Matthew Simpson [Fri, 24 Feb 2017 18:20:12 +0000 (18:20 +0000)]
[LV] Merge floating-point and integer induction widening code
This patch merges the existing floating-point induction variable widening code
into the integer induction variable widening code, creating a single set of
functions for both kinds of inductions. The primary motivation for doing this
is to enable vector phi node creation for floating-point induction variables.
Extra const in the StringRef argument meant that MSVC complained about it not correctly overriding from OperandPredicateMatcher::emitCxxPredicateExpr (which didn't have the const)
Sanjay Patel [Fri, 24 Feb 2017 17:17:33 +0000 (17:17 +0000)]
[DAGCombiner] add missing folds for scalar select of {-1,0,1}
The motivation for filling out these select-of-constants cases goes back to D24480,
where we discussed removing an IR fold from add(zext) --> select. And that goes back to:
https://reviews.llvm.org/rL75531
https://reviews.llvm.org/rL159230
The idea is that we should always canonicalize patterns like this to a select-of-constants
in IR because that's the smallest IR and the best for value tracking. Note that we currently
do the opposite in some cases (like the cases in *this* patch). Ie, the proposed folds in
this patch already exist in InstCombine today:
https://github.com/llvm-mirror/llvm/blob/master/lib/Transforms/InstCombine/InstCombineSelect.cpp#L1151
As this patch shows, most targets generate better machine code for simple ext/add/not ops
rather than a select of constants. So the follow-up steps to make this less of a patchwork
of special-case folds and missing IR canonicalization:
1. Have DAGCombiner convert any select of constants into ext/add/not ops.
2 Have InstCombine canonicalize in the other direction (create more selects).
Simon Dardis [Fri, 24 Feb 2017 16:32:18 +0000 (16:32 +0000)]
Recommit "[mips] Fix atomic compare and swap at O0."
This time with the missing files.
Similar to PR/25526, fast-regalloc introduces spills at the end of basic
blocks. When this occurs in between an ll and sc, the store can cause the
atomic sequence to fail.
This patch fixes the issue by introducing more pseudos to represent atomic
operations and moving their lowering to after the expansion of postRA
pseudos.
Simon Dardis [Fri, 24 Feb 2017 16:27:45 +0000 (16:27 +0000)]
[mips] Fix atomic compare and swap at O0.
Similar to PR/25526, fast-regalloc introduces spills at the end of basic
blocks. When this occurs in between an ll and sc, the store can cause the
atomic sequence to fail.
This patch fixes the issue by introducing more pseudos to represent atomic
operations and moving their lowering to after the expansion of postRA
pseudos.
Daniel Sanders [Fri, 24 Feb 2017 15:43:30 +0000 (15:43 +0000)]
[globalisel] Decouple src pattern operands from dst pattern operands.
Summary:
This isn't testable for AArch64 by itself so this patch also adds
support for constant immediates in the pattern and physical
register uses in the result.
The new IntOperandMatcher matches the constant in patterns such as
'(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold
immediates into an instruction so this is the first rule that will match
across multiple BB's.
The Renderer hierarchy is responsible for adding operands to the result
instruction. Renderers can copy operands (CopyRenderer) or add physical
registers (in particular %wzr and %xzr) to the result instruction
in any order (OperandMatchers now import the operand names from
SelectionDAG to allow renderers to access any operand). This allows us to
emit the result instruction for:
%1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0
%1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0
although the latter is untested since the matcher/importer has not been
taught about commutativity yet.
Added BuildMIAction which can build new instructions and mutate them where
possible. W.r.t the mutation aspect, MatchActions are now told the name of
an instruction they can recycle and BuildMIAction will emit mutation code
when the renderers are appropriate. They are appropriate when all operands
are rendered using CopyRenderer and the indices are the same as the matcher.
This currently assumes that all operands have at least one matcher.
Finally, this change also fixes a crash in
AArch64InstructionSelector::select() caused by an immediate operand
passing isImm() rather than isCImm(). This was uncovered by the other
changes and was detected by existing tests.
Simon Pilgrim [Fri, 24 Feb 2017 15:35:52 +0000 (15:35 +0000)]
[X86][SSE] Target shuffle combine can try to combine up to 16 vectors
Noticed while profiling PR32037, the target shuffle ops were being stored in SmallVector<*,8> types but the combiner could store as many as 16 ops at maximum depth (2 per depth).
