Eric Christopher [Tue, 10 Jun 2014 20:39:38 +0000 (20:39 +0000)]
Have isInTailCallPosition take the DAG so that we can use the
version of TargetLowering/Machine from there on the way to avoiding
TargetMachine in TargetLowering.
Reid Kleckner [Tue, 10 Jun 2014 20:16:36 +0000 (20:16 +0000)]
Revert "Patch by Ray Donnelly to print register names instead of numbers."
This reverts commit r206683.
The code was confusing SEH register numbers with DWARF register numbers.
The test case it was committed with was obviously incorrect. The
disassembler was roundtripping '.seh_pushreg %rsi' as '.seh_pushreg
%rbp', and other exciting things.
Andrea Di Biagio [Tue, 10 Jun 2014 16:42:57 +0000 (16:42 +0000)]
[X86] Improved target combine rules for selecting horizontal add/sub.
This patch slightly changes the algorithm introduced at revision 210477
to fix a problem where the algorithm was producing incorrect code for
the VEX.256 encoded versions of horizontal add/sub.
For these cases, we now try to split the two 256-bit vectors into
128-bit chunks before emitting horizontal add/sub dag nodes.
Tom Stellard [Tue, 10 Jun 2014 16:42:41 +0000 (16:42 +0000)]
Hexagon: Expand i1 SELECT_CC
il is legal for Hexagon, so I should have marked this as Expand for
SELECT_CC when I removed setOperationAction(ISD::SELECT_CC, MVT::Other,
Expand); in r210541.
Renato Golin [Tue, 10 Jun 2014 16:39:21 +0000 (16:39 +0000)]
Fix a bug in the Thumb1 ARM Load/Store optimizer
Previously, the basic block was searched for future uses of the base register,
and if necessary any writeback to the base register was reset using a SUB
instruction (e.g. before calling a function) just before such a use. However,
this step happened *before* the merged LDM/STM instruction was built. So if
there was (e.g.) a function call directly after the not-yet-formed LDM/STM,
the pass would first insert a SUB instruction to reset the base register,
and then (at the same location, incorrectly) insert the LDM/STM itself.
to indicate that they wanted to expand ISD::SELECT_CC for all types.
This wasn't applied correctly everywhere, and it makes writing new
DAG patterns with ISD::SELECT_CC difficult.
Bill Schmidt [Tue, 10 Jun 2014 14:35:01 +0000 (14:35 +0000)]
[PPC64LE] Recognize shufflevector patterns for little endian
Various masks on shufflevector instructions are recognizable as
specific PowerPC instructions (vector pack, vector merge, etc.).
There is existing code in PPCISelLowering.cpp to recognize the correct
patterns for big endian code. The masks for these instructions are
different for little endian code due to the big-endian numbering
employed by these instructions. This patch adds the recognition code
for little endian.
I've added a new test case test/CodeGen/PowerPC/vec_shuffle_le.ll for
this. The existing recognizer test (vec_shuffle.ll) is unnecessarily
verbose and difficult to read, so I felt it was better to add a new
test rather than modify the old one.
Add detection of OS X relocatable SDK to compiler-rt as a lit.util function
Clang's lit cfg already detects the currently selected SDK via
"xcrun --show-sdk-path". The same thing should be done for compiler-rt tests,
to make them work on recent OS X versions. Instead of duplicating the detection
code, this patch extracts the detection function into a lit.util method.
Patch by Kuba Brecka (kuba.brecka@gmail.com),
reviewed at http://reviews.llvm.org/D4072
Artyom Skrobov [Tue, 10 Jun 2014 13:11:35 +0000 (13:11 +0000)]
Condition codes AL and NV are invalid in the aliases that use
inverted condition codes (CINC, CINV, CNEG, CSET, and CSETM).
Matching aliases based on "immediate classes", when disassembling,
wasn't previously supported, hence adding MCOperandPredicate
into class Operand, and implementing the support for it
in AsmWriterEmitter.
The parsing for those aliases was already custom, so just adding
the missing condition into AArch64AsmParser::parseCondCode.
Tim Northover [Tue, 10 Jun 2014 10:50:24 +0000 (10:50 +0000)]
AArch64: disallow x30 & x29 as the destination for indirect tail calls
As Ana Pazos pointed out, these have to be restored to their incoming values
before a function returns; i.e. before the tail call. So they can't be used
correctly as the destination register.
