2004-10-18 21:08:22 +00:00
|
|
|
//===- LoopStrengthReduce.cpp - Strength Reduce GEPs in Loops -------------===//
|
2005-04-21 23:48:37 +00:00
|
|
|
//
|
2004-10-18 21:08:22 +00:00
|
|
|
// The LLVM Compiler Infrastructure
|
|
|
|
//
|
|
|
|
// This file was developed by Nate Begeman and is distributed under the
|
|
|
|
// University of Illinois Open Source License. See LICENSE.TXT for details.
|
2005-04-21 23:48:37 +00:00
|
|
|
//
|
2004-10-18 21:08:22 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
//
|
|
|
|
// This pass performs a strength reduction on array references inside loops that
|
|
|
|
// have as one or more of their components the loop induction variable. This is
|
|
|
|
// accomplished by creating a new Value to hold the initial value of the array
|
|
|
|
// access for the first iteration, and then creating a new GEP instruction in
|
|
|
|
// the loop to increment the value by the appropriate amount.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2005-08-03 23:30:08 +00:00
|
|
|
#define DEBUG_TYPE "loop-reduce"
|
2004-10-18 21:08:22 +00:00
|
|
|
#include "llvm/Transforms/Scalar.h"
|
|
|
|
#include "llvm/Constants.h"
|
|
|
|
#include "llvm/Instructions.h"
|
|
|
|
#include "llvm/Type.h"
|
2005-03-04 04:04:26 +00:00
|
|
|
#include "llvm/DerivedTypes.h"
|
2004-10-18 21:08:22 +00:00
|
|
|
#include "llvm/Analysis/Dominators.h"
|
|
|
|
#include "llvm/Analysis/LoopInfo.h"
|
2005-07-30 00:15:07 +00:00
|
|
|
#include "llvm/Analysis/ScalarEvolutionExpander.h"
|
2004-10-18 21:08:22 +00:00
|
|
|
#include "llvm/Support/CFG.h"
|
2005-07-30 00:15:07 +00:00
|
|
|
#include "llvm/Support/GetElementPtrTypeIterator.h"
|
Fix a FIXME: if we are inserting code for a PHI argument, split the critical
edge so that the code is not always executed for both operands. This
prevents LSR from inserting code into loops whose exit blocks contain
PHI uses of IV expressions (which are outside of loops). On gzip, for
example, we turn this ugly code:
.LBB_test_1: ; loopentry
add r27, r3, r28
lhz r27, 3(r27)
add r26, r4, r28
lhz r26, 3(r26)
add r25, r30, r28 ;; Only live if exiting the loop
add r24, r29, r28 ;; Only live if exiting the loop
cmpw cr0, r27, r26
bne .LBB_test_5 ; loopexit
into this:
.LBB_test_1: ; loopentry
or r27, r28, r28
add r28, r3, r27
lhz r28, 3(r28)
add r26, r4, r27
lhz r26, 3(r26)
cmpw cr0, r28, r26
beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2: ; loopentry.loopexit_crit_edge
add r2, r30, r27
add r8, r29, r27
b .LBB_test_9 ; loopexit
.LBB_test_2: ; shortcirc_next.0
...
blt .LBB_test_1
into this:
.LBB_test_1: ; loopentry
or r27, r28, r28
add r28, r3, r27
lhz r28, 3(r28)
add r26, r4, r27
lhz r26, 3(r26)
cmpw cr0, r28, r26
beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2: ; loopentry.loopexit_crit_edge
add r2, r30, r27
add r8, r29, r27
b .LBB_t_3: ; shortcirc_next.0
.LBB_test_3: ; shortcirc_next.0
...
blt .LBB_test_1
Next step: get the block out of the loop so that the loop is all
fall-throughs again.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22766 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-12 22:06:11 +00:00
|
|
|
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
|
2004-10-18 21:08:22 +00:00
|
|
|
#include "llvm/Transforms/Utils/Local.h"
|
2005-03-04 04:04:26 +00:00
|
|
|
#include "llvm/Target/TargetData.h"
|
2004-10-18 21:08:22 +00:00
|
|
|
#include "llvm/ADT/Statistic.h"
|
2005-07-30 00:15:07 +00:00
|
|
|
#include "llvm/Support/Debug.h"
|
2005-07-30 18:22:27 +00:00
|
|
|
#include <algorithm>
|
2004-10-18 21:08:22 +00:00
|
|
|
#include <set>
|
|
|
|
using namespace llvm;
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
Statistic<> NumReduced ("loop-reduce", "Number of GEPs strength reduced");
|
2005-08-04 22:34:05 +00:00
|
|
|
Statistic<> NumInserted("loop-reduce", "Number of PHIs inserted");
|
Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
For code like this:
void foo(float *a, float *b, int n, int stride_a, int stride_b) {
int i;
for (i=0; i<n; i++)
a[i*stride_a] = b[i*stride_b];
}
we now emit:
.LBB_foo2_2: ; no_exit
lfs f0, 0(r4)
stfs f0, 0(r3)
addi r7, r7, 1
add r4, r2, r4
add r3, r6, r3
cmpw cr0, r7, r5
blt .LBB_foo2_2 ; no_exit
instead of:
.LBB_foo_2: ; no_exit
mullw r8, r2, r7 ;; multiply!
slwi r8, r8, 2
lfsx f0, r4, r8
mullw r8, r2, r6 ;; multiply!
slwi r8, r8, 2
stfsx f0, r3, r8
addi r2, r2, 1
cmpw cr0, r2, r5
blt .LBB_foo_2 ; no_exit
loops with variable strides occur pretty often. For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.
Now we can allow indvars to turn functions written like this:
void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
int i, ai = 0, bi = 0;
for (i=0; i<n; i++)
{
a[ai] = b[bi];
ai += stride_a;
bi += stride_b;
}
}
into code like the above for better analysis. With this patch, they generate
identical code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22740 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-10 00:45:21 +00:00
|
|
|
Statistic<> NumVariable("loop-reduce","Number of PHIs with variable strides");
|
2004-10-18 21:08:22 +00:00
|
|
|
|
2005-08-03 22:21:05 +00:00
|
|
|
/// IVStrideUse - Keep track of one use of a strided induction variable, where
|
|
|
|
/// the stride is stored externally. The Offset member keeps track of the
|
|
|
|
/// offset from the IV, User is the actual user of the operand, and 'Operand'
|
|
|
|
/// is the operand # of the User that is the use.
|
|
|
|
struct IVStrideUse {
|
|
|
|
SCEVHandle Offset;
|
|
|
|
Instruction *User;
|
|
|
|
Value *OperandValToReplace;
|
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-08 05:28:22 +00:00
|
|
|
|
|
|
|
// isUseOfPostIncrementedValue - True if this should use the
|
|
|
|
// post-incremented version of this IV, not the preincremented version.
|
|
|
|
// This can only be set in special cases, such as the terminating setcc
|
_test:
li r2, 0
LBB_test_1: ; no_exit.2
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
addi r3, r3, 4
cmpwi cr0, r2, 701
blt cr0, LBB_test_1 ; no_exit.2
LBB_test_2: ; loopexit.2.loopexit
addi r2, r2, 1
stw r2, 0(r4)
blr
[zion ~/llvm]$ cat > ~/xx
Uses of IV's outside of the loop should use hte post-incremented version
of the IV, not the preincremented version. This helps many loops (e.g. in sixtrack)
which used to generate code like this (this is the code from the
dont-hoist-simple-loop-constants.ll testcase):
_test:
li r2, 0 **** IV starts at 0
LBB_test_1: ; no_exit.2
or r5, r2, r2 **** Copy for loop exit
li r2, 0
stw r2, 0(r3)
addi r3, r3, 4
addi r2, r5, 1
addi r6, r5, 2 **** IV+2
cmpwi cr0, r6, 701
blt cr0, LBB_test_1 ; no_exit.2
LBB_test_2: ; loopexit.2.loopexit
addi r2, r5, 2 **** IV+2
stw r2, 0(r4)
blr
And now generated code like this:
_test:
li r2, 1 *** IV starts at 1
LBB_test_1: ; no_exit.2
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
addi r3, r3, 4
cmpwi cr0, r2, 701 *** IV.postinc + 0
blt cr0, LBB_test_1
LBB_test_2: ; loopexit.2.loopexit
stw r2, 0(r4) *** IV.postinc + 0
blr
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23313 91177308-0d34-0410-b5e6-96231b3b80d8
2005-09-12 06:04:47 +00:00
|
|
|
// instruction for a loop or uses dominated by the loop.
|
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-08 05:28:22 +00:00
|
|
|
bool isUseOfPostIncrementedValue;
|
2005-08-03 22:21:05 +00:00
|
|
|
|
|
|
|
IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
|
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-08 05:28:22 +00:00
|
|
|
: Offset(Offs), User(U), OperandValToReplace(O),
|
|
|
|
isUseOfPostIncrementedValue(false) {}
|
2005-08-03 22:21:05 +00:00
|
|
|
};
|
|
|
|
|
|
|
|
/// IVUsersOfOneStride - This structure keeps track of all instructions that
|
|
|
|
/// have an operand that is based on the trip count multiplied by some stride.
