to catch cases like:
%reg1024<def> = MOV r1
%reg1025<def> = MOV r0
%reg1026<def> = ADD %reg1024, %reg1025
r0 = MOV %reg1026
By commuting ADD, it let coalescer eliminate all of the copies. However, there
was a bug in the heuristics where it ended up commuting the ADD in:
%reg1024<def> = MOV r0
%reg1025<def> = MOV 0
%reg1026<def> = ADD %reg1024, %reg1025
r0 = MOV %reg1026
That did no benefit but rather ensure the last MOV would not be coalesced.
rdar://11355268
llvm-svn: 156048
This patch will optimize the following cases on X86
(a > b) ? (a-b) : 0
(a >= b) ? (a-b) : 0
(b < a) ? (a-b) : 0
(b <= a) ? (a-b) : 0
FROM
movl %edi, %ecx
subl %esi, %ecx
cmpl %edi, %esi
movl $0, %eax
cmovll %ecx, %eax
TO
xorl %eax, %eax
subl %esi, %edi
cmovll %eax, %edi
movl %edi, %eax
rdar: 10734411
llvm-svn: 155919
On x86-32, structure return via sret lets the callee pop the hidden
pointer argument off the stack, which the caller then re-pushes.
However if the calling convention is fastcc, then a register is used
instead, and the caller should not adjust the stack. This is
implemented with a check of IsTailCallConvention
X86TargetLowering::LowerCall but is now checked properly in
X86FastISel::DoSelectCall.
(this time, actually commit what was reviewed!)
llvm-svn: 155825
This time, also fix the caller of AddGlue to properly handle
incomplete chains. AddGlue had failure modes, but shamefully hid them
from its caller. It's luck ran out.
Fixes rdar://11314175: BuildSchedUnits assert.
llvm-svn: 155749
On x86-32, structure return via sret lets the callee pop the hidden
pointer argument off the stack, which the caller then re-pushes.
However if the calling convention is fastcc, then a register is used
instead, and the caller should not adjust the stack. This is
implemented with a check of IsTailCallConvention
X86TargetLowering::LowerCall but is now checked properly in
X86FastISel::DoSelectCall.
llvm-svn: 155745
x == -y --> x+y == 0
x != -y --> x+y != 0
On x86, the generated code goes from
negl %esi
cmpl %esi, %edi
je .LBB0_2
to
addl %esi, %edi
je .L4
This case is correctly handled for ARM with "cmn".
Patch by Manman Ren.
rdar://11245199
PR12545
llvm-svn: 155739
* Model FPSW (the FPU status word) as a register.
* Add ISel patterns for the FUCOM*, FNSTSW and SAHF instructions.
* During Legalize/Lowering, build a node sequence to transfer the comparison
result from FPSW into EFLAGS. If you're wondering about the right-shift: That's
an implicit sub-register extraction (%ax -> %ah) which is handled later on by
the instruction selector.
Fixes PR6679. Patch by Christoph Erhardt!
llvm-svn: 155704
DAGCombine strangeness may result in multiple loads from the same
offset. They both may try to glue themselves to another load. We could
insist that the redundant loads glue themselves to each other, but the
beter fix is to bail out from bad gluing at the time we detect it.
Fixes rdar://11314175: BuildSchedUnits assert.
llvm-svn: 155668
using the pattern (vbroadcast (i32load src)). In some cases, after we generate
this pattern new users are added to the load node, which prevent the selection
of the blend pattern. This commit provides fallback patterns which perform
in-vector broadcast (using in-vector vbroadcast in AVX2 and pshufd on AVX1).
llvm-svn: 155437
on X86 Atom. Some of our tests failed because the tail merging part of
the BranchFolding pass was creating new basic blocks which did not
contain live-in information. When the anti-dependency code in the Post-RA
scheduler ran, it would sometimes rename the register containing
the function return value because the fact that the return value was
live-in to the subsequent block had been lost. To fix this, it is necessary
to run the RegisterScavenging code in the BranchFolding pass.
This patch makes sure that the register scavenging code is invoked
in the X86 subtarget only when post-RA scheduling is being done.
Post RA scheduling in the X86 subtarget is only done for Atom.
This patch adds a new function to the TargetRegisterClass to control
whether or not live-ins should be preserved during branch folding.
This is necessary in order for the anti-dependency optimizations done
during the PostRASchedulerList pass to work properly when doing
Post-RA scheduling for the X86 in general and for the Intel Atom in particular.
