Added getTargetLowering() - returns DAG lowering info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26739 91177308-0d34-0410-b5e6-96231b3b80d8

Added target lowering hooks which LSR consults to make more intelligent
transformation decisions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26738 91177308-0d34-0410-b5e6-96231b3b80d8

Handle builtins that directly correspond to GCC builtins.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26737 91177308-0d34-0410-b5e6-96231b3b80d8

emit a mapping from LLVM intrinsic -> GCC builtins.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26736 91177308-0d34-0410-b5e6-96231b3b80d8

Verify that packed type operands have the right size and base type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26735 91177308-0d34-0410-b5e6-96231b3b80d8

Add a first ppc altivec intrinsic.  Add packed type support.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26734 91177308-0d34-0410-b5e6-96231b3b80d8

For functions that use vector registers, save VRSAVE, mark used
registers, and update it on entry to each function, then restore it on exit.

This compiles:

void func(vfloat *a, vfloat *b, vfloat *c) {
        *a = *b * *c + *c;
}

to this:

_func:
        mfspr r2, 256
        oris r6, r2, 49152
        mtspr 256, r6
        lvx v0, 0, r5
        lvx v1, 0, r4
        vmaddfp v0, v1, v0, v0
        stvx v0, 0, r3
        mtspr 256, r2
        blr

GCC produces this (which has additional stack accesses):

_func:
        mfspr r0,256
        stw r0,-4(r1)
        oris r0,r0,0xc000
        mtspr 256,r0
        lvx v0,0,r5
        lvx v1,0,r4
        lwz r12,-4(r1)
        vmaddfp v0,v0,v1,v0
        stvx v0,0,r3
        mtspr 256,r12
        blr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26733 91177308-0d34-0410-b5e6-96231b3b80d8

Optimizing to calloc() to malloc() + memset() can be bad on some platforms.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26732 91177308-0d34-0410-b5e6-96231b3b80d8

make sure dead token factor nodes are removed by the dag combiner.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26731 91177308-0d34-0410-b5e6-96231b3b80d8

Bring makefile back into compliance with standard by using
$(Echo) instead of @echo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26730 91177308-0d34-0410-b5e6-96231b3b80d8

Handle the removal of the debug chain.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26729 91177308-0d34-0410-b5e6-96231b3b80d8

remove two implemented items

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26728 91177308-0d34-0410-b5e6-96231b3b80d8

Fold X+Y -> X|Y when safe.  This implements:
Regression/CodeGen/PowerPC/and_add.ll

a case that occurs with dynamic allocas of constant size.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26727 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26726 91177308-0d34-0410-b5e6-96231b3b80d8

I can't convince myself that this is safe, remove the recursive call.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26725 91177308-0d34-0410-b5e6-96231b3b80d8

add a couple of missing folds

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26724 91177308-0d34-0410-b5e6-96231b3b80d8

For targets with FABS/FNEG support, lower copysign to an integer load,
a select and FABS/FNEG.

This speeds up a trivial (aka stupid) copysign benchmark I wrote from 6.73s
to 2.64s, woo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26723 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a couple of bugs that broke the alpha tester build

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26722 91177308-0d34-0410-b5e6-96231b3b80d8

Handle cracked instructions in dispatch group formation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26721 91177308-0d34-0410-b5e6-96231b3b80d8

Mark instructions that are cracked by the PPC970 decoder as such.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26720 91177308-0d34-0410-b5e6-96231b3b80d8

Several big changes:
1. Use flags on the instructions in the .td file to indicate the PPC970 unit
   type instead of a table in the .cpp file.  Much cleaner.
2. Change the hazard recognizer to build d-groups according to the actual
   algorithm used, not my flawed understanding of it.
3. Model "must be in the first slot" and "must be the only instr in a group"
   accurately.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26719 91177308-0d34-0410-b5e6-96231b3b80d8

Don't advance the hazard recognizer when there are no hazards and no instructions
to be emitted.

Don't add one to the latency of a completed instruction if the latency of the
op is 0.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26718 91177308-0d34-0410-b5e6-96231b3b80d8

Chain operands aren't real uses: they don't require the full latency of the
predecessor to finish before they can start.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26717 91177308-0d34-0410-b5e6-96231b3b80d8

As a pending queue data structure to keep track of instructions whose
operands have all issued, but whose results are not yet available.  This
allows us to compile:

int G;
int test(int A, int B, int* P) {
   return (G+A)*(B+1);
}

to:

_test:
        lis r2, ha16(L_G$non_lazy_ptr)
        addi r4, r4, 1
        lwz r2, lo16(L_G$non_lazy_ptr)(r2)
        lwz r2, 0(r2)
        add r2, r2, r3
        mullw r3, r2, r4
        blr

instead of this, which has a stall between the lis/lwz:

_test:
        lis r2, ha16(L_G$non_lazy_ptr)
        lwz r2, lo16(L_G$non_lazy_ptr)(r2)
        addi r4, r4, 1
        lwz r2, 0(r2)
        add r2, r2, r3
        mullw r3, r2, r4
        blr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26716 91177308-0d34-0410-b5e6-96231b3b80d8