Sanjay Patel [Fri, 24 Feb 2017 15:18:42 +0000 (15:18 +0000)]
[InstCombine] don't try SimplifyDemandedInstructionBits from zext/sext because it's slow and unnecessary
This one seems more obvious than D30270 that it can't make improvements because an extension always needs
all of the incoming bits. There's one specific transform in SimplifyDemandedInstructionBits of converting
a sext to a zext when the sign-bit is known zero, but that is handled explicitly in visitSext() with
ComputeSignBit().
Like D30270, there are no IR differences (other than instruction names) for the case in PR32037:
https://bugs.llvm.org//show_bug.cgi?id=32037
...and no regression test differences.
Zext/sext are a smaller part of the profile, but this still appears to shave off another 0.5% or so from
'opt -O2'.
Simon Dardis [Fri, 24 Feb 2017 14:34:32 +0000 (14:34 +0000)]
[mips] Handle 64 bit immediate in and/or/xor pseudo instructions on mips64
Previously LLVM was assuming 32-bit signed immediates which results in and with
a bitmask that has bit 31 set to incorrectly include bits 63-32 in the result.
After applying this patch I can now compile all of the FreeBSD mips assembly
code with clang.
This issue also affects the nor, slt and sltu macros and I will fix those in a
separate review.
Daniel Sanders [Fri, 24 Feb 2017 13:58:11 +0000 (13:58 +0000)]
[globalisel] Sort RuleMatchers by priority.
Summary:
This makes more important rules have priority over less important rules.
For example, '%a = G_ADD $b:s64, $c:s64' has priority over
'%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the
correct order by chance.
NFC in this patch but it is required to make the next patch work correctly.
Simon Dardis [Fri, 24 Feb 2017 10:50:27 +0000 (10:50 +0000)]
[mips][mc] Fix a crash when disassembling odd sized sections
Make the MIPS disassembler consistent with the other targets in returning
a Size of zero when the input buffer cannot contain an instruction due
to it's size. Previously it reported the minimum instruction size when
it failed due to the buffer not being big enough for an instruction
causing llvm-objdump to crash when disassembling all sections.
Which can be particularly slow for large APInts (MaskSizeInBits > 64) as they require the allocation memory for the temporary variable.
This is one of the key compile time issues identified in PR32037.
This patch adds the APInt::setBits() helper method which avoids the temporary memory allocation completely, this first implementation uses setBit() internally instead but already significantly reduces the regression in PR32037 (~10% drop). Additional optimization may be possible.
I investigated whether there is need for APInt::clearBits() and APInt::flipBits() equivalents but haven't seen these patterns to be particularly common, but reusing the code would be trivial.
Craig Topper [Fri, 24 Feb 2017 07:21:10 +0000 (07:21 +0000)]
[AVX-512] Separate the fadd/fsub/fmul/fdiv/fmax/fmin with rounding mode ISD opcodes into separate packed and scalar opcodes. This is more consistent with the rest of the ISD opcodes. NFC
Petr Hosek [Fri, 24 Feb 2017 03:10:10 +0000 (03:10 +0000)]
[Fuchsia] Use thread-pointer ABI slots for stack-protector and safe-stack
The Fuchsia ABI defines slots from the thread pointer where the
stack-guard value for stack-protector, and the unsafe stack pointer
for safe-stack, are stored. This parallels the Android ABI support.
Xin Tong [Fri, 24 Feb 2017 01:43:36 +0000 (01:43 +0000)]
Fix an iterator invalidation bug when simplifying LIC user.
LoopUnswitch/simplify-with-nonvalness.ll is the test case for this.
The LIC has 2 users and deleting the 1st user when it can be simplified
invalidated the iterator for the 2nd user.
Adam Nemet [Fri, 24 Feb 2017 01:22:55 +0000 (01:22 +0000)]
[LazyMachineBFI] Add testcase
This is based on Justin's testcase and checking whether BFI is not populated
in case hotness is off.
This is a patch meant on top of Justin's patch to enable Machine opt-remarks
in the
AsmPrinter (http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20170130/426595.html)
[CGP] Split some critical edges coming out of indirect branches
Splitting critical edges when one of the source edges is an indirectbr
is hard in general (because it requires changing the memory the indirectbr
reads). But if a block only has a single indirectbr predecessor (which is
the common case), we can simulate splitting that edge by splitting
the destination block, and retargeting the *direct* branches.
