This saves us some unnecessary copies.
If the inputs to a G_SELECT are floating point, we should use fcsel rather than
csel.
Changes here are...
- Teach selectCopy about s1-to-s1 copies across register banks.
- AArch64RegisterBankInfo about G_SELECT in general.
- Teach the instruction selector about the FCSEL instructions.
Also add two tests:
- select-select.mir to show that we get the expected FCSEL
- regbank-select.mir (unfortunately named) to show the register banks on
G_SELECT are properly preserved
And update fast-isel-select.ll to show that we do the same thing as other
instruction selectors in these cases.
llvm-svn: 359940
The VOP3 form should always be the preferred selection form to be
shrunk later.
The r600 sub test needs to be split out because it asserts on the
arguments in the new test during the calling convention lowering.
llvm-svn: 359899
This was broken if the original operand was killed. The kill flag
would appear on both instructions, and fail the verifier. Keep the
kill flag, but remove the operands from the old instruction. This has
an added benefit of really reducing the use count for future folds.
Ideally the pass would be structured more like what PeepholeOptimizer
does to avoid this hack to avoid breaking instruction iterators.
llvm-svn: 359891
When a fold of an immediate into a sub/subrev required shrinking the
instruction, the wrong VOP2 opcode was used. This was using the VOP2
equivalent of the original instruction, not the commuted instruction
with the inverted opcode.
llvm-svn: 359883
Using/updating a dominator tree to match math overflow patterns may be very
expensive in compile-time (because of the way CGP uses a DT), so just handle
the single-block case.
Also, we were restarting the iterator loops when doing the overflow intrinsic
transforms by marking the dominator tree for update. That was done to prevent
iterating over a removed instruction. But we can postpone the deletion using
the existing "RemovedInsts" structure, and that means we don't need to update
the DT.
See post-commit thread for rL354298 for more details:
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20190422/646276.html
Differential Revision: https://reviews.llvm.org/D61075
llvm-svn: 359879
This is the second part of the commit fixing PR38917 (hoisting
partitially redundant machine instruction). Most of PRE (partitial
redundancy elimination) and CSE work is done on LLVM IR, but some of
redundancy arises during DAG legalization. Machine CSE is not enough
to deal with it. This simple PRE implementation works a little bit
intricately: it passes before CSE, looking for partitial redundancy
and transforming it to fully redundancy, anticipating that the next
CSE step will eliminate this created redundancy. If CSE doesn't
eliminate this, than created instruction will remain dead and eliminated
later by Remove Dead Machine Instructions pass.
The third part of the commit is supposed to refactor MachineCSE,
to make it more clear and to merge MachinePRE with MachineCSE,
so one need no rely on further Remove Dead pass to clear instrs
not eliminated by CSE.
First step: https://reviews.llvm.org/D54839
Fixes llvm.org/PR38917
Reviewers: RKSimon
Subscribers: hfinkel, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D56772
llvm-svn: 359870
We use to incorrectly use the store size instead of the alloc size when
creating the stack slot for allocas.
On aarch64 this can be demonstrated by allocating weirdly sized types.
For instance, in the added test case, we use an alloca for i19. We used
to allocate a slot of size 24-bit (19 rounded up to the next byte),
whereas we really want to use a full 32-bit slot for this type.
llvm-svn: 359856
The primary fix here is to WinException.cpp: we need to exclude jump
tables when computing the length of a function, or else we fail to
correctly compute the length. (We can only compute the number of bytes
consumed by certain assembler directives after the entire file is
parsed. ".p2align" is one of those directives, and is used by jump table
generation.)
The secondary fix, to MCWin64EH, is to make sure we don't silently
miscompile if we hit a similar situation in the future.
It's possible we could extend ARM64EmitUnwindInfo so it allows function
bodies that contain assembler directives, but that's a lot more
complicated; see the FIXME in MCWin64EH.cpp.
Fixes https://bugs.llvm.org/show_bug.cgi?id=41581 .
Differential Revision: https://reviews.llvm.org/D61095
llvm-svn: 359849
The default impementation in the base class for TargetLowering::getRegForInlineAsmConstraint doesn't work for mask registers when the VT is a scalar type integer types since the only legal mask types are vXi1. So we end up just getting whatever the first register class that contains the register. Currently this appears to be VK1, but its really dependent on the order tablegen outputs the register classes.