|
|
|
|
/// The stride for all of these users is common and kept external to this
|
|
|
|
/// structure.
|
|
|
|
struct IVUsersOfOneStride {
|
2005-07-30 00:15:07 +00:00
|
|
|
/// Users - Keep track of all of the users of this stride as well as the
|
2005-08-03 22:21:05 +00:00
|
|
|
/// initial value and the operand that uses the IV.
|
|
|
|
std::vector<IVStrideUse> Users;
|
|
|
|
|
|
|
|
void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
|
|
|
|
Users.push_back(IVStrideUse(Offset, User, Operand));
|
2005-07-30 00:15:07 +00:00
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
|
2004-10-18 21:08:22 +00:00
|
|
|
class LoopStrengthReduce : public FunctionPass {
|
|
|
|
LoopInfo *LI;
|
|
|
|
DominatorSet *DS;
|
2005-07-30 00:15:07 +00:00
|
|
|
ScalarEvolution *SE;
|
|
|
|
const TargetData *TD;
|
|
|
|
const Type *UIntPtrTy;
|
2004-10-18 21:08:22 +00:00
|
|
|
bool Changed;
|
2005-08-02 02:52:02 +00:00
|
|
|
|
|
|
|
/// MaxTargetAMSize - This is the maximum power-of-two scale value that the
|
|
|
|
/// target can handle for free with its addressing modes.
|
2005-03-04 04:04:26 +00:00
|
|
|
unsigned MaxTargetAMSize;
|
2005-07-30 00:15:07 +00:00
|
|
|
|
|
|
|
/// IVUsesByStride - Keep track of all uses of induction variables that we
|
|
|
|
/// are interested in. The key of the map is the stride of the access.
|
Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
For code like this:
void foo(float *a, float *b, int n, int stride_a, int stride_b) {
int i;
for (i=0; i<n; i++)
a[i*stride_a] = b[i*stride_b];
}
we now emit:
.LBB_foo2_2: ; no_exit
lfs f0, 0(r4)
stfs f0, 0(r3)
addi r7, r7, 1
add r4, r2, r4
add r3, r6, r3
cmpw cr0, r7, r5
blt .LBB_foo2_2 ; no_exit
instead of:
.LBB_foo_2: ; no_exit
mullw r8, r2, r7 ;; multiply!
slwi r8, r8, 2
lfsx f0, r4, r8
mullw r8, r2, r6 ;; multiply!
slwi r8, r8, 2
stfsx f0, r3, r8
addi r2, r2, 1
cmpw cr0, r2, r5
blt .LBB_foo_2 ; no_exit
loops with variable strides occur pretty often. For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.
Now we can allow indvars to turn functions written like this:
void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
int i, ai = 0, bi = 0;
for (i=0; i<n; i++)
{
a[ai] = b[bi];
ai += stride_a;
bi += stride_b;
}
}
into code like the above for better analysis. With this patch, they generate
identical code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22740 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-10 00:45:21 +00:00
|
|
|
std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
|
2005-07-30 00:15:07 +00:00
|
|
|
|
2005-10-09 06:20:55 +00:00
|
|
|
/// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
|
|
|
|
/// We use this to iterate over the IVUsesByStride collection without being
|
|
|
|
/// dependent on random ordering of pointers in the process.
|
|
|
|
std::vector<SCEVHandle> StrideOrder;
|
|
|
|
|
2005-08-04 01:19:13 +00:00
|
|
|
/// CastedValues - As we need to cast values to uintptr_t, this keeps track
|
|
|
|
/// of the casted version of each value. This is accessed by
|
|
|
|
/// getCastedVersionOf.
|
|
|
|
std::map<Value*, Value*> CastedPointers;
|
2005-07-30 00:15:07 +00:00
|
|
|
|
|
|
|
/// DeadInsts - Keep track of instructions we may have made dead, so that
|
|
|
|
/// we can remove them after we are done working.
|
|
|
|
std::set<Instruction*> DeadInsts;
|
2004-10-18 21:08:22 +00:00
|
|
|
public:
|
2005-03-04 04:04:26 +00:00
|
|
|
LoopStrengthReduce(unsigned MTAMS = 1)
|
|
|
|
: MaxTargetAMSize(MTAMS) {
|
|
|
|
}
|
|
|
|
|
2004-10-18 21:08:22 +00:00
|
|
|
virtual bool runOnFunction(Function &) {
|
|
|
|
LI = &getAnalysis<LoopInfo>();
|
|
|
|
DS = &getAnalysis<DominatorSet>();
|
2005-07-30 00:15:07 +00:00
|
|
|
SE = &getAnalysis<ScalarEvolution>();
|
|
|
|
TD = &getAnalysis<TargetData>();
|
|
|
|
UIntPtrTy = TD->getIntPtrType();
|
2004-10-18 21:08:22 +00:00
|
|
|
Changed = false;
|
|
|
|
|
|
|
|
for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
|
|
|
|
runOnLoop(*I);
|
2005-08-04 01:19:13 +00:00
|
|
|
|
2004-10-18 21:08:22 +00:00
|
|
|
return Changed;
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
|
2005-08-17 06:35:16 +00:00
|
|
|
// We split critical edges, so we change the CFG. However, we do update
|
|
|
|
// many analyses if they are around.
|
|
|
|
AU.addPreservedID(LoopSimplifyID);
|
|
|
|
AU.addPreserved<LoopInfo>();
|
|
|
|
AU.addPreserved<DominatorSet>();
|
|
|
|
AU.addPreserved<ImmediateDominators>();
|
|
|
|
AU.addPreserved<DominanceFrontier>();
|
|
|
|
AU.addPreserved<DominatorTree>();
|
|
|
|
|
2005-02-27 19:37:07 +00:00
|
|
|
AU.addRequiredID(LoopSimplifyID);
|
2004-10-18 21:08:22 +00:00
|
|
|
AU.addRequired<LoopInfo>();
|
|
|
|
AU.addRequired<DominatorSet>();
|
2005-03-04 04:04:26 +00:00
|
|
|
AU.addRequired<TargetData>();
|
2005-07-30 00:15:07 +00:00
|
|
|
AU.addRequired<ScalarEvolution>();
|
2004-10-18 21:08:22 +00:00
|
|
|
}
|
2005-08-04 01:19:13 +00:00
|
|
|
|
|
|
|
/// getCastedVersionOf - Return the specified value casted to uintptr_t.
|
|
|
|
///
|
|
|
|
Value *getCastedVersionOf(Value *V);
|
|
|
|
private:
|
2004-10-18 21:08:22 +00:00
|
|
|
void runOnLoop(Loop *L);
|
2005-08-04 17:40:30 +00:00
|
|
|
bool AddUsersIfInteresting(Instruction *I, Loop *L,
|
|
|
|
std::set<Instruction*> &Processed);
|
|
|
|
SCEVHandle GetExpressionSCEV(Instruction *E, Loop *L);
|
|
|
|
|
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-08 05:28:22 +00:00
|
|
|
void OptimizeIndvars(Loop *L);
|
2005-07-30 00:15:07 +00:00
|
|
|
|
Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
For code like this:
void foo(float *a, float *b, int n, int stride_a, int stride_b) {
int i;
for (i=0; i<n; i++)
a[i*stride_a] = b[i*stride_b];
}
we now emit:
.LBB_foo2_2: ; no_exit
lfs f0, 0(r4)
stfs f0, 0(r3)
addi r7, r7, 1
add r4, r2, r4
add r3, r6, r3
cmpw cr0, r7, r5
blt .LBB_foo2_2 ; no_exit
instead of:
.LBB_foo_2: ; no_exit
mullw r8, r2, r7 ;; multiply!
slwi r8, r8, 2
lfsx f0, r4, r8
mullw r8, r2, r6 ;; multiply!
slwi r8, r8, 2
stfsx f0, r3, r8
addi r2, r2, 1
cmpw cr0, r2, r5
blt .LBB_foo_2 ; no_exit
loops with variable strides occur pretty often. For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.
