Moritz Roth [Tue, 16 Sep 2014 16:25:07 +0000 (16:25 +0000)]
ARM load/store optimizer: Don't materialize a new base register with
ADDS/SUBS unless it's safe to clobber the condition flags.
If the merged instructions are in a range where the CPSR is live,
e.g. between a CMP -> Bcc, we can't safely materialize a new base
register.
This problem is quite rare, I couldn't come up with a test case and I've
never actually seen this happen in the tests I'm running - there is a
potential trigger for this in LNT/oggenc (spills being inserted between
a CMP/Bcc), but at the moment this isn't being merged. I'll try to
reduce that into a small test case once I've committed my upcoming patch
to make merging less conservative.
[mips] Improve the error messages given by MipsAsmParser.
Summary: Changed error messages to be more informative and to resemble other clang/llvm error messages (first letter is lower case, no ending punctuation) and updated corresponding tests.
Joe Abbey [Tue, 16 Sep 2014 09:18:23 +0000 (09:18 +0000)]
ARMAsmBackend uses a factory method to generate binary file format specific
objects. There were a few FIXMEs in ARMAsmBackend.cpp suggesting the class
definitions should be in a separate file. Starting with ARMAsmBackend, the
class definition has been put in a header file, and #includes reduced. Each
sub-type of ARMAsmBackend is now in its own header file.
Derived types have been painted with a different color of bike-shed:
Hal Finkel [Tue, 16 Sep 2014 04:35:50 +0000 (04:35 +0000)]
Fix BasicTTI::getCmpSelInstrCost to deal with illegal vector types
The default implementation of getCmpSelInstrCost, which provides the cost of
icmp/fcmp/select instructions, did not deal sensibly with illegal vector types
that were scalarized. We'd ask for the legalization cost of the vector type,
which would return something like (4, f64) given an input of <4 x double>, and
we'd then check the TLI status of the ISD opcode on that scalar type. This would
result in querying (ISD::VSELECT, f64), for example. Amusingly enough,
ISD::VSELECT on scalar types is marked as Legal by default (as with most other
operations), and most backends never change this because VSELECT is never
generated on scalars. However, seeing the resulting operation as Legal, we'd
neglect to add the scalarization cost before returning. The result is that we'd
grossly under-estimate the cost of cmps/selects on illegal vector types.
Now, if type legalization clearly results in scalarization, we skip the early
return and add the scalarization cost.
David Majnemer [Tue, 16 Sep 2014 03:52:46 +0000 (03:52 +0000)]
yaml2obj: Support bigobj
Teach yaml2obj how to make a bigobj COFF file. Like the rest of LLVM,
we automatically decide whether or not to use regular COFF or bigobj
COFF on the fly depending on how many sections the resulting object
would have.
This ends the task of adding bigobj support to LLVM.
N.B. This was tested by forcing yaml2obj to be used in bigobj mode
regardless of the number of sections. While a dedicated test was
written, the smallest I could make it was 36 MB (!) of yaml and it still
took a significant amount of time to execute on a powerful machine.
[x86] Remove a FIXME that doesn't make any sense. Only the lanes feeding
the blend that is matched by this are "used" in any sense, and so any
build_vector or other nodes feeding these will already drop other lanes.
[x86] Remove the last vestiges of the BLENDI-based ADDSUB pattern
matching. This design just fundamentally didn't work because ADDSUB is
available prior to any legal lowerings of BLENDI nodes. Instead, we have
a dedicated ADDSUB synthetic ISD node which is pattern matched trivially
into the instructions. These nodes are then recognized by both the
existing and a trivial new lowering combine in the backend. Removing
these patterns required adding 2 missing shuffle masks to the DAG
combine, without which tests would have failed. Added the masks and
a helpful assert as well to catch if anything ever goes wrong here.
[x86] As a follow-up to r217819, don't check for VSELECT legality now
that we don't use VSELECT and directly emit an addsub synthetic node.
Also remove a stale comment referencing VSELECT.
The test case is updated to use 'core2' which only has SSE3, not SSE4.1,
and it still passes. Previously it would not because we lacked
sufficient blend support to legalize the VSELECT.
[x86] Add the beginnings of a proper DAG combine to match ADDSUBPS and
ADDSUBPD nodes out of blends of adds and subs.
This allows us to actually form these instructions with SSE3 rather than
only forming them when we had both SSE3 for the ADDSUB instructions and
SSE4.1 for the blend instructions. ;] Kind-of important.
I've adjusted the CPU requirements on one of the tests to demonstrate
this kicking in nicely for an SSE3 cpu configuration.
[FastISel][AArch64] Allow handling of vectors during return lowering for little endian machines.
Allow handling of vectors during return lowering at least for little endian machines.
This was restricted in r208200 to fix it for big endian machines (according to
the comment), but it also disabled it for little endian too.
llvm-cov: Fix an issue with showing regions but not counts
In r217746, though it was supposed to be NFC, I broke llvm-cov's
handling of showing regions without showing counts. This should've
shown up in the existing tests, except they were checking debug output
that was displayed regardless of what was actually output. I've moved
the relevant debug output to a more appropriate place so that the
tests catch this kind of thing.
Nick Kledzik [Mon, 15 Sep 2014 21:51:49 +0000 (21:51 +0000)]
[Support] add decodeSLEB128()
We already have routines to encode SLEB128 as well as encode/decode ULEB128.
