This adds the missing test case for the previous commit:
Allow handling of vectors during return lowering for little endian machines.
Sorry for the noise.
llvm-svn: 217847
This lowers frem to a runtime libcall inside fast-isel.
The test case also checks the CallLoweringInfo bug that was exposed by this
change.
This fixes rdar://problem/18342783.
llvm-svn: 217833
Summary:
Expand list of supported targets for Mips to include mips32 r1.
Previously it only include r2. More patches are coming where there is
a difference but in the current patches as pushed upstream, r1 and r2
are equivalent.
Test Plan:
simplestorefp1.ll
add new build bots at mips to test this flavor at both -O0 and -O2
Reviewers: dsanders
Reviewed By: dsanders
Differential Revision: http://reviews.llvm.org/D5306
llvm-svn: 217821
On MachO, and MachO only, we cannot have a truly empty function since that
breaks the linker logic for atomizing the section.
When we are emitting a frame pointer, the presence of an unreachable will
create a cfi instruction pointing past the last instruction. This is perfectly
fine. The FDE information encodes the pc range it applies to. If some tool
cannot handle this, we should explicitly say which bug we are working around
and only work around it when it is actually relevant (not for ELF for example).
Given the unreachable we could omit the .cfi_def_cfa_register, but then
again, we could also omit the entire function prologue if we wanted to.
llvm-svn: 217801
introduced in r217629.
We were returning the old sext instead of the new zext as the promoted instruction!
Thanks Joerg Sonnenberger for the test case.
llvm-svn: 217800
Peephole optimization was folding MOVSDrm, which is a zero-extending double
precision floating point load, into ADDPDrr, which is a SIMD add of two packed
double precision floating point values.
(before)
%vreg21<def> = MOVSDrm <fi#0>, 1, %noreg, 0, %noreg; mem:LD8[%7](align=16)(tbaa=<badref>) VR128:%vreg21
%vreg23<def,tied1> = ADDPDrr %vreg20<tied0>, %vreg21; VR128:%vreg23,%vreg20,%vreg21
(after)
%vreg23<def,tied1> = ADDPDrm %vreg20<tied0>, <fi#0>, 1, %noreg, 0, %noreg; mem:LD8[%7](align=16)(tbaa=<badref>) VR128:%vreg23,%vreg20
X86InstrInfo::foldMemoryOperandImpl already had the logic that prevented this
from happening. However the check wasn't being conducted for loads from stack
objects. This commit factors out the logic into a new function and uses it for
checking loads from stack slots are not zero-extending loads.
rdar://problem/18236850
llvm-svn: 217799
Add some more tests to make sure better operand
choices are still made. Leave some cases that seem
to have no reason to ever be e64 alone.
llvm-svn: 217789
I noticed some odd looking cases where addr64 wasn't set
when storing to a pointer in an SGPR. This seems to be intentional,
and partially tested already.
The documentation seems to describe addr64 in terms of which registers
addressing modifiers come from, but I would expect to always need
addr64 when using 64-bit pointers. If no offset is applied,
it makes sense to not need to worry about doing a 64-bit add
for the final address. A small immediate offset can be applied,
so is it OK to not have addr64 set if a carry is necessary when adding
the base pointer in the resource to the offset?
llvm-svn: 217785
when SSE4.1 is available.
This removes a ton of domain crossing from blend code paths that were
ending up in the floating point code path.
This is just the tip of the iceberg though. The real switch is for
integer blend lowering to more actively rely on this instruction being
available so we don't hit shufps at all any longer. =] That will come in
a follow-up patch.
Another place where we need better support is for using PBLENDVB when
doing so avoids the need to have two complementary PSHUFB masks.
llvm-svn: 217767
missing specific checks.
While there is a lot of redundancy here where all-but-one mode use the
same code generation, I'd rather have each variant spelled out and
checked so that readers aren't misled by an omission in the test suite.
llvm-svn: 217765
instructions from the relevant shuffle patterns.
This is the last tweak I'm aware of to generate essentially perfect
v4f32 and v2f64 shuffles with the new vector shuffle lowering up through
SSE4.1. I'm sure I've missed some and it'd be nice to check since v4f32
is amenable to exhaustive exploration, but this is all of the tricks I'm
aware of.
With AVX there is a new trick to use the VPERMILPS instruction, that's
coming up in a subsequent patch.
llvm-svn: 217761
instructions when it finds an appropriate pattern.
These are lovely instructions, and its a shame to not use them. =] They
are fast, and can hand loads folded into their operands, etc.
I've also plumbed the comment shuffle decoding through the various
layers so that the test cases are printed nicely.
llvm-svn: 217758
AVX is available, and generally tidy up things surrounding UNPCK
formation.
Originally, I was thinking that the only advantage of PSHUFD over UNPCK
instruction variants was its free copy, and otherwise we should use the
shorter encoding UNPCK instructions. This isn't right though, there is
a larger advantage of being able to fold a load into the operand of
a PSHUFD. For UNPCK, the operand *must* be in a register so it can be
the second input.
This removes the UNPCK formation in the target-specific DAG combine for
v4i32 shuffles. It also lifts the v8 and v16 cases out of the
AVX-specific check as they are potentially replacing multiple
instructions with a single instruction and so should always be valuable.
The floating point checks are simplified accordingly.
This also adjusts the formation of PSHUFD instructions to attempt to
match the shuffle mask to one which would fit an UNPCK instruction
variant. This was originally motivated to allow it to match the UNPCK
instructions in the combiner, but clearly won't now.
Eventually, we should add a MachineCombiner pass that can form UNPCK
instructions post-RA when the operand is known to be in a register and
thus there is no loss.
llvm-svn: 217755
'punpckhwd' instructions when suitable rather than falling back to the
generic algorithm.
While we could canonicalize to these patterns late in the process, that
wouldn't help when the freedom to use them is only visible during
initial lowering when undef lanes are well understood. This, it turns
out, is very important for matching the shuffle patterns that are used
to lower sign extension. Fixes a small but relevant regression in
gcc-loops with the new lowering.
