Dan Liew [Fri, 8 Jan 2016 22:36:22 +0000 (22:36 +0000)]
Teach the CMake build system to run lit's test suite. These can be run
directy with ``make check-lit`` and are run as part of
``make check-all``.
In principle we should run lit's testsuite before testing LLVM using lit
so that any problems with lit get discovered before testing LLVM so we
can bail out early. However this implementation (``check-all`` runs all
tests together) seemed simpler and will still report failing lit tests.
Note that the tests and the configured ``lit.site.cfg`` have to be
copied into the build directory to avoid polluting the source tree.
And expand the select into a branch structure. This later enables
jump-threading over bb in this pass.
Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
select if the associated PHI has at least one constant. If the unfolded
select is not jump-threaded, it will be folded again in the later
optimizations.
Justin Bogner [Fri, 8 Jan 2016 19:08:53 +0000 (19:08 +0000)]
LoopInfo: Simplify ownership of Loop objects
It's strange that LoopInfo mostly owns the Loop objects, but that it
defers deleting them to the loop pass manager. Instead, change the
oddly named "updateUnloop" to "markAsRemoved" and have it queue the
Loop object for deletion. We can't delete the Loop immediately when we
remove it, since we need its pointer identity still, so we'll mark the
object as "invalid" so that clients can see what's going on.
Weiming Zhao [Fri, 8 Jan 2016 18:37:43 +0000 (18:37 +0000)]
Disable shrink-wrap for Thumb1
Summary: In ARMConstantIslandPass, which runs after Shrink Wrap pass, long jumps will be fixed up as BL (tBfar) which depends on spilling LR in epilogue. However, shrink-wrap may remove the LR, which causes issues when the function returns.
Do not ASSERTZEXT for i16 result of bitcast from f16 operand
Summary:
During legalization if i16, do not ASSERTZEXT the result of FP_TO_FP16.
Directly return an FP_TO_FP16 node with return type as the
promote-to-type of i16.
This patch also removes extraneous length check. This legalization
should be valid even if integer and float types are of different
lengths.
This patch breaks a hard-float test for fp16 args. The test is changed
to allow a vmov to zero-out the top bits, and also ensure that the
return value is in an FP register.
David Majnemer [Fri, 8 Jan 2016 17:24:47 +0000 (17:24 +0000)]
[WinEH] CatchHandler which don't have catch objects in StackColoring
StackColoring rewrites the frame indicies of operations involving
allocas if it can find that the life time of two objects do not overlap.
MSVC EH needs to be kept aware of this if happens in the event that a
catch object has moved around. However, we represent the non-existance
of a catch object with a sentinel frame index (INT_MAX). This sentinel
also happens to be the EmptyKey of the SlotRemap DenseMap. Testing for
whether or not we need to translate the frame index fails in this case
because we call the count method on the DenseMap with the EmptyKey,
leading to assertions. Instead, check if it is our sentinel value
before trying to look into the DenseMap.
Teresa Johnson [Fri, 8 Jan 2016 17:06:29 +0000 (17:06 +0000)]
[ThinLTO] Use new in-place symbol changes for exporting module
Due to the new in-place ThinLTO symbol handling support added in
r257174, we now invoke renameModuleForThinLTO on the current
module from within the FunctionImport pass.
Additionally, renameModuleForThinLTO no longer needs to return the
Module as it is performing the renaming in place on the one provided.
This commit will be immediately preceeded by a companion clang patch to
remove its invocation of renameModuleForThinLTO.
Teresa Johnson [Fri, 8 Jan 2016 15:00:00 +0000 (15:00 +0000)]
[ThinLTO] Enable in-place symbol changes for exporting module
Summary:
Move ThinLTO global value processing functions out of ModuleLinker and
into a new ThinLTOGlobalProcessor class, which performs any necessary
linkage and naming changes on the given module in place.
As a result, renameModuleForThinLTO no longer needs to create a new
Module when performing any necessary local to global promotion on a
module that we are possibly exporting from during a ThinLTO backend
compilation.
During function importing the ThinLTO processing is still invoked from
the ModuleLinker (via the new class), as it needs to perform renaming and
linkage changes on the source module, e.g. in order to get the correct
renaming during local to global promotion.
Teresa Johnson [Fri, 8 Jan 2016 14:17:41 +0000 (14:17 +0000)]
[ThinLTO] Delay metadata materializtion in function importer
The function importer was still materializing metadata when modules were
loaded for function importing. We only want to materialize it when we
are going to invoke the metadata linking postpass. Materializing it
before function importing is not only unnecessary, but also causes
metadata referenced by imported functions to be mapped in early, and
then not connected to the rest of the module level metadata when it is
ultimately linked in.
