- X86 unfolding should check if the instructions being unfolded has memoperands.
If there is no memoperands, then it must assume conservative alignment. If this
would introduce an expensive sse unaligned load / store, then unfoldMemoryOperand
etc. should not unfold the instruction.
llvm-svn: 107509
PrologEpilog code, and use it to determine whether
the asm forces stack alignment or not. gcc consistently
does not do this for GCC-style asms; Apple gcc inconsistently
sometimes does it for asm blocks. There is no
convenient place to put a bit in either the SDNode or
the MachineInstr form, so I've added an extra operand
to each; unlovely, but it does allow for expansion for
more bits, should we need it. PR 5125. Some
existing testcases are affected.
The operand lists of the SDNode and MachineInstr forms
are indexed with awesome mnemonics, like "2"; I may
fix this someday, but not now. I'm not making it any
worse. If anyone is inspired I think you can find all
the right places from this patch.
llvm-svn: 107506
have any effect, and second, deleting stores can potentially invalidate
an AliasAnalysis, and there's currently no notification for this.
llvm-svn: 107496
the noalias argument on function attributes be usable to model the
C99 restrict keyword on arguments, and to allow AliasAnalysis to
consider a noalias-attributed argument to be an "identified object".
To support this, refactor a new "based on" concept out of the current
pointer aliasing "associated" concept. This "based on" concept is very
similar to (though it is not identical with) the "based on" concept
in C99.
Also, reword the definition of NoAlias to more closely describe the
concept that the optimizer uses.
llvm-svn: 107495
This allows us to recognize the common case where all uses could be
rematerialized, and no stack slot allocation is necessary.
If some values could be fully rematerialized, remove them from the live range
before allocating a stack slot for the rest.
llvm-svn: 107492
second round of low-level interface squeeze-out:
making all of CallInst's low-level operand accessors
private
If you get compile errors I strongly urge you to
update your code.
I tried to write the necessary clues into the
header where the compiler may point to, but no
guarantees. It works for my GCC.
You have several options to update your code:
- you can use the v2.8 ArgOperand accessors
- you can go via a temporary CallSite
- you can upcast to, say, User and call its
low-level accessors if your code is definitely
operand-order agnostic.
If you run into serious problems, please
comment in below thread (and back out this
revision only if absolutely necessary):
<http://groups.google.com/group/llvm-dev/browse_thread/thread/64650cf343b28271>
llvm-svn: 107480
getFunctionAlignment and the corresponding use of that value in the ARM
asm printer, but now we're using the standard asm printer. The result of
this was that function alignments were dropped completely for Thumb functions.
Radar 8143571.
llvm-svn: 107435
Objective-C metadata types which should be marked as "weak", but which the
linker will remove upon final linkage. However, this linkage isn't specific to
Objective-C.
For example, the "objc_msgSend_fixup_alloc" symbol is defined like this:
.globl l_objc_msgSend_fixup_alloc
.weak_definition l_objc_msgSend_fixup_alloc
.section __DATA, __objc_msgrefs, coalesced
.align 3
l_objc_msgSend_fixup_alloc:
.quad _objc_msgSend_fixup
.quad L_OBJC_METH_VAR_NAME_1
This is different from the "linker_private" linkage type, because it can't have
the metadata defined with ".weak_definition".
Currently only supported on Darwin platforms.
llvm-svn: 107433
to update their code to high-level interfaces
If you get compile errors in your project
please update your code according to the
comments.
This is a re-commit of r107396 which causes
compile errors for the indicated usage patterns
instead of link errors (which are less easy to
fix because of missing source location).
If you get compile errors please perform
following functionally equivalent transformations:
- getOperand(0) ---> getCalledValue()
- setOperand(0, V) ---> setCalledFunction(V)
This will make your code more future-proof
and avoid potentially hard-to-debug bugs.
please refer to this thread on llvm-dev:
<http://groups.google.com/group/llvm-dev/browse_thread/thread/64650cf343b28271>
llvm-svn: 107432
such a way that debug info for symbols preserved even if symbols are
optimized away by the optimizer.
Add new special pass to remove debug info for such symbols.
llvm-svn: 107416
to update their code to high-level interfaces
If you get compile errors in your project
please update your code according to the
comments.
llvm-svn: 107396
- Add encode bits for VEX_W
- All 128-bit SSE 1 & SSE2 instructions that are described
in the .td file now have a AVX encoded form already working.
llvm-svn: 107365
entries associated with the value being erased in the
folding set map. These entries used to be harmless, because
a SCEVUnknown doesn't store any information about its Value*,
so having a new Value allocated at the old Value's address
wasn't a problem. But now that ScalarEvolution is storing more
information about values, this is no longer safe.
llvm-svn: 107316
LocalRewriter::runOnMachineFunction uses this information to mark dead spill
slots.
This means that InlineSpiller now also works for functions that spill.
llvm-svn: 107302
replaced by a bigger array in SmallPtrSet (by overridding it), instead just use a
pointer to the start of the storage, and have SmallPtrSet pass in the value to use.