Some code in the caller ends up looking up the type for this register class and find v1i1 then generates a copyfromreg from the physical k-register with the v1i1 type. Then it generates an any_extend from v1i1 to the scalar VT which isn't legal. This bad any_extend sticks around until isel where it selects a MOVZX32rr8 with a v1i1 input or maybe a i8 input. Not sure but eventually we pick up a copy from VK1 to GR8 in MachineIR which isn't supported. This leads to a failure in physical register copying.
This patch uses the scalar type to find a VK class of the right size. In the attached test case this will be VK16. This causes a bitcast from vk16 to i16 to be generated instead of an any_extend. This will be properly iseled to a VK16 to GR32 copy and a GR32->GR16 extract_subreg.
Fixes PR41678
Differential Revision: https://reviews.llvm.org/D61453
llvm-svn: 359837
The original patch was committed at rL359398 and reverted at rL359695 because of
infinite looping.
This includes a fix to check for a vector splat of "1.0" to avoid the infinite loop.
Original commit message:
This was originally part of D61028, but it's an independent diff.
If we try the repeated divisor reciprocal transform before producing an estimate sequence,
then we have an opportunity to use scalar fdiv. On x86, the trade-off is 1 divss vs. 5
vector FP ops in the default estimate sequence. On recent chips (Skylake, Ryzen), the
full-precision division is only 3 cycle throughput, so that's probably the better perf
default option and avoids problems from x86's inaccurate estimates.
The last 2 tests show that users still have the option to override the defaults by using
the function attributes for reciprocal estimates, but those patterns are potentially made
faster by converting the vector ops (including ymm ops) to scalar math.
Differential Revision: https://reviews.llvm.org/D61149
llvm-svn: 359793
We don't have FP exception limits in the IR constant folder for the binops (apart from strict ops),
so it does not make sense to have them here in the DAG either. Nothing else in the backend tries
to preserve exceptions (again outside of strict ops), so I don't see how this could have ever
worked for real code that cares about FP exceptions.
There are still cases (examples: unary opcodes in SDAG, FMA in IR) where we are trying (at least
partially) to preserve exceptions without even asking if the target supports FP exceptions. Those
should be corrected in subsequent patches.
Real support for FP exceptions requires several changes to handle the constrained/strict FP ops.
Differential Revision: https://reviews.llvm.org/D61331
llvm-svn: 359791
Limiting scalar hadd/hsub generation to the lowest xmm looks to be unnecessary - we will be extracting one upper xmm whatever, and we can remove a shuffle by using the hop which is inline with what shouldUseHorizontalOp expects to happen anyway.
Testing on btver2 (the main target for fast-hops) shows this is beneficial even for float ops where we have a 'shuffle' to extract the float result:
https://godbolt.org/z/0R-U-K
Differential Revision: https://reviews.llvm.org/D61426
llvm-svn: 359786
Select G_SEXT and G_ZEXT with destination types smaller than 32 bits in
the exact same way as 32 bits. This overwrites the higher bits, but that
should be ok since all legal users of types smaller than 32 bits ignore
those bits anyway.
llvm-svn: 359768
Accidentally put a fast-isel-abort=2 instead of the GISel abort line.
This test doesn't actually fall back at all for GISel though, so remove the
fallback checks entirely.
llvm-svn: 359737
This adds support for using fmov rather than a standard mov to materialize
G_FCONSTANT when it's safe to do so.
Update arm64-fast-isel-materialize.ll and select-constant.mir to show that the
selection is correct.
llvm-svn: 359734
This is a slightly reduced version of the test from D61384.
Adding this as a preliminary step, so I can update D61149 with
the proposed fix.
llvm-svn: 359709
We already perform horizontal add/sub if we extract from elements 0 and 1, this patch extends it to non-0/1 element extraction indices (as long as they are from the lowest 128-bit vector).
Differential Revision: https://reviews.llvm.org/D61263
llvm-svn: 359707
This reverts commit fb9a5307a94e6f1f850e4d89f79103b123f16279 (rL359398)
because it can cause an infinite loop due to opposing combines.
llvm-svn: 359695