When I changed this I noticed that several 'pshufd' lowerings became
unpck variants. This is bad because it removes the ability to freely
copy in the same instruction. I've adjusted the widening test to handle
undef lanes correctly and now those will correctly continue to use
'pshufd' to lower. However, this caused a bunch of churn in the test
cases. No functional change, just churn.
Both of these changes are part of addressing a general weakness in the
new lowering -- it doesn't sufficiently leverage undef lanes. I've at
least a couple of patches that will help there at least in an academic
sense.
llvm-svn: 217752
These are super simple. They even take precedence over crazy
instructions like INSERTPS because they have very high throughput on
modern x86 chips.
I still have to teach the integer shuffle variants about this to avoid
so many domain crossings. However, due to the particular instructions
available, that's a touch more complex and so a separate patch.
Also, the backend doesn't seem to realize it can commute blend
instructions by negating the mask. That would help remove a number of
copies here. Suggestions on how to do this welcome, it's an area I'm
less familiar with.
llvm-svn: 217744
support transforming the forms from the new vector shuffle lowering to
use 'movddup' when appropriate.
A bunch of the cases where we actually form 'movddup' don't actually
show up in the test results because something even later than DAG
legalization maps them back to 'unpcklpd'. If this shows back up as
a performance problem, I'll probably chase it down, but it is at least
an encoded size loss. =/
To make this work, also always do this canonicalizing step for floating
point vectors where the baseline shuffle instructions don't provide any
free copies of their inputs. This also causes us to canonicalize
unpck[hl]pd into mov{hl,lh}ps (resp.) which is a nice encoding space
win.
There is one test which is "regressed" by this: extractelement-load.
There, the test case where the optimization it is testing *fails*, the
exact instruction pattern which results is slightly different. This
should probably be fixed by having the appropriate extract formed
earlier in the DAG, but that would defeat the purpose of the test.... If
this test case is critically important for anyone, please let me know
and I'll try to work on it. The prior behavior was actually contrary to
the comment in the test case and seems likely to have been an accident.
llvm-svn: 217738
Inline asm may specify 'U' and 'X' constraints to print a 'u' for an
update-form memory reference, or an 'x' for an indexed-form memory
reference. However, these are really only useful in GCC internal code
generation. In inline asm the operand of the memory constraint is
typically just a register containing the address, so 'U' and 'X' make
no sense.
This patch quietly accepts 'U' and 'X' in inline asm patterns, but
otherwise does nothing. If we ever unexpectedly see a non-register,
we'll assert and sort it out afterwards.
I've added a new test for these constraints; the test case should be
used for other asm-constraints changes down the road.
llvm-svn: 217622
Do
(shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
This is already done for multiplies, but since multiplies
by powers of two are turned into shifts, we also need
to handle it here.
This might want checks for isLegalAddImmediate to avoid
transforming an add of a legal immediate with one that isn't.
llvm-svn: 217610
r189189 implemented AVX512 unpack by essentially performing a 256-bit unpack
between the low and the high 256 bits of src1 into the low part of the
destination and another unpack of the low and high 256 bits of src2 into the
high part of the destination.
I don't think that's how unpack works. AVX512 unpack simply has more 128-bit
lanes but other than it works the same way as AVX. So in each 128-bit lane,
we're always interleaving certain parts of both operands rather different
parts of one of the operands.
E.g. for this:
__v16sf a = { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 };
__v16sf b = { 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31 };
__v16sf c = __builtin_shufflevector(a, b, 0, 8, 1, 9, 4, 12, 5, 13, 16,
24, 17, 25, 20, 28, 21, 29);
we generated punpcklps (notice how the elements of a and b are not interleaved
in the shuffle). In turn, c was set to this:
0 16 1 17 4 20 5 21 8 24 9 25 12 28 13 29
Obviously this should have just returned the mask vector of the shuffle
vector.
I mostly reverted this change and made sure the original AVX code worked
for 512-bit vectors as well.
Also updated the tests because they matched the logic from the code.
llvm-svn: 217602
This is an extension of the change made with r215820:
http://llvm.org/viewvc/llvm-project?view=revision&revision=215820
That patch allowed combining of splatted vector FP constants that are multiplied.
This patch allows combining non-uniform vector FP constants too by relaxing the
check on the type of vector. Also, canonicalize a vector fmul in the
same way that we already do for scalars - if only one operand of the fmul is a
constant, make it operand 1. Otherwise, we miss potential folds.
This fold is also done by -instcombine, but it's possible that extra
fmuls may have been generated during lowering.
Differential Revision: http://reviews.llvm.org/D5254
llvm-svn: 217599
Now that the operations are all implemented, we can test this sub-arch here.
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Matt Arsenault <matthew.arsenault@amd.com>
llvm-svn: 217595
David Blaikie's commits r217563 & r217564, which added shared_ptr to the
CostPool have fixed some memory leak issues exposed by the PBQP with
coalescing constraints.
The sanitizer bot was failing because of those leaks. Now that the leaks
are gone, we can reenable the aarch64/pbqp test.
llvm-svn: 217580
This adds target specific support for using the PBQP register allocator on the
AArch64, for the A57 cpu.
By default, the PBQP allocator is not used, unless explicitely required
on the command line with "-aarch64-pbqp".
llvm-svn: 217504
using static relocation model and small code model.
Summary: currently we generate GOT based relocations for weak symbol
references regardless of the underlying relocation model. This should
be change so that in static relocation model we use a constant pool
load instead.
Patch from: Keith Walker
Reviewers: Renato Golin, Tim Northover
llvm-svn: 217503
The only Thumb-1 multi-store capable of using LR is the PUSH instruction, which
translates to STMDB, so we shouldn't convert STMIAs.
Patch by Sergey Dmitrouk.
llvm-svn: 217498
Summary:
In AT&T annotation for both x86_64 and x32 calls should be printed as
callq in assembly. It's only a matter of correct mnemonic, object output
is ok.
