Matt Arsenault [Thu, 23 Feb 2017 03:58:53 +0000 (03:58 +0000)]
LoadStoreVectorizer: Split even sized illegal chains properly
Implement isLegalToVectorizeLoadChain for AMDGPU to avoid
producing private address spaces accesses that will need to be
split up later. This was doing the wrong thing in the case
where the queried chain was an even number of elements.
A possible <4 x i32> store was being split into
store <2 x i32>
store i32
store i32
Craig Topper [Thu, 23 Feb 2017 03:22:14 +0000 (03:22 +0000)]
[X86][IR] In AutoUpgrade, check explicitly for xop.vpcmov and xop.vpcmov.256 instead of anything starting with xop.vpcmov
There were some older intrinsics that only existed for less than a month in 2012 that still exist in some out of tree test files that start with this string, but aren't able to be handled by the current upgrade code and fire an assert. Now we'll go back to treating them as not intrinsics at all and just passing them through to output.
Matt Arsenault [Wed, 22 Feb 2017 23:27:53 +0000 (23:27 +0000)]
AMDGPU: Fold FP clamp as modifier bit
The manual is unclear on the details of this. It's not
clear to me if denormals are not allowed with clamp,
or if that is only omod. Not allowing denorms for
fp16 or fp64 isn't useful so I also question if that
is really a restriction. Same with whether this is valid
without IEEE mode enabled.
Sanjay Patel [Wed, 22 Feb 2017 23:01:12 +0000 (23:01 +0000)]
[InstCombine] don't try SimplifyDemandedInstructionBits from add/sub because it's slow and unlikely to succeed
Notably, no regression tests change when we remove these calls, and these are expensive calls.
The motivation comes from the general acknowledgement that the compiler is getting slower:
http://lists.llvm.org/pipermail/llvm-dev/2017-January/109188.html
http://lists.llvm.org/pipermail/llvm-dev/2016-December/108279.html
And specifically the test case attached to PR32037:
https://bugs.llvm.org//show_bug.cgi?id=32037
Profiling the middle-end (opt) part of the compile:
$ ./opt -O2 row_common.bc -o /dev/null
...visitAdd and visitSub are near the top of the instcombine list, and the calls to SimplifyDemandedInstructionBits()
are high within each of those. Those calls account for 1%+ of the opt time in either debug or release profiles. And
that's the rough win I see from this patch when testing opt built release from r295864 on an iMac with Haswell 4GHz
(model 4790K).
It seems unlikely that we'd be able to eliminate add/sub or change their operands given that add/sub normally affect
all bits, and the PR32037 example shows no IR difference after this change using -O2.
Also worth noting - the code comment in visitAdd:
// This handles stuff like (X & 254)+1 -> (X&254)|1
...isn't true. That transform is handled later with a call to haveNoCommonBitsSet().
Matt Arsenault [Wed, 22 Feb 2017 22:23:32 +0000 (22:23 +0000)]
AMDGPU: Don't add emergency stack slot if all spills are SGPR->VGPR
This should avoid reporting any stack needs to be allocated in the
case where no stack is truly used. An unused stack slot is still
left around in other cases where there are real stack objects
but no spilling occurs.
Daniel Berlin [Wed, 22 Feb 2017 22:20:58 +0000 (22:20 +0000)]
PredicateInfo: Support switch statements
Summary:
Depends on D29606 and D29682
Makes us pass GVN's edge.ll (we also will pass a few other testcases
they just need cleaning up).
Thoughts on the Predicate* hiearchy of classes especially welcome :)
(it's not clear to me how best to organize it, and currently, the getBlock* seems ... uglier than maybe wasting a field somewhere or something).
Wei Mi [Wed, 22 Feb 2017 21:47:08 +0000 (21:47 +0000)]
[LSR] Canonicalize formula and put recursive Reg related with current loop in ScaledReg.
After rL294814, LSR formula can have multiple SCEVAddRecExprs inside of its BaseRegs.
Previous canonicalization will swap the first SCEVAddRecExpr in BaseRegs with ScaledReg.
But now we want to swap the SCEVAddRecExpr Reg related with current loop with ScaledReg.
Otherwise, we may generate code like this: RegA + lsr.iv + RegB, where loop invariant
parts RegA and RegB are not grouped together and cannot be promoted outside of loop.
With this patch, it will ensure lsr.iv to be generated later in the expr:
RegA + RegB + lsr.iv, so that RegA + RegB can be promoted outside of loop.
Matt Arsenault [Wed, 22 Feb 2017 21:05:25 +0000 (21:05 +0000)]
AMDGPU: Always allocate emergency stack slot at offset 0
This allows us to ensure that 0 is never a valid pointer
to a user object, and ensures that the offset is always legal
without needing a register to access it. This comes at the cost
of usable offsets and wasted stack space.
Alexey Bataev [Wed, 22 Feb 2017 20:06:40 +0000 (20:06 +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.
