When we load the v12i32 type, the GenWidenVectorLoads method generates two loads: v8i32 and v4i32
and attempts to use CONCAT_VECTORS to join them. In this fix I concat undef values to widen
the smaller value. The test "widen_load-2.ll" also exposes this bug on AVX.
llvm-svn: 147964
hoped this would revive one of the llvm-gcc selfhost build bots, but it
didn't so it doesn't appear that my transform is the culprit.
If anyone else is seeing failures, please let me know!
llvm-svn: 147957
strange build bot failures that look like a miscompile into an infloop.
I'll investigate this tomorrow, but I'd both like to know whether my
patch is the culprit, and get the bots back to green.
llvm-svn: 147945
detect a pattern which can be implemented with a small 'shl' embedded in
the addressing mode scale. This happens in real code as follows:
unsigned x = my_accelerator_table[input >> 11];
Here we have some lookup table that we look into using the high bits of
'input'. Each entity in the table is 4-bytes, which means this
implicitly gets turned into (once lowered out of a GEP):
*(unsigned*)((char*)my_accelerator_table + ((input >> 11) << 2));
The shift right followed by a shift left is canonicalized to a smaller
shift right and masking off the low bits. That hides the shift right
which x86 has an addressing mode designed to support. We now detect
masks of this form, and produce the longer shift right followed by the
proper addressing mode. In addition to saving a (rather large)
instruction, this also reduces stalls in Intel chips on benchmarks I've
measured.
In order for all of this to work, one part of the DAG needs to be
canonicalized *still further* than it currently is. This involves
removing pointless 'trunc' nodes between a zextload and a zext. Without
that, we end up generating spurious masks and hiding the pattern.
llvm-svn: 147936
Consider this code:
int h() {
int x;
try {
x = f();
g();
} catch (...) {
return x+1;
}
return x;
}
The variable x is undefined on the first edge to the landing pad, but it
has the f() return value on the second edge to the landing pad.
SplitAnalysis::getLastSplitPoint() would assume that the return value
from f() was live into the landing pad when f() throws, which is of
course impossible.
Detect these cases, and treat them as if the landing pad wasn't there.
This allows spill code to be inserted after the function call to f().
<rdar://problem/10664933>
llvm-svn: 147912
define physical registers. It's currently very restrictive, only catching
cases where the CE is in an immediate (and only) predecessor. But it catches
a surprising large number of cases.
rdar://10660865
llvm-svn: 147827
opportunities that only present themselves after late optimizations
such as tail duplication .e.g.
## BB#1:
movl %eax, %ecx
movl %ecx, %eax
ret
The register allocator also leaves some of them around (due to false
dep between copies from phi-elimination, etc.)
This required some changes in codegen passes. Post-ra scheduler and the
pseudo-instruction expansion passes have been moved after branch folding
and tail merging. They were before branch folding before because it did
not always update block livein's. That's fixed now. The pass change makes
independently since we want to properly schedule instructions after
branch folding / tail duplication.
rdar://10428165
rdar://10640363
llvm-svn: 147716
a combined-away node and the result of the combine isn't substantially
smaller than the input, it's just canonicalized. This is the first part
of a significant (7%) performance gain for Snappy's hot decompression
loop.
llvm-svn: 147604
Testing: passed 'make check' including LIT tests for all sequences being handled (both SSE and AVX)
Reviewers: Evan Cheng, David Blaikie, Bruno Lopes, Elena Demikhovsky, Chad Rosier, Anton Korobeynikov
llvm-svn: 147601
(x > y) ? x : y
=>
(x >= y) ? x : y
So for something like
(x - y) > 0 : (x - y) ? 0
It will be
(x - y) >= 0 : (x - y) ? 0
This makes is possible to test sign-bit and eliminate a comparison against
zero. e.g.
subl %esi, %edi
testl %edi, %edi
movl $0, %eax
cmovgl %edi, %eax
=>
xorl %eax, %eax
subl %esi, $edi
cmovsl %eax, %edi
rdar://10633221
llvm-svn: 147512
The failure seen on win32, when i64 type is illegal.
It happens on stage of conversion VECTOR_SHUFFLE to BUILD_VECTOR.
The failure message is:
llc: SelectionDAG.cpp:784: void VerifyNodeCommon(llvm::SDNode*): Assertion `(I->getValueType() == EltVT || (EltVT.isInteger() && I->getValueType().isInteger() && EltVT.bitsLE(I->getValueType()))) && "Wrong operand type!"' failed.