Test Plan: trivial test added
Reviewers: nadav, dschuff, craig.topper
Subscribers: llvm-commits, zinovy.nis
Differential Revision: http://reviews.llvm.org/D5213
llvm-svn: 217435
When compiling without SSE2, isTruncStoreLegal(F64, F32) would return Legal, whereas with SSE2 it would return Expand. And since the Target doesn't seem to actually handle a truncstore for double -> float, it would just output a store of a full double in the space for a float hence overwriting other bits on the stack.
Patch by Luqman Aden!
llvm-svn: 217410
Previously, fast-isel would not clean up after failing to select a call
instruction, because it would have called flushLocalValueMap() which moves
the insertion point, making SavedInsertPt in selectInstruction() invalid.
Fixing this by making SavedInsertPt a member variable, and having
flushLocalValueMap() update it.
This removes some redundant code at -O0, and more importantly fixes PR20863.
Differential Revision: http://reviews.llvm.org/D5249
llvm-svn: 217401
support for MOVDDUP which is really important for matrix multiply style
operations that do lots of non-vector-aligned load and splats.
The original motivation was to add support for MOVDDUP as the lack of it
regresses matmul_f64_4x4 by 5% or so. However, all of the rules here
were somewhat suspicious.
First, we should always be using the floating point domain shuffles,
regardless of how many copies we have to make as a movapd is *crazy*
faster than the domain switching cost on some chips. (Mostly because
movapd is crazy cheap.) Because SHUFPD can't do the copy-for-free trick
of the PSHUF instructions, there is no need to avoid canonicalizing on
UNPCK variants, so do that canonicalizing. This also ensures we have the
chance to form MOVDDUP. =]
Second, we assume SSE2 support when doing any vector lowering, and given
that we should just use UNPCKLPD and UNPCKHPD as they can operate on
registers or memory. If vectors get spilled or come from memory at all
this is going to allow the load to be folded into the operation. If we
want to optimize for encoding size (the only difference, and only
a 2 byte difference) it should be done *much* later, likely after RA.
llvm-svn: 217332
parsing (and latent bug in the instruction definitions).
This is effectively a revert of r136287 which tried to address
a specific and narrow case of immediate operands failing to be accepted
by x86 instructions with a pretty heavy hammer: it introduced a new kind
of operand that behaved differently. All of that is removed with this
commit, but the test cases are both preserved and enhanced.
The core problem that r136287 and this commit are trying to handle is
that gas accepts both of the following instructions:
insertps $192, %xmm0, %xmm1
insertps $-64, %xmm0, %xmm1
These will encode to the same byte sequence, with the immediate
occupying an 8-bit entry. The first form was fixed by r136287 but that
broke the prior handling of the second form! =[ Ironically, we would
still emit the second form in some cases and then be unable to
re-assemble the output.
The reason why the first instruction failed to be handled is because
prior to r136287 the operands ere marked 'i32i8imm' which forces them to
be sign-extenable. Clearly, that won't work for 192 in a single byte.
However, making thim zero-extended or "unsigned" doesn't really address
the core issue either because it breaks negative immediates. The correct
fix is to make these operands 'i8imm' reflecting that they can be either
signed or unsigned but must be 8-bit immediates. This patch backs out
r136287 and then changes those places as well as some others to use
'i8imm' rather than one of the extended variants.
Naturally, this broke something else. The custom DAG nodes had to be
updated to have a much more accurate type constraint of an i8 node, and
a bunch of Pat immediates needed to be specified as i8 values.
The fallout didn't end there though. We also then ceased to be able to
match the instruction-specific intrinsics to the instructions so
modified. Digging, this is because they too used i32 rather than i8 in
their signature. So I've also switched those intrinsics to i8 arguments
in line with the instructions.
In order to make the intrinsic adjustments of course, I also had to add
auto upgrading for the intrinsics.
I suspect that the intrinsic argument types may have led everything down
this rabbit hole. Pretty happy with the result.
llvm-svn: 217310
computation was totally wrong, but somehow it didn't really show up with
llc.
I've added an assert that triggers on multiple existing test cases and
updated one of them to show the correct value.
There appear to still be more bugs lurking around insertps's mask. =/
However, note that this only really impacts the new vector shuffle
lowering.
llvm-svn: 217289
This problem is bigger than just fsub, but this is the minimum fix to solve
fneg for PR20556 ( http://llvm.org/bugs/show_bug.cgi?id=20556 ), and we solve
zero subtraction with the same change.
llvm-svn: 217286
This reverts commit r217211.
Both the bfd ld and gold outputs were valid. They were using a Rela relocation,
so the value present in the relocated location was not used, which caused me
to misread the output.
llvm-svn: 217264
shuffle lowering for integer vectors and share it from v4i32, v8i16, and
v16i8 code paths.
Ironically, the SSE2 v16i8 code for this is now better than the SSSE3!
=] Will have to fix the SSSE3 code next to just using a single pshufb.
llvm-svn: 217240
Patched by Sergey Dmitrouk.
This pass tries to make consecutive compares of values use same operands to
allow CSE pass to remove duplicated instructions. For this it analyzes
branches and adjusts comparisons with immediate values by converting:
GE -> GT
GT -> GE
LT -> LE
LE -> LT
and adjusting immediate values appropriately. It basically corrects two
immediate values towards each other to make them equal.
llvm-svn: 217220
Follow up to r217138, extending the logic to other NEON-immediate instructions.
As before, the instruction already performs the correct operation and we're
just using a different type for convenience, so we want a true nop-cast.
Patch by Asiri Rathnayake.
llvm-svn: 217159
We were materialising big-endian constants using DAG nodes with types different
from what was requested, followed by a bitcast. This is fine on little-endian
machines where bitcasting is a nop, but we need a slightly different
representation for big-endian. This adds a new set of NVCAST (natural-vector
cast) operations which are always nops.
Patch by Asiri Rathnayake.
llvm-svn: 217138
vzext patterns and insert-element patterns that for SSE4 have dedicated
instructions.
With this we can enable the experimental mode in a regression test that
happens to cover some of the past set of issues. You can see that the
new logic does significantly better here on the floating point cases.
A follow-up to this change and the previous ones will hoist the logic
into helpers so it can be shared across element type sizes as in this
particular case it generalizes cleanly.
llvm-svn: 217136
abilities of INSERTPS which are really powerful and come up in very
important contexts such as forming diagonal matrices, etc.