The patch adds and invokes the new function trackLivenessAfterRegAlloc()
instead of using the existing requiresRegisterScavenging().
It changes BranchFolding.cpp to call trackLivenessAfterRegAlloc() instead of
requiresRegisterScavenging(). It changes the all the targets that
implemented requiresRegisterScavenging() to also implement
trackLivenessAfterRegAlloc().
It adds an assertion in the Post RA scheduler to make sure that post RA
liveness information is available when it is needed.
It changes the X86 break-anti-dependencies test to use –mcpu=atom, in order
to avoid running into the added assertion.
Finally, this patch restores the use of anti-dependency checking
(which was turned off temporarily for the 3.1 release) for
Intel Atom in the Post RA scheduler.
Patch by Andy Zhang!
Thanks to Jakob and Anton for their reviews.
llvm-svn: 155395
The X86 target is editing the selection DAG while isel is selecting
nodes following a topological ordering. When the DAG hacking triggers
CSE, nodes can be deleted and bad things happen.
llvm-svn: 155257
when the set bits aren't the same for both args of the xor.
This transformation is in the function TargetLowering::SimplifyDemandedBits
in the file lib/CodeGen/SelectionDAG/TargetLowering.cpp.
I have tested this test using a previous version of llc which the defect and
the a version of llc which does not. I got the expected fail and pass,
respectively.
This test goes with rdar://11195364 and the check in with the fix: svn r154955
llvm-svn: 155156
also fix SimplifyLibCalls to use TLI rather than compile-time conditionals to enable optimizations on floor, ceil, round, rint, and nearbyint
llvm-svn: 154960
both fallthrough and a conditional branch target the same successor.
Gracefully delete the conditional branch and introduce any unconditional
branch needed to reach the actual successor. This fixes memory
corruption in 2009-06-15-RegScavengerAssert.ll and possibly other tests.
Also, while I'm here fix a latent bug I spotted by inspection. I never
applied the same fundamental fix to this fallthrough successor finding
logic that I did to the logic used when there are no conditional
branches. As a consequence it would have selected landing pads had they
be aligned in just the right way here. I don't have a test case as
I spotted this by inspection, and the previous time I found this
required have of TableGen's source code to produce it. =/ I hate backend
bugs. ;]
Thanks to Jim Grosbach for helping me reason through this and reviewing
the fix.
llvm-svn: 154867
This is mostly to test the waters. I'd like to get results from FNT
build bots and other bots running on non-x86 platforms.
This feature has been pretty heavily tested over the last few months by
me, and it fixes several of the execution time regressions caused by the
inlining work by preventing inlining decisions from radically impacting
block layout.
I've seen very large improvements in yacr2 and ackermann benchmarks,
along with the expected noise across all of the benchmark suite whenever
code layout changes. I've analyzed all of the regressions and fixed
them, or found them to be impossible to fix. See my email to llvmdev for
more details.
I'd like for this to be in 3.1 as it complements the inliner changes,
but if any failures are showing up or anyone has concerns, it is just
a flag flip and so can be easily turned off.
I'm switching it on tonight to try and get at least one run through
various folks' performance suites in case SPEC or something else has
serious issues with it. I'll watch bots and revert if anything shows up.
llvm-svn: 154816
once we start changing the block layout, so just nuke it. If anyone has
ideas about how to craft a code layout agnostic form of the test please
let me know.
llvm-svn: 154815
rotation. When there is a loop backedge which is an unconditional
branch, we will end up with a branch somewhere no matter what. Try
placing this backedge in a fallthrough position above the loop header as
that will definitely remove at least one branch from the loop iteration,
where whole loop rotation may not.
I haven't seen any benchmarks where this is important but loop-blocks.ll
tests for it, and so this will be covered when I flip the default.
llvm-svn: 154812
laid out in a form with a fallthrough into the header and a fallthrough
out of the bottom. In that case, leave the loop alone because any
rotation will introduce unnecessary branches. If either side looks like
it will require an explicit branch, then the rotation won't add any, do
it to ensure the branch occurs outside of the loop (if possible) and
maximize the benefit of the fallthrough in the bottom.
llvm-svn: 154806
This is a complex change that resulted from a great deal of
experimentation with several different benchmarks. The one which proved
the most useful is included as a test case, but I don't know that it
captures all of the relevant changes, as I didn't have specific
regression tests for each, they were more the result of reasoning about
what the old algorithm would possibly do wrong. I'm also failing at the
moment to craft more targeted regression tests for these changes, if
anyone has ideas, it would be welcome.