This has the disadvantage that SmallPtrSet becomes bigger by one pointer. It has
the advantage that it no longer uses tricky C++ rules, and is clearly correct while
I'm not sure the previous version was. This was inspired by g++-4.6 pointing out
that SmallPtrSetImpl was writing off the end of SmallArray, which it was. Since
SmallArray is replaced with a bigger array in SmallPtrSet, the write was still to
valid memory. But it was writing off the end of the declared array type - sounds
kind of dubious to me, like it sounded dubious to g++-4.6. Maybe g++-4.6 is wrong
and this construct is perfectly valid and correctly compiled by all compilers, but
I think it is better to avoid the whole can of worms by avoiding this construct.
llvm-svn: 107285
on ScalarEvolution successfully folding and preserving
range information for both A-B and B-A. Now, if it gets
either one, it's sufficient.
llvm-svn: 107249
InlineSpiller inserts loads and spills immediately instead of deferring to
VirtRegMap. This is possible now because SlotIndexes allows instructions to be
inserted and renumbered.
This is work in progress, and is mostly a copy of TrivialSpiller so far. It
works very well for functions that don't require spilling.
llvm-svn: 107227
metadata types which should be marked as "weak", but which the linker will
remove upon final linkage. For example, the "objc_msgSend_fixup_alloc" symbol is
defined like this:
.globl l_objc_msgSend_fixup_alloc
.weak_definition l_objc_msgSend_fixup_alloc
.section __DATA, __objc_msgrefs, coalesced
.align 3
l_objc_msgSend_fixup_alloc:
.quad _objc_msgSend_fixup
.quad L_OBJC_METH_VAR_NAME_1
This is different from the "linker_private" linkage type, because it can't have
the metadata defined with ".weak_definition".
llvm-svn: 107205
SmallArray[SmallSize] in the SmallPtrSetIteratorImpl, and this is
one off the end of the array. For those who care, right now gcc
warns about writing off the end because it is confused about the
declaration of SmallArray as having length 1 in the parent class
SmallPtrSetIteratorImpl. However if you tweak code to unconfuse
it, then it still warns about writing off the end of the array,
because of this buffer overflow. In short, even with this fix
gcc-4.6 will warn about writing off the end of the array, but now
that is only because it is confused.
llvm-svn: 107200
and make PATypeHolder work with null pointers.
The implicitly generated one didn't work on numerous levels, but was still
accepted, allowing all sorts of bugs with default constructed pa type holders.
Previously, they "sort of" worked if they were default constructed and then
destructed. Now they really work, and you can even default construct one,
then assign to it, amazing.
llvm-svn: 107195
A partial redefine needs to be treated like a tied operand, and the register
must be reloaded while processing use operands.
This fixes a bug where partially redefined registers were processed as normal
defs with a reload added. The reload could clobber another use operand if it was
a kill that allowed register reuse.
llvm-svn: 107193
The LowerSubregs pass needs to preserve implicit def operands attached to
EXTRACT_SUBREG instructions when it replaces those instructions with copies.
llvm-svn: 107189
is stripped off. Currently set unconditionally, since the API
does not provide a way of working out if anything was actually
stripped off.
llvm-svn: 107142
of getPhysicalRegisterRegClass with it.
If we want to make a copy (or estimate its cost), it is better to use the
smallest class as more efficient operations might be possible.
llvm-svn: 107140
in terms of Op<> and ArgOffset. This works for
values of {0, 1} for ArgOffset.
Please note that ArgOffset will become 0 soon and
will go away eventually.
llvm-svn: 107129
instruction to an add scev, it's not safe to blindly transfer the
inbounds flag from a gep instruction to an nsw on the scev for the
gep.
llvm-svn: 107117
There are 2 changes relative to the previous version of the patch:
1) For the "simple" if-conversion case, there's no need to worry about
RemoveExtraEdges not handling an unanalyzable branch. Predicated terminators
are ignored in this context, so RemoveExtraEdges does the right thing.
This might break someday if we ever treat indirect branches (BRIND) as
predicable, but for now, I just removed this part of the patch, because
in the case where we do not add an unconditional branch, we rely on keeping
the fall-through edge to CvtBBI (which is empty after this transformation).
The change relative to the previous patch is:
@@ -1036,10 +1036,6 @@
IterIfcvt = false;
}
- // RemoveExtraEdges won't work if the block has an unanalyzable branch,
- // which is typically the case for IfConvertSimple, so explicitly remove
- // CvtBBI as a successor.
- BBI.BB->removeSuccessor(CvtBBI->BB);
RemoveExtraEdges(BBI);
// Update block info. BB can be iteratively if-converted.