With this I ended up being able to remove the somewhat weird helper
I added for INSERTPS because we can collapse the entire state to a no-op
mask. Added a bunch of tests for inserting into a zero-ish vector.
llvm-svn: 217117
'insertps' patterns.
This replaces two shuffles with a single insertps in very common cases.
My next patch will extend this to leverage the zeroing capabilities of
insertps which will allow it to be used in a much wider set of cases.
llvm-svn: 217100
This CL replaces the constant DarwinX86AsmBackend.PushInstrSize with a method
that lets the backend account for different sizes of "push %reg" instruction
sizes.
llvm-svn: 217020
This reapplies r216805 with a fix to a copy-past error, which resulted in an
incorrect register class.
Original commit message:
Select the correct register class for the various instructions that are
generated when combining instructions and constrain the registers to the
appropriate register class.
This fixes rdar://problem/18183707.
llvm-svn: 217019
There is already target-dependent instruction selection support for Adds/Subs to
support compares and the intrinsics with overflow check. This takes advantage of
the existing infrastructure to also support Add/Sub, which allows the folding of
immediates, sign-/zero-extends, and shifts.
This fixes rdar://problem/18207316.
llvm-svn: 217007
This uses the target-dependent selection code for shifts first, which allows us
to create better code for shifts with immediates and sign-/zero-extend folding.
Vector type are not handled yet and the code falls back to target-independent
instruction selection for these cases.
This fixes rdar://problem/17907920.
llvm-svn: 216985
The only valid lowering of atomic stores in the X86 backend was mov from
register to memory. As a result, storing an immediate required a useless copy
of the immediate in a register. Now these can be compiled as a simple mov.
Similarily, adding/and-ing/or-ing/xor-ing an
immediate to an atomic location (but through an atomic_store/atomic_load,
not a fetch_whatever intrinsic) can now make use of an 'add $imm, x(%rip)'
instead of using a register. And the same applies to inc/dec.
This second point matches the first issue identified in
http://llvm.org/bugs/show_bug.cgi?id=17281
llvm-svn: 216980
If an fmul was introduced by lowering, it wouldn't be folded
into a multiply by a constant since the earlier combine would
have replaced the fmul with the fadd.
llvm-svn: 216932
When I recommitted r208640 (in r216898) I added an exclusion for TargetConstant
offsets, as there is no guarantee that a backend can handle them on generic
ADDs (even if it generates them during address-mode matching) -- and,
specifically, applying this transformation directly with TargetConstants caused
a self-hosting failure on PPC64. Ignoring all TargetConstants, however, is less
than ideal. Instead, for non-opaque constants, we can convert them into regular
constants for use with the generated ADD (or SUB).
llvm-svn: 216908
We have been using .init-array for most systems for quiet some time,
but tools like llc are still defaulting to .ctors because the old
option was never changed.
This patch makes llc default to .init-array and changes the option to
be -use-ctors.
Clang is not affected by this. It has its own fancier logic.
llvm-svn: 216905
I reverted r208640 in r209747 because r208640 broke self-hosting on PPC64. The
underlying cause of the failure is that pre-inc loads with increments
represented by ISD::TargetConstants were being transformed into ISD:::ADDs with
ISD::TargetConstant operands. PPC doesn't have a pattern for those, and so they
were selected as invalid r+r adds.
This recommits r208640, rebased and with an exclusion for ISD::TargetConstant
increments. This behavior seems correct, although in the future we might want
to ask the target to split out the indexing that uses ISD::TargetConstants.
Unfortunately, I don't yet have small test case where the relevant invalid
'add' instruction is not itself dead (and thus eliminated by
DeadMachineInstructionElim -- sometimes bugpoint is too good at removing things)
Original commit message (by Adam Nemet):
Right now the load may not get DCE'd because of the side-effect of updating
the base pointer.
This can happen if we lower a read-modify-write of an illegal larger type
(e.g. i48) such that the modification only affects one of the subparts (the
lower i32 part but not the higher i16 part). See the testcase.
In order to spot the dead load we need to revisit it when SimplifyDemandedBits
decided that the value of the load is masked off. This is the
CommitTargetLoweringOpt piece.
I checked compile time with ARM64 by sending SPEC bitcode files through llc.
No measurable change.
Fixes <rdar://problem/16031651>
llvm-svn: 216898
Summary:
Fixes a FIXME in MachineSinking. Instead of using the simple heuristics
in isPostDominatedBy, use the real MachinePostDominatorTree. The old
heuristics caused instructions to sink unnecessarily, and might create
register pressure.
Test Plan:
Added a NVPTX codegen test to verify that our change is in effect. It also
shows the unnecessary register pressure caused by over-sinking. Updated
affected tests in AArch64 and X86.
Reviewers: eliben, meheff, Jiangning
Reviewed By: Jiangning
Subscribers: jholewinski, aemerson, mcrosier, llvm-commits
Differential Revision: http://reviews.llvm.org/D4814
llvm-svn: 216862
Select the correct register class for the various instructions that are
generated when combining instructions and constrain the registers to the
appropriate register class.
This fixes rdar://problem/18183707.
llvm-svn: 216805
When sinking an instruction it might be moved past the original last use of one
of its operands. This last use has the kill flag set and the verifier will
obviously complain about this.
Before Machine Sinking (AArch64):
%vreg3<def> = ASRVXr %vreg1, %vreg2<kill>
%XZR<def> = SUBSXrs %vreg4, %vreg1<kill>, 160, %NZCV<imp-def>
...
After Machine Sinking:
%XZR<def> = SUBSXrs %vreg4, %vreg1<kill>, 160, %NZCV<imp-def>
...
%vreg3<def> = ASRVXr %vreg1, %vreg2<kill>
This fix clears all the kill flags in all instruction that use the same operands
as the instruction that is being sunk.