Mark llvm.stacksave as only reading memory, this fixes
Regression/Transforms/InstCombine/stacksaverestore.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26715 91177308-0d34-0410-b5e6-96231b3b80d8

rename priorityqueue -> availablequeue.  When a node is scheduled, remember
which cycle it lands on.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26714 91177308-0d34-0410-b5e6-96231b3b80d8

Make CurrCycle a local var instead of an instance var

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26713 91177308-0d34-0410-b5e6-96231b3b80d8

Move some methods around so that BU specific code is together, TD specific code
is together, and direction independent code is together.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26712 91177308-0d34-0410-b5e6-96231b3b80d8

merge preds/chainpreds -> preds set
merge succs/chainsuccs -> succs set

This has no functionality change, simplifies the code, and reduces the size
of sunits.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26711 91177308-0d34-0410-b5e6-96231b3b80d8

blr is a branch too

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26710 91177308-0d34-0410-b5e6-96231b3b80d8

add an example

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26709 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26708 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR681 by using the standard Lengauer and Tarjan algorithm for dominator
set construction, rather than intersecting various std::sets.  This reduces
the memory usage for the testcase in PR681 from 496 to 26MB of ram on my
darwin system, and reduces the runtime from 32.8 to 0.8 seconds on a
2.5GHz G5.  This also enables future code sharing between Dom and PostDom
now that they share near-identical implementations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26707 91177308-0d34-0410-b5e6-96231b3b80d8

fix pasto in generate assertion msg

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26706 91177308-0d34-0410-b5e6-96231b3b80d8

Doh!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26705 91177308-0d34-0410-b5e6-96231b3b80d8

Added a parameter to control whether Constant::getStringValue() would chop
off the result string at the first null terminator.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26704 91177308-0d34-0410-b5e6-96231b3b80d8

remove some dead code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26703 91177308-0d34-0410-b5e6-96231b3b80d8

remove ShouldEmitDebugFunctions, a hack to support llvm-db via the jit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26702 91177308-0d34-0410-b5e6-96231b3b80d8

scrape out bits of llvm-db

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26701 91177308-0d34-0410-b5e6-96231b3b80d8

Remove UnixLocalInferiorProcess: debugging via the JIT isn't a good idea.

Perhaps llvm-db will turn into a read debugger someday.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26700 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify this testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26699 91177308-0d34-0410-b5e6-96231b3b80d8

reorder these to make it work with static libraries

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26698 91177308-0d34-0410-b5e6-96231b3b80d8

teach the JIT to encode vector registers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26697 91177308-0d34-0410-b5e6-96231b3b80d8

Fix another broken intrinsic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26696 91177308-0d34-0410-b5e6-96231b3b80d8

weak globals on darwin require an extra load, breaking this test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26695 91177308-0d34-0410-b5e6-96231b3b80d8

Fix incorrect definitions of these intrinsics, which broke a bunch of
stuff last night.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26694 91177308-0d34-0410-b5e6-96231b3b80d8

Move simple-selector-specific types to the simple selector.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26693 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify the interface to the schedulers, to not pass the selected heuristicin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26692 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify the interface to the schedulers, to not pass the selected heuristic
in.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26691 91177308-0d34-0410-b5e6-96231b3b80d8

Move some simple-sched-specific instance vars to the simple scheduler.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26690 91177308-0d34-0410-b5e6-96231b3b80d8

prune #includes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26689 91177308-0d34-0410-b5e6-96231b3b80d8

move some simple scheduler methods into the simple scheduler

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26688 91177308-0d34-0410-b5e6-96231b3b80d8

Make EmitNode take a SDNode instead of a NodeInfo*

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26687 91177308-0d34-0410-b5e6-96231b3b80d8

Move the VRBase field from NodeInfo to being a separate, explicit, map.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26686 91177308-0d34-0410-b5e6-96231b3b80d8

Store VRBase in a map, not in NodeInfo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26685 91177308-0d34-0410-b5e6-96231b3b80d8

no need to build groups anymore

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26684 91177308-0d34-0410-b5e6-96231b3b80d8

Create SUnits directly from the SelectionDAG.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26683 91177308-0d34-0410-b5e6-96231b3b80d8

Push PrepareNodeInfo/IdentifyGroups down the inheritance hierarchy

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26682 91177308-0d34-0410-b5e6-96231b3b80d8

make some methods protected instead of private

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26681 91177308-0d34-0410-b5e6-96231b3b80d8

Teach the latency scheduler some new tricks.  In particular, to break ties,
keep track of a sense of "mobility", i.e. how many other nodes scheduling one
node will free up.  For something like this:

float testadd(float *X, float *Y, float *Z, float *W, float *V) {
  return (*X+*Y)*(*Z+*W)+*V;
}

For example, this makes us schedule *X then *Y, not *X then *Z.  The former
allows us to issue the add, the later only lets us issue other loads.