This last function fills out the matrix. I'll need this for some llvm-objdump
work I am doing.
Add mips32 r1 to the list of supported targets for Mips fast-isel
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
[x86] Start fixing our emission of ADDSUBPS and ADDSUBPD instructions by
introducing a synthetic X86 ISD node representing this generic
operation.
The relevant patterns for mapping these nodes into the concrete
instructions are also added, and a gnarly bit of C++ code in the
target-specific DAG combiner is replaced with simple code emitting this
primitive.
The next step is to generically combine blends of adds and subs into
this node so that we can drop the reliance on an SSE4.1 ISD node
(BLENDI) when matching an SSE3 feature (ADDSUB).
David Majnemer [Mon, 15 Sep 2014 19:42:42 +0000 (19:42 +0000)]
MC: Add support for BigObj
Teach WinCOFFObjectWriter how to write -mbig-obj style object files;
these object files allow for more sections inside an object file.
Our support for BigObj is notably different from binutils and cl: we
implicitly upgrade object files to BigObj instead of asking the user to
compile the same file *again* but with another flag. This matches up
with how LLVM treats ELF variants.
This was tested by forcing LLVM to always emit BigObj files and running
the entire test suite. A specific test has also been added.
I've lowered the maximum number of sections in a normal COFF file,
VS "14" CTP 3 supports no more than 65279 sections. This is important
otherwise we might not switch to BigObj quickly enough, leaving us with
a COFF file that we couldn't link.
yaml2obj support is all that remains to implement.
Rafael Espindola [Mon, 15 Sep 2014 18:32:58 +0000 (18:32 +0000)]
Fix a lot of confusion around inserting nops on empty functions.
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.
[CodeGenPrepare][AddressingModeMatcher] Fix a think-o for the sext(zext) -> zext promotion
introduced in r217629.
We were returning the old sext instead of the new zext as the promoted instruction!
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.
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.
Matt Arsenault [Mon, 15 Sep 2014 16:48:01 +0000 (16:48 +0000)]
R600/SI: Add some mubuf testcases.
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?
[x86] Begin emitting PBLENDW instructions for integer blend operations
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.
[x86] Add an explicit SSE3 run to this test and flesh out a bunch of
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.
[x86] Teach the x86 DAG combiner to form UNPCKLPS and UNPCKHPS
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.
[x86] Undo a flawed transform I added to form UNPCK instructions when
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.
[x86] Teach the new vector shuffle lowering to use 'punpcklwd' and
'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.
Use fully qualified name inside a typedef from llvm::iterator_range<...> to
iterator_range. This is reported (rightly I think) by GCC as an
ambiguous name redefinition. Hope this fixes the buildbots.
David Majnemer [Mon, 15 Sep 2014 08:15:28 +0000 (08:15 +0000)]
InstSimplify: Simplify trivial and/or of icmps
Some ICmpInsts when anded/ored with another ICmpInst trivially reduces
to true or false depending on whether or not all integers or no integers
satisfy the intersected/unioned range.
This sort of trivial looking code can come about when InstCombine
performs a range reduction-type operation on sdiv and the like.
Summary:
replaceAllUsesWith had been modified to allow a DbgNode value to be
replaced by itself. In that case a new node is created by copying the
current DbgNode and the copy is used as replacement value.
When that copying happens, the value stored in this->DbgNode at the end
of RAUW would be a reference to the Node that has just been deleted.
This doesn't produce any bug right now, because the DI node on which we
call RAUW won't be used again.
Move replaceAllUsesWith() from DIType to DIDescriptor.
RAUW was only used on DIType to merge declarations and full definitions
of types. In order to support the same functionality for functions and
global variables, move the function up type DI type hierarchy to the
common parent of DIType, DISubprogram and DIVariable which is
DIDescriptor.
This functionality will be exercized when we add the code to emit
imported declarations for forward declared function/variables.
A DWARFUnitSection is the collection of Units that have been extracted from
the same debug section.
By embeding a reference to their DWARFUnitSection in each unit, the DIEs
will be able to resolve inter-unit references by interrogating their Unit's
DWARFUnitSection.
This is a minimal patch where the DWARFUnitSection is-a SmallVector of Units,
thus exposing exactly the same interface as before. Followup-up patches might
change from inheritance to composition in order to expose only the wanted
DWARFUnitSection abstraction.
llvm-cov: Simplify CounterMappingRegion, pushing logic to its user
A single function in SourceCoverageDataManager was the only user of
some of the comparisons in CounterMappingRegion, and at this point we
know that only one file is relevant. This lets us use slightly simpler
logic directly in the client.
[x86] Teach the new vector shuffle lowering to use BLENDPS and BLENDPD.
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.
[x86] Teach the vector combiner that picks a canonical shuffle from to
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.
In DwarfEHPrepare, after all passes are run, RewindFunction may be a dangling
pointer to a dead function. To make sure it's valid, doFinalization nullptrs
RewindFunction just like the constructor and so it will be found on next run.
James Molloy [Sun, 14 Sep 2014 18:24:26 +0000 (18:24 +0000)]
[A57FPLoadBalancing] Modify r217689 - actually we do need to check defs
... Just make sure we check uses first so we see the kill first. It
turns out ignoring defs gives some pretty nasty runtime failures.
I'm certain this is the fix but I'm still reducing a testcase.
projects / llvm / log