This fixes rdar://problem/18180996.
llvm-svn: 216803
Summary:
If a variadic function body contains a musttail call, then we copy all
of the remaining register parameters into virtual registers in the
function prologue. We track the virtual registers through the function
body, and add them as additional registers to pass to the call. Because
this is all done in virtual registers, the register allocator usually
gives us good code. If the function does a call, however, it will have
to spill and reload all argument registers (ew).
Forwarding regparms on x86_32 is not implemented because most compilers
don't support varargs in 32-bit with regparms.
Reviewers: majnemer
Subscribers: aemerson, llvm-commits
Differential Revision: http://reviews.llvm.org/D5060
llvm-svn: 216780
We've rejected these kinds of functions since r28405 in 2006 because
it's impossible to lower the return of a callee cleanup varargs
function. However there are lots of legal ways to leave such a function
without returning, such as aborting. Today we can leave a function with
a musttail call to another function with the correct prototype, and
everything works out.
I'm removing the verifier check declaring that a normal return from such
a function is UB.
Reviewed By: nlewycky
Differential Revision: http://reviews.llvm.org/D5059
llvm-svn: 216779
This patch checks for DAG patterns that are an add or a sub followed by a
compare on 16 and 8 bit inputs. Since AArch64 does not support those types
natively they are legalized into 32 bit values, which means that mask operations
are inserted into the DAG to emulate overflow behaviour. In many cases those
masks do not change the result of the processing and just introduce a dependent
operation, often in the middle of a hot loop.
This patch detects the relevent DAG patterns and then tests to see if the
transforms are equivalent with and without the mask, removing the mask if
possible. The exact mechanism of this patch was discusses in
http://lists.cs.uiuc.edu/pipermail/llvmdev/2014-July/074444.html
There is a reasonably good chance there are missed oppurtunities due to similiar
(but not identical) DAG patterns that could be funneled into this test, adding
them should be simple if we see test cases.
Tests included.
rdar://13754426
llvm-svn: 216776
The new solution is to not use this lowering if there are any dynamic
allocas in the current function. We know up front if there are dynamic
allocas, but we don't know if we'll need to create stack temporaries
with large alignment during lowering. Conservatively assume that we will
need such temporaries.
Reviewed By: hans
Differential Revision: http://reviews.llvm.org/D5128
llvm-svn: 216775
When we select a trunc instruction we don't emit any code if the type is already
i32 or smaller. This is because the instruction that uses the truncated value
will deal with it.
This behavior can incorrectly transfer a kill flag, which was meant for the
result of the truncate, onto the source register.
%2 = trunc i32 %1 to i16
... = ... %2 -> ... = ... vreg1 <kill>
... = ... %1 ... = ... vreg1
This commit fixes this by emitting a COPY instruction, so that the result and
source register are distinct virtual registers.
This fixes rdar://problem/18178188.
llvm-svn: 216750
Summary:
Instead of specifying the alignment as metadata which may be destroyed by
transformation passes, make the alignment the second argument to ldu/ldg
intrinsic calls.
Test Plan:
ldu-ldg.ll
ldu-i8.ll
ldu-reg-plus-offset.ll
Reviewers: eliben, meheff, jholewinski
Reviewed By: meheff, jholewinski
Subscribers: jholewinski, llvm-commits
Differential Revision: http://reviews.llvm.org/D5093
llvm-svn: 216731
While working on a Thumb-2 code size optimization I just realized that we don't have any regression tests for it.
So here's a first test case, I plan to increase the coverage over time.
llvm-svn: 216728
In an llvm-stress generated test, we were trying to create a v0iN type and
asserting when that failed. This case could probably be handled by the
function, but not without added complexity and the situation it arises in is
sufficiently odd that there's probably no benefit anyway.
Should fix PR20775.
llvm-svn: 216725
The code in SelectionDAG::getMemset for some reason assumes the value passed to
memset is an i32. This breaks the generated code for targets that only have
registers smaller than 32 bits because the value might get split into multiple
registers by the calling convention. See the test for the MSP430 target included
in the patch for an example.
This patch ensures that nothing is assumed about the type of the value. Instead,
the type is taken from the selected overload of the llvm.memset intrinsic.
llvm-svn: 216716
This fix checks first if the instruction to be folded (e.g. sign-/zero-extend,
or shift) is in the same machine basic block as the instruction we are folding
into.
Not doing so can result in incorrect code, because the value might not be
live-out of the basic block, where the value is defined.
This fixes rdar://problem/18169495.
llvm-svn: 216700
The AArch64 target lowering for [zs]ext of vectors is set up to handle
input simple types and expects the generic SDag path to do something reasonable
with anything that's not a simple type. The code, however, was only
checking that the result type was a simple type and assuming that
implied that the source type would also be a simple type. That's not a
valid assumption, as operations like "zext <1 x i1> %0 to <1 x i32>"
demonstrate. The fix is to simply explicitly validate the source type
as well as the result type.
PR20791
llvm-svn: 216689
On MachO, putting a symbol that doesn't start with a 'L' or 'l' in one of the
__TEXT,__literal* sections prevents the linker from merging the context of the
section.
Since private GVs are the ones the get mangled to start with 'L' or 'l', we now
only put those on the __TEXT,__literal* sections.
llvm-svn: 216682
file.
Changing code that is covered by these tests is just too hard to debug
currently, and now it will be clear the nature of the changes.
llvm-svn: 216643
The included test case would fail, because the MI PHI node would have two
operands from the same predecessor.
This problem occurs when a switch instruction couldn't be selected. This happens
always, because there is no default switch support for FastISel to begin with.
The problem was that FastISel would first add the operand to the PHI nodes and
then fall-back to SelectionDAG, which would then in turn add the same operands
to the PHI nodes again.
This fix removes these duplicate PHI node operands by reseting the
PHINodesToUpdate to its original state before FastISel tried to select the
instruction.
This fixes <rdar://problem/18155224>.
llvm-svn: 216640
Currently instructions are folded very aggressively for AArch64 into the memory
operation, which can lead to the use of killed operands:
%vreg1<def> = ADDXri %vreg0<kill>, 2
%vreg2<def> = LDRBBui %vreg0, 2
... = ... %vreg1 ...
This usually happens when the result is also used by another non-memory
instruction in the same basic block, or any instruction in another basic block.
