Fix schedule model for BFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85809 91177308-0d34-0410-b5e6-96231b3b80d8

Hyphenate some comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85808 91177308-0d34-0410-b5e6-96231b3b80d8

Chain dependencies used to enforce memory order should have latency of 0 (except for true dependency of Store followed by aliased Load... we estimate that case with a single cycle of latency assuming the hardware will bypass)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85807 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for BlockAddress values in ARM constant pools.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85806 91177308-0d34-0410-b5e6-96231b3b80d8

Prune unnecessary include.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85805 91177308-0d34-0410-b5e6-96231b3b80d8

Initilize the machine LICM CSE map upon the first time an instruction is hoisted to
the loop preheader. Add instructions which are already in the preheader block that
may be common expressions of those that are hoisted out. These does get a few more
instructions CSE'ed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85799 91177308-0d34-0410-b5e6-96231b3b80d8

These are done / no longer care.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85798 91177308-0d34-0410-b5e6-96231b3b80d8

Add an entry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85797 91177308-0d34-0410-b5e6-96231b3b80d8

now that ip sccp *really* subsumes ipcp, remove ipcp again.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85796 91177308-0d34-0410-b5e6-96231b3b80d8

improve IPSCCP to be able to propagate the result of "!mayBeOverridden"
function to calls of that function, regardless of whether it has local
linkage or has its address taken.  Not escaping should only affect
whether we make an aggressive assumption about the arguments to a
function, not whether we can track the result of it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85795 91177308-0d34-0410-b5e6-96231b3b80d8

Remove an irrelevant and poorly reduced test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85794 91177308-0d34-0410-b5e6-96231b3b80d8

don't mark the arguments of prototype overdefined, they will never be queried.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85793 91177308-0d34-0410-b5e6-96231b3b80d8

restore some code I removed in r85788, refactor it into
a shared place instead of duplicating it 4 times.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85792 91177308-0d34-0410-b5e6-96231b3b80d8

remove some confused code that dates from when we had
"multiple return values" but not "first class aggregates"

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85791 91177308-0d34-0410-b5e6-96231b3b80d8

avoid redundant lookups in BBExecutable, and make it a SmallPtrSet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85790 91177308-0d34-0410-b5e6-96231b3b80d8

Use the libanalysis 'ConstantFoldLoadFromConstPtr' function
instead of reinventing SCCP-specific logic.  This gives us
new powers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85789 91177308-0d34-0410-b5e6-96231b3b80d8

switch the main 'ValueState' map from being an std::map to being
a DenseMap.  Doing this required being aware of subtle iterator
invalidation issues, but it provides a big speedup.  In a
release-asserts build, this sped up optimizing 403.gcc from
1.34s -> 0.79s (IPSCCP) and 1.11s -> 0.44s (SCCP).

This commit also conflates in a bunch of general cleanups, sorry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85788 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak ARMBaseRegisterInfo::copyRegToReg.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85787 91177308-0d34-0410-b5e6-96231b3b80d8

fix a bug exposed by moving SRoA earlier which caused a crash building kc++

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85786 91177308-0d34-0410-b5e6-96231b3b80d8

Missing bit of universal built + hosted

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85785 91177308-0d34-0410-b5e6-96231b3b80d8

only IPSCCP incoming arguments if the function is executable, this fixes
an assertion on the buildbot.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85784 91177308-0d34-0410-b5e6-96231b3b80d8

add a new ValueState::getConstantInt() helper, use it to
simplify some code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85783 91177308-0d34-0410-b5e6-96231b3b80d8

Fix malloc.h is deprecated warning on DragonFly BSD.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85782 91177308-0d34-0410-b5e6-96231b3b80d8

Fix for warning seen on DF-BSD, Victor, please fix this to use a shift instead of pow()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85781 91177308-0d34-0410-b5e6-96231b3b80d8

tidy up some more: remove some extraneous inline specifiers, return harder.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85780 91177308-0d34-0410-b5e6-96231b3b80d8

Apply fix for PR5135, Credit to Andreas Neustifter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85779 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate the SCCPSolver::getValueMapping method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85778 91177308-0d34-0410-b5e6-96231b3b80d8

fix failures introduced in r85774

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85777 91177308-0d34-0410-b5e6-96231b3b80d8

factor duplicated code into a new DeleteInstructionInBlock
function, eliminate temporary (and pointless) smallvector.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85776 91177308-0d34-0410-b5e6-96231b3b80d8