Tim Northover [Tue, 10 Jun 2014 10:49:07 +0000 (10:49 +0000)]
X86: elide comparisons after cmpxchg instructions.
The C++ and C semantics of the compare_and_swap operations actually
require us to return a boolean "success" value. In LLVM terms this
means a second comparison of the output of "cmpxchg" against the input
desired value.
However, x86's "cmpxchg" instruction sets all flags for the comparison
formed, so we can skip any secondary comparison. (N.b. this isn't true
for cmpxchg8b/16b, which only set ZF).
Tim Northover [Tue, 10 Jun 2014 09:52:44 +0000 (09:52 +0000)]
AArch64: teach FastISel how to handle offset FrameIndices
Previously we were abandonning the attempt, leading to some combination of
extra work (when selection of a load/store fails completely) and inferior code
(when this leads to a real memcpy call instead of inlining).
Tim Northover [Tue, 10 Jun 2014 09:52:40 +0000 (09:52 +0000)]
AArch64: make FastISel memcpy emission more robust.
We were hitting an assert if FastISel couldn't create the load or store we
requested. Currently this happens for large frame-local addresses, though
CodeGen could be improved there.
Richard Trieu [Mon, 9 Jun 2014 22:53:16 +0000 (22:53 +0000)]
Removing an "if (!this)" check from two print methods. The condition will
never be true in a well-defined context. The checking for null pointers
has been moved into the caller logic so it does not rely on undefined behavior.
David Blaikie [Mon, 9 Jun 2014 22:26:20 +0000 (22:26 +0000)]
SmallVector: support resize(N) with move-only types
Unfortunately there's no way to elegantly do this with pre-canned
algorithms. Using a generating iterator doesn't work because you default
construct for each element, then move construct into the actual slot
(bad for copy but non-movable types, and a little unneeded overhead even
in the move-only case), so just write it out manually.
This solution isn't exception safe (if one of the element's ctors calls
we don't fall back, destroy the constructed elements, and throw on -
which std::uninitialized_fill does do) but SmallVector (and LLVM) isn't
exception safe anyway.
Bill Schmidt [Mon, 9 Jun 2014 22:00:52 +0000 (22:00 +0000)]
[PPC64LE] Generate correct code for unaligned little-endian vector loads
The code in PPCTargetLowering::PerformDAGCombine() that handles
unaligned Altivec vector loads generates a lvsl followed by a vperm.
As we've seen in numerous other places, the vperm instruction has a
big-endian bias, and this is fixed for little endian by complementing
the permute control vector and swapping the input operands. In this
case the lvsl is providing the permute control vector. Rather than
generating an lvsl and a complement operation, it is sufficient to
generate an lvsr instruction instead. Thus for LE code generation we
will generate an lvsr rather than an lvsl, and swap the other input
arguments on the vperm.
The existing test/CodeGen/PowerPC/vec_misalign.ll is updated to test
the code generation for PPC64 and PPC64LE, in addition to the existing
PPC32/G5 testing.
Alexey Samsonov [Mon, 9 Jun 2014 21:53:47 +0000 (21:53 +0000)]
Generate better location ranges for some register-described variables.
Don't terminate location ranges for register-described variables
at the end of machine basic block if this register is never modified
in the function body, except for the prologue and epilogue. Prologue
location is guessed by FrameSetup flags on MachineInstructions, while
epilogue location is deduced from debug locations of instructions
in the basic blocks ending with return instructions.
This patch is mostly targeted to fix non-trivial debug locations for
variables addressed via stack and frame pointers.
It is not really a generic fix. We can still produce poor debug info
for register-described variables if this register *is* modified somewhere
in the function, but in unrelated places. This might be the case for the debug
info in optimized binaries (e.g. for local variables in inlined functions).
LiveDebugVariables pass in CodeGen attempts to fix this problem by adjusting
DBG_VALUE instructions, but this pass is tied to greedy register allocator,
which is used in optimized builds only. Proper fix would likely involve
generalizing LiveDebugVariables to all register allocators. See more discussion
in http://reviews.llvm.org/D3933 review thread.
I'm proceeding with this patch to fix immediate severe problems and
important cases, e.g. fix completely broken debug info with AddressSanitizer
and fix PR19307 (missing debug info for by-value std::string arguments).