The first big thing broken with the old algorithm is the idea that we
can take a basic block which has a loop-exiting successor and a looping
successor and use the looping successor as the layout top in order to
get that particular block to be the bottom of the loop after layout.
This happens to work in many cases, but not in all.
The second big thing broken was that we didn't try to select the exit
which fell into the nearest enclosing loop (to which we exit at all). As
a consequence, even if the rotation worked perfectly, it would result in
one of two bad layouts. Either the bottom of the loop would get
fallthrough, skipping across a nearer enclosing loop and thereby making
it discontiguous, or it would be forced to take an explicit jump over
the nearest enclosing loop to earch its successor. The point of the
rotation is to get fallthrough, so we need it to fallthrough to the
nearest loop it can.
The fix to the first issue is to actually layout the loop from the loop
header, and then rotate the loop such that the correct exiting edge can
be a fallthrough edge. This is actually much easier than I anticipated
because we can handle all the hard parts of finding a viable rotation
before we do the layout. We just store that, and then rotate after
layout is finished. No inner loops get split across the post-rotation
backedge because we check for them when selecting the rotation.
That fix exposed a latent problem with our exitting block selection --
we should allow the backedge to point into the middle of some inner-loop
chain as there is no real penalty to it, the whole point is that it
*won't* be a fallthrough edge. This may have blocked the rotation at all
in some cases, I have no idea and no test case as I've never seen it in
practice, it was just noticed by inspection.
Finally, all of these fixes, and studying the loops they produce,
highlighted another problem: in rotating loops like this, we sometimes
fail to align the destination of these backwards jumping edges. Fix this
by actually walking the backwards edges rather than relying on loopinfo.
This fixes regressions on heapsort if block placement is enabled as well
as lots of other cases where the previous logic would introduce an
abundance of unnecessary branches into the execution.
llvm-svn: 154783
Original message:
Modify the code that lowers shuffles to blends from using blendvXX to vblendXX.
blendV uses a register for the selection while Vblend uses an immediate.
On sandybridge they still have the same latency and execute on the same execution ports.
llvm-svn: 154483
blendv uses a register for the selection while vblend uses an immediate.
On sandybridge they still have the same latency and execute on the same execution ports.
llvm-svn: 154396
x86 addressing modes. This allows PIE-based TLS offsets to fit directly
into an addressing mode immediate offset, which is the last remaining
code quality issue from PR12380. With this patch, that PR is completely
fixed.
To understand why this patch is correct to match these offsets into
addressing mode immediates, break it down by cases:
1) 32-bit is trivially correct, and unmodified here.
2) 64-bit non-small mode is unchanged and never matches.
3) 64-bit small PIC code which is RIP-relative is handled specially in
the match to try to fit RIP into the base register. If it fails, it
now early exits. This behavior is unchanged by the patch.
4) 64-bit small non-PIC code which is not RIP-relative continues to work
as it did before. The reason these immediates are safe is because the
ABI ensures they fit in small mode. This behavior is unchanged.
5) 64-bit small PIC code which is *not* using RIP-relative addressing.
This is the only case changed by the patch, and the primary place you
see it is in TLS, either the win64 section offset TLS or Linux
local-exec TLS model in a PIC compilation. Here the ABI again ensures
that the immediates fit because we are in small mode, and any other
operations required due to the PIC relocation model have been handled
externally to the Wrapper node (extra loads etc are made around the
wrapper node in ISelLowering).
I've tested this as much as I can comparing it with GCC's output, and
everything appears safe. I discussed this with Anton and it made sense
to him at least at face value. That said, if there are issues with PIC
code after this patch, yell and we can revert it.
llvm-svn: 154304
comprehensive testing of TLS codegen for x86. Convert all of the ones
that were still using grep to use FileCheck. Remove some redundancies
between them.
Perhaps most interestingly expand the test cases so that they actually
fully list the instruction snippet being tested. TLS operations are
*very* narrowly defined, and so these seem reasonably stable. More
importantly, the existing test cases already were crazy fine grained,
expecting specific registers to be allocated. This just clarifies that
no *other* instructions are expected, and fills in some crucial gaps
that weren't being tested at all.
This will make any subsequent changes to TLS much more clear during
review.
llvm-svn: 154303
when -ffast-math, i.e. don't just always do it if the reciprocal can
be formed exactly. There is already an IR level transform that does
that, and it does it more carefully.
llvm-svn: 154296
optimizations which are valid for position independent code being linked
into a single executable, but not for such code being linked into
a shared library.