Geoff Berry [Wed, 22 Feb 2017 19:10:45 +0000 (19:10 +0000)]
[AArch64] Extend AArch64RedundantCopyElimination to do simple copy propagation.
Summary:
Extend AArch64RedundantCopyElimination to catch cases where the register
that is known to be zero is COPY'd in the predecessor block. Before
this change, this pass would catch cases like:
CBZW %W0, <BB#1>
BB#1:
%W0 = COPY %WZR // removed
After this change, cases like the one below are also caught:
This change results in a 4% increase in static copies removed by this
pass when compiling the llvm test-suite. It also fixes regressions
caused by doing post-RA copy propagation (a separate change to be put up
for review shortly).
Dan Gohman [Wed, 22 Feb 2017 18:34:16 +0000 (18:34 +0000)]
[WebAssembly] Define a table of function signatures for runtime library calls.
LLVM CodeGen emits references to external symbols that are never declared in
LLVM IR level, so they have no declared signature. However, WebAssembly requires
all functions be declared with signatures. This patch adds a table for providing
signatures for known runtime libcalls that will be used in subsequent patches to
emit declarations for such functions.
Michal Gorny [Wed, 22 Feb 2017 18:09:15 +0000 (18:09 +0000)]
[Support] Provide linux/magic.h fallback for older kernels
The function for distinguishing local and remote files added in r295768
unconditionally uses linux/magic.h header to provide necessary
filesystem magic numbers. However, in kernel headers predating 2.6.18
the magic numbers are spread throughout multiple include files.
Furthermore, LLVM did not require kernel headers being installed so far.
To increase the portability across different versions of Linux kernel
and different Linux systems, add CMake header checks for linux/magic.h
and -- if it is missing -- the linux/nfs_fs.h and linux/smb.h headers
which contained the numbers previously.
Furthermore, since the numbers are static and the feature does not seem
critical enough to make LLVM require kernel headers at all, add fallback
constants for the case when none of the necessary headers is available.
Bill Seurer [Wed, 22 Feb 2017 16:27:33 +0000 (16:27 +0000)]
[DAGCombiner] revert r295336
r295336 causes a bootstrapped clang to fail for many compilations on
powerpc BE. See
http://lab.llvm.org:8011/builders/clang-ppc64be-linux-multistage/builds/2315
for example.
The pass tries to fix a spill of LR that turns out to be unnecessary.
So it removes the tPOP but forgets to remove tPUSH.
This causes the stack be misaligned upon returning the function.
Javed Absar [Wed, 22 Feb 2017 07:22:57 +0000 (07:22 +0000)]
[ARM] Classification Improvements to ARM Sched-Models. NFCI.
This patch adds missing sched classes for Thumb2 instructions.
This has been missing so far, and as a consequence, machine
scheduler models for individual sub-targets have tended to
be larger than they needed to be. These patches should help
write schedulers better and faster in the future
for ARM sub-targets.
Craig Topper [Wed, 22 Feb 2017 06:54:18 +0000 (06:54 +0000)]
[AVX-512] Allow legacy scalar min/max intrinsics to select EVEX instructions when available
This patch introduces new X86ISD::FMAXS and X86ISD::FMINS opcodes. The legacy intrinsics now lower to this node. As do the AVX-512 masked intrinsics when the rounding mode is CUR_DIRECTION.
I've merged a copy of the tablegen multiclass avx512_fp_scalar into avx512_fp_scalar_sae. avx512_fp_scalar still needs to support CUR_DIRECTION appearing as a rounding mode for X86ISD::FADD_ROUND and others.
Sanjoy Das [Wed, 22 Feb 2017 06:52:32 +0000 (06:52 +0000)]
[ValueTracking] Make poison propagation more aggressive
Summary:
Motivation: fix PR31181 without regression (the actual fix is still in
progress). However, the actual content of PR31181 is not relevant
here.
This change makes poison propagation more aggressive in the following
cases:
1. poision * Val == poison, for any Val. In particular, this changes
existing intentional and documented behavior in these two cases:
a. Val is 0
b. Val is 2^k * N
2. poison << Val == poison, for any Val
3. getelementptr is poison if any input is poison
I think all of these are justified (and are axiomatically true in the
new poison / undef model):
1a: we need poison * 0 to be poison to allow transforms like these:
A * (B + C) ==> A * B + A * C
If poison * 0 were 0 then the above transform could not be allowed
since e.g. we could have A = poison, B = 1, C = -1, making the LHS
1b: we need e.g. poison * 4 to be poison since we want to allow
A * 4 ==> A + A + A + A
If poison * 4 were a value with all of their bits poison except the
last four; then we'd not be able to do this transform since then if A
were poison the LHS would only be "partially" poison while the RHS
would be "full" poison.
2: Same reasoning as (1b), we'd like have the following kinds
transforms be legal:
Matt Arsenault [Tue, 21 Feb 2017 23:35:48 +0000 (23:35 +0000)]
AMDGPU: Redefine clamp node as clamp 0.0-1.0
Change implementation to use max instead of add.
min/max/med3 do not flush denormals regardless of the mode,
so it is OK to use it whether or not they are enabled.