The armv7-windows-itanium environment is nearly identical to the MSVC ABI. It
has a few divergences, mostly revolving around the use of the Itanium ABI for
C++. VLA support is one of the extensions that are amongst the set of the
extensions.
This adds support for proper VLA emission for this environment. This is
somewhat similar to the handling for __chkstk emission on X86 and the large
stack frame emission for ARM. The invocation style for chkstk is still
controlled via the -mcmodel flag to clang.
Added functions cross-reference test.
Originally this similar was initiated by Björn Steinbrink here:
http://reviews.llvm.org/D3437
Bug itself has been fixed by principal changes in MergeFunctions. Though
special checks for functions merging are still actual. And the test has
been accepted with slight modifications.
Alp Toker [Mon, 9 Jun 2014 18:28:53 +0000 (18:28 +0000)]
Fold FEnv.h into the implementation
Support headers shouldn't use config.h definitions, and they should never be
undefined like this.
ConstantFolding.cpp was the only user of this facility and already includes
config.h for other math features, so it makes sense to move the checks there at
point of use.
(The implicit config.h was also quite dangerous -- removing the FEnv.h include
would have silently disabled math constant folding without causing any tests to
fail. Need to investigate -Wundef once the cleanup is done.)
This eliminates the last config.h include from LLVM headers, paving the way for
more consistent configuration checks.
[X86] Add target combine rules for horizontal add/sub.
This patch adds new target specific combine rules to identify horizontal
add/sub idioms from BUILD_VECTOR dag nodes.
This patch also teaches the DAGCombiner how to canonicalize sequences of
insert_vector_elt dag nodes according to the following rule:
(insert_vector_elt (insert_vector_elt A, I0), I1) ->
(insert_vecto_elt (insert_vector_elt A, I1), I0)
This new canonicalization rule only triggers if the inner insert_vector
dag node has exactly one use; also, both indices must be known constants,
and I1 < I0.
This last rule made it possible to write a simpler algorithm to identify
horizontal add/sub patterns because now we don't have to worry about the
ordering of insert_vector_elt dag nodes.
Bill Schmidt [Mon, 9 Jun 2014 16:06:29 +0000 (16:06 +0000)]
[PPC64LE] Generate correct little-endian code for v16i8 multiply
The existing code in PPCTargetLowering::LowerMUL() for multiplying two
v16i8 values assumes that vector elements are numbered in big-endian
order. For little-endian targets, the vector element numbering is
reversed, but the vmuleub, vmuloub, and vperm instructions still
assume big-endian numbering. To account for this, we must adjust the
permute control vector and reverse the order of the input registers on
the vperm instruction.
The existing test/CodeGen/PowerPC/vec_mul.ll is updated to be executed
on powerpc64 and powerpc64le targets as well as the original powerpc
(32-bit) target.
[X86] Avoid emitting unnecessary test instructions.
This patch teaches the backend how to check for the 'NoSignedWrap' flag on
binary operations to improve the emission of 'test' instructions.
If the result of a binary operation is known not to overflow we know that
resetting the Overflow flag is unnecessary and so we can avoid emitting
the test instruction.
[DAG] Expose NoSignedWrap, NoUnsignedWrap and Exact flags to SelectionDAG.
This patch modifies SelectionDAGBuilder to construct SDNodes with associated
NoSignedWrap, NoUnsignedWrap and Exact flags coming from IR BinaryOperator
instructions.
Added a new SDNode type called 'BinaryWithFlagsSDNode' to allow accessing
nsw/nuw/exact flags during codegen.
Alexey Volkov [Mon, 9 Jun 2014 11:40:41 +0000 (11:40 +0000)]
[X86] Use ADD/SUB instead of INC/DEC for Silvermont
According to Intel Software Optimization Manual
on Silvermont INC or DEC instructions require
an additional uop to merge the flags.
As a result, a branch instruction depending
on an INC or a DEC instruction incurs a 1 cycle penalty.
Fix line numbers for code inlined from __nodebug__ functions.
Instructions from __nodebug__ functions don't have file:line
information even when inlined into no-nodebug functions. As a result,
intrinsics (SSE and other) from <*intrin.h> clang headers _never_
have file:line information.
With this change, an instruction without !dbg metadata gets one from
the call instruction when inlined.