I discussed the design of this with Eric Christopher, and the decision
was to support an optional bit rather than a completely separate
relocation model. Fundamentally, this is still PIC relocation, its just
that certain optimizations are only valid under a PIC relocation model
when the resulting code won't be in a shared library. The simplest path
to here is to expose a single bit option in the TargetOptions. If folks
have different/better designs, I'm all ears. =]
I've included the first optimization based upon this: changing TLS
models to the *Exec models when PIE is enabled. This is the LLVM
component of PR12380 and is all of the hard work.
llvm-svn: 154294
Previously we used three instructions to broadcast an immediate value into a
vector register.
On Sandybridge we continue to load the broadcasted value from the constant pool.
llvm-svn: 154284
shuffle node because it could introduce new shuffle nodes that were not
supported efficiently by the target.
2. Add a more restrictive shuffle-of-shuffle optimization for cases where the
second shuffle reverses the transformation of the first shuffle.
llvm-svn: 154266
reciprocal if converting to the reciprocal is exact. Do it even if inexact
if -ffast-math. This substantially speeds up ac.f90 from the polyhedron
benchmarks.
llvm-svn: 154265
by default.
This is a behaviour configurable in the MCAsmInfo. I've decided to turn
it on by default in (possibly optimistic) hopes that most assemblers are
reasonably sane. If this proves a problem, switching to default seems
reasonable.
I'm not sure if this is the opportune place to test, but it seemed good
to make sure it was tested somewhere.
llvm-svn: 154235
LSR always tries to make the ICmp in the loop latch use the incremented
induction variable. This allows the induction variable to be kept in a
single register.
When the induction variable limit is equal to the stride,
SimplifySetCC() would break LSR's hard work by transforming:
(icmp (add iv, stride), stride) --> (cmp iv, 0)
This forced us to use lea for the IC update, preventing the simpler
incl+cmp.
<rdar://problem/7643606>
<rdar://problem/11184260>
llvm-svn: 154119
Do not try to optimize swizzles of shuffles if the source shuffle has more than
a single user, except when the source shuffle is also a swizzle.
llvm-svn: 153864
This is the CodeGen equivalent of r153747. I tested that there is not noticeable
performance difference with any combination of -O0/-O2 /-g when compiling
gcc as a single compilation unit.
llvm-svn: 153817
This is a code change to add support for changing instruction sequences of the form:
load
inc/dec of 8/16/32/64 bits
store
into the appropriate X86 inc/dec through memory instruction:
inc[qlwb] / dec[qlwb]
The checks that were in X86DAGToDAGISel::Select(SDNode *Node)>>ISD::STORE have been extracted to isLoadIncOrDecStore and reworked to use the better
named wrappers for getOperand(unsigned) (e.g. getOffset()) and replaced Chain.getNode() with LoadNode. The comments have also been expanded.
llvm-svn: 153635
This is a code change to add support for changing instruction sequences of the form:
load
inc/dec of 8/16/32/64 bits
store
into the appropriate X86 inc/dec through memory instruction:
inc[qlwb] / dec[qlwb]
The checks that were in X86DAGToDAGISel::Select(SDNode *Node)>>ISD::STORE have been extracted to isLoadIncOrDecStore and reworked to use the better
named wrappers for getOperand(unsigned) (e.g. getOffset()) and replaced Chain.getNode() with LoadNode. The comments have also been expanded.
llvm-svn: 153617
* Removed test/lib/llvm.exp - it is no longer needed
* Deleted the dg.exp reading code from test/lit.cfg. There are no dg.exp files
left in the test suite so this code is no longer required. test/lit.cfg is
now much shorter and clearer
* Removed a lot of duplicate code in lit.local.cfg files that need access to
the root configuration, by adding a "root" attribute to the TestingConfig
object. This attribute is dynamically computed to provide the same
information as was previously provided by the custom getRoot functions.
* Documented the config.root attribute in docs/CommandGuide/lit.pod
llvm-svn: 153408
This results in things such as
vmovups 16(%rdi), %xmm0
vinsertf128 $1, %xmm0, %ymm0, %ymm0
to be combined to
vinsertf128 $1, 16(%rdi), %ymm0, %ymm0
rdar://11076953
llvm-svn: 153092
i128). In that case, we may not be able to print out the MCExpr as an
expression. For instance, we could have an MCExpr like this:
0xBEEF0000BEEF0000 | (0xBEEF0000BEEF0000 << 64)
The MCExpr printer handles sizes up to 64-bits, but this expression would
require 128-bits. In this situation, try to evaluate the constant expression and
emit that as the value into 64-bit chunks.