This fix teaches hasTrivialKill to not only check the LLVM IR that the value has
a single use, but also to check if the register that represents that value has
already been used. This can happen when the instruction with the use was folded
into another instruction (in this particular case a load instruction).
This fixes rdar://problem/18142857.
llvm-svn: 216634
Currently instructions are folded very aggressively into the memory operation,
which can lead to the use of killed operands:
%vreg1<def> = ADDXri %vreg0<kill>, 2
%vreg2<def> = LDRBBui %vreg0, 2
... = ... %vreg1 ...
This usually happens when the result is also used by another non-memory
instruction in the same basic block, or any instruction in another basic block.
If the computed address is used by only memory operations in the same basic
block, then it is safe to fold them. This is because all memory operations will
fold the address computation and the original computation will never be emitted.
This fixes rdar://problem/18142857.
llvm-svn: 216629
When the address comes directly from a shift instruction then the address
computation cannot be folded into the memory instruction, because the zero
register is not available as a base register. Simplify addess needs to emit the
shift instruction and use the result as base register.
llvm-svn: 216621
Use the zero register directly when possible to avoid an unnecessary register
copy and a wasted register at -O0. This also uses integer stores to store a
positive floating-point zero. This saves us from materializing the positive zero
in a register and then storing it.
llvm-svn: 216617
This teaches the AArch64 backend to deal with the operations required
to deal with the operations on v4f16 and v8f16 which are exposed by
NEON intrinsics, plus the add, sub, mul and div operations.
llvm-svn: 216555
we stopped efficiently lowering sextload using the SSE41 instructions
for that operation.
This is a consequence of a bad predicate I used thinking of the memory
access needs. The code actually handles the cases where the predicate
doesn't apply, and handles them much better. =] Simple fix and a test
case added. Fixes PR20767.
llvm-svn: 216538
This combine is essentially combining target-specific nodes back into target
independent nodes that it "knows" will be combined yet again by a target
independent DAG combine into a different set of target-independent nodes that
are legal (not custom though!) and thus "ok". This seems... deeply flawed. The
crux of the problem is that we don't combine un-legalized shuffles that are
introduced by legalizing other operations, and thus we don't see a very
profitable combine opportunity. So the backend just forces the input to that
combine to re-appear.
However, for this to work, the conditions detected to re-form the unlegalized
nodes must be *exactly* right. Previously, failing this would have caused poor
code (if you're lucky) or a crasher when we failed to select instructions.
After r215611 we would fall back into the legalizer. In some cases, this just
"fixed" the crasher by produces bad code. But in the test case added it caused
the legalizer and the dag combiner to iterate forever.
The fix is to make the alignment checking in the x86 side of things match the
alignment checking in the generic DAG combine exactly. This isn't really a
satisfying or principled fix, but it at least make the code work as intended.
It also highlights that it would be nice to detect the availability of under
aligned loads for a given type rather than bailing on this optimization. I've
left a FIXME to document this.
Original commit message for r215611 which covers the rest of the chang:
[SDAG] Fix a case where we would iteratively legalize a node during
combining by replacing it with something else but not re-process the
node afterward to remove it.
In a truly remarkable stroke of bad luck, this would (in the test case
attached) end up getting some other node combined into it without ever
getting re-processed. By adding it back on to the worklist, in addition
to deleting the dead nodes more quickly we also ensure that if it
*stops* being dead for any reason it makes it back through the
legalizer. Without this, the test case will end up failing during
instruction selection due to an and node with a type we don't have an
instruction pattern for.
It took many million runs of the shuffle fuzz tester to find this.
llvm-svn: 216537
When a shift with extension or an add with shift and extension cannot be folded
into the memory operation, then the address calculation has to be materialized
separately. While doing so the code forgot to consider a possible sign-/zero-
extension. This fix folds now also the sign-/zero-extension into the add or
shift instruction which is used to materialize the address.
This fixes rdar://problem/18141718.
llvm-svn: 216511
Adds code generation support for dcbtst (data cache prefetch for write) and
icbt (instruction cache prefetch for read - Book E cores only).
We still end up with a 'cannot select' error for the non-supported prefetch
intrinsic forms. This will be fixed in a later commit.
Fixes PR20692.
llvm-svn: 216339
This reverts commit r215862 due to nightly failures. Will work on getting a
reduced test case, but I wanted to get our bots green in the meantime.
llvm-svn: 216325
these DAG combines.
The DAG auto-CSE thing is truly terrible. Due to it, when RAUW-ing
a node with its operand, you can cause its uses to CSE to itself, which
then causes their uses to become your uses which causes them to be
picked up by the RAUW. For nodes that are determined to be "no-ops",
this is "fine". But if the RAUW is one of several steps to enact
a transformation, this causes the DAG to really silently eat an discard
nodes that you would never expect. It took days for me to actually
pinpoint a test case triggering this and a really frustrating amount of
time to even comprehend the bug because I never even thought about the
ability of RAUW to iteratively consume nodes due to CSE-ing them into
itself.
To fix this, we have to build up a brand-new chain of operations any
time we are combining across (potentially) intervening nodes. But once
the logic is added to do this, another issue surfaces: CombineTo eagerly
deletes the one node combined, *but no others*. This is... really
frustrating. If deleting it makes its operands become dead, those
operand nodes often won't go onto the worklist in the
order you would want -- they're already on it and not near the top. That
means things higher on the worklist will get combined prior to these
dead nodes being GCed out of the worklist, and if the chain is long, the
immediate users won't be enough to re-detect where the root of the chain
is that became single-use again after deleting the dead nodes. The
better way to do this is to never immediately delete nodes, and instead
to just enqueue them so we can recursively delete them. The
combined-from node is typically not on the worklist anyways by virtue of
having been popped off.... But that in turn breaks other tests that
*require* CombineTo to delete unused nodes. :: sigh ::
Fortunately, there is a better way. This whole routine should have been
returning the replacement rather than using CombineTo which is quite
hacky. Switch to that, and all the pieces fall together.
