass to give to the instruction since the memory form of the instruction only reads 32 or 64 bits.
llvm-svn: 288781
The sqrtsd instruction only loads 64-bits and writes bits 63:0 with the sqrt result. Bits 127:64 are preserved in the destination register. The semantics of the intrinsic indicate bits 127:64 should come from the intrinsic argument which in this case is a 128-bit load. So the generated code should have a 128-bit load and use a register form of sqrtsd.
llvm-svn: 288780
The intrinsic takes one argument, the lower bits are affected by the operation and the upper bits should be passed through. The instruction itself takes two operands, the high bits of the first operand are passed through and the low bits of the second operand are modified by the operation. To match this to the intrinsic we should pass the single intrinsic input to both operands.
I had to remove the stack folding test for these instructions since they depended on the incorrect behavior. The same register is now used for both inputs so the load can't be folded.
llvm-svn: 288779
Summary:
This patch removes the scalar logical operation alias instructions. We can just use reg class copies and use the normal packed instructions instead. This removes the need for putting these instructions in the execution domain fixing tables as was done recently.
I removed the loadf64_128 and loadf32_128 patterns as DAG combine creates a narrower load for (extractelt (loadv4f32)) before we ever get to isel.
I plan to add similar patterns for AVX512DQ in a future commit to allow use of the larger register class when available.
Reviewers: spatel, delena, zvi, RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D27401
llvm-svn: 288771
This changes the scalar non-intrinsic non-avx roundss/sd instruction
definitions not to read their destination register - allowing partial dependency
breaking.
This fixes PR31143.
Differential Revision: https://reviews.llvm.org/D27323
llvm-svn: 288703
so we can stop using DW_OP_bit_piece with the wrong semantics.
The entire back story can be found here:
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20161114/405934.html
The gist is that in LLVM we've been misinterpreting DW_OP_bit_piece's
offset field to mean the offset into the source variable rather than
the offset into the location at the top the DWARF expression stack. In
order to be able to fix this in a subsequent patch, this patch
introduces a dedicated DW_OP_LLVM_fragment operation with the
semantics that we used to apply to DW_OP_bit_piece, which is what we
actually need while inside of LLVM. This patch is complete with a
bitcode upgrade for expressions using the old format. It does not yet
fix the DWARF backend to use DW_OP_bit_piece correctly.
Implementation note: We discussed several options for implementing
this, including reserving a dedicated field in DIExpression for the
fragment size and offset, but using an custom operator at the end of
the expression works just fine and is more efficient because we then
only pay for it when we need it.
Differential Revision: https://reviews.llvm.org/D27361
rdar://problem/29335809
llvm-svn: 288683
We treat bitwise 'not' as a special operation and try not to reduce its all-ones mask.
Presumably, this is because a 'not' may be cheaper than a generic 'xor' or it may get
folded into another logic op if the target has those. However, if we can remove a logic
instruction by changing the xor's constant mask value, that should always be a win.
Note that the IR version of SimplifyDemandedBits() does not treat 'not' as a special-case
currently (although that's marked with a FIXME). So if you run this IR through -instcombine,
you should get the same end result. I'm hoping to add a different backend transform that
will expose this problem though, so I need to solve this first.
Differential Revision: https://reviews.llvm.org/D27356
llvm-svn: 288676
Currently the fast isel code emits an avx1 instruction sequence even with avx512. This is different than normal isel. A follow up commit will fix this.
llvm-svn: 288635
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
This change fixes a regression in r279537 and
makes getRawSubclassData behave like r279536.
Without this change, the fp128-g.ll test case will have an
infinite loop involving SoftenFloatRes_LOAD.
Differential Revision: http://reviews.llvm.org/D26942
llvm-svn: 288420
Choosing a "cfi" name makes the intend a bit clearer in an assembly dump
and more importantly the assembly dumps are slightly more stable as the
numbers don't move around anymore when unrelated code calls
createTempSymbol() more or less often.
As they are temp labels the name doesn't influence the generated object
code.
Differential Revision: https://reviews.llvm.org/D27244
llvm-svn: 288290
The LLDB tests are now ready for this patch.
DWARF specifies that "line 0" really means "no appropriate source
location" in the line table. Use this for branch targets and some
other cases that have no specified source location, to prevent
inheriting unfortunate line numbers from physically preceding
instructions (which might be from completely unrelated source).
Differential Revision: http://reviews.llvm.org/D24180
llvm-svn: 288283
Initial support for target shuffle constant folding in cases where all shuffle inputs are constant. We may be able to relax this and merge shuffles with only some constant inputs in the future.
I've added the helper function getTargetConstantBitsFromNode (based off a similar function in X86ShuffleDecodeConstantPool.cpp) that could be reused for other cases requiring constant vector extraction.
Differential Revision: https://reviews.llvm.org/D27220
llvm-svn: 288250
DWARF specifies that "line 0" really means "no appropriate source
location" in the line table. Use this for branch targets and some
other cases that have no specified source location, to prevent
inheriting unfortunate line numbers from physically preceding
instructions (which might be from completely unrelated source).
Differential Revision: http://reviews.llvm.org/D24180
llvm-svn: 288212
It looks like this logic was duplicated long ago and the GCC side of
things has grown additional functionality. We need ${:uid} at least to
generate unique MS inline asm labels (PR23715), so expose these.
llvm-svn: 288092
Bit-shifts by a whole number of bytes can be represented as a shuffle mask suitable for combining.
