Instead, just create one, and make sure everything that needs
it can access it. Previously most of the SelectionDAGISel
subclasses all had their own TargetLowering object, which was
redundant with the TargetLowering object in the TargetMachine
subclasses, except on Sparc, where SparcTargetMachine
didn't have a TargetLowering object. Change Sparc to work
more like the other targets here.
llvm-svn: 57016
process up to a higher level. This allows FastISel to leverage
more of SelectionDAGISel's infastructure, such as updating Machine
PHI nodes.
Also, implement transitioning from SDISel back to FastISel in
the middle of a block, so it's now possible to go back and
forth. This allows FastISel to hand individual CallInsts and other
complicated things off to SDISel to handle, while handling the rest
of the block itself.
To help support this, reorganize the SelectionDAG class so that it
is allocated once and reused throughout a function, instead of
being completely reallocated for each block.
llvm-svn: 55219
replacement of multiple values. This is slightly more efficient
than doing multiple ReplaceAllUsesOfValueWith calls, and theoretically
could be optimized even further. However, an important property of this
new function is that it handles the case where the source value set and
destination value set overlap. This makes it feasible for isel to use
SelectNodeTo in many very common cases, which is advantageous because
SelectNodeTo avoids a temporary node and it doesn't require CSEMap
updates for users of values that don't change position.
Revamp MorphNodeTo, which is what does all the work of SelectNodeTo, to
handle operand lists more efficiently, and to correctly handle a number
of corner cases to which its new wider use exposes it.
This commit also includes a change to the encoding of post-isel opcodes
in SDNodes; now instead of being sandwiched between the target-independent
pre-isel opcodes and the target-dependent pre-isel opcodes, post-isel
opcodes are now represented as negative values. This makes it possible
to test if an opcode is pre-isel or post-isel without having to know
the size of the current target's post-isel instruction set.
These changes speed up llc overall by 3% and reduce memory usage by 10%
on the InstructionCombining.cpp testcase with -fast and -regalloc=local.
llvm-svn: 53728
and better control the abstraction. Rename the type
to MVT. To update out-of-tree patches, the main
thing to do is to rename MVT::ValueType to MVT, and
rewrite expressions like MVT::getSizeInBits(VT) in
the form VT.getSizeInBits(). Use VT.getSimpleVT()
to extract a MVT::SimpleValueType for use in switch
statements (you will get an assert failure if VT is
an extended value type - these shouldn't exist after
type legalization).
This results in a small speedup of codegen and no
new testsuite failures (x86-64 linux).
llvm-svn: 52044
Added ISD::DECLARE node type to represent llvm.dbg.declare intrinsic. Now the intrinsic calls are lowered into a SDNode and lives on through out the codegen passes.
For now, since all the debugging information recording is done at isel time, when a ISD::DECLARE node is selected, it has the side effect of also recording the variable. This is a short term solution that should be fixed in time.
llvm-svn: 46659
and StoreSDNode into their common base class LSBaseSDNode. Member
functions getLoadedVT and getStoredVT are replaced with the common
getMemoryVT to simplify code that will handle both loads and stores.
llvm-svn: 46538
the code generated is not wonderful. This turns a miscompilation into
a code quality bug (noted in the ppc readme). This fixes PR642, which
is over 2 years old (!). Nate, please review this.
llvm-svn: 45742
that it is cheap and efficient to get.
Move a variety of predicates from TargetInstrInfo into
TargetInstrDescriptor, which makes it much easier to query a predicate
when you don't have TII around. Now you can use MI->getDesc()->isBranch()
instead of going through TII, and this is much more efficient anyway. Not
all of the predicates have been moved over yet.
Update old code that used MI->getInstrDescriptor()->Flags to use the
new predicates in many places.
llvm-svn: 45674
that "machine" classes are used to represent the current state of
the code being compiled. Given this expanded name, we can start
moving other stuff into it. For now, move the UsedPhysRegs and
LiveIn/LoveOuts vectors from MachineFunction into it.