<rdar://problem/11070338>
llvm-svn: 153081
X86InstrCompiler.td.
It also adds –mcpu-generic to the legalize-shift-64.ll test so the test
will pass if run on an Intel Atom CPU, which would otherwise
produce an instruction schedule which differs from that which the test expects.
llvm-svn: 153033
This results in things such as
vmovaps -96(%rbx), %xmm1
vinsertf128 $1, %xmm1, %ymm0, %ymm0
to be combined to
vinsertf128 $1, -96(%rbx), %ymm0, %ymm0
rdar://10643481
llvm-svn: 152762
Original commit message from r147481:
DAGCombine for transforming 128->256 casts into a vmovaps, rather
then a vxorps + vinsertf128 pair if the original vector came from a load.
Fix:
Unaligned loads need to generate a vmovups.
rdar://10974078
llvm-svn: 152366
This was testing the handling of sub-register coalescing followed by
remat. The original problem was caused by the extra <imp-def> operands
added by sub-register coalescing. Those <imp-def> operands are not
added any longer, and the test case passes even when the original patch
is reverted.
llvm-svn: 152040
Some BBs can become dead after codegen preparation. If we delete them here, it
could help enable tail-call optimizations later on.
<rdar://problem/10256573>
llvm-svn: 152002
In this instance we are generating the tail-call during legalizeDAG. The 2nd
floor call can't be a tail call because it clobbers %xmm1, which is defined by
the first floor call. The first floor call can't be a tail-call because it's
not in the tail position. The only reasonable way I could think to fix this
in a target-independent manner was to check for glue logic on the copy reg.
rdar://10930395
llvm-svn: 151877
so that the test will not fail when run on an Intel Atom
processor, due to the Atom scheduler producing an instruction sequence that is
different from that which is normally expected.
llvm-svn: 151832
To avoid problems with zero shifts when getting the bits that move between words
we use a trick: first shift the by amount-1, then do another shift by one. When
amount is 0 (and size 32) we first shift by 31, then by one, instead of by 32.
Also fix a latent bug that emitted the low and high words in the wrong order
when shifting right.
Fixes PR12113.
llvm-svn: 151637
When the GEP index is a vector of pointers, the code that calculated the size
of the element started from the vector type, and not the contained pointer type.
As a result, instead of looking at the data element pointed by the vector, this
code used the size of the vector. This works for 32bit members (on 32bit
systems), but not for other types. Added code to peel the vector type and
added a test.
llvm-svn: 151626
[Joe Groff] Hi everyone. My previous patch applied as r151382 had a few problems:
Clang raised a warning, and X86 LowerOperation would assert out for
fptoui f64 to i32 because it improperly lowered to an illegal
BUILD_PAIR. Here's a patch that addresses these issues. Let me know if
any other changes are necessary. Thanks.
llvm-svn: 151432
ecx = mov eax
al = mov ch
The second copy is not a nop because the sub-indices of ecx,ch is not the
same of that of eax/al.
Re-enabled machine-cp.
PR11940
llvm-svn: 151002
processor, due to the Atom scheduler producing an instruction sequence that is
different from that which is expected.
Patch by Michael Spencer!
llvm-svn: 150736
that are greater than the vector element type. For example BUILD_VECTOR
of type <1 x i1> with a constant i8 operand.
This patch fixes the assertion.
llvm-svn: 150477
If the DEC node had more than one user, it was doing this lowering but
leaving the original DEC node around and so decrementing twice.
Fixes PR11964.
llvm-svn: 150356
v8i8 -> v8i32 on AVX machines. The codegen often scalarizes ANY_EXTEND nodes.
The DAGCombiner has two optimizations that can mitigate the problem. First,
if all of the operands of a BUILD_VECTOR node are extracted from an ZEXT/ANYEXT
nodes, then it is possible to create a new simplified BUILD_VECTOR which uses
UNDEFS/ZERO values to eliminate the scalar ZEXT/ANYEXT nodes.
Second, another dag combine optimization lowers BUILD_VECTOR into a shuffle
vector instruction.
In the case of zext v8i8->v8i32 on AVX, a value in an XMM register is to be
shuffled into a wide YMM register.
This patch modifes the second optimization and allows the creation of
shuffle vectors even when the newly generated vector and the original vector
from which we extract the values are of different types.
llvm-svn: 150340
Creates a configurable regalloc pipeline.