Added a 'getFauxShuffleMask' function to allow us to create shuffle masks from other suitable operations.
llvm-svn: 288040
Summary: When selectScalarSSELoad is looking for a scalar_to_vector of a scalar load, it makes sure the load is only used by the scalar_to_vector. But it doesn't make sure the scalar_to_vector is only used once. This can cause the same load to be folded multiple times. This can be bad for performance. This also causes the chain output to be duplicated, but not connected to anything so chain dependencies will not be satisfied.
Reviewers: RKSimon, zvi, delena, spatel
Subscribers: andreadb, llvm-commits
Differential Revision: https://reviews.llvm.org/D26790
llvm-svn: 287983
Summary:
Shuffle lowering may have widened the element size of a i32 shuffle to i64 before selecting X86ISD::SHUF128. If this shuffle was used by a vselect this can prevent us from selecting masked operations.
This patch detects this and changes the element size to match the vselect.
I don't handle changing integer to floating point or vice versa as its not clear if its better to push such a bitcast to the inputs of the shuffle or to the user of the vselect. So I'm ignoring that case for now.
Reviewers: delena, zvi, RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D27087
llvm-svn: 287939
Vectorize UINT_TO_FP v2i32 -> v2f64 instead of scalarization (albeit still on the SIMD unit).
The codegen matches that generated by legalization (and is in fact used by AVX for UINT_TO_FP v4i32 -> v4f64), but has to be done in the x86 backend to account for legalization via 4i32.
Differential Revision: https://reviews.llvm.org/D26938
llvm-svn: 287886
The bug arises during register allocation on i686 for
CMPXCHG8B instruction when base pointer is needed. CMPXCHG8B
needs 4 implicit registers (EAX, EBX, ECX, EDX) and a memory address,
plus ESI is reserved as the base pointer. With such constraints the only
way register allocator would do its job successfully is when the addressing
mode of the instruction requires only one register. If that is not the case
- we are emitting additional LEA instruction to compute the address.
It fixes PR28755.
Patch by Alexander Ivchenko <alexander.ivchenko@intel.com>
Differential Revision: https://reviews.llvm.org/D25088
llvm-svn: 287875
We did not support subregs in InlineSpiller:foldMemoryOperand() because targets
may not deal with them correctly.
This adds a target hook to let the spiller know that a target can handle
subregs, and actually enables it for x86 for the case of stack slot reloads.
This fixes PR30832.
Differential Revision: https://reviews.llvm.org/D26521
llvm-svn: 287792
We have the following DAGCombiner transformations:
(mul (shl X, c1), c2) -> (mul X, c2 << c1)
(mul (shl X, C), Y) -> (shl (mul X, Y), C)
(shl (mul x, c1), c2) -> (mul x, c1 << c2)
Usually the constant shift is optimised by SelectionDAG::getNode when it is
constructed, by SelectionDAG::FoldConstantArithmetic, but when we're dealing
with vectors and one of those vector constants contains an undef element
FoldConstantArithmetic does not fold and we enter an infinite loop.
Fix this by making FoldConstantArithmetic use getNode to decide how to fold each
vector element, the same as FoldConstantVectorArithmetic does, and rather than
adding the constant shift to the work list instead only apply the transformation
if it's already been folded into a constant, as if it's not we're going to loop
endlessly. Additionally add missing NoOpaques to one of those transformations,
which I noticed when writing the tests for this.
Differential Revision: https://reviews.llvm.org/D26605
llvm-svn: 287766
Implemented widening (v2f32) and splitting (v16f64).
On splitting, I use "popcnt" to calculate memory increment.
More type legalization work will come in the next patches.
llvm-svn: 287761
This occurs during UINT_TO_FP v2f64 lowering.
We can easily generalize this to other horizontal ops (FHSUB, PACKSS, PACKUS) as required - we are doing something similar with PACKUS in lowerV2I64VectorShuffle
llvm-svn: 287676
Add basic ComputeNumSignBits support for TRUNCATE ops for cases where the source's number of sign bits overlaps with the truncated size.
Improves X86 SIGN_EXTEND_IN_REG vector cases which were needlessly sign extending boolean vector results.
Differential Revision: https://reviews.llvm.org/D26851
llvm-svn: 287635
Summary:
The index and one of the table operands can be swapped by changing the opcode to the other version. Neither of these operands are the one that can load from memory so this can't be used to increase memory folding opportunities.
We need to handle the unmasked forms and the kz forms. Since the load operand isn't being commuted we can commute the load and broadcast instructions too.
Reviewers: igorb, delena, Ayal, Farhana, RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D25652
llvm-svn: 287621
Summary:
Shuffle lowering widens the element size of a shuffle if elements are contiguous. This is sometimes help because wider element types have more shuffle options. If the shuffle is one of the arguments to a vselect this shuffle widening can introduce a bitcast between the vselect and the shuffle. This will prevent isel from selecting a masked operation. If the shuffle can be written equally efficiently with a different element size to match the vselect type we should change the shuffle type to allow masking.
This patch does this conversion for all VALIGND/VALIGNQ sizes. It also supports turning 128-bit PALIGNR into VALIGND/VALIGNQ. This fixes the case shown in PR31018.
I plan to add support for more operations in future patches.
Reviewers: RKSimon, zvi, delena
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D26902
llvm-svn: 287612
Summary: Merging an empty case block into the header block of switch could cause
ISel to add COPY instructions in the header of switch, instead of the case
block, if the case block is used as an incoming block of a PHI. This could
potentially increase dynamic instructions, especially when the switch is in a
loop. I added a test case which was reduced from the benchmark I was targetting.
Reviewers: t.p.northover, mcrosier, manmanren, wmi, davidxl
Subscribers: qcolombet, danielcdh, hfinkel, mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D22696
llvm-svn: 287553
Craig Topper