I suspect the same kind of miscompile is possible in the half-shuffle
folding code, and potentially the recursive folding code. I'll be
switching those over to a pattern more like this one for safety's sake
even though I don't immediately have any test cases for them. Note that
the only way I got a test case for this instance was with *heavily* DAG
combined 256-bit shuffle sequences generated by my fuzzer. ;]
llvm-svn: 216319
There's no need to do this if the user doesn't call va_start. In the
future, we're going to have thunks that forward these register
parameters with musttail calls, and they won't need these spills for
handling va_start.
Most of the test suite changes are adding va_start calls to existing
tests to keep things working.
llvm-svn: 216294
instruction from ARMInstrInfo to ARMBaseInstrInfo.
That way, thumb mode can also benefit from the advanced copy optimization.
<rdar://problem/12702965>
llvm-svn: 216274
This is mostly achieved by providing the correct register class manually,
because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and
MVT::i64.
Also cleanup the code to use the FastEmitInst_* method whenever possible. This
makes sure that the operands' register class is properly constrained. For all
the remaining cases this adds the missing constrainOperandRegClass calls for
each operand.
llvm-svn: 216225
The AdvSIMD pass may produce copies that are not coalescer-friendly. The
peephole optimizer knows how to fix that as demonstrated in the test case.
<rdar://problem/12702965>
llvm-svn: 216200
There are two add-immediate instructions in Thumb1: tADDi8 and tADDi3. Only
the latter supports using different source and destination registers, so
whenever we materialize a new base register (at a certain offset) we'd do
so by moving the base register value to the new register and then adding in
place. This patch changes the code to use a single tADDi3 if the offset is
small enough to fit in 3 bits.
Differential Revision: http://reviews.llvm.org/D5006
llvm-svn: 216193
The FPv4-SP floating-point unit is generally referred to as
single-precision only, but it does have double-precision registers and
load, store and GPR<->DPR move instructions which operate on them.
This patch enables the use of these registers, the main advantage of
which is that we now comply with the AAPCS-VFP calling convention.
This partially reverts r209650, which added some AAPCS-VFP support,
but did not handle return values or alignment of double arguments in
registers.
This patch also adds tests for Thumb2 code generation for
floating-point instructions and intrinsics, which previously only
existed for ARM.
llvm-svn: 216172
advanced copy optimization.
This is the final step patch toward transforming:
udiv r0, r0, r2
udiv r1, r1, r3
vmov.32 d16[0], r0
vmov.32 d16[1], r1
vmov r0, r1, d16
bx lr
into:
udiv r0, r0, r2
udiv r1, r1, r3
bx lr
Indeed, thanks to this patch, this optimization is able to look through
vmov.32 d16[0], r0
vmov.32 d16[1], r1
and is able to rewrite the following sequence:
vmov.32 d16[0], r0
vmov.32 d16[1], r1
vmov r0, r1, d16
into simple generic GPR copies that the coalescer managed to remove.
<rdar://problem/12702965>
llvm-svn: 216144
On pre-v6 hardware, 'MOV lo, lo' gives undefined results, so such copies need to
be avoided. This patch trades simplicity for implementation time at the expense
of performance... As they say: correctness first, then performance.
See http://lists.cs.uiuc.edu/pipermail/llvmdev/2014-August/075998.html for a few
ideas on how to make this better.
llvm-svn: 216138
Fix for PR20648 - http://llvm.org/bugs/show_bug.cgi?id=20648
This patch checks the operands of a vselect to see if all values are constants.
If yes, bail out of any further attempts to create a blend or shuffle because
SelectionDAGLegalize knows how to turn this kind of vselect into a single load.
This already happens for machines without SSE4.1, so the added checks just send
more targets down that path.
Differential Revision: http://reviews.llvm.org/D4934
llvm-svn: 216121
The goal of the patch is to implement section 3.2.3 of the AMD64 ABI
correctly. The controlling sentence is, "The size of each argument gets
rounded up to eightbytes. Therefore the stack will always be eightbyte
aligned." The equivalent sentence in the i386 ABI page 37 says, "At all
times, the stack pointer should point to a word-aligned area." For both
architectures, the stack pointer is not being rounded up to the nearest
eightbyte or word between the last normal argument and the first
variadic argument.
Patch by Thomas Jablin!
llvm-svn: 216119
Summary: This fixes http://llvm.org/bugs/show_bug.cgi?id=19530.
The problem is that X86ISelLowering erroneously thought the third call
was eligible for tail call elimination.
It would have been if it's return value was actually the one returned
by the calling function, but here that is not the case and
additional values are being returned.
Test Plan: Test case from the original bug report is included.
Reviewers: rafael
Reviewed By: rafael
Subscribers: rafael, llvm-commits
Differential Revision: http://reviews.llvm.org/D4968
llvm-svn: 216117
In PR20308 ( http://llvm.org/bugs/show_bug.cgi?id=20308 ), the critical-anti-dependency breaker
caused a miscompile because it broke a WAR hazard using a register that it thinks is available
based on info from a kill inst. Until PR18663 is solved, we shouldn't use any def/use info from
a kill because they are really just nops.
This patch adds guard checks for kills around calls to ScanInstruction() where the DefIndices
array is set. For good measure, add an assert in ScanInstruction() so we don't hit this bug again.
The test case is a reduced version of the code from the bug report.
Differential Revision: http://reviews.llvm.org/D4977
llvm-svn: 216114
the isRegSequence property.
This is a follow-up of r215394 and r215404, which respectively introduces the
isRegSequence property and uses it for ARM.
Thanks to the property introduced by the previous commits, this patch is able
to optimize the following sequence:
vmov d0, r2, r3
vmov d1, r0, r1
vmov r0, s0
vmov r1, s2
udiv r0, r1, r0
vmov r1, s1
vmov r2, s3
udiv r1, r2, r1
vmov.32 d16[0], r0
vmov.32 d16[1], r1
vmov r0, r1, d16
bx lr
into:
udiv r0, r0, r2
udiv r1, r1, r3
vmov.32 d16[0], r0
vmov.32 d16[1], r1
vmov r0, r1, d16
bx lr
This patch refactors how the copy optimizations are done in the peephole
optimizer. Prior to this patch, we had one copy-related optimization that
replaced a copy or bitcast by a generic, more suitable (in terms of register
file), copy.