I added a special test that checks vector shuffle on win32.
llvm-svn: 147445
The failure seen on win32, when i64 type is illegal.
It happens on stage of conversion VECTOR_SHUFFLE to BUILD_VECTOR.
The failure message is:
llc: SelectionDAG.cpp:784: void VerifyNodeCommon(llvm::SDNode*): Assertion `(I->getValueType() == EltVT || (EltVT.isInteger() && I->getValueType().isInteger() && EltVT.bitsLE(I->getValueType()))) && "Wrong operand type!"' failed.
I added a special test that checks vector shuffle on win32.
llvm-svn: 147399
LZCNT instructions are available. Force promotion to i32 to get
a smaller encoding since the fix-ups necessary are just as complex for
either promoted type
We can't do standard promotion for CTLZ when lowering through BSR
because it results in poor code surrounding the 'xor' at the end of this
instruction. Essentially, if we promote the entire CTLZ node to i32, we
end up doing the xor on a 32-bit CTLZ implementation, and then
subtracting appropriately to get back to an i8 value. Instead, our
custom logic just uses the knowledge of the incoming size to compute
a perfect xor. I'd love to know of a way to fix this, but so far I'm
drawing a blank. I suspect the legalizer could be more clever and/or it
could collude with the DAG combiner, but how... ;]
llvm-svn: 147251
my C-brain happy. Remove the unnecessary bits of pedantic IR fluff like
nounwind. Remove stray uses comments. Name things semantically rather
than tN so that adding a new test in the middle doesn't cause pain, and
so that new tests can be grouped semantically.
This exposes how little systematic testing is going on here. I noticed
this by finding several bugs via inspection and wondering why this test
wasn't catching any of them. =[
llvm-svn: 147248
'bsf' instructions here.
This one is actually debatable to my eyes. It's not clear that any chip
implementing 'tzcnt' would have a slow 'bsf' for any reason, and unless
EFLAGS or a zero input matters, 'tzcnt' is just a longer encoding.
Still, this restores the old behavior with 'tzcnt' enabled for now.
llvm-svn: 147246
X86ISelLowering C++ code. Because this is lowered via an xor wrapped
around a bsr, we want the dagcombine which runs after isel lowering to
have a chance to clean things up. In particular, it is very common to
see code which looks like:
(sizeof(x)*8 - 1) ^ __builtin_clz(x)
Which is trying to compute the most significant bit of 'x'. That's
actually the value computed directly by the 'bsr' instruction, but if we
match it too late, we'll get completely redundant xor instructions.
The more naive code for the above (subtracting rather than using an xor)
still isn't handled correctly due to the dagcombine getting confused.
Also, while here fix an issue spotted by inspection: we should have been
expanding the zero-undef variants to the normal variants when there is
an 'lzcnt' instruction. Do so, and test for this. We don't want to
generate unnecessary 'bsr' instructions.
These two changes fix some regressions in encoding and decoding
benchmarks. However, there is still a *lot* to be improve on in this
type of code.
llvm-svn: 147244
use the zero-undefined variants of CTTZ and CTLZ. These are just simple
patterns for now, there is more to be done to make real world code using
these constructs be optimized and codegen'ed properly on X86.
The existing tests are spiffed up to check that we no longer generate
unnecessary cmov instructions, and that we generate the very important
'xor' to transform bsr which counts the index of the most significant
one bit to the number of leading (most significant) zero bits. Also they
now check that when the variant with defined zero result is used, the
cmov is still produced.
llvm-svn: 146974
I followed three heuristics for deciding whether to set 'true' or
'false':
- Everything target independent got 'true' as that is the expected
common output of the GCC builtins.
- If the target arch only has one way of implementing this operation,
set the flag in the way that exercises the most of codegen. For most
architectures this is also the likely path from a GCC builtin, with
'true' being set. It will (eventually) require lowering away that
difference, and then lowering to the architecture's operation.
- Otherwise, set the flag differently dependending on which target
operation should be tested.