This is motivated by the use of computed gotos in python 2.7: PyEval_EvalFrame()
ends up using an indirect branch with ~100 successors, and passing a constant to
each of those. Since MachineSink can't break indirect critical edges on demand
(and doing this in MIR doesn't look feasible), this causes us to emit about ~100
defs of registers containing constants, which we in the predecessor block, where
only one of those constants is used in each successor. So, at each computed goto,
we needlessly spill about a 100 constants to stack. The end result is that a
clang-compiled python interpreter can be about ~2.5x slower on a simple python
reduction loop than a gcc-compiled interpreter.
Adrian McCarthy [Fri, 24 Feb 2017 00:10:47 +0000 (00:10 +0000)]
Implement some methods for NativeRawSymbol
This allows the ability to call IPDBSession::getGlobalScope with a NativeSession and
to then query it for some basic fields from the PDB's InfoStream.
Note that the symbols now have non-const references back to the Session so that
NativeRawSymbol can access the PDBFile through the Session.
Ahmed Bougacha [Thu, 23 Feb 2017 23:57:36 +0000 (23:57 +0000)]
[GlobalISel] Don't translate other blocks when one failed.
We were stopping the translation of the parent block when the
translation of an instruction failed, but we were still trying to
translate the other blocks of the parent function.
Xin Tong [Thu, 23 Feb 2017 23:42:19 +0000 (23:42 +0000)]
LoopUnswitch - Simplify based on known not to a be constant.
Summary: In case we do not know what the condition is in an unswitched loop, but we know its definitely NOT a known constant. We can perform simplifcations based on this information.
Bryant Wong [Thu, 23 Feb 2017 23:00:46 +0000 (23:00 +0000)]
[ADT] Fix zip iterator interface.
This commit provides `zip_{first,shortest}` with the standard member types and
methods expected of iterators (e.g., `difference_type`), in order for zip to be
used with other adaptors, such as `make_filter_range`.
Support for reverse iteration has also been added.
Tim Northover [Thu, 23 Feb 2017 22:35:00 +0000 (22:35 +0000)]
ARM: make sure FastISel bails on f64 operations for Cortex-M4.
FastISel wasn't checking the isFPOnlySP subtarget feature before emitting
double-precision operations, so it got completely invalid CodeGen for doubles
on Cortex-M4F.
The normal ISel testing wasn't spectacular either so I added a second RUN line
to improve that while I was in the area.
Vedant Kumar [Thu, 23 Feb 2017 22:20:32 +0000 (22:20 +0000)]
[llvm-cov] Strip redundant path components from filenames (fix PR31982)
Instead of stripping the longest common prefix off of the filenames in a
report, strip out the longest chain of redundant path components. This
fixes the case in PR31982, where there are two files with the same
prefix, and stripping out the LCP makes things less intelligible.
Dehao Chen [Thu, 23 Feb 2017 22:15:18 +0000 (22:15 +0000)]
Add call branch annotation for ICP promoted direct call in SamplePGO mode.
Summary: SamplePGO uses branch_weight annotation to represent callsite hotness. When ICP promotes an indirect call to direct call, we need to make sure the direct call is annotated with branch_weight in SamplePGO mode, so that downstream function inliner can use hot callsite heuristic.
[Hexagon] Allow setting register in BitVal without storing into map
In the bit tracker, references to other bit values in which the register
is 0 are prohibited. This means that generating self-referential register
cells like { w:32 [0-15]:s[0-15] [16-31]:s[15] } is impossible. In order
to get a self-referential cell, it had to be stored into a map and then
reloaded from it. To avoid this step, add a function that will set the
register to a given value without going through the map.
Ahmed Bougacha [Thu, 23 Feb 2017 21:05:42 +0000 (21:05 +0000)]
[GlobalISel] Emit opt remarks on isel fallbacks.
Having more fine-grained information on the specific construct that
caused us to fallback is valuable for large-scale data collection.
We still have the fallback warning, that's also used for FastISel.
We still need to remove the fallback warning, and teach FastISel to also
emit remarks (it currently has a combination of the warning, stats, and
debug prints: the remarks could unify all three).