Update all the clients to match.
This also reduces some needless #includes, such as MachineModuleInfo
from MachineFunction.
llvm-svn: 45467
TargetLowering to SelectionDAG so that they have more convenient
access to the current DAG, in preparation for the ValueType routines
being changed from standalone functions to members of SelectionDAG for
the pre-legalize vector type changes.
llvm-svn: 37704
value and CR reg #. This requires swapping the order of these everywhere
that touches BCC and requires us to write custom matching logic for
PPCcondbranch :(
llvm-svn: 31835
(because the 64-bit reg target versions aren't implemented yet), doesn't
support r+r addr modes, and doesn't handle stores, but it works otherwise. :)
This is disabled unless -enable-ppc-preinc is passed to llc for now.
llvm-svn: 31621
Turn on -Wunused and -Wno-unused-parameter. Clean up most of the resulting
fall out by removing unused variables. Remaining warnings have to do with
unused functions (I didn't want to delete code without review) and unused
variables in generated code. Maintainers should clean up the remaining
issues when they see them. All changes pass DejaGnu tests and Olden.
llvm-svn: 31380
in the start of an array and a count of operands where applicable. In many
cases, the number of operands is known, so this static array can be allocated
on the stack, avoiding the heap. In many other cases, a SmallVector can be
used, which has the same benefit in the common cases.
I updated a lot of code calling getNode that takes a vector, but ran out of
time. The rest of the code should be updated, and these methods should be
removed.
We should also do the same thing to eliminate the methods that take a
vector of MVT::ValueTypes.
It would be extra nice to convert the dagiselemitter to avoid creating vectors
for operands when calling getTargetNode.
llvm-svn: 29566
the copyto/fromregs instead of making the PPCISD::CALL selection code create
them. This vastly simplifies the selection code, and moves the ABI handling
parts into one place.
llvm-svn: 28346
still a couple missed optimizations, but we now generate all the possible
rlwimis for multiple inserts into the same bitfield. More regression tests
to come.
llvm-svn: 28156
x86 and ppc for 100% dense switch statements when relocations are non-PIC.
This support will be extended and enhanced in the coming days to support
PIC, and less dense forms of jump tables.
llvm-svn: 27947
disp field of the load/store if possible. This compiles
CodeGen/PowerPC/load-constant-addr.ll to:
_test:
lis r2, 2838
lfs f1, 26848(r2)
blr
instead of:
_test:
lis r2, 2838
ori r2, r2, 26848
lfs f1, 0(r2)
blr
llvm-svn: 26908
constant pool load. This generates significantly nicer code for splats.
When tblgen gets bugfixed, we can remove the custom selection code.
llvm-svn: 26898
Make the PPC backend not dependent on BRTWOWAY_CC and make the branch
selector smarter about the code it generates, fixing a case in the
readme.
llvm-svn: 26814
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.
llvm-svn: 26719
We do not want to emit "Loop: ... brcond Out; br Loop", as it adds an extra
instruction in the loop. Instead, invert the condition and emit
"Loop: ... br!cond Loop; br Out.
Generalize the fix by moving it from PPCDAGToDAGISel to SelectionDAGLowering.
llvm-svn: 26231
us to load and store vectors directly at a pointer (offset of zero) by
using r0 as the base register. This also requires some asm printer work
to satisfy the darwin assembler.
For
void %foo(<4 x float> * %a) {
entry:
%tmp1 = load <4 x float> * %a;
%tmp2 = add <4 x float> %tmp1, %tmp1
store <4 x float> %tmp2, <4 x float> *%a
ret void
}
We now produce:
_foo:
lvx v0, 0, r3
vaddfp v0, v0, v0
stvx v0, 0, r3
blr
Instead of:
_foo:
li r2, 0
lvx v0, r2, r3
vaddfp v0, v0, v0
stvx v0, r2, r3
blr
llvm-svn: 24872
from the DAGToDAG cpp file. This adds pattern support for vector and
scalar fma, which passes test/Regression/CodeGen/PowerPC/fma.ll, and
does the right thing in the presence of -disable-excess-fp-precision.