With this patch, the peephole optimizer features two copy-related optimizations:
1. One for rewriting generic copies to generic copies:
PeepholeOptimizer::optimizeCoalescableCopy.
2. One for replacing non-generic copies with generic copies:
PeepholeOptimizer::optimizeUncoalescableCopy.
The goals of these two optimizations are slightly different: one rewrite the
operand of the instruction (#1), the other kills off the non-generic instruction
and replace it by a (sequence of) generic instruction(s).
Both optimizations rely on the ValueTracker introduced in r212100.
The ValueTracker has been refactored to use the information from the
TargetInstrInfo for non-generic instruction. As part of the refactoring, we
switched the tracking from the index of the definition to the actual register
(virtual or physical). This one change is to provide better consistency with
register related APIs and to ease the use of the TargetInstrInfo.
Moreover, this patch introduces a new helper class CopyRewriter used to ease the
rewriting of generic copies (i.e., #1).
Finally, this patch adds a dead code elimination pass right after the peephole
optimizer to get rid of dead code that may appear after rewriting.
This is related to <rdar://problem/12702965>.
Review: http://reviews.llvm.org/D4874
llvm-svn: 216088
This fixes a bug I introduced in a previous commit (r216033). Sign-/Zero-
extension from i1 cannot be folded into the ADDS/SUBS instructions. Instead both
operands have to be sign-/zero-extended with separate instructions.
Related to <rdar://problem/17913111>.
llvm-svn: 216073
legalization stage. With those two optimizations, fewer signed/zero extension
instructions can be inserted, and then we can expose more opportunities to
Machine CSE pass in back-end.
llvm-svn: 216066
Summary:
Fixes http://llvm.org/bugs/show_bug.cgi?id=20016 reproducible on new
lea-5.ll case.
Also use RSP/RBP for x32 lea to save 1 byte used for 0x67 prefix in
ESP/EBP case.
Test Plan: lea tests modified to include x32/nacl and new test added
Reviewers: nadav, dschuff, t.p.northover
Subscribers: llvm-commits, zinovy.nis
Differential Revision: http://reviews.llvm.org/D4929
llvm-svn: 216065
LLVM generates illegal `rbit r0, #352` instruction for rbit intrinsic.
According to ARM ARM, rbit only takes register as argument, not immediate.
The correct instruction should be rbit <Rd>, <Rm>.
The bug was originally introduced in r211057.
Differential Revision: http://reviews.llvm.org/D4980
llvm-svn: 216064
Use FMOVWSr/FMOVXDr instead of FMOVSr/FMOVDr, which have the proper register
class to be used with the zero register. This makes the MachineInstruction
verifier happy again.
This is related to <rdar://problem/18027157>.
llvm-svn: 216040
Factor out the ADDS/SUBS instruction emission code into helper functions and
make the helper functions more clever to support most of the different ADDS/SUBS
instructions the architecture support. This includes better immedediate support,
shift folding, and sign-/zero-extend folding.
This fixes <rdar://problem/17913111>.
llvm-svn: 216033
This adds the missing test that I promised for r215753 to test the
materialization of the floating-point value +0.0.
Related to <rdar://problem/18027157>.
llvm-svn: 216019
Note: This was originally reverted to track down a buildbot error. Reapply
without any modifications.
Original commit message:
FastISel didn't take much advantage of the different addressing modes available
to it on AArch64. This commit allows the ComputeAddress method to recognize more
addressing modes that allows shifts and sign-/zero-extensions to be folded into
the memory operation itself.
For Example:
lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3]
ldr x0, [x0, x1]
sxtw x1, w1
lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3]
ldr x0, [x0, x1]
llvm-svn: 216013
Note: This was originally reverted to track down a buildbot error. Reapply
without any modifications.
Original commit message:
In the large code model for X86 floating-point constants are placed in the
constant pool and materialized by loading from it. Since the constant pool
could be far away, a PC relative load might not work. Therefore we first
materialize the address of the constant pool with a movabsq and then load
from there the floating-point value.
Fixes <rdar://problem/17674628>.
llvm-svn: 216012
Note: This was originally reverted to track down a buildbot error. Reapply
without any modifications.
Original commit message:
This mostly affects the i64 value type, which always resulted in an 15byte
mobavsq instruction to materialize any constant. The custom code checks the
value of the immediate and tries to use a different and smaller mov
instruction when possible.
This fixes <rdar://problem/17420988>.
llvm-svn: 216010
Note: This was originally reverted to track down a buildbot error. Reapply
without any modifications.
Original commit message:
This change materializes now the value "0" from the zero register.
The zero register can be folded by several instruction, so no
materialization is need at all.
Fixes <rdar://problem/17924413>.
llvm-svn: 216009
Note: This was originally reverted to track down a buildbot error. This commit
exposed a latent bug that was fixed in r215753. Therefore it is reapplied
without any modifications.
I run it through SPEC2k and SPEC2k6 for AArch64 and it didn't introduce any new
regeressions.
Original commit message:
This changes the order in which FastISel tries to materialize a constant.
Originally it would try to use a simple target-independent approach, which
can lead to the generation of inefficient code.
On X86 this would result in the use of movabsq to materialize any 64bit
integer constant - even for simple and small values such as 0 and 1. Also
some very funny floating-point materialization could be observed too.
On AArch64 it would materialize the constant 0 in a register even the
architecture has an actual "zero" register.
On ARM it would generate unnecessary mov instructions or not use mvn.
This change simply changes the order and always asks the target first if it
likes to materialize the constant. This doesn't fix all the issues
mentioned above, but it enables the targets to implement such
optimizations.
Related to <rdar://problem/17420988>.
llvm-svn: 216006
This fixes a few BuildMI callsites where the result register was added by
using addReg, which is per default a use and therefore an operand register.
Also use the zero register as result register when emitting a compare
instruction (SUBS with unused result register).
llvm-svn: 215997
Juergen Ributzka