2) My patch exposed a bug in the code for merging the tail of a "diamond",
which had previously never been exercised. The code was simply checking that
the tail had a single predecessor, but there was a case in
MultiSource/Benchmarks/VersaBench/dbms where that single predecessor was
neither edge of the diamond. I added the following change to check for
that:
@@ -1276,7 +1276,18 @@
// tail, add a unconditional branch to it.
if (TailBB) {
BBInfo TailBBI = BBAnalysis[TailBB->getNumber()];
- if (TailBB->pred_size() == 1 && !TailBBI.HasFallThrough) {
+ bool CanMergeTail = !TailBBI.HasFallThrough;
+ // There may still be a fall-through edge from BBI1 or BBI2 to TailBB;
+ // check if there are any other predecessors besides those.
+ unsigned NumPreds = TailBB->pred_size();
+ if (NumPreds > 1)
+ CanMergeTail = false;
+ else if (NumPreds == 1 && CanMergeTail) {
+ MachineBasicBlock::pred_iterator PI = TailBB->pred_begin();
+ if (*PI != BBI1->BB && *PI != BBI2->BB)
+ CanMergeTail = false;
+ }
+ if (CanMergeTail) {
MergeBlocks(BBI, TailBBI);
TailBBI.IsDone = true;
} else {
With these fixes, I was able to run all the SingleSource and MultiSource
tests successfully.
llvm-svn: 107110
have to be registers, per gcc documentation. This affects
the logic for determining what "g" should lower to. PR 7393.
A couple of existing testcases are affected.
llvm-svn: 107079
When an instruction has tied operands and physreg defines, we must take extra
care that the tied operands conflict with neither physreg defs nor uses.
The special treatment is given to inline asm and instructions with tied operands
/ early clobbers and physreg defines.
This fixes PR7509.
llvm-svn: 107043
(in both CallInst and InvokeInst)
also add a (short-lived) constant to CallInst, that names
the operand index of the first call argument. This is
strictly transitional and should not be used for new code.
llvm-svn: 107001
large integers, the first inserted value would always create
an 'or X, 0'. Even though this is trivially zapped by
instcombine, don't bother creating this pointless instruction.
llvm-svn: 106979
the returned value after the tail call if it differs from other return
values. The optimal thing to do would be to introduce a phi node for
the return value, but for the moment just fix the miscompile.
llvm-svn: 106947
if-conversion. The RemoveExtraEdges function doesn't work for blocks that
end with unanalyzable branches, so in those cases, the "extra" edges must
be explicitly removed. The CopyAndPredicateBlock and MergeBlocks methods
can also avoid copying successor edges due to branches that have already
been removed. The latter case is especially helpful when MergeBlocks is
called for handling "diamond" if-conversions, where otherwise you can end
up with some weird intermediate states in the CFG. Unfortunately I've
been unable to find cases where this cleanup actually makes a significant
difference in the code. There is one test where we manage to remove an
empty block at the end of a function. Radar 6911268.
llvm-svn: 106939
CopyFromReg nodes for aliasing registers (AX and AL). This confuses the fast
register allocator.
Instead of CopyFromReg(AL), use ExtractSubReg(CopyFromReg(AX), sub_8bit).
This fixes PR7312.
llvm-svn: 106934
introduced in r106343, but only showed up recently (with a particular compiler &
linker combination) because of the particular check, and because we have no
builtin checking for dereferencing the end of an array, which is truly
unfortunate.
llvm-svn: 106908
The VNInfo.kills vector was almost unused except for all the code keeping it
updated. The few places using it were easily rewritten to check for interval
ends instead.
The two new methods LiveInterval::killedAt and killedInRange are replacements.
This brings us down to 3 independent data structures tracking kills.
llvm-svn: 106905
SCEVUnknown values which are loop-variant, as LSR can't do anything
interesting with these values in any case. This fixes very slow compile
times on loops which have large numbers of such values.
llvm-svn: 106897
for an "i" constraint should get lowered; PR 6309. While
this argument was passed around a lot, this is the only
place it was used, so it goes away from a lot of other
places.
llvm-svn: 106893
are dead, not just the def of this register. I.e., a register could be dead, but
it's subreg isn't.
Testcase to follow with a subsequent patch.
llvm-svn: 106878
with the following instructions. This is done via trickery by considering the
instruction preceding the IT to be the hazard. Care must be taken to ensure
it's the first non-debug instruction, or the presence of debug info will
affect codegen.
Part of the continuing work for rdar://7797940, making ARM code-gen unaffected
by the presence of debug information.
llvm-svn: 106871
buffer in the same chunk of memory.
2 less mallocs for every uninitialized MemoryBuffer and 1 less malloc for every
MemoryBuffer pointing to a memory range translate into 20% less mallocs on
clang -cc1 -Eonly Cocoa_h.m.
llvm-svn: 106839
and CallInst for getting hold
of the intrinsic's arguments
simplify along the way (at least for me this is much more legible now)
Bill, Baldrick or Anton, please review\!
llvm-svn: 106838
This fixes PR7479 and PR7485. The test cases from those PRs are big, so not
included. However, PR7485 comes from self hosting on FreeBSD, so we will surely
hear about any regression.
llvm-svn: 106811
Evan Cheng