Allows us to match:
void %foo(<4 x float> * %a) {
entry:
%tmp1 = load <4 x float> * %a;
%tmp2 = mul <4 x float> %tmp1, %tmp1
%tmp3 = add <4 x float> %tmp2, %tmp1
store <4 x float> %tmp3, <4 x float> *%a
ret void
}
As:
_foo:
li r2, 0
lvx v0, r2, r3
vmaddfp v0, v0, v0, v0
stvx v0, r2, r3
blr
Or, with llc -disable-excess-fp-precision,
_foo:
li r2, 0
lvx v0, r2, r3
vxor v1, v1, v1
vmaddfp v1, v0, v0, v1
vaddfp v0, v1, v0
stvx v0, r2, r3
blr
llvm-svn: 24719
changes allow us to generate the following code:
_foo:
li r2, 0
lvx v0, r2, r3
vaddfp v0, v0, v0
stvx v0, r2, r3
blr
for this llvm:
void %foo(<4 x float>* %a) {
entry:
%tmp1 = load <4 x float>* %a
%tmp2 = add <4 x float> %tmp1, %tmp1
store <4 x float> %tmp2, <4 x float>* %a
ret void
}
llvm-svn: 24534
of some code. This exposes the implicit load from the stubs to the DAG, allowing
them to be optimized by the dag combiner. It also moves darwin specific stuff
out of the isel into the legalizer, and allows more to be moved to the .td file.
llvm-svn: 24397
on Darwin to remove smarts from the isel. This is currently disabled by
default (uncomment setOperationAction(ISD::GlobalAddress to enable it).
tblgen needs to become smarter about tglobaladdr nodes and bigger patterns
needed to be added to the .td file. However, we can currently emit stuff like
this: :)
li r2, lo16(L_x$non_lazy_ptr)
lis r3, ha16(L_x$non_lazy_ptr)
lwzx r2, r3, r2
The obvious improvements will follow.
llvm-svn: 24390
Add a new flag to TargetLowering indicating if the target has really cheap
signed division by powers of two, make ppc use it. This will probably go
away in the future.
Implement some more ISD::SDIV folds in the dag combiner
Remove now dead code in the x86 backend.
llvm-svn: 23853
code for long long foo(long long a, long long b) { return a + b; }
_foo:
or r2, r3, r3
or r3, r4, r4
or r4, r5, r5
or r5, r6, r6
rldicr r2, r2, 32, 31
rldicl r3, r3, 0, 32
rldicr r4, r4, 32, 31
rldicl r5, r5, 0, 32
or r2, r3, r2
or r3, r5, r4
add r4, r3, r2
rldicl r2, r4, 32, 32
or r4, r4, r4
or r3, r2, r2
blr
llvm-svn: 23809
will have to tide us over until we get real subreg support, but it prevents
the PrologEpilogInserter from spilling 8 byte GPRs on a G4 processor.
Add some initial support for TRUNCATE and ANY_EXTEND, but they don't
currently work due to issues with ScheduleDAG. Something wll have to be
figured out.
llvm-svn: 23803
allows us to lower legal return types to something else, to meet ABI
requirements (such as that i64 be returned in two i32 regs on Darwin/ppc).
llvm-svn: 23802
the 177.mesa failure from last night, and fixes the
CodeGen/PowerPC/2005-10-08-ArithmeticRotate.ll regression test I added.
If this code cannot be fixed, it should be removed for good, but I'll leave
it to Nate to decide its fate.
llvm-svn: 23670
is faster and uses less stack space. This reduces our stack requirement
enough to compile sixtrack, and though it's a hack, should be enough until
we switch to iterative isel
llvm-svn: 23664
helps but not enough.