Now we can allow indvars to turn functions written like this:
void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
int i, ai = 0, bi = 0;
for (i=0; i<n; i++)
{
a[ai] = b[bi];
ai += stride_a;
bi += stride_b;
}
}
into code like the above for better analysis. With this patch, they generate
identical code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22740 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-10 00:45:21 +00:00
|
|
|
void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
|
|
|
|
IVUsersOfOneStride &Uses,
|
2005-08-03 22:21:05 +00:00
|
|
|
Loop *L, bool isOnlyStride);
|
2004-10-18 21:08:22 +00:00
|
|
|
void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
|
|
|
|
};
|
2005-04-21 23:48:37 +00:00
|
|
|
RegisterOpt<LoopStrengthReduce> X("loop-reduce",
|
2005-09-27 21:10:32 +00:00
|
|
|
"Loop Strength Reduction");
|
2004-10-18 21:08:22 +00:00
|
|
|
}
|
|
|
|
|
2005-03-04 04:04:26 +00:00
|
|
|
FunctionPass *llvm::createLoopStrengthReducePass(unsigned MaxTargetAMSize) {
|
|
|
|
return new LoopStrengthReduce(MaxTargetAMSize);
|
2004-10-18 21:08:22 +00:00
|
|
|
}
|
|
|
|
|
2005-08-04 01:19:13 +00:00
|
|
|
/// getCastedVersionOf - Return the specified value casted to uintptr_t.
|
|
|
|
///
|
|
|
|
Value *LoopStrengthReduce::getCastedVersionOf(Value *V) {
|
|
|
|
if (V->getType() == UIntPtrTy) return V;
|
|
|
|
if (Constant *CB = dyn_cast<Constant>(V))
|
|
|
|
return ConstantExpr::getCast(CB, UIntPtrTy);
|
|
|
|
|
|
|
|
Value *&New = CastedPointers[V];
|
|
|
|
if (New) return New;
|
|
|
|
|
|
|
|
BasicBlock::iterator InsertPt;
|
|
|
|
if (Argument *Arg = dyn_cast<Argument>(V)) {
|
|
|
|
// Insert into the entry of the function, after any allocas.
|
|
|
|
InsertPt = Arg->getParent()->begin()->begin();
|
|
|
|
while (isa<AllocaInst>(InsertPt)) ++InsertPt;
|
|
|
|
} else {
|
|
|
|
if (InvokeInst *II = dyn_cast<InvokeInst>(V)) {
|
|
|
|
InsertPt = II->getNormalDest()->begin();
|
|
|
|
} else {
|
|
|
|
InsertPt = cast<Instruction>(V);
|
|
|
|
++InsertPt;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Do not insert casts into the middle of PHI node blocks.
|
|
|
|
while (isa<PHINode>(InsertPt)) ++InsertPt;
|
|
|
|
}
|
2005-08-04 19:08:16 +00:00
|
|
|
|
|
|
|
New = new CastInst(V, UIntPtrTy, V->getName(), InsertPt);
|
|
|
|
DeadInsts.insert(cast<Instruction>(New));
|
|
|
|
return New;
|
2005-08-04 01:19:13 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2004-10-18 21:08:22 +00:00
|
|
|
/// DeleteTriviallyDeadInstructions - If any of the instructions is the
|
|
|
|
/// specified set are trivially dead, delete them and see if this makes any of
|
|
|
|
/// their operands subsequently dead.
|
|
|
|
void LoopStrengthReduce::
|
|
|
|
DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
|
|
|
|
while (!Insts.empty()) {
|
|
|
|
Instruction *I = *Insts.begin();
|
|
|
|
Insts.erase(Insts.begin());
|
|
|
|
if (isInstructionTriviallyDead(I)) {
|
2005-03-01 03:46:11 +00:00
|
|
|
for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
|
|
|
|
if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
|
|
|
|
Insts.insert(U);
|
2005-08-03 21:36:09 +00:00
|
|
|
SE->deleteInstructionFromRecords(I);
|
|
|
|
I->eraseFromParent();
|
2004-10-18 21:08:22 +00:00
|
|
|
Changed = true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2005-07-30 00:15:07 +00:00
|
|
|
|
2005-08-04 17:40:30 +00:00
|
|
|
/// GetExpressionSCEV - Compute and return the SCEV for the specified
|
|
|
|
/// instruction.
|
|
|
|
SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp, Loop *L) {
|
Fix some 80 column violations.
Once we compute the evolution for a GEP, tell SE about it. This allows users
of the GEP to know it, if the users are not direct. This allows us to compile
this testcase:
void fbSolidFillmmx(int w, unsigned char *d) {
while (w >= 64) {
*(unsigned long long *) (d + 0) = 0;
*(unsigned long long *) (d + 8) = 0;
*(unsigned long long *) (d + 16) = 0;
*(unsigned long long *) (d + 24) = 0;
*(unsigned long long *) (d + 32) = 0;
*(unsigned long long *) (d + 40) = 0;
*(unsigned long long *) (d + 48) = 0;
*(unsigned long long *) (d + 56) = 0;
w -= 64;
d += 64;
}
}
into:
.LBB_fbSolidFillmmx_2: ; no_exit
li r2, 0
stw r2, 0(r4)
stw r2, 4(r4)
stw r2, 8(r4)
stw r2, 12(r4)
stw r2, 16(r4)
stw r2, 20(r4)
stw r2, 24(r4)
stw r2, 28(r4)
stw r2, 32(r4)
stw r2, 36(r4)
stw r2, 40(r4)
stw r2, 44(r4)
stw r2, 48(r4)
stw r2, 52(r4)
stw r2, 56(r4)
stw r2, 60(r4)
addi r4, r4, 64
addi r3, r3, -64
cmpwi cr0, r3, 63
bgt .LBB_fbSolidFillmmx_2 ; no_exit
instead of:
.LBB_fbSolidFillmmx_2: ; no_exit
li r11, 0
stw r11, 0(r4)
stw r11, 4(r4)
stwx r11, r10, r4
add r12, r10, r4
stw r11, 4(r12)
stwx r11, r9, r4
add r12, r9, r4
stw r11, 4(r12)
stwx r11, r8, r4
add r12, r8, r4
stw r11, 4(r12)
stwx r11, r7, r4
add r12, r7, r4
stw r11, 4(r12)
stwx r11, r6, r4
add r12, r6, r4
stw r11, 4(r12)
stwx r11, r5, r4
add r12, r5, r4
stw r11, 4(r12)
stwx r11, r2, r4
add r12, r2, r4
stw r11, 4(r12)
addi r4, r4, 64
addi r3, r3, -64
cmpwi cr0, r3, 63
bgt .LBB_fbSolidFillmmx_2 ; no_exit
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22737 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 23:39:36 +00:00
|
|
|
// Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
|
|
|
|
// If this is a GEP that SE doesn't know about, compute it now and insert it.
|
|
|
|
// If this is not a GEP, or if we have already done this computation, just let
|
|
|
|
// SE figure it out.
|
2005-08-04 17:40:30 +00:00
|
|
|
GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
|
Fix some 80 column violations.
Once we compute the evolution for a GEP, tell SE about it. This allows users
of the GEP to know it, if the users are not direct. This allows us to compile
this testcase:
void fbSolidFillmmx(int w, unsigned char *d) {
while (w >= 64) {
*(unsigned long long *) (d + 0) = 0;
*(unsigned long long *) (d + 8) = 0;
*(unsigned long long *) (d + 16) = 0;
*(unsigned long long *) (d + 24) = 0;
*(unsigned long long *) (d + 32) = 0;
*(unsigned long long *) (d + 40) = 0;
*(unsigned long long *) (d + 48) = 0;
*(unsigned long long *) (d + 56) = 0;
w -= 64;
d += 64;
}
}
into:
.LBB_fbSolidFillmmx_2: ; no_exit
li r2, 0
stw r2, 0(r4)
stw r2, 4(r4)
stw r2, 8(r4)
stw r2, 12(r4)
stw r2, 16(r4)
stw r2, 20(r4)
stw r2, 24(r4)
stw r2, 28(r4)
stw r2, 32(r4)
stw r2, 36(r4)
stw r2, 40(r4)
stw r2, 44(r4)
stw r2, 48(r4)
stw r2, 52(r4)
stw r2, 56(r4)
stw r2, 60(r4)
addi r4, r4, 64
addi r3, r3, -64
cmpwi cr0, r3, 63
bgt .LBB_fbSolidFillmmx_2 ; no_exit
instead of:
.LBB_fbSolidFillmmx_2: ; no_exit
li r11, 0
stw r11, 0(r4)
stw r11, 4(r4)
stwx r11, r10, r4
add r12, r10, r4
stw r11, 4(r12)
stwx r11, r9, r4
add r12, r9, r4
stw r11, 4(r12)
stwx r11, r8, r4
add r12, r8, r4
stw r11, 4(r12)
stwx r11, r7, r4
add r12, r7, r4
stw r11, 4(r12)
stwx r11, r6, r4
add r12, r6, r4
stw r11, 4(r12)
stwx r11, r5, r4
add r12, r5, r4
stw r11, 4(r12)
stwx r11, r2, r4
add r12, r2, r4
stw r11, 4(r12)
addi r4, r4, 64
addi r3, r3, -64
cmpwi cr0, r3, 63
bgt .LBB_fbSolidFillmmx_2 ; no_exit
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22737 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 23:39:36 +00:00
|
|
|
if (!GEP || SE->hasSCEV(GEP))
|
2005-08-04 17:40:30 +00:00
|
|
|
return SE->getSCEV(Exp);
|
|
|
|
|
2005-07-30 00:15:07 +00:00
|
|
|
// Analyze all of the subscripts of this getelementptr instruction, looking
|
|
|
|
// for uses that are determined by the trip count of L. First, skip all
|
|
|
|
// operands the are not dependent on the IV.
|
|
|
|
|
|
|
|
// Build up the base expression. Insert an LLVM cast of the pointer to
|
|
|
|
// uintptr_t first.
|
2005-08-04 17:40:30 +00:00
|
|
|
SCEVHandle GEPVal = SCEVUnknown::get(getCastedVersionOf(GEP->getOperand(0)));
|
2005-07-30 00:15:07 +00:00
|
|
|
|
|
|
|
gep_type_iterator GTI = gep_type_begin(GEP);
|
2005-08-04 17:40:30 +00:00
|
|
|
|
|
|
|
for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
|
2005-07-30 00:15:07 +00:00
|
|
|
// If this is a use of a recurrence that we can analyze, and it comes before
|
|
|
|
// Op does in the GEP operand list, we will handle this when we process this
|
|
|
|
// operand.