Let me know if anyone has any issue with this pattern or would like
specific tests of another form. This should allow the x86 codegen to
just iteratively improve as I teach the backend how to differentiate
between the two forms, and everything else should remain exactly the
same.
llvm-svn: 146370
We must not issue a bitcast operation for integer-promotion of vector types, because the
location of the values in the vector may be different.
llvm-svn: 146150
libgcc sets the stack limit field in TCB to 256 bytes above the actual
allocated stack limit. This means if the function's stack frame needs
less than 256 bytes, we can just compare the stack pointer with the
stack limit. This should result in lesser calls to __morestack.
llvm-svn: 145766
Currently LLVM pads the call to __morestack with a add and sub of 8
bytes to esp. This isn't correct since __morestack expects the call
to be followed directly by a ret.
This commit also adjusts the relevant test-case.
llvm-svn: 145765
Like V_SET0, these instructions are expanded by ExpandPostRA to xorps /
vxorps so they can participate in execution domain swizzling.
This also makes the AVX variants redundant.
llvm-svn: 145440
Conservatively returns zero when the GV does not specify an alignment nor is it
initialized. Previously it returns ABI alignment for type of the GV. However, if
the type is a "packed" type, then the under-specified alignments is attached to
the load / store instructions. In that case, the alignment of the type cannot be
trusted.
rdar://10464621
llvm-svn: 145300
than ABI alignment. These are loads / stores from / to "packed" data structures.
Their alignments are intentionally under-specified.
rdar://10301431
llvm-svn: 145273
was centered around the premise of laying out a loop in a chain, and
then rotating that chain. This is good for preserving contiguous layout,
but bad for actually making sane rotations. In order to keep it safe,
I had to essentially make it impossible to rotate deeply nested loops.
The information needed to correctly reason about a deeply nested loop is
actually available -- *before* we layout the loop. We know the inner
loops are already fused into chains, etc. We lose information the moment
we actually lay out the loop.
The solution was the other alternative for this algorithm I discussed
with Benjamin and some others: rather than rotating the loop
after-the-fact, try to pick a profitable starting block for the loop's
layout, and then use our existing layout logic. I was worried about the
complexity of this "pick" step, but it turns out such complexity is
needed to handle all the important cases I keep teasing out of benchmarks.
This is, I'm afraid, a bit of a work-in-progress. It is still
misbehaving on some likely important cases I'm investigating in Olden.
It also isn't really tested. I'm going to try to craft some interesting
nested-loop test cases, but it's likely to be extremely time consuming
and I don't want to go there until I'm sure I'm testing the correct
behavior. Sadly I can't come up with a way of getting simple, fine
grained test cases for this logic. We need complex loop structures to
even trigger much of it.
llvm-svn: 145183
heavily on AnalyzeBranch. That routine doesn't behave as we want given
that rotation occurs mid-way through re-ordering the function. Instead
merely check that there are not unanalyzable branching constructs
present, and then reason about the CFG via successor lists. This
actually simplifies my mental model for all of this as well.
The concrete result is that we now will rotate more loop chains. I've
added a test case from Olden highlighting the effect. There is still
a bit more to do here though in order to regain all of the performance
in Olden.
llvm-svn: 145179
pass. This is designed to achieve one of the important optimizations
that the old code placement pass did, but more simply.
This is a somewhat rough and *very* conservative version of the
transform. We could get a lot fancier here if there are profitable cases
to do so. In particular, this only looks for a single pattern, it
insists that the loop backedge being rotated away is the last backedge
in the chain, and it doesn't provide any means of doing better in-loop
placement due to the rotation. However, it appears that it will handle
the important loops I am finding in the LLVM test suite.
llvm-svn: 145158
was returning incorrect values in rare cases, and incorrectly marking
exact conversions as inexact in some more common cases. Fixes PR11406, and a
missed optimization in test/CodeGen/X86/fp-stack-O0.ll.
llvm-svn: 145141
need lots of fanciness around retaining a reference to a Chain's slot in
the BlockToChain map, but that's all gone now. We can just go directly
to allocating the new chain (which will update the mapping for us) and
using it.
Somewhat gross mechanically generated test case replicates the issue
Duncan spotted when actually testing this out.
llvm-svn: 145120
conflicts, we should only be adding the first block of the chain to the
list, lest we try to merge into the middle of that chain. Most of the
places we were doing this we already happened to be looking at the first
block, but there is no reason to assume that, and in some cases it was
clearly wrong.