Chad Rosier [Mon, 9 Jun 2014 01:25:51 +0000 (01:25 +0000)]
[AArch64] When combining constant mul of power of 2 plus/minus 1, prefer shift
plus add. The shift can be folded into the add. This only effects codegen
when the constant is 3.
Jingyue Wu [Sun, 8 Jun 2014 23:49:34 +0000 (23:49 +0000)]
[SeparateConstOffsetFromGEP] inbounds zext => sext for better splitting
For each array index that is in the form of zext(a), convert it to sext(a)
if we can prove zext(a) <= max signed value of typeof(a). The conversion
helps to split zext(x + y) into sext(x) + sext(y).
David Blaikie [Sun, 8 Jun 2014 19:33:40 +0000 (19:33 +0000)]
SmallVector: Improve test coverage for insert with repetition
To test cases that involve actual repetition (> 1 elements), at least
one element before the insertion point, and some elements of the
original range that still fit in that range space after insertion.
Actually we need coverage for the inverse case too (where no elements
after the insertion point fit into the previously allocated space), but
this'll do for now, and I might end up rewriting bits of SmallVector to
avoid that special case anyway.
Tobias Grosser [Sun, 8 Jun 2014 19:21:20 +0000 (19:21 +0000)]
ScalarEvolution: Derive element size from the type of the loaded element
Before, we where looking at the size of the pointer type that specifies the
location from which to load the element. This did not make any sense at all.
This change fixes a bug in the delinearization where we failed to delinerize
certain load instructions.
X86Subtarget::isTargetCygMing || X86Subtarget::isTargetKnownWindowsMSVC is
equivalent to all Windows environments. Simplify the check to isOSWindows.
NFC.
David Blaikie [Sun, 8 Jun 2014 17:33:47 +0000 (17:33 +0000)]
SmallVectorTest: Remove some more robust checks added in r210429 since they caught some bugs I haven't fixed yet.
Specifically this caused inserting an element from a SmallVector into
itself when such an insertion would cause a reallocation. We have code
to handle this for non-reallocating cases, but it's not robust against
reallocation.
David Blaikie [Sun, 8 Jun 2014 16:55:13 +0000 (16:55 +0000)]
Fix some more moving-from-moved-from objects issues in SmallVector
(& because it makes it easier to test, this also improves
correctness/performance slightly by moving the last element in an insert
operation, rather than copying it)
Stephen Canon [Sun, 8 Jun 2014 16:53:31 +0000 (16:53 +0000)]
APFloat: x - NaN needs to flip the signbit of NaN when x is a number.
Because we don't have a separate negate( ) function, 0 - NaN does double-duty as the IEEE-754 negate( ) operation, which (unlike most FP ops) *does* attach semantic meaning to the signbit of NaN.
David Blaikie [Sun, 8 Jun 2014 16:18:35 +0000 (16:18 +0000)]
AsmMatchers: Use unique_ptr to manage ownership of MCParsedAsmOperand
I saw at least a memory leak or two from inspection (on probably
untested error paths) and r206991, which was the original inspiration
for this change.
I ran this idea by Jim Grosbach a few weeks ago & he was OK with it.
Since it's a basically mechanical patch that seemed sufficient - usual
post-commit review, revert, etc, as needed.
David Blaikie [Sun, 8 Jun 2014 16:00:02 +0000 (16:00 +0000)]
Ensure SmallVector::insert doesn't overwrite the last element in the range with the already-moved-from value
This would cause the last element in a range to be in a moved-from state
after an insert at a non-end position, losing that value entirely in the
process.
Side note: move_backward is subtle. It copies [A, B) to C-1 and down.
(the fact that it decrements both the second and third iterators before
the first movement is the subtle part... kind of surprising, anyway)
Alp Toker [Sun, 8 Jun 2014 09:13:42 +0000 (09:13 +0000)]
Revert "Do materialize for floating point"
1) The commit was made despite profound lack of understanding:
"I did not understand the comment about using dyn_cast instead of isa. I will
commit as is and make the update after. You can explain what you meant to me."
Commit first, understand later isn't OK.
2) Review comments were simply ignored:
"Can you edit the summary to describe what the patch is for? It appears to be
a list of commits at the moment."
3) The patch got LGTM'd off-list without any indication of readiness.
4) The public mailing list was excluded from patch review so all of this was
hidden from the community.