Augmented the test case to specifically check for the metadata being
properly connected, which it wasn't before this fix.
Prevent renaming of CR fields in AADB when a CR restore is present
This patch corresponds to review:
http://reviews.llvm.org/D15930
Moves to and from CR fields depend on shifts/masks that depend on the
target/source CR field. Thus, post-ra anti-dep breaking must not later
change that CR register assignment.
Silviu Baranga [Fri, 8 Jan 2016 11:11:04 +0000 (11:11 +0000)]
Re-commit r257064, this time with a fixed assert
In setInsertionPoint if the value is not a PHI, Instruction or
Argument it should be a Constant, not a ConstantExpr.
Original commit message:
[InstCombine] Look through PHIs, GEPs, IntToPtrs and PtrToInts to expose more constants when comparing GEPs
Summary:
When comparing two GEP instructions which have the same base pointer
and one of them has a constant index, it is possible to only compare
indices, transforming it to a compare with a constant. This removes
one use for the GEP instruction with the constant index, can reduce
register pressure and can sometimes lead to removing the comparisson
entirely.
InstCombine was already doing this when comparing two GEPs if the base
pointers were the same. However, in the case where we have complex
pointer arithmetic (GEPs applied to GEPs, PHIs of GEPs, conversions to
or from integers, etc) the value of the original base pointer will be
hidden to the optimizer and this transformation will be disabled.
This change detects when the two sides of the comparison can be
expressed as GEPs with the same base pointer, even if they don't
appear as such in the IR. The transformation will convert all the
pointer arithmetic to arithmetic done on indices and all the relevant
uses of GEPs to GEPs with a common base pointer. The GEP comparison
will be converted to a comparison done on indices.
[attrs] Split the late-revisit pattern for deducing norecurse in
a top-down manner into a true top-down or RPO pass over the call graph.
There are specific patterns of function attributes, notably the
norecurse attribute, which are most effectively propagated top-down
because all they us caller information.
Walk in RPO over the call graph SCCs takes the form of a module pass run
immediately after the CGSCC pass managers postorder walk of the SCCs,
trying again to deduce norerucrse for each singular SCC in the call
graph.
This removes a very legacy pass manager specific trick of using a lazy
revisit list traversed during finalization of the CGSCC pass. There is
no analogous finalization step in the new pass manager, and a lazy
revisit list is just trying to produce an RPO iteration of the call
graph. We can do that more directly if more expensively. It seems
unlikely that this will be the expensive part of any compilation though
as we never examine the function bodies here. Even in an LTO run over
a very large module, this should be a reasonable fast set of operations
over a reasonably small working set -- the function call graph itself.
In the future, if this really is a compile time performance issue, we
can look at building support for both post order and RPO traversals
directly into a pass manager that builds and maintains the PO list of
SCCs.
[LCG] Re-order the lazy node iterator below the node type to make some
subsequent work I'm doing not have its delta obscured by boring code
motion. NFC.
David Majnemer [Fri, 8 Jan 2016 08:03:55 +0000 (08:03 +0000)]
[WinEH] Update WinEHFuncInfo if StackColoring merges allocas
Windows EH keeping track of which frame index corresponds to a catchpad
in order to inform the runtime where the catch parameter should be
initialized. LLVM's optimizations are able to prove that the memory
used by the catch parameter can be reused with another memory
optimization, changing it's frame index.
We need to keep WinEHFuncInfo up to date with respect to this or we will
miscompile/assert.
[PGO] Ensure vp data in indexed profile always sorted
Done in InstrProfWriter to eliminate the need for client
code to do the sorting. The operation is done once and reused
many times so it is more efficient. Update unit test to remove
sorting. Also update expected output of affected tests.
Kyle Butt [Fri, 8 Jan 2016 02:06:19 +0000 (02:06 +0000)]
Add call sequence start and end for __tls_get_addr
This is a fix for bug http://llvm.org/bugs/show_bug.cgi?id=25839.
For a PIC TLS variable access in a function, prologue (mflr followed by std and
stdu) gets scheduled after a tls_get_addr call. tls_get_addr messed up LR but
no one saves/restores it.
Also added a test for save/restore clobbered registers during calling __tls_get_addr.
Kyle Butt [Fri, 8 Jan 2016 01:55:13 +0000 (01:55 +0000)]
[Vectorization] Actually return from error case in isStridedPtr
The early return seems to be missed. This causes a radical and wrong loop
optimization on powerpc. It isn't reproducible on x86_64, because
"UseInterleaved" is false.
Sanjay Patel [Fri, 8 Jan 2016 01:39:16 +0000 (01:39 +0000)]
[InstCombine] insert a new shuffle in a safe place (PR25999)
Limit this transform to a basic block and guard against PHIs.