I've added a couple of tests here. One already worked, but I like having
an explicit test for it. The other is reduced from a test case Duncan
reduced for me and used to crash. Now it is handled correctly.
llvm-svn: 145119
Before:
movabsq $4294967296, %rax ## encoding: [0x48,0xb8,0x00,0x00,0x00,0x00,0x01,0x00,0x00,0x00]
testq %rax, %rdi ## encoding: [0x48,0x85,0xf8]
jne LBB0_2 ## encoding: [0x75,A]
After:
btq $32, %rdi ## encoding: [0x48,0x0f,0xba,0xe7,0x20]
jb LBB0_2 ## encoding: [0x72,A]
btq is usually slower than testq because it doesn't fuse with the jump, but here we're better off
saving one register and a giant movabsq.
llvm-svn: 145103
further. This invariant just wasn't going to work in the face of
unanalyzable branches; we need to be resillient to the phenomenon of
chains poking into a loop and poking out of a loop. In fact, we already
were, we just needed to not assert on it.
This was found during a bootstrap with block placement turned on.
llvm-svn: 145100
successors, they just are all landing pad successors. We handle this the
same way as no successors. Comments attached for the next person to wade
through here and another lovely test case courtesy of Benjamin Kramer's
bugpoint reduction.
llvm-svn: 145098
This was a bug in keeping track of the available domains when merging
domain values.
The wrong domain mask caused ExecutionDepsFix to try to move VANDPSYrr
to the integer domain which is only available in AVX2.
Also add an assertion to catch future attempts at emitting AVX2
instructions.
llvm-svn: 145096
reversed in the function's original ordering, and we happened to
encounter it while handling an outer unnatural CFG structure.
Thanks to the test case reduced from GCC's source by Benjamin Kramer.
This may also fix a crasher in gzip that Duncan reduced for me, but
I haven't yet gotten to testing that one.
llvm-svn: 145094
updateTerminator code didn't correctly handle EH terminators in one very
specific case. AnalyzeBranch would find no terminator instruction, and
so the fallback in updateTerminator is to assume fallthrough. This is
correct, but the destination of the fallthrough was assumed to be the
first successor.
This is *almost always* true, but in certain cases the loop
transformations will cause the landing pad to be the first successor!
Instead of this brittle logic, actually look through the successors for
a non-landing-pad accessor, and to assert if more than one is found.
This will hopefully fix some (if not all) of the self host miscompiles
with block placement. Thanks to Benjamin Kramer for reporting, Nick
Lewycky for an initial stab at a reduction, and Duncan for endless
advice on EH (which I know nothing about) as well as reviewing the
actual fix.
llvm-svn: 145062
properly account for the *global* probability of the edge being taken.
This manifested as a very large number of unconditional branches to
blocks being merged against the CFG even though they weren't
particularly hot within the CFG.
The fix is to check whether the edge being merged is both locally hot
relative to other successors for the source block, and globally hot
compared to other (unmerged) predecessors of the destination block.
This introduces a new crasher on GCC single-source, but it's currently
behind a flag, and Ben has offered to work on the reduction. =]
llvm-svn: 145010
is actually being tested. Also add some FileCheck goodness to much more
carefully ensure that the result is the desired result. Before this test
would only have failed through an assert failure if the underlying fix
were reverted.
Also, add some weight metadata and a comment explaining exactly what is
going on to a trick section of the test case. Originally, we were
getting very unlucky and trying to form a block chain that isn't
actually profitable. I'm working on a fix to avoid forming these
unprofitable chains, and that would also have masked any failure from
this test case. The easy solution is to add some metadata that makes it
*really* profitable to form the bad chain here.
llvm-svn: 145006
formation phase and into the initial walk of the basic blocks. We
essentially pre-merge all blocks where unanalyzable fallthrough exists,
as we won't be able to update the terminators effectively after any
reorderings. This is quite a bit more principled as there may be CFGs
where the second half of the unanalyzable pair has some analyzable
predecessor that gets placed first. Then it may get placed next,
implicitly breaking the unanalyzable branch even though we never even
looked at the part that isn't analyzable. I've included a test case that
triggers this (thanks Benjamin yet again!), and I'm hoping to synthesize
some more general ones as I dig into related issues.
Also, to make this new scheme work we have to be able to handle branches
into the middle of a chain, so add this check. We always fallback on the
incoming ordering.
Finally, this starts to really underscore a known limitation of the
current implementation -- we don't consider broken predecessors when
merging successors. This can caused major missed opportunities, and is
something I'm planning on looking at next (modulo more bug reports).
llvm-svn: 144994