|
|
|
|
if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
|
|
|
|
const StructLayout *SL = TD->getStructLayout(STy);
|
|
|
|
unsigned Idx = cast<ConstantUInt>(GEP->getOperand(i))->getValue();
|
|
|
|
uint64_t Offset = SL->MemberOffsets[Idx];
|
2005-08-04 17:40:30 +00:00
|
|
|
GEPVal = SCEVAddExpr::get(GEPVal,
|
|
|
|
SCEVUnknown::getIntegerSCEV(Offset, UIntPtrTy));
|
2005-03-06 22:52:29 +00:00
|
|
|
} else {
|
2005-08-04 19:08:16 +00:00
|
|
|
Value *OpVal = getCastedVersionOf(GEP->getOperand(i));
|
|
|
|
SCEVHandle Idx = SE->getSCEV(OpVal);
|
|
|
|
|
2005-08-04 17:40:30 +00:00
|
|
|
uint64_t TypeSize = TD->getTypeSize(GTI.getIndexedType());
|
|
|
|
if (TypeSize != 1)
|
|
|
|
Idx = SCEVMulExpr::get(Idx,
|
|
|
|
SCEVConstant::get(ConstantUInt::get(UIntPtrTy,
|
|
|
|
TypeSize)));
|
|
|
|
GEPVal = SCEVAddExpr::get(GEPVal, Idx);
|
2005-03-06 22:52:29 +00:00
|
|
|
}
|
2004-10-18 21:08:22 +00:00
|
|
|
}
|
2005-07-30 00:15:07 +00:00
|
|
|
|
Fix some 80 column violations.
Once we compute the evolution for a GEP, tell SE about it. This allows users
of the GEP to know it, if the users are not direct. This allows us to compile
this testcase:
void fbSolidFillmmx(int w, unsigned char *d) {
while (w >= 64) {
*(unsigned long long *) (d + 0) = 0;
*(unsigned long long *) (d + 8) = 0;
*(unsigned long long *) (d + 16) = 0;
*(unsigned long long *) (d + 24) = 0;
*(unsigned long long *) (d + 32) = 0;
*(unsigned long long *) (d + 40) = 0;
*(unsigned long long *) (d + 48) = 0;
*(unsigned long long *) (d + 56) = 0;
w -= 64;
d += 64;
}
}
into:
.LBB_fbSolidFillmmx_2: ; no_exit
li r2, 0
stw r2, 0(r4)
stw r2, 4(r4)
stw r2, 8(r4)
stw r2, 12(r4)
stw r2, 16(r4)
stw r2, 20(r4)
stw r2, 24(r4)
stw r2, 28(r4)
stw r2, 32(r4)
stw r2, 36(r4)
stw r2, 40(r4)
stw r2, 44(r4)
stw r2, 48(r4)
stw r2, 52(r4)
stw r2, 56(r4)
stw r2, 60(r4)
addi r4, r4, 64
addi r3, r3, -64
cmpwi cr0, r3, 63
bgt .LBB_fbSolidFillmmx_2 ; no_exit
instead of:
.LBB_fbSolidFillmmx_2: ; no_exit
li r11, 0
stw r11, 0(r4)
stw r11, 4(r4)
stwx r11, r10, r4
add r12, r10, r4
stw r11, 4(r12)
stwx r11, r9, r4
add r12, r9, r4
stw r11, 4(r12)
stwx r11, r8, r4
add r12, r8, r4
stw r11, 4(r12)
stwx r11, r7, r4
add r12, r7, r4
stw r11, 4(r12)
stwx r11, r6, r4
add r12, r6, r4
stw r11, 4(r12)
stwx r11, r5, r4
add r12, r5, r4
stw r11, 4(r12)
stwx r11, r2, r4
add r12, r2, r4
stw r11, 4(r12)
addi r4, r4, 64
addi r3, r3, -64
cmpwi cr0, r3, 63
bgt .LBB_fbSolidFillmmx_2 ; no_exit
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22737 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 23:39:36 +00:00
|
|
|
SE->setSCEV(GEP, GEPVal);
|
2005-08-04 17:40:30 +00:00
|
|
|
return GEPVal;
|
2005-07-30 00:15:07 +00:00
|
|
|
}
|
|
|
|
|
2005-08-04 19:08:16 +00:00
|
|
|
/// getSCEVStartAndStride - Compute the start and stride of this expression,
|
|
|
|
/// returning false if the expression is not a start/stride pair, or true if it
|
|
|
|
/// is. The stride must be a loop invariant expression, but the start may be
|
|
|
|
/// a mix of loop invariant and loop variant expressions.
|
|
|
|
static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
|
Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
For code like this:
void foo(float *a, float *b, int n, int stride_a, int stride_b) {
int i;
for (i=0; i<n; i++)
a[i*stride_a] = b[i*stride_b];
}
we now emit:
.LBB_foo2_2: ; no_exit
lfs f0, 0(r4)
stfs f0, 0(r3)
addi r7, r7, 1
add r4, r2, r4
add r3, r6, r3
cmpw cr0, r7, r5
blt .LBB_foo2_2 ; no_exit
instead of:
.LBB_foo_2: ; no_exit
mullw r8, r2, r7 ;; multiply!
slwi r8, r8, 2
lfsx f0, r4, r8
mullw r8, r2, r6 ;; multiply!
slwi r8, r8, 2
stfsx f0, r3, r8
addi r2, r2, 1
cmpw cr0, r2, r5
blt .LBB_foo_2 ; no_exit
loops with variable strides occur pretty often. For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.
Now we can allow indvars to turn functions written like this:
void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
int i, ai = 0, bi = 0;
for (i=0; i<n; i++)
{
a[ai] = b[bi];
ai += stride_a;
bi += stride_b;
}
}
into code like the above for better analysis. With this patch, they generate
identical code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22740 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-10 00:45:21 +00:00
|
|
|
SCEVHandle &Start, SCEVHandle &Stride) {
|
2005-08-04 19:08:16 +00:00
|
|
|
SCEVHandle TheAddRec = Start; // Initialize to zero.
|
|
|
|
|
|
|
|
// If the outer level is an AddExpr, the operands are all start values except
|
|
|
|
// for a nested AddRecExpr.
|
|
|
|
if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
|
|
|
|
for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
|
|
|
|
if (SCEVAddRecExpr *AddRec =
|
|
|
|
dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
|
|
|
|
if (AddRec->getLoop() == L)
|
|
|
|
TheAddRec = SCEVAddExpr::get(AddRec, TheAddRec);
|
|
|
|
else
|
|
|
|
return false; // Nested IV of some sort?
|
|
|
|
} else {
|
|
|
|
Start = SCEVAddExpr::get(Start, AE->getOperand(i));
|
|
|
|
}
|
|
|
|
|
|
|
|
} else if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SH)) {
|
|
|
|
TheAddRec = SH;
|
|
|
|
} else {
|
|
|
|
return false; // not analyzable.
|
|
|
|
}
|
|
|
|
|
|
|
|
SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
|
|
|
|
if (!AddRec || AddRec->getLoop() != L) return false;
|
|
|
|
|
|
|
|
// FIXME: Generalize to non-affine IV's.
|
|
|
|
if (!AddRec->isAffine()) return false;
|
|
|
|
|
|
|
|
Start = SCEVAddExpr::get(Start, AddRec->getOperand(0));
|
|
|
|
|
|
|
|
if (!isa<SCEVConstant>(AddRec->getOperand(1)))
|
Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
For code like this:
void foo(float *a, float *b, int n, int stride_a, int stride_b) {
int i;
for (i=0; i<n; i++)
a[i*stride_a] = b[i*stride_b];
}
we now emit:
.LBB_foo2_2: ; no_exit
lfs f0, 0(r4)
stfs f0, 0(r3)
addi r7, r7, 1
add r4, r2, r4
add r3, r6, r3
cmpw cr0, r7, r5
blt .LBB_foo2_2 ; no_exit
instead of:
.LBB_foo_2: ; no_exit
mullw r8, r2, r7 ;; multiply!
slwi r8, r8, 2
lfsx f0, r4, r8
mullw r8, r2, r6 ;; multiply!
slwi r8, r8, 2
stfsx f0, r3, r8
addi r2, r2, 1
cmpw cr0, r2, r5
blt .LBB_foo_2 ; no_exit
loops with variable strides occur pretty often. For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.