Hopefully, this fixes the remaining failures in PR25999:
https://llvm.org/bugs/show_bug.cgi?id=25999
JF Bastien [Thu, 7 Jan 2016 23:18:29 +0000 (23:18 +0000)]
WebAssembly: use .skip instead of .zero directive
.zero is confusing when used with two arguments. Documentation:
This directive emits SIZE 0-valued bytes. SIZE must be an absolute
expression. This directive is actually an alias for the '.skip'
directive so in can take an optional second argument of the value to
store in the bytes instead of zero. Using '.zero' in this way would be
confusing however.
Keno Fischer [Thu, 7 Jan 2016 22:18:37 +0000 (22:18 +0000)]
[Verifier] Check that debug values have proper size
Summary:
Teach the Verifier to make sure that the storage size given to llvm.dbg.declare
or the value size given to llvm.dbg.value agree with what is declared in
DebugInfo. This is implicitly assumed in a number of passes (e.g. in SROA).
Additionally this catches a number of common mistakes, such as passing a
pointer when a value was intended or vice versa.
One complication comes from stack coloring which modifies the original IR when
it merges allocas in order to make sure that if AA falls back to the IR it gets
the correct result. However, given this new invariant, indiscriminately
replacing one alloca by a different (differently sized one) is no longer valid.
Fix this by just undefing out any use of the alloca in a dbg.declare in this
case.
Additionally, I had to fix a number of test cases. Of particular note:
- I regenerated dbg-changes-codegen-branch-folding.ll from the given source as
it was affected by the bug fixed in r256077
- two-cus-from-same-file.ll was changed to avoid having a variable-typed debug
variable as that would depend on the target, even though this test is
supposed to be generic
- I had to manually declared size/align for reference type. See also the
discussion for D14275/r253186.
- fpstack-debuginstr-kill.ll required changing `double` to `long double`
- most others were just a question of adding OP_deref
Dimitry Andric [Thu, 7 Jan 2016 22:09:12 +0000 (22:09 +0000)]
Turn off lldb debug tuning by default for FreeBSD
Summary:
In rL242338, debugger tuning was introduced, and the tuning for FreeBSD
was set to lldb by default. However, for the foreseeable future we
still need to default to gdb tuning, since lldb is not ready for all of
FreeBSD's architectures, and some system tools (like objcopy, etc) have
not yet been adapted to cope with the lldb tuned format, which has
.apple sections.
Therefore, let FreeBSD use gdb by default for now.
David Majnemer [Thu, 7 Jan 2016 21:36:16 +0000 (21:36 +0000)]
[SCCP] Don't violate the lattice invariants
We marked values which are 'undef' as constant instead of undefined
which violates SCCP's invariants. If we can figure out that a
computation results in 'undef', leave it in the undefined state.
Coverage mapping data may reference names of functions
that are skipped by FE (e.g, unused inline functions). Since
those functions are skipped, normal instr-prof function lowering
pass won't put those names in the right section, so special
handling is needed to walk through coverage mapping structure
and recollect the references.
With this patch, only names that are skipped are processed. This
simplifies the lowering code and it no longer needs to make
assumptions coverage mapping data layout. It should also be
more efficient.
David Majnemer [Thu, 7 Jan 2016 19:25:39 +0000 (19:25 +0000)]
[SCCP] Can't go from overdefined to constant
The fix for PR23999 made us mark loads of null as producing the constant
undef which upsets the lattice. Instead, keep the load as "undefined".
This fixes PR26044.
Dan Gohman [Thu, 7 Jan 2016 18:49:53 +0000 (18:49 +0000)]
[WebAssembly] Use the default private label prefixes.
The MC assembler doesn't like using the empty string as a private label
prefix because then it treats all labels as private. This commit reverts
back to the default prefix, which is .L, which is common in ELF targets
and consistent with the LLVM name mangler.
Nicolai Haehnle [Thu, 7 Jan 2016 17:10:29 +0000 (17:10 +0000)]
AMDGPU/SI: Fold operands with sub-registers
Summary:
Multi-dword constant loads generated unnecessary moves from SGPRs into VGPRs,
increasing the code size and VGPR pressure. These moves are now folded away.
Note that this lack of operand folding was not a problem for VMEM loads,
because COPY nodes from VReg_Nnn to VGPR32 are eliminated by the register
coalescer.
Some tests are updated, note that the fsub.ll test explicitly checks that
the move is elided.