Chris used to use '...' instead of proper grammar.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85775 91177308-0d34-0410-b5e6-96231b3b80d8

remove some extraneous llvmcontext stuff.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85774 91177308-0d34-0410-b5e6-96231b3b80d8

change LatticeVal to use PointerIntPair to save some space.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85773 91177308-0d34-0410-b5e6-96231b3b80d8

fix instcombine to only do store sinking when the alignments
of the two loads agree.  Propagate that onto the new store.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85772 91177308-0d34-0410-b5e6-96231b3b80d8

merge a test into store.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85771 91177308-0d34-0410-b5e6-96231b3b80d8

convert to filecheck

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85770 91177308-0d34-0410-b5e6-96231b3b80d8

Add missing end-tag.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85769 91177308-0d34-0410-b5e6-96231b3b80d8

Some formatting changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85768 91177308-0d34-0410-b5e6-96231b3b80d8

Handle splats of undefs properly. This includes the testcase for PR5364 as well.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85767 91177308-0d34-0410-b5e6-96231b3b80d8

Do not infer the target type for COPY_TO_REGCLASS from dest regclass, this won't work if it can contain several types. Require explicit result type for the node for now. This fixes PR5364.

PS: It seems that blackfin usage of copy_to_regclass is completely bogus!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85766 91177308-0d34-0410-b5e6-96231b3b80d8

64-bit FP loads & stores operate on both NEON and VFP pipelines.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85765 91177308-0d34-0410-b5e6-96231b3b80d8

Use NEON reg-reg moves, where profitable. This reduces "domain-cross" stalls, when we used to mix vfp and neon code (the former were used for reg-reg moves)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85764 91177308-0d34-0410-b5e6-96231b3b80d8

Add PseudoSourceValue::mayAlias. It returns true if the object can ever alias any LLVM IR value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85762 91177308-0d34-0410-b5e6-96231b3b80d8

Line this up as well.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85748 91177308-0d34-0410-b5e6-96231b3b80d8

Fix whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85747 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a couple more places where we are creating ld / st instructions without memoperands.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85746 91177308-0d34-0410-b5e6-96231b3b80d8

Make use of imm12 version of Thumb2 ldr / str instructions more aggressively.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85743 91177308-0d34-0410-b5e6-96231b3b80d8

fix two strange things in the default passmgr:

1. we'd run simplifycfg at the very start, even though
the per function passes have already cleaned this up.

2. In the main per-function pipeline that is interlaced with inlining
   etc, we would do instcombine, jump threading, simplifycfg *before*
   doing SROA.  SROA is much more likely to expose opportunities for
   these passes than they are for SROA, so move SRoA up earlier.

also add some comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85742 91177308-0d34-0410-b5e6-96231b3b80d8

merge phi-merge.ll into phi.ll

I don't know what Dan wants to do with phi-merge-gep.ll, I'll let
him deal with it because instcombine may end up sinking these.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85739 91177308-0d34-0410-b5e6-96231b3b80d8

when merging two loads, make sure to take the min of their alignment,
not the max.  This didn't matter until the previous patch because
instcombine would refuse to sink loads with differenting alignments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85738 91177308-0d34-0410-b5e6-96231b3b80d8

split load sinking out to its own function, like gep sinking.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85737 91177308-0d34-0410-b5e6-96231b3b80d8

fix a bug noticed by inspection: when instcombine sinks loads through
phis, it didn't preserve the alignment of the load.  This is a missed
optimization of the alignment is high and a miscompilation when the
alignment is low.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85736 91177308-0d34-0410-b5e6-96231b3b80d8

IPSCCP apparently is not a superset of IPCP, this is bad,
but I'll investigate it separately.  This unbreaks
test/FrontendC/weak_constant.c

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85735 91177308-0d34-0410-b5e6-96231b3b80d8

convert to filecheck.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85734 91177308-0d34-0410-b5e6-96231b3b80d8

Improve the other instance of the comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85733 91177308-0d34-0410-b5e6-96231b3b80d8

Add a missing closing parenthesis, and tweak to fit in 80
columns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85732 91177308-0d34-0410-b5e6-96231b3b80d8

only run GlobalDCE at -O3 and run it late instead of early.
GlobalOpt already deletes trivially dead functions/globals,
so GlobalDCE only adds values for cycles of dead things.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85731 91177308-0d34-0410-b5e6-96231b3b80d8

cleanups, switch GlobalDCE to SmallPtrSet instead of std::set

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85730 91177308-0d34-0410-b5e6-96231b3b80d8

We currently only run ipsccp at LTO time, which is silly.  It subsumes
ipconstprop and doesn't take much time.  Just run it in its place.