Start pulling cases out of PPC32DAGToDAGISel::Select. With GCC 4, this function
required 8512 bytes of stack space for each invocation (GCC 3 required less
than 700 bytes). Pulling this first function out gets us down to 8224. More
to come :(
llvm-svn: 23647
we're in a single-mbb loop, make sure to emit the backwards branch as the
conditional branch instead of the uncond branch. For example, emit this:
LBBl29_z__44:
stw r9, 0(r15)
stw r9, 4(r15)
stw r9, 8(r15)
stw r9, 12(r15)
addi r15, r15, 16
addi r8, r8, 1
cmpw cr0, r8, r28
ble cr0, LBBl29_z__44
b LBBl29_z__48 *** NOT PART OF LOOP
Instead of:
LBBl29_z__44:
stw r9, 0(r15)
stw r9, 4(r15)
stw r9, 8(r15)
stw r9, 12(r15)
addi r15, r15, 16
addi r8, r8, 1
cmpw cr0, r8, r28
bgt cr0, LBBl29_z__48 *** PART OF LOOP!
b LBBl29_z__44
The former sequence has one fewer dispatch group for the loop body.
llvm-svn: 23582
are allowed to generate 64-bit-only PowerPC instructions for 32 bit hosts,
such as the PowerPC 970.
This speeds up 189.lucas from 81.99 to 32.64 seconds.
llvm-svn: 23250
directly out of R1 (without using a CopyFromReg, which uses a chain), multiple
allocas were getting CSE'd together, producing bogus code. For this:
int %foo(bool %X, int %A, int %B) {
br bool %X, label %T, label %F
F:
%G = alloca int
%H = alloca int
store int %A, int* %G
store int %B, int* %H
%R = load int* %G
ret int %R
T:
ret int 0
}
We were generating:
_foo:
stwu r1, -16(r1)
stw r31, 4(r1)
or r31, r1, r1
stw r1, 12(r31)
cmpwi cr0, r3, 0
bne cr0, .LBB_foo_2 ; T
.LBB_foo_1: ; F
li r2, 16
subf r2, r2, r1 ;; One alloca
or r1, r2, r2
or r3, r1, r1
or r1, r2, r2
or r2, r1, r1
stw r4, 0(r3)
stw r5, 0(r2)
lwz r3, 0(r3)
lwz r1, 12(r31)
lwz r31, 4(r31)
lwz r1, 0(r1)
blr
.LBB_foo_2: ; T
li r3, 0
lwz r1, 12(r31)
lwz r31, 4(r31)
lwz r1, 0(r1)
blr
Now we generate:
_foo:
stwu r1, -16(r1)
stw r31, 4(r1)
or r31, r1, r1
stw r1, 12(r31)
cmpwi cr0, r3, 0
bne cr0, .LBB_foo_2 ; T
.LBB_foo_1: ; F
or r2, r1, r1
li r3, 16
subf r2, r3, r2 ;; Alloca 1
or r1, r2, r2
or r2, r1, r1
or r6, r1, r1
subf r3, r3, r6 ;; Alloca 2
or r1, r3, r3
or r3, r1, r1
stw r4, 0(r2)
stw r5, 0(r3)
lwz r3, 0(r2)
lwz r1, 12(r31)
lwz r31, 4(r31)
lwz r1, 0(r1)
blr
.LBB_foo_2: ; T
li r3, 0
lwz r1, 12(r31)
lwz r31, 4(r31)
lwz r1, 0(r1)
blr
This fixes Povray and SPASS with the dag isel, the last two failing cases.
Tommorow we will hopefully turn it on by default! :)
llvm-svn: 23190
could cause a miscompile. Fixing this didn't fix the two programs that fail
though. :(
This also changes the implementation to follow the pattern selector more
closely, causing us to select 0 to li instead of lis.
llvm-svn: 23189
Remove code (last hunk) that miscompiled immediate and's, such as
and uint %tmp.30, 4294958079
into
andi. r8, r8, 56319
andis. r8, r8, 65535
instead of:
li r9, -9217
and r8, r8, r9
The first always generates zero.
This fixes espresso.
llvm-svn: 23155
Dan Gohman