Also allow using clamp with f16, and use knowledge
of dx10_clamp.
Artem Belevich [Tue, 21 Feb 2017 22:56:05 +0000 (22:56 +0000)]
[NVPTX] Unify vectorization of load/stores of aggregate arguments and return values.
Original code only used vector loads/stores for explicit vector arguments.
It could also do more loads/stores than necessary (e.g v5f32 would
touch 8 f32 values). Aggregate types were loaded one element at a time,
even the vectors contained within.
This change attempts to generalize (and simplify) parameter space
loads/stores so that vector loads/stores can be used more broadly.
Functionality of the patch has been verified by compiling thrust
test suite and manually checking the differences between PTX
generated by llvm with and without the patch.
General algorithm:
* ComputePTXValueVTs() flattens input/output argument into a flat list
of scalars to load/store and returns their types and offsets.
* VectorizePTXValueVTs() uses that data to create vectorization plan
which returns an array of flags marking boundaries of vectorized
load/stores. Scalars are represented as 1-element vectors.
* Code that generates loads/stores implements a simple state machine
that constructs a vector according to the plan.
Zachary Turner [Tue, 21 Feb 2017 21:31:28 +0000 (21:31 +0000)]
Try to fix the buildbot on OSX.
Since I'm only seeing failures on OSX, and it's saying
permission denied, I'm suspecting this is due to the addition
of the MAP_RESILIENT_CODESIGN and/or MAP_RESILIENT_MEDIA flags.
Speculatively trying to remove those to get the bots working.
Rafael Espindola [Tue, 21 Feb 2017 20:40:54 +0000 (20:40 +0000)]
Don't modify archive members unless really needed.
For whatever reason ld64 requires that member headers (not the member
themselves) should be aligned. The only way to do that is to edit the
previous member so that it ends at an aligned boundary.
Since modifying data put in an archive is an undesirable property,
llvm-ar should only do it when it is absolutely necessary.
Sanjay Patel [Tue, 21 Feb 2017 19:33:53 +0000 (19:33 +0000)]
[InstCombine] canonicalize non-obivous forms of integer min/max
This is part of trying to clean up our handling of min/max patterns in IR.
By converting these to canonical form, we're more likely to recognize them
because there are various places in InstCombine that don't use
matchSelectPattern or m_SMax and friends.
The backend fixups referenced in the now deleted TODO comment were added with:
https://reviews.llvm.org/rL291392
https://reviews.llvm.org/rL289738
If there's any codegen fallout from this change, we should be able to address
it in DAGCombiner or target-specific lowering.
Zachary Turner [Tue, 21 Feb 2017 19:29:56 +0000 (19:29 +0000)]
Remove svn:eol-style property from 2 files.
There are still over 3400 files remaining with this property set, but there are tens of thousands more with the property not set. Until we decide what to do on a global scale, this at least unblocks me temporarily.
Matt Arsenault [Tue, 21 Feb 2017 19:12:08 +0000 (19:12 +0000)]
AMDGPU: Don't use stack space for SGPR->VGPR spills
Before frame offsets are calculated, try to eliminate the
frame indexes used by SGPR spills. Then we can delete them
after.
I think for now we can be sure that no other instruction
will be re-using the same frame indexes. It should be easy
to notice if this assumption ever breaks since everything
asserts if it tries to use a dead frame index later.
The unused emergency stack slot seems to still be left behind,
so an additional 4 bytes is still wasted.
Adrian Prantl [Tue, 21 Feb 2017 19:03:15 +0000 (19:03 +0000)]
Teach the IR verifier to reject conflicting debug info for function arguments.
Conflicting debug info for function arguments causes hard-to-debug
assertions in the DWARF backend, so the Verifier should reject it.
For performance reasons this only checks function arguments from
non-inlined debug intrinsics for now.
Geoff Berry [Tue, 21 Feb 2017 18:53:14 +0000 (18:53 +0000)]
[CodeGenPrepare] Sink and duplicate more 'and' instructions.
Summary:
Rework the code that was sinking/duplicating (icmp and, 0) sequences
into blocks where they were being used by conditional branches to form
more tbz instructions on AArch64. The new code is more general in that
it just looks for 'and's that have all icmp 0's as users, with a target
hook used to select which subset of 'and' instructions to consider.
This change also enables 'and' sinking for X86, where it is more widely
beneficial than on AArch64.
The 'and' sinking/duplicating code is moved into the optimizeInst phase
of CodeGenPrepare, where it can take advantage of the fact the
OptimizeCmpExpression has already sunk/duplicated any icmps into the
blocks where they are used. One minor complication from this change is
that optimizeLoadExt needed to be updated to always mark 'and's it has
determined should be in the same block as their feeding load in the
InsertedInsts set to avoid an infinite loop of hoisting and sinking the
same 'and'.
This change fixes a regression on X86 in the tsan runtime caused by
moving GVNHoist to a later place in the optimization pipeline (see
PR31382).
projects / llvm / log