Ensure specific llc options do what they say and nothing more: -reglloc=... has no effect other than selecting the allocator pass itself. This patch introduces a new umbrella flag, "-optimize-regalloc", to enable/disable the optimizing regalloc "superpass". This allows for example testing coalscing and scheduling under -O0 or vice-versa.
When a CodeGen pass requires the MachineFunction to have a particular property, we need to explicitly define that property so it can be directly queried rather than naming a specific Pass. For example, to check for SSA, use MRI->isSSA, not addRequired<PHIElimination>.
CodeGen transformation passes are never "required" as an analysis
ProcessImplicitDefs does not require LiveVariables.
We have a plan to massively simplify some of the early passes within the regalloc superpass.
llvm-svn: 150226
In this patch we optimize this pattern and convert the sequence into extract op of a narrow type.
This allows the BUILD_VECTOR dag optimizations to construct efficient shuffle operations in many cases.
llvm-svn: 149692
Adds an instruction itinerary to all x86 instructions, giving each a default latency of 1, using the InstrItinClass IIC_DEFAULT.
Sets specific latencies for Atom for the instructions in files X86InstrCMovSetCC.td, X86InstrArithmetic.td, X86InstrControl.td, and X86InstrShiftRotate.td. The Atom latencies for the remainder of the x86 instructions will be set in subsequent patches.
Adds a test to verify that the scheduler is working.
Also changes the scheduling preference to "Hybrid" for i386 Atom, while leaving x86_64 as ILP.
Patch by Preston Gurd!
llvm-svn: 149558
vectors of all one bits to be printed more cleverly in the AsmPrinter.
Unfortunately, the byte value for all one bits is the same with
-fsigned-char as the error return of '-1'. Force this to be the unsigned
byte value when returning it to avoid this problem, and update the test
case for the shiny new behavior.
Yay for building LLVM and Clang with -funsigned-char.
Chris, please review, and let me know if there is any reason to not
desire this change. It seems good on the surface, and certainly intended
based on the code written.
llvm-svn: 149299
The Win64 calling convention has xmm6-15 as callee-saved while still
clobbering all ymm registers.
Add a YMM_HI_6_15 pseudo-register that aliases the clobbered part of the
ymm registers, and mark that as call-clobbered. This allows live xmm
registers across calls.
This hack wouldn't be necessary with RegisterMask operands representing
the call clobbers, but they are not quite operational yet.
llvm-svn: 149088
. "fptosi" and "fptoui" IR instructions are defined with round-to-zero rounding mode.
. Currently for AVX mode for <4xdouble> and <8xdouble> the "VCVTPD2DQ.128" and "VCVTPD2DQ.256" instructions are selected (for .fp_to_sint. DAG node operation ) by AVX codegen. However they use round-to-nearest-even rounding mode.
. Consequently, the conversion produces incorrect numbers.
The fix is to replace selection of VCVTPD2DQ instructions with VCVTTPD2DQ instructions. The latter use truncate (i.e. round-to-zero) rounding mode.
As .fp_to_sint. DAG node operation is used only for lowering of "fptosi" and "fptoui" IR instructions, the fix in X86InstrSSE.td definition file doesn.t have an impact on other LLVM flows.
The patch includes changes in the .td file, LIT test for the changes and a fix in a legacy LIT test (which produced asm code conflicting with LLVN IR spec).
llvm-svn: 149056
In CanXFormVExtractWithShuffleIntoLoad we assumed that EXTRACT_VECTOR_ELT can be later handled by the DAGCombiner.
However, in some cases on AVX, the EXTRACT_VECTOR_ELT is legalized to EXTRACT_SUBVECTOR + EXTRACT_VECTOR_ELT, which
currently is not handled by the DAGCombiner. In this patch I added a check that we only extract from the XMM part.
llvm-svn: 148298
We know that the blend instructions only use the MSB, so if the mask is
sign-extended then we can convert it into a SHL instruction. This is a
common pattern because the type-legalizer sign-extends the i1 type which
is used by the LLVM-IR for the condition.
Added a new optimization in SimplifyDemandedBits for SIGN_EXTEND_INREG -> SHL.
llvm-svn: 148225
lc: X86ISelLowering.cpp:6480: llvm::SDValue llvm::X86TargetLowering::LowerVECTOR_SHUFFLE(llvm::SDValue, llvm::SelectionDAG&) const: Assertion `V1.getOpcode() != ISD::UNDEF&& "Op 1 of shuffle should not be undef"' failed.
Added a test.
llvm-svn: 148044
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
Craig Topper