Now we can allow indvars to turn functions written like this:
void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
int i, ai = 0, bi = 0;
for (i=0; i<n; i++)
{
a[ai] = b[bi];
ai += stride_a;
bi += stride_b;
}
}
into code like the above for better analysis. With this patch, they generate
identical code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22740 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-10 00:45:21 +00:00
|
|
|
DEBUG(std::cerr << "[" << L->getHeader()->getName()
|
|
|
|
<< "] Variable stride: " << *AddRec << "\n");
|
|
|
|
|
|
|
|
Stride = AddRec->getOperand(1);
|
|
|
|
// Check that all constant strides are the unsigned type, we don't want to
|
|
|
|
// have two IV's one of signed stride 4 and one of unsigned stride 4 to not be
|
|
|
|
// merged.
|
|
|
|
assert((!isa<SCEVConstant>(Stride) || Stride->getType()->isUnsigned()) &&
|
2005-08-04 19:08:16 +00:00
|
|
|
"Constants should be canonicalized to unsigned!");
|
Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
For code like this:
void foo(float *a, float *b, int n, int stride_a, int stride_b) {
int i;
for (i=0; i<n; i++)
a[i*stride_a] = b[i*stride_b];
}
we now emit:
.LBB_foo2_2: ; no_exit
lfs f0, 0(r4)
stfs f0, 0(r3)
addi r7, r7, 1
add r4, r2, r4
add r3, r6, r3
cmpw cr0, r7, r5
blt .LBB_foo2_2 ; no_exit
instead of:
.LBB_foo_2: ; no_exit
mullw r8, r2, r7 ;; multiply!
slwi r8, r8, 2
lfsx f0, r4, r8
mullw r8, r2, r6 ;; multiply!
slwi r8, r8, 2
stfsx f0, r3, r8
addi r2, r2, 1
cmpw cr0, r2, r5
blt .LBB_foo_2 ; no_exit
loops with variable strides occur pretty often. For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.
Now we can allow indvars to turn functions written like this:
void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
int i, ai = 0, bi = 0;
for (i=0; i<n; i++)
{
a[ai] = b[bi];
ai += stride_a;
bi += stride_b;
}
}
into code like the above for better analysis. With this patch, they generate
identical code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22740 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-10 00:45:21 +00:00
|
|
|
|
2005-08-04 19:08:16 +00:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2005-10-03 01:04:44 +00:00
|
|
|
/// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
|
|
|
|
/// and now we need to decide whether the user should use the preinc or post-inc
|
|
|
|
/// value. If this user should use the post-inc version of the IV, return true.
|
|
|
|
///
|
|
|
|
/// Choosing wrong here can break dominance properties (if we choose to use the
|
|
|
|
/// post-inc value when we cannot) or it can end up adding extra live-ranges to
|
|
|
|
/// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
|
|
|
|
/// should use the post-inc value).
|
|
|
|
static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
|
Make IVUseShouldUsePostIncValue more aggressive when the use is a PHI. In
particular, it should realize that phi's use their values in the pred block
not the phi block itself. This change turns our em3d loop from this:
_test:
cmpwi cr0, r4, 0
bgt cr0, LBB_test_2 ; entry.no_exit_crit_edge
LBB_test_1: ; entry.loopexit_crit_edge
li r2, 0
b LBB_test_6 ; loopexit
LBB_test_2: ; entry.no_exit_crit_edge
li r6, 0
LBB_test_3: ; no_exit
or r2, r6, r6
lwz r6, 0(r3)
cmpw cr0, r6, r5
beq cr0, LBB_test_6 ; loopexit
LBB_test_4: ; endif
addi r3, r3, 4
addi r6, r2, 1
cmpw cr0, r6, r4
blt cr0, LBB_test_3 ; no_exit
LBB_test_5: ; endif.loopexit.loopexit_crit_edge
addi r3, r2, 1
blr
LBB_test_6: ; loopexit
or r3, r2, r2
blr
into:
_test:
cmpwi cr0, r4, 0
bgt cr0, LBB_test_2 ; entry.no_exit_crit_edge
LBB_test_1: ; entry.loopexit_crit_edge
li r2, 0
b LBB_test_5 ; loopexit
LBB_test_2: ; entry.no_exit_crit_edge
li r6, 0
LBB_test_3: ; no_exit
lwz r2, 0(r3)
cmpw cr0, r2, r5
or r2, r6, r6
beq cr0, LBB_test_5 ; loopexit
LBB_test_4: ; endif
addi r3, r3, 4
addi r6, r6, 1
cmpw cr0, r6, r4
or r2, r6, r6
blt cr0, LBB_test_3 ; no_exit
LBB_test_5: ; loopexit
or r3, r2, r2
blr
Unfortunately, this is actually worse code, because the register coallescer
is getting confused somehow. If it were doing its job right, it could turn the
code into this:
_test:
cmpwi cr0, r4, 0
bgt cr0, LBB_test_2 ; entry.no_exit_crit_edge
LBB_test_1: ; entry.loopexit_crit_edge
li r6, 0
b LBB_test_5 ; loopexit
LBB_test_2: ; entry.no_exit_crit_edge
li r6, 0
LBB_test_3: ; no_exit
lwz r2, 0(r3)
cmpw cr0, r2, r5
beq cr0, LBB_test_5 ; loopexit
LBB_test_4: ; endif
addi r3, r3, 4
addi r6, r6, 1
cmpw cr0, r6, r4
blt cr0, LBB_test_3 ; no_exit
LBB_test_5: ; loopexit
or r3, r6, r6
blr
... which I'll work on next. :)
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23604 91177308-0d34-0410-b5e6-96231b3b80d8
2005-10-03 02:50:05 +00:00
|
|
|
Loop *L, DominatorSet *DS, Pass *P) {
|
2005-10-03 01:04:44 +00:00
|
|
|
// If the user is in the loop, use the preinc value.
|
|
|
|
if (L->contains(User->getParent())) return false;
|
|
|
|
|
Make IVUseShouldUsePostIncValue more aggressive when the use is a PHI. In
particular, it should realize that phi's use their values in the pred block
not the phi block itself. This change turns our em3d loop from this:
_test:
cmpwi cr0, r4, 0
bgt cr0, LBB_test_2 ; entry.no_exit_crit_edge
LBB_test_1: ; entry.loopexit_crit_edge
li r2, 0
b LBB_test_6 ; loopexit
LBB_test_2: ; entry.no_exit_crit_edge
li r6, 0
LBB_test_3: ; no_exit
or r2, r6, r6
lwz r6, 0(r3)
cmpw cr0, r6, r5
beq cr0, LBB_test_6 ; loopexit
LBB_test_4: ; endif
addi r3, r3, 4
addi r6, r2, 1
cmpw cr0, r6, r4
blt cr0, LBB_test_3 ; no_exit
LBB_test_5: ; endif.loopexit.loopexit_crit_edge
addi r3, r2, 1
blr
LBB_test_6: ; loopexit
or r3, r2, r2
blr
into:
_test:
cmpwi cr0, r4, 0
bgt cr0, LBB_test_2 ; entry.no_exit_crit_edge
LBB_test_1: ; entry.loopexit_crit_edge
li r2, 0
b LBB_test_5 ; loopexit
LBB_test_2: ; entry.no_exit_crit_edge
li r6, 0
LBB_test_3: ; no_exit
lwz r2, 0(r3)
cmpw cr0, r2, r5
or r2, r6, r6
beq cr0, LBB_test_5 ; loopexit
LBB_test_4: ; endif
addi r3, r3, 4
addi r6, r6, 1
cmpw cr0, r6, r4
or r2, r6, r6
blt cr0, LBB_test_3 ; no_exit
LBB_test_5: ; loopexit
or r3, r2, r2
blr
Unfortunately, this is actually worse code, because the register coallescer
is getting confused somehow. If it were doing its job right, it could turn the
code into this:
_test:
cmpwi cr0, r4, 0
bgt cr0, LBB_test_2 ; entry.no_exit_crit_edge
LBB_test_1: ; entry.loopexit_crit_edge
li r6, 0
b LBB_test_5 ; loopexit
LBB_test_2: ; entry.no_exit_crit_edge
li r6, 0
LBB_test_3: ; no_exit
lwz r2, 0(r3)
cmpw cr0, r2, r5
beq cr0, LBB_test_5 ; loopexit
LBB_test_4: ; endif
addi r3, r3, 4
addi r6, r6, 1
cmpw cr0, r6, r4
blt cr0, LBB_test_3 ; no_exit
LBB_test_5: ; loopexit
or r3, r6, r6
blr
... which I'll work on next. :)
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23604 91177308-0d34-0410-b5e6-96231b3b80d8
2005-10-03 02:50:05 +00:00
|
|
|
BasicBlock *LatchBlock = L->getLoopLatch();
|
|
|
|
|
|
|
|
// Ok, the user is outside of the loop. If it is dominated by the latch
|
|
|
|
// block, use the post-inc value.
|
|
|
|
if (DS->dominates(LatchBlock, User->getParent()))
|
|
|
|
return true;
|
|
|
|
|
|
|
|
// There is one case we have to be careful of: PHI nodes. These little guys
|
|
|
|
// can live in blocks that do not dominate the latch block, but (since their
|
|
|
|
// uses occur in the predecessor block, not the block the PHI lives in) should
|
|
|
|
// still use the post-inc value. Check for this case now.
|
|
|
|
PHINode *PN = dyn_cast<PHINode>(User);
|
|
|
|
if (!PN) return false; // not a phi, not dominated by latch block.
|
|
|
|
|
|
|
|
// Look at all of the uses of IV by the PHI node. If any use corresponds to
|
|
|
|
// a block that is not dominated by the latch block, give up and use the
|
|
|
|
// preincremented value.