The abort-on-fallback path could also be better handled using remarks:
one could imagine a "-Rpass-error", analoguous to "-Werror", which would
promote missed/failed remarks to errors. It's not clear whether that
would be useful for other remarks though, so we're not there yet.
Correct register pressure calculation in presence of subregs
If a subreg is used in an instruction it counts as a whole superreg
for the purpose of register pressure calculation. This patch corrects
improper register pressure calculation by examining operand's lane mask.
Chad Rosier [Thu, 23 Feb 2017 18:49:03 +0000 (18:49 +0000)]
[Reassociate] Add negated value of negative constant to the Duplicates list.
In OptimizeAdd, we scan the operand list to see if there are any common factors
between operands that can be factored out to reduce the number of multiplies
(e.g., 'A*A+A*B*C+D' -> 'A*(A+B*C)+D'). For each operand of the operand list, we
only consider unique factors (which is tracked by the Duplicate set). Now if we
find a factor that is a negative constant, we add the negated value as a factor
as well, because we can percolate the negate out. However, we mistakenly don't
add this negated constant to the Duplicates set.
Consider the expression A*2*-2 + B. Obviously, nothing to factor.
For the added value A*2*-2 we over count 2 as a factor without this change,
which causes the assert reported in PR30256. The problem is that this code is
assuming that all the multiply operands of the add are already reassociated.
This change avoids the issue by making OptimizeAdd tolerate multiplies which
haven't been completely optimized; this sort of works, but we're doing wasted
work: we'll end up revisiting the add later anyway.
Another possible approach would be to enforce RPO iteration order more strongly.
If we have RedoInsts, we process them immediately in RPO order, rather than
waiting until we've finished processing the whole function. Intuitively, it
seems like the natural approach: reassociation works on expression trees, so
the optimization only works in one direction. That said, I'm not sure how
practical that is given the current Reassociate; the "optimal" form for an
expression depends on its use list (see all the uses of "user_back()"), so
Reassociate is really an iterative optimization of sorts, so any changes here
would probably get messy.
Adam Nemet [Thu, 23 Feb 2017 17:30:01 +0000 (17:30 +0000)]
[LazyMachineBFI] Reimplement with getAnalysisIfAvailable
Since LoopInfo is not available in machine passes as universally as in IR
passes, using the same approach for OptimizationRemarkEmitter as we did for IR
will run LoopInfo and DominatorTree unnecessarily. (LoopInfo is not used
lazily by ORE.)
To fix this, I am modifying the approach I took in D29836. LazyMachineBFI now
uses its client passes including MachineBFI itself that are available or
otherwise compute them on the fly.
So for example GreedyRegAlloc, since it's already using MBFI, will reuse that
instance. On the other hand, AsmPrinter in Justin's patch will generate DT,
LI and finally BFI on the fly.
(I am of course wondering now if the simplicity of this approach is even
preferable in IR. I will do some experiments.)
Testing is provided by an updated version of D29837 which requires Justin's
patch to bring ORE to the AsmPrinter.
Diana Picus [Thu, 23 Feb 2017 14:18:41 +0000 (14:18 +0000)]
[ARM] GlobalISel: Lower call returns
Introduce a common ValueHandler for call returns and formal arguments, and
inherit two different versions for handling the differences (at the moment the
only difference is the way physical registers are marked as used).
Alexey Bataev [Thu, 23 Feb 2017 13:37:09 +0000 (13:37 +0000)]
[SLP] Fix for PR32036: Vectorized horizontal reduction returning wrong
result
Summary:
If the same value is used several times as an extra value, SLP
vectorizer takes it into account only once instead of actual number of
using.
For example:
```
int val = 1;
for (int y = 0; y < 8; y++) {
for (int x = 0; x < 8; x++) {
val = val + input[y * 8 + x] + 3;
}
}
```
We have 2 extra rguments: `1` - initial value of horizontal reduction
and `3`, which is added 8*8 times to the reduction. Before the patch we
added `1` to the reduction value and added once `3`, though it must be
added 64 times.
Diana Picus [Thu, 23 Feb 2017 13:25:43 +0000 (13:25 +0000)]
[ARM] GlobalISel: Lower call parameters in regs
Add support for lowering calls with parameters than can fit into regs. Use the
same ValueHandler that we used for function returns, but rename it to match its
new, extended purpose.
projects / llvm / log