This turns the above code from this:

_testadd:
        lfs f0, 0(r3)
        lfs f1, 0(r6)
        lfs f2, 0(r4)
        lfs f3, 0(r5)
        fadds f0, f0, f2
        fadds f1, f3, f1
        lfs f2, 0(r7)
        fmadds f1, f0, f1, f2
        blr

into this:

_testadd:
        lfs f0, 0(r6)
        lfs f1, 0(r5)
        fadds f0, f1, f0
        lfs f1, 0(r4)
        lfs f2, 0(r3)
        fadds f1, f2, f1
        lfs f2, 0(r7)
        fmadds f1, f1, f0, f2
        blr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26680 91177308-0d34-0410-b5e6-96231b3b80d8

Put intrinsics.gen in its proper place.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26679 91177308-0d34-0410-b5e6-96231b3b80d8

add an aggregate method for reinserting scheduled nodes, add a callback for
priority impls that want to be notified when a node is scheduled

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26678 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an incorrect intrinsic description

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26677 91177308-0d34-0410-b5e6-96231b3b80d8

Fix VC++ build breakage.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26676 91177308-0d34-0410-b5e6-96231b3b80d8

Adding an intrinsic is simpler still.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26674 91177308-0d34-0410-b5e6-96231b3b80d8

use autogenerated side-effect information

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26673 91177308-0d34-0410-b5e6-96231b3b80d8

generate side-effect info

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26672 91177308-0d34-0410-b5e6-96231b3b80d8

Simpler still

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26671 91177308-0d34-0410-b5e6-96231b3b80d8

Use autogenerated mod/ref info for intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26670 91177308-0d34-0410-b5e6-96231b3b80d8

Parse mod/ref properties, autogen mod/ref information

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26669 91177308-0d34-0410-b5e6-96231b3b80d8

Intrinsic adding is a little bit simpler now

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26668 91177308-0d34-0410-b5e6-96231b3b80d8

Use the autogenerated intrinsic verifier

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26667 91177308-0d34-0410-b5e6-96231b3b80d8

parse intrinsic types
autogenerate an intrinsic verifier

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26666 91177308-0d34-0410-b5e6-96231b3b80d8

Add option -enable-x86-lsr to enable x86 loop strength reduction pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26665 91177308-0d34-0410-b5e6-96231b3b80d8

Use the function name matcher autogenerated from the .td file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26664 91177308-0d34-0410-b5e6-96231b3b80d8

autogenerate the function name recognizer

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26663 91177308-0d34-0410-b5e6-96231b3b80d8

This rule also depends on tblgen

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26662 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26661 91177308-0d34-0410-b5e6-96231b3b80d8

use the enum list autogen'd from Intrinsics.td

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26660 91177308-0d34-0410-b5e6-96231b3b80d8

remove dbg_declare, it's not used yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26659 91177308-0d34-0410-b5e6-96231b3b80d8

silly case insensitive file systems...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26658 91177308-0d34-0410-b5e6-96231b3b80d8

Build intrinsics.gen from intrinsics.td

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26657 91177308-0d34-0410-b5e6-96231b3b80d8

Add a test case for (store (op (load ..) ..) ..) folding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26656 91177308-0d34-0410-b5e6-96231b3b80d8

Update these tests (which use autoupgrade) to run constprop and check
that the file parses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26655 91177308-0d34-0410-b5e6-96231b3b80d8

autoupgrade memcpy/memmove/memset with signed counts.
 CVS: ----------------------------------------------------------------------

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26654 91177308-0d34-0410-b5e6-96231b3b80d8

these are copies too

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26653 91177308-0d34-0410-b5e6-96231b3b80d8

remove some now-dead code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26652 91177308-0d34-0410-b5e6-96231b3b80d8

fcopysign for mixed mode

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26651 91177308-0d34-0410-b5e6-96231b3b80d8

back out my previous hack

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26650 91177308-0d34-0410-b5e6-96231b3b80d8

relax fcopysign

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26649 91177308-0d34-0410-b5e6-96231b3b80d8

alpha and llvm have different oppinions on which arg is the sign bit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26647 91177308-0d34-0410-b5e6-96231b3b80d8

remove temporary option

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26646 91177308-0d34-0410-b5e6-96231b3b80d8

Forgot this on last check in.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26645 91177308-0d34-0410-b5e6-96231b3b80d8

Alpha Scheduling classes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26643 91177308-0d34-0410-b5e6-96231b3b80d8

temporary hack to get the build working again, apparently a header
commit was forgotten

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26642 91177308-0d34-0410-b5e6-96231b3b80d8

fcopysign and get rid of dsnode cruft.  custom PA runtimes make this better in some senses

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26641 91177308-0d34-0410-b5e6-96231b3b80d8

fcopysign support

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26640 91177308-0d34-0410-b5e6-96231b3b80d8

Move bit field endianness to backend.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26639 91177308-0d34-0410-b5e6-96231b3b80d8

Temporary hack to enable more (store (op (load ...))) folding. This makes
it possible when a TokenFactor is between the load and store. But is still
missing some cases due to ordering issue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26638 91177308-0d34-0410-b5e6-96231b3b80d8

yes yes, enabled debug output is bad

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26637 91177308-0d34-0410-b5e6-96231b3b80d8