|
|
|
|
unsigned NumUses = 0;
|
|
|
|
for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
|
|
|
|
if (PN->getIncomingValue(i) == IV) {
|
|
|
|
++NumUses;
|
|
|
|
if (!DS->dominates(LatchBlock, PN->getIncomingBlock(i)))
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Okay, all uses of IV by PN are in predecessor blocks that really are
|
|
|
|
// dominated by the latch block. Split the critical edges and use the
|
|
|
|
// post-incremented value.
|
|
|
|
for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
|
|
|
|
if (PN->getIncomingValue(i) == IV) {
|
|
|
|
SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P);
|
|
|
|
if (--NumUses == 0) break;
|
|
|
|
}
|
|
|
|
|
|
|
|
return true;
|
2005-10-03 01:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-07-30 00:15:07 +00:00
|
|
|
/// AddUsersIfInteresting - Inspect the specified instruction. If it is a
|
|
|
|
/// reducible SCEV, recursively add its users to the IVUsesByStride set and
|
|
|
|
/// return true. Otherwise, return false.
|
2005-08-04 17:40:30 +00:00
|
|
|
bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
|
|
|
|
std::set<Instruction*> &Processed) {
|
2005-07-30 00:21:31 +00:00
|
|
|
if (I->getType() == Type::VoidTy) return false;
|
2005-08-04 17:40:30 +00:00
|
|
|
if (!Processed.insert(I).second)
|
|
|
|
return true; // Instruction already handled.
|
|
|
|
|
2005-08-04 19:08:16 +00:00
|
|
|
// Get the symbolic expression for this instruction.
|
2005-08-04 17:40:30 +00:00
|
|
|
SCEVHandle ISE = GetExpressionSCEV(I, L);
|
2005-08-04 19:08:16 +00:00
|
|
|
if (isa<SCEVCouldNotCompute>(ISE)) return false;
|
|
|
|
|
|
|
|
// Get the start and stride for this expression.
|
|
|
|
SCEVHandle Start = SCEVUnknown::getIntegerSCEV(0, ISE->getType());
|
Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
For code like this:
void foo(float *a, float *b, int n, int stride_a, int stride_b) {
int i;
for (i=0; i<n; i++)
a[i*stride_a] = b[i*stride_b];
}
we now emit:
.LBB_foo2_2: ; no_exit
lfs f0, 0(r4)
stfs f0, 0(r3)
addi r7, r7, 1
add r4, r2, r4
add r3, r6, r3
cmpw cr0, r7, r5
blt .LBB_foo2_2 ; no_exit
instead of:
.LBB_foo_2: ; no_exit
mullw r8, r2, r7 ;; multiply!
slwi r8, r8, 2
lfsx f0, r4, r8
mullw r8, r2, r6 ;; multiply!
slwi r8, r8, 2
stfsx f0, r3, r8
addi r2, r2, 1
cmpw cr0, r2, r5
blt .LBB_foo_2 ; no_exit
loops with variable strides occur pretty often. For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.
Now we can allow indvars to turn functions written like this:
void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
int i, ai = 0, bi = 0;
for (i=0; i<n; i++)
{
a[ai] = b[bi];
ai += stride_a;
bi += stride_b;
}
}
into code like the above for better analysis. With this patch, they generate
identical code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22740 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-10 00:45:21 +00:00
|
|
|
SCEVHandle Stride = Start;
|
2005-08-04 19:08:16 +00:00
|
|
|
if (!getSCEVStartAndStride(ISE, L, Start, Stride))
|
|
|
|
return false; // Non-reducible symbolic expression, bail out.
|
2005-08-04 17:40:30 +00:00
|
|
|
|
2005-07-30 00:15:07 +00:00
|
|
|
for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;++UI){
|
|
|
|
Instruction *User = cast<Instruction>(*UI);
|
|
|
|
|
|
|
|
// Do not infinitely recurse on PHI nodes.
|
2005-09-13 02:09:55 +00:00
|
|
|
if (isa<PHINode>(User) && Processed.count(User))
|
2005-07-30 00:15:07 +00:00
|
|
|
continue;
|
|
|
|
|
|
|
|
// If this is an instruction defined in a nested loop, or outside this loop,
|
When processing outer loops and we find uses of an IV in inner loops, make
sure to handle the use, just don't recurse into it.
This permits us to generate this code for a simple nested loop case:
.LBB_foo_0: ; entry
stwu r1, -48(r1)
stw r29, 44(r1)
stw r30, 40(r1)
mflr r11
stw r11, 56(r1)
lis r2, ha16(L_A$non_lazy_ptr)
lwz r30, lo16(L_A$non_lazy_ptr)(r2)
li r29, 1
.LBB_foo_1: ; no_exit.0
bl L_bar$stub
li r2, 1
or r3, r30, r30
.LBB_foo_2: ; no_exit.1
lfd f0, 8(r3)
stfd f0, 0(r3)
addi r4, r2, 1
addi r3, r3, 8
cmpwi cr0, r2, 100
or r2, r4, r4
bne .LBB_foo_2 ; no_exit.1
.LBB_foo_3: ; loopexit.1
addi r30, r30, 800
addi r2, r29, 1
cmpwi cr0, r29, 100
or r29, r2, r2
bne .LBB_foo_1 ; no_exit.0
.LBB_foo_4: ; return
lwz r11, 56(r1)
mtlr r11
lwz r30, 40(r1)
lwz r29, 44(r1)
lwz r1, 0(r1)
blr
instead of this:
_foo:
.LBB_foo_0: ; entry
stwu r1, -48(r1)
stw r28, 44(r1) ;; uses an extra register.
stw r29, 40(r1)
stw r30, 36(r1)
mflr r11
stw r11, 56(r1)
li r30, 1
li r29, 0
or r28, r29, r29
.LBB_foo_1: ; no_exit.0
bl L_bar$stub
mulli r2, r28, 800 ;; unstrength-reduced multiply
lis r3, ha16(L_A$non_lazy_ptr) ;; loop invariant address computation
lwz r3, lo16(L_A$non_lazy_ptr)(r3)
add r2, r2, r3
mulli r4, r29, 800 ;; unstrength-reduced multiply
addi r3, r3, 8
add r3, r4, r3
li r4, 1
.LBB_foo_2: ; no_exit.1
lfd f0, 0(r3)
stfd f0, 0(r2)
addi r5, r4, 1
addi r2, r2, 8 ;; multiple stride 8 IV's
addi r3, r3, 8
cmpwi cr0, r4, 100
or r4, r5, r5
bne .LBB_foo_2 ; no_exit.1
.LBB_foo_3: ; loopexit.1
addi r28, r28, 1 ;;; Many IV's with stride 1
addi r29, r29, 1
addi r2, r30, 1
cmpwi cr0, r30, 100
or r30, r2, r2
bne .LBB_foo_1 ; no_exit.0
.LBB_foo_4: ; return
lwz r11, 56(r1)
mtlr r11
lwz r30, 36(r1)
lwz r29, 40(r1)
lwz r28, 44(r1)
lwz r1, 0(r1)
blr
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22640 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-04 00:14:11 +00:00
|
|
|
// don't recurse into it.
|
2005-08-04 19:08:16 +00:00
|
|
|
bool AddUserToIVUsers = false;
|
When processing outer loops and we find uses of an IV in inner loops, make
sure to handle the use, just don't recurse into it.
This permits us to generate this code for a simple nested loop case:
.LBB_foo_0: ; entry
stwu r1, -48(r1)
stw r29, 44(r1)
stw r30, 40(r1)
mflr r11
stw r11, 56(r1)
lis r2, ha16(L_A$non_lazy_ptr)
lwz r30, lo16(L_A$non_lazy_ptr)(r2)
li r29, 1
.LBB_foo_1: ; no_exit.0
bl L_bar$stub
li r2, 1
or r3, r30, r30
.LBB_foo_2: ; no_exit.1
lfd f0, 8(r3)
stfd f0, 0(r3)
addi r4, r2, 1
addi r3, r3, 8
cmpwi cr0, r2, 100
or r2, r4, r4
bne .LBB_foo_2 ; no_exit.1
.LBB_foo_3: ; loopexit.1
addi r30, r30, 800
addi r2, r29, 1
cmpwi cr0, r29, 100
or r29, r2, r2
bne .LBB_foo_1 ; no_exit.0
.LBB_foo_4: ; return
lwz r11, 56(r1)
mtlr r11
lwz r30, 40(r1)
lwz r29, 44(r1)
lwz r1, 0(r1)
blr
instead of this:
_foo:
.LBB_foo_0: ; entry
stwu r1, -48(r1)
stw r28, 44(r1) ;; uses an extra register.
stw r29, 40(r1)
stw r30, 36(r1)
mflr r11
stw r11, 56(r1)
li r30, 1
li r29, 0
or r28, r29, r29
.LBB_foo_1: ; no_exit.0
bl L_bar$stub
mulli r2, r28, 800 ;; unstrength-reduced multiply
lis r3, ha16(L_A$non_lazy_ptr) ;; loop invariant address computation
lwz r3, lo16(L_A$non_lazy_ptr)(r3)
add r2, r2, r3
mulli r4, r29, 800 ;; unstrength-reduced multiply
addi r3, r3, 8
add r3, r4, r3
li r4, 1
.LBB_foo_2: ; no_exit.1
lfd f0, 0(r3)
stfd f0, 0(r2)
addi r5, r4, 1
addi r2, r2, 8 ;; multiple stride 8 IV's
addi r3, r3, 8
cmpwi cr0, r4, 100
or r4, r5, r5
bne .LBB_foo_2 ; no_exit.1
.LBB_foo_3: ; loopexit.1
addi r28, r28, 1 ;;; Many IV's with stride 1
addi r29, r29, 1
addi r2, r30, 1
cmpwi cr0, r30, 100
or r30, r2, r2
bne .LBB_foo_1 ; no_exit.0
.LBB_foo_4: ; return
lwz r11, 56(r1)
mtlr r11
lwz r30, 36(r1)
lwz r29, 40(r1)
lwz r28, 44(r1)
lwz r1, 0(r1)
blr
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22640 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-04 00:14:11 +00:00
|
|
|
if (LI->getLoopFor(User->getParent()) != L) {
|
2005-09-13 02:09:55 +00:00
|
|
|
DEBUG(std::cerr << "FOUND USER in other loop: " << *User
|
When processing outer loops and we find uses of an IV in inner loops, make
sure to handle the use, just don't recurse into it.