With the IR generated by current Mesa, the changes are obviously relatively
minor:
Nicolai Haehnle [Thu, 7 Jan 2016 17:10:20 +0000 (17:10 +0000)]
AMDGPU/SI: xnack_mask is always reserved on VI
Summary:
Somehow, I first interpreted the docs as saying space for xnack_mask is only
reserved when XNACK is enabled via SH_MEM_CONFIG. I felt uneasy about this and
went back to actually test what is happening, and it turns out that xnack_mask
is always reserved at least on Tonga and Carrizo, in the sense that flat_scr
is always fixed below the SGPRs that are used to implement xnack_mask, whether
or not they are actually used.
I confirmed this by writing a shader using inline assembly to tease out the
aliasing between flat_scratch and regular SGPRs. For example, on Tonga, where
we fix the number of SGPRs to 80, s[74:75] aliases flat_scratch (so
xnack_mask is s[76:77] and vcc is s[78:79]).
This patch changes both the calculation of the total number of SGPRs and the
various register reservations to account for this.
It ought to be possible to use the gap left by xnack_mask when the feature
isn't used, but this patch doesn't try to do that. (Note that the same applies
to vcc.)
Note that previously, even before my earlier change in r256794, the SGPRs that
alias to xnack_mask could end up being used as well when flat_scr was unused
and the total number of SGPRs happened to fall on the right alignment
(e.g. highest regular SGPR being used s29 and VCC used would lead to number
of SGPRs being 32, where s28 and s29 alias with xnack_mask). So if there
were some conflict due to such aliasing, we should have noticed that already.
Silviu Baranga [Thu, 7 Jan 2016 14:56:08 +0000 (14:56 +0000)]
[InstCombine] Look through PHIs, GEPs, IntToPtrs and PtrToInts to expose more constants when comparing GEPs
Summary:
When comparing two GEP instructions which have the same base pointer
and one of them has a constant index, it is possible to only compare
indices, transforming it to a compare with a constant. This removes
one use for the GEP instruction with the constant index, can reduce
register pressure and can sometimes lead to removing the comparisson
entirely.
InstCombine was already doing this when comparing two GEPs if the
base pointers were the same. However, in the case where we have
complex pointer arithmetic (GEPs applied to GEPs, PHIs of GEPs,
conversions to or from integers, etc) the value of the original
base pointer will be hidden to the optimizer and this transformation
will be disabled.
This change detects when the two sides of the comparison can be
expressed as GEPs with the same base pointer, even if they don't
appear as such in the IR. The transformation will convert all the
pointer arithmetic to arithmetic done on indices and all the
relevant uses of GEPs to GEPs with a common base pointer. The
GEP comparison will be converted to a comparison done on indices.
James Molloy [Thu, 7 Jan 2016 13:33:28 +0000 (13:33 +0000)]
[GlobalsAA] Partially back out r248576
See PR25822 for a more full summary, but we were conflating the concepts of "capture" and "escape". We were proving nocapture and using that proof to infer noescape, which is not true. Escaped-ness is a function-local property - as soon as a value is used in a call argument it escapes. Capturedness is a related but distinct property. It implies a *temporally limited* escape. Consider:
static int a;
int b;
int g(int * nocapture arg);
int f() {
a = 2; // Even though a escapes to g, it is not captured so can be treated as non-escaping here.
g(&a); // But here it must be treated as escaping.
g(&b); // Now that g(&a) has returned we know it was not captured so we can treat it as non-escaping again.
}
The original commit did not sufficiently understand this nuance and so caused PR25822 and PR26046.
r248576 included both a performance improvement (which has been backed out) and a related conformance fix (which has been kept along with its testcase).
Mark arm as the 32bit variant of aarch64 in Triple
Change Triple::get32BitArchVariant to return arm/armeb as the 32bit
variant of aarch64/aarch64_be and do the same change for the oppoiste
direction in Triple::get64BitArchVariant.
Simon Pilgrim [Thu, 7 Jan 2016 10:24:19 +0000 (10:24 +0000)]
[X86][SSE} Add INSERTPS as a target shuffle
Follow up to D15378, added INSERTPS to the list of decodable target shuffles and enabled XFormVExtractWithShuffleIntoLoad to handle target shuffles with SentinelZero and tested this with INSERTPS.
Tim Northover [Thu, 7 Jan 2016 09:03:03 +0000 (09:03 +0000)]
ARM: support TLS accesses on Darwin platforms
Darwin TLS accesses most closely resemble ELF's general-dynamic situation,
since they have to be able to handle all possible situations. The descriptors
and so on are obviously slightly different though.
Jonas Paulsson [Thu, 7 Jan 2016 07:20:55 +0000 (07:20 +0000)]
[SystemZ] Add hasSideEffects flag on Serialize instruction.
Serialize will perform a hardware serialization operation, and is
acting as a memory barrier. Therefore it must have the hasSideEffects
flag set so it will be treated as a global memory object.
projects / llvm / log