This adds a testcase for it, which I plan to expand to cover other
"integration" cases, where we expect the optimizer to be able to
eliminate various things.  Due to phase order issues we've regressed
in a number of areas and integration tests are the only way I see to
prevent this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85729 91177308-0d34-0410-b5e6-96231b3b80d8

remove a bunch of locking from LLVMContextImpl.  Since only one thread
can be banging on a context at a time, this isn't needed.  Owen, please
review.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85728 91177308-0d34-0410-b5e6-96231b3b80d8

improve comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85725 91177308-0d34-0410-b5e6-96231b3b80d8

add a comment about why we don't allow inlining indbr.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85724 91177308-0d34-0410-b5e6-96231b3b80d8

Fix tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85723 91177308-0d34-0410-b5e6-96231b3b80d8

the verifier shouldn't modify the IR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85722 91177308-0d34-0410-b5e6-96231b3b80d8

Reverting 85714, 85715, 85716, which are breaking the build

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85717 91177308-0d34-0410-b5e6-96231b3b80d8

Add a function to Passes.h to allow clients to create instances
of the ScalarEvolution pass without needing to #include ScalarEvolution.h.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85716 91177308-0d34-0410-b5e6-96231b3b80d8

Don't #include Pass.h from CallGraph.h.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85715 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the #include of Pass.h from PassManager.h. This breaks a significant
#include dependency, as frontends commonly pull in PassManager.h.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85714 91177308-0d34-0410-b5e6-96231b3b80d8

teach ipsccp and ipconstprop that a blockaddress doesn't 'take the address' of a function
in a way that should prevent ip constprop.  This allows clang/test/CodeGen/indirect-goto.c
to pass with the new indirect goto lowering.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85709 91177308-0d34-0410-b5e6-96231b3b80d8

change llvm::MergeBlockIntoPredecessor to not merge two blocks BB1->BB2
when BB2 has its address taken.  Since it ends up doing BB2->rauw(BB1),
this can cause the address of the entry block to be taken.  Since it is
generally undesirable to nuke blocks whose address is taken, even when
we can, just unconditionally stop this xform.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85708 91177308-0d34-0410-b5e6-96231b3b80d8

strengthen an assumption: RevectorBlockTo knows that PredBB
ended in an uncond branch because the pass requires BreakCriticalEdges.

However, BCE doesn't eliminate critical adges from indbrs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85707 91177308-0d34-0410-b5e6-96231b3b80d8

fix an issue where the verifier would reject a function whose entry
block had its address taken even if the blockaddress was dead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85706 91177308-0d34-0410-b5e6-96231b3b80d8

if CostMetrics says to never duplicate some code, don't unswitch a loop.
This prevents unswitching from duplicating indbr's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85705 91177308-0d34-0410-b5e6-96231b3b80d8

constant fold indirectbr(blockaddress(%bb)) -> br label %bb.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85704 91177308-0d34-0410-b5e6-96231b3b80d8

improve x86 codegen support for blockaddress.  We now compile
the testcase into:

_test1:                                                     ## @test1
## BB#0:                                                    ## %entry
leaq L_test1_bb6(%rip), %rax
jmpq *%rax
L_test1_bb:                                                 ## Address Taken
LBB1_1:                                                     ## %bb
movb $1, %al
ret
L_test1_bb6:                                                ## Address Taken
LBB1_2:                                                     ## %bb6
movb $2, %al
ret

Note, it is very very strange that BlockAddressSDNode doesn't carry
around TargetFlags.  Dan, please fix this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85703 91177308-0d34-0410-b5e6-96231b3b80d8

pull check for return inst out of loop, never inline a callee that contains
an indirectbr.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85702 91177308-0d34-0410-b5e6-96231b3b80d8