This permits us to generate this code for a simple nested loop case:
.LBB_foo_0: ; entry
stwu r1, -48(r1)
stw r29, 44(r1)
stw r30, 40(r1)
mflr r11
stw r11, 56(r1)
lis r2, ha16(L_A$non_lazy_ptr)
lwz r30, lo16(L_A$non_lazy_ptr)(r2)
li r29, 1
.LBB_foo_1: ; no_exit.0
bl L_bar$stub
li r2, 1
or r3, r30, r30
.LBB_foo_2: ; no_exit.1
lfd f0, 8(r3)
stfd f0, 0(r3)
addi r4, r2, 1
addi r3, r3, 8
cmpwi cr0, r2, 100
or r2, r4, r4
bne .LBB_foo_2 ; no_exit.1
.LBB_foo_3: ; loopexit.1
addi r30, r30, 800
addi r2, r29, 1
cmpwi cr0, r29, 100
or r29, r2, r2
bne .LBB_foo_1 ; no_exit.0
.LBB_foo_4: ; return
lwz r11, 56(r1)
mtlr r11
lwz r30, 40(r1)
lwz r29, 44(r1)
lwz r1, 0(r1)
blr
instead of this:
_foo:
.LBB_foo_0: ; entry
stwu r1, -48(r1)
stw r28, 44(r1) ;; uses an extra register.
stw r29, 40(r1)
stw r30, 36(r1)
mflr r11
stw r11, 56(r1)
li r30, 1
li r29, 0
or r28, r29, r29
.LBB_foo_1: ; no_exit.0
bl L_bar$stub
mulli r2, r28, 800 ;; unstrength-reduced multiply
lis r3, ha16(L_A$non_lazy_ptr) ;; loop invariant address computation
lwz r3, lo16(L_A$non_lazy_ptr)(r3)
add r2, r2, r3
mulli r4, r29, 800 ;; unstrength-reduced multiply
addi r3, r3, 8
add r3, r4, r3
li r4, 1
.LBB_foo_2: ; no_exit.1
lfd f0, 0(r3)
stfd f0, 0(r2)
addi r5, r4, 1
addi r2, r2, 8 ;; multiple stride 8 IV's
addi r3, r3, 8
cmpwi cr0, r4, 100
or r4, r5, r5
bne .LBB_foo_2 ; no_exit.1
.LBB_foo_3: ; loopexit.1
addi r28, r28, 1 ;;; Many IV's with stride 1
addi r29, r29, 1
addi r2, r30, 1
cmpwi cr0, r30, 100
or r30, r2, r2
bne .LBB_foo_1 ; no_exit.0
.LBB_foo_4: ; return
lwz r11, 56(r1)
mtlr r11
lwz r30, 36(r1)
lwz r29, 40(r1)
lwz r28, 44(r1)
lwz r1, 0(r1)
blr
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22640 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-04 00:14:11 +00:00
|
|
|
<< " OF SCEV: " << *ISE << "\n");
|
2005-08-04 19:08:16 +00:00
|
|
|
AddUserToIVUsers = true;
|
2005-08-04 17:40:30 +00:00
|
|
|
} else if (!AddUsersIfInteresting(User, L, Processed)) {
|
2005-08-04 00:40:47 +00:00
|
|
|
DEBUG(std::cerr << "FOUND USER: " << *User
|
|
|
|
<< " OF SCEV: " << *ISE << "\n");
|
2005-08-04 19:08:16 +00:00
|
|
|
AddUserToIVUsers = true;
|
|
|
|
}
|
2005-04-21 23:48:37 +00:00
|
|
|
|
2005-08-04 19:08:16 +00:00
|
|
|
if (AddUserToIVUsers) {
|
2005-10-09 06:20:55 +00:00
|
|
|
IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
|
|
|
|
if (StrideUses.Users.empty()) // First occurance of this stride?
|
|
|
|
StrideOrder.push_back(Stride);
|
|
|
|
|
2005-08-04 00:40:47 +00:00
|
|
|
// Okay, we found a user that we cannot reduce. Analyze the instruction
|
_test:
li r2, 0
LBB_test_1: ; no_exit.2
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
addi r3, r3, 4
cmpwi cr0, r2, 701
blt cr0, LBB_test_1 ; no_exit.2
LBB_test_2: ; loopexit.2.loopexit
addi r2, r2, 1
stw r2, 0(r4)
blr
[zion ~/llvm]$ cat > ~/xx
Uses of IV's outside of the loop should use hte post-incremented version
of the IV, not the preincremented version. This helps many loops (e.g. in sixtrack)
which used to generate code like this (this is the code from the
dont-hoist-simple-loop-constants.ll testcase):
_test:
li r2, 0 **** IV starts at 0
LBB_test_1: ; no_exit.2
or r5, r2, r2 **** Copy for loop exit
li r2, 0
stw r2, 0(r3)
addi r3, r3, 4
addi r2, r5, 1
addi r6, r5, 2 **** IV+2
cmpwi cr0, r6, 701
blt cr0, LBB_test_1 ; no_exit.2
LBB_test_2: ; loopexit.2.loopexit
addi r2, r5, 2 **** IV+2
stw r2, 0(r4)
blr
And now generated code like this:
_test:
li r2, 1 *** IV starts at 1
LBB_test_1: ; no_exit.2
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
addi r3, r3, 4
cmpwi cr0, r2, 701 *** IV.postinc + 0
blt cr0, LBB_test_1
LBB_test_2: ; loopexit.2.loopexit
stw r2, 0(r4) *** IV.postinc + 0
blr
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23313 91177308-0d34-0410-b5e6-96231b3b80d8
2005-09-12 06:04:47 +00:00
|
|
|
// and decide what to do with it. If we are a use inside of the loop, use
|
|
|
|
// the value before incrementation, otherwise use it after incrementation.