Fix BlockAddress::replaceUsesOfWithOnConstant to correctly
maintain the block use count in SubclassData.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85701 91177308-0d34-0410-b5e6-96231b3b80d8

implement linker support for BlockAddress.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85700 91177308-0d34-0410-b5e6-96231b3b80d8

Revert 85678/85680.  The decision is to stay with the current form of
indirectbr, thus we don't need "blockaddr(@func, null)".  Eliminate it
for simplicity.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85699 91177308-0d34-0410-b5e6-96231b3b80d8

Use cbz and cbnz instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85698 91177308-0d34-0410-b5e6-96231b3b80d8

vml[as].f32 cause stalls in following advanced SIMD instructions. Avoid using
them for scalar floating point operations for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85697 91177308-0d34-0410-b5e6-96231b3b80d8

Consolidate test files

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85696 91177308-0d34-0410-b5e6-96231b3b80d8

Change to use FileCheck

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85695 91177308-0d34-0410-b5e6-96231b3b80d8

Make tests more explicit about which instructions are expected.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85694 91177308-0d34-0410-b5e6-96231b3b80d8

Grammar tweak to comments

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85693 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure PRE doesn't split crit edges from indirectbr.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85692 91177308-0d34-0410-b5e6-96231b3b80d8

Update test to be more explicit about what instruction sequences are expected for each operation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85691 91177308-0d34-0410-b5e6-96231b3b80d8

llvm::SplitEdge should refuse to split an edge from an indirectbr.
Fix CodeGenPrepare to not try to split edges from indirectbr.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85690 91177308-0d34-0410-b5e6-96231b3b80d8

Update test to be more explicit about what instruction sequences are expected for each operation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85689 91177308-0d34-0410-b5e6-96231b3b80d8

update the comment above llvm::SplitCriticalEdge, and make
it abort on IndirectBrInst as describe in the comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85688 91177308-0d34-0410-b5e6-96231b3b80d8

Expand 64-bit logical shift right inline

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85687 91177308-0d34-0410-b5e6-96231b3b80d8

Expand 64-bit arithmetic shift right inline

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85685 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a missing newline in the dwarf output code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85684 91177308-0d34-0410-b5e6-96231b3b80d8

Make -print-machineinstrs more readable.
 - Be consistent when referring to MachineBasicBlocks: BB#0.
 - Be consistent when referring to virtual registers: %reg1024.
 - Be consistent when referring to unknown physical registers: %physreg10.
 - Be consistent when referring to known physical registers: %RAX
 - Be consistent when referring to register 0: %reg0
 - Be consistent when printing alignments: align=16
 - Print jump table contents.
 - Don't print host addresses, in general.
 - and various other cleanups.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85682 91177308-0d34-0410-b5e6-96231b3b80d8

Factor out more code into addCommonCodeGenPasses. The JIT wasn't
previously running CodePlacementOpt. Also print headers before
each dump in -print-machineinstrs mode, so that it's clear which
dump is which.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85681 91177308-0d34-0410-b5e6-96231b3b80d8

adjust a couple xforms to work with null bb's in BlockAddress.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85680 91177308-0d34-0410-b5e6-96231b3b80d8

Make blockaddress(@func, null) be valid, and make 'deleting a basic
block with a blockaddress still referring to it' replace the invalid
blockaddress with a new blockaddress(@func, null) instead of a
inttoptr(1).

This changes the bitcode encoding format, and still needs codegen
support (this should produce a non-zero value, referring to the entry
block of the function would also be quite reasonable).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85678 91177308-0d34-0410-b5e6-96231b3b80d8

Force triple; darwin's ASM syntax differs from linux's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85676 91177308-0d34-0410-b5e6-96231b3b80d8

Expand 64 bit left shift inline rather than using the libcall. For now, this
is unconditional. Making it still use the libcall when optimizing for size
would be a good adjustment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85675 91177308-0d34-0410-b5e6-96231b3b80d8

Add missing colons for FileCheck.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85674 91177308-0d34-0410-b5e6-96231b3b80d8

Convert to FileCheck

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85673 91177308-0d34-0410-b5e6-96231b3b80d8