|
Make IVUseShouldUsePostIncValue more aggressive when the use is a PHI. In
particular, it should realize that phi's use their values in the pred block
not the phi block itself. This change turns our em3d loop from this:
_test:
cmpwi cr0, r4, 0
bgt cr0, LBB_test_2 ; entry.no_exit_crit_edge
LBB_test_1: ; entry.loopexit_crit_edge
li r2, 0
b LBB_test_6 ; loopexit
LBB_test_2: ; entry.no_exit_crit_edge
li r6, 0
LBB_test_3: ; no_exit
or r2, r6, r6
lwz r6, 0(r3)
cmpw cr0, r6, r5
beq cr0, LBB_test_6 ; loopexit
LBB_test_4: ; endif
addi r3, r3, 4
addi r6, r2, 1
cmpw cr0, r6, r4
blt cr0, LBB_test_3 ; no_exit
LBB_test_5: ; endif.loopexit.loopexit_crit_edge
addi r3, r2, 1
blr
LBB_test_6: ; loopexit
or r3, r2, r2
blr
into:
_test:
cmpwi cr0, r4, 0
bgt cr0, LBB_test_2 ; entry.no_exit_crit_edge
LBB_test_1: ; entry.loopexit_crit_edge
li r2, 0
b LBB_test_5 ; loopexit
LBB_test_2: ; entry.no_exit_crit_edge
li r6, 0
LBB_test_3: ; no_exit
lwz r2, 0(r3)
cmpw cr0, r2, r5
or r2, r6, r6
beq cr0, LBB_test_5 ; loopexit
LBB_test_4: ; endif
addi r3, r3, 4
addi r6, r6, 1
cmpw cr0, r6, r4
or r2, r6, r6
blt cr0, LBB_test_3 ; no_exit
LBB_test_5: ; loopexit
or r3, r2, r2
blr
Unfortunately, this is actually worse code, because the register coallescer
is getting confused somehow. If it were doing its job right, it could turn the
code into this:
_test:
cmpwi cr0, r4, 0
bgt cr0, LBB_test_2 ; entry.no_exit_crit_edge
LBB_test_1: ; entry.loopexit_crit_edge
li r6, 0
b LBB_test_5 ; loopexit
LBB_test_2: ; entry.no_exit_crit_edge
li r6, 0
LBB_test_3: ; no_exit
lwz r2, 0(r3)
cmpw cr0, r2, r5
beq cr0, LBB_test_5 ; loopexit
LBB_test_4: ; endif
addi r3, r3, 4
addi r6, r6, 1
cmpw cr0, r6, r4
blt cr0, LBB_test_3 ; no_exit
LBB_test_5: ; loopexit
or r3, r6, r6
blr
... which I'll work on next. :)
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23604 91177308-0d34-0410-b5e6-96231b3b80d8
2005-10-03 02:50:05 +00:00
|
|
|
if (IVUseShouldUsePostIncValue(User, I, L, DS, this)) {
|
_test:
li r2, 0
LBB_test_1: ; no_exit.2
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
addi r3, r3, 4
cmpwi cr0, r2, 701
blt cr0, LBB_test_1 ; no_exit.2
LBB_test_2: ; loopexit.2.loopexit
addi r2, r2, 1
stw r2, 0(r4)
blr
[zion ~/llvm]$ cat > ~/xx
Uses of IV's outside of the loop should use hte post-incremented version
of the IV, not the preincremented version. This helps many loops (e.g. in sixtrack)
which used to generate code like this (this is the code from the
dont-hoist-simple-loop-constants.ll testcase):
_test:
li r2, 0 **** IV starts at 0
LBB_test_1: ; no_exit.2
or r5, r2, r2 **** Copy for loop exit
li r2, 0
stw r2, 0(r3)
addi r3, r3, 4
addi r2, r5, 1
addi r6, r5, 2 **** IV+2
cmpwi cr0, r6, 701
blt cr0, LBB_test_1 ; no_exit.2
LBB_test_2: ; loopexit.2.loopexit
addi r2, r5, 2 **** IV+2
stw r2, 0(r4)
blr
And now generated code like this:
_test:
li r2, 1 *** IV starts at 1
LBB_test_1: ; no_exit.2
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
addi r3, r3, 4
cmpwi cr0, r2, 701 *** IV.postinc + 0
blt cr0, LBB_test_1
LBB_test_2: ; loopexit.2.loopexit
stw r2, 0(r4) *** IV.postinc + 0
blr
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23313 91177308-0d34-0410-b5e6-96231b3b80d8
2005-09-12 06:04:47 +00:00
|
|
|
// The value used will be incremented by the stride more than we are
|
|
|
|
// expecting, so subtract this off.
|
|
|
|
SCEVHandle NewStart = SCEV::getMinusSCEV(Start, Stride);
|
2005-10-09 06:20:55 +00:00
|
|
|
StrideUses.addUser(NewStart, User, I);
|
|
|
|
StrideUses.Users.back().isUseOfPostIncrementedValue = true;
|
Make IVUseShouldUsePostIncValue more aggressive when the use is a PHI. In
particular, it should realize that phi's use their values in the pred block
not the phi block itself. This change turns our em3d loop from this:
_test:
cmpwi cr0, r4, 0
bgt cr0, LBB_test_2 ; entry.no_exit_crit_edge
LBB_test_1: ; entry.loopexit_crit_edge
li r2, 0
b LBB_test_6 ; loopexit
LBB_test_2: ; entry.no_exit_crit_edge
li r6, 0
LBB_test_3: ; no_exit
or r2, r6, r6
lwz r6, 0(r3)
cmpw cr0, r6, r5
beq cr0, LBB_test_6 ; loopexit
LBB_test_4: ; endif
addi r3, r3, 4
addi r6, r2, 1
cmpw cr0, r6, r4
blt cr0, LBB_test_3 ; no_exit
LBB_test_5: ; endif.loopexit.loopexit_crit_edge
addi r3, r2, 1
blr
LBB_test_6: ; loopexit
or r3, r2, r2
blr
into:
_test:
cmpwi cr0, r4, 0
bgt cr0, LBB_test_2 ; entry.no_exit_crit_edge
LBB_test_1: ; entry.loopexit_crit_edge
li r2, 0
b LBB_test_5 ; loopexit
LBB_test_2: ; entry.no_exit_crit_edge
li r6, 0
LBB_test_3: ; no_exit
lwz r2, 0(r3)
cmpw cr0, r2, r5
or r2, r6, r6
beq cr0, LBB_test_5 ; loopexit
LBB_test_4: ; endif
addi r3, r3, 4
addi r6, r6, 1
cmpw cr0, r6, r4
or r2, r6, r6
blt cr0, LBB_test_3 ; no_exit
LBB_test_5: ; loopexit
or r3, r2, r2
blr
Unfortunately, this is actually worse code, because the register coallescer
is getting confused somehow. If it were doing its job right, it could turn the
code into this:
_test:
cmpwi cr0, r4, 0
bgt cr0, LBB_test_2 ; entry.no_exit_crit_edge
LBB_test_1: ; entry.loopexit_crit_edge
li r6, 0
b LBB_test_5 ; loopexit
LBB_test_2: ; entry.no_exit_crit_edge
li r6, 0
LBB_test_3: ; no_exit
lwz r2, 0(r3)
cmpw cr0, r2, r5
beq cr0, LBB_test_5 ; loopexit
LBB_test_4: ; endif
addi r3, r3, 4
addi r6, r6, 1
cmpw cr0, r6, r4
blt cr0, LBB_test_3 ; no_exit
LBB_test_5: ; loopexit
or r3, r6, r6
blr
... which I'll work on next. :)
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23604 91177308-0d34-0410-b5e6-96231b3b80d8
2005-10-03 02:50:05 +00:00
|
|
|
DEBUG(std::cerr << " USING POSTINC SCEV, START=" << *NewStart<< "\n");
|
2005-10-03 01:04:44 +00:00
|
|
|
} else {
|
2005-10-09 06:20:55 +00:00
|
|
|
StrideUses.addUser(Start, User, I);
|
_test:
li r2, 0
LBB_test_1: ; no_exit.2
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
addi r3, r3, 4
cmpwi cr0, r2, 701
blt cr0, LBB_test_1 ; no_exit.2
LBB_test_2: ; loopexit.2.loopexit
addi r2, r2, 1
stw r2, 0(r4)
blr
[zion ~/llvm]$ cat > ~/xx
Uses of IV's outside of the loop should use hte post-incremented version
of the IV, not the preincremented version. This helps many loops (e.g. in sixtrack)
which used to generate code like this (this is the code from the
dont-hoist-simple-loop-constants.ll testcase):
_test:
li r2, 0 **** IV starts at 0
LBB_test_1: ; no_exit.2
or r5, r2, r2 **** Copy for loop exit
li r2, 0
stw r2, 0(r3)
addi r3, r3, 4
addi r2, r5, 1
addi r6, r5, 2 **** IV+2
cmpwi cr0, r6, 701
blt cr0, LBB_test_1 ; no_exit.2
LBB_test_2: ; loopexit.2.loopexit
addi r2, r5, 2 **** IV+2
stw r2, 0(r4)
blr
And now generated code like this:
_test:
li r2, 1 *** IV starts at 1
LBB_test_1: ; no_exit.2
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
addi r3, r3, 4
cmpwi cr0, r2, 701 *** IV.postinc + 0
blt cr0, LBB_test_1
LBB_test_2: ; loopexit.2.loopexit
stw r2, 0(r4) *** IV.postinc + 0
blr
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23313 91177308-0d34-0410-b5e6-96231b3b80d8
2005-09-12 06:04:47 +00:00
|
|
|
}
|
2005-07-30 00:15:07 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
/// BasedUser - For a particular base value, keep information about how we've
|
|
|
|
/// partitioned the expression so far.
|
|
|
|
struct BasedUser {
|
2005-08-08 22:56:21 +00:00
|
|
|
/// Base - The Base value for the PHI node that needs to be inserted for
|
|
|
|
/// this use. As the use is processed, information gets moved from this
|
|
|
|
/// field to the Imm field (below). BasedUser values are sorted by this
|
|
|
|
/// field.
|
|
|
|
SCEVHandle Base;
|
|
|
|
|
2005-07-30 00:15:07 +00:00
|
|
|
/// Inst - The instruction using the induction variable.
|
|
|
|
Instruction *Inst;
|
|
|
|
|
2005-08-03 22:21:05 +00:00
|
|
|
/// OperandValToReplace - The operand value of Inst to replace with the
|
|
|
|
/// EmittedBase.
|
|
|
|
Value *OperandValToReplace;
|
2005-07-30 00:15:07 +00:00
|
|
|
|
|
|
|
/// Imm - The immediate value that should be added to the base immediately
|
|
|
|
/// before Inst, because it will be folded into the imm field of the
|
|
|
|
/// instruction.
|
|
|
|
SCEVHandle Imm;
|
|
|
|
|
|
|
|
/// EmittedBase - The actual value* to use for the base value of this
|
|
|
|
/// operation. This is null if we should just use zero so far.
|
|
|
|
Value *EmittedBase;
|
|
|
|
|
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-08 05:28:22 +00:00
|
|
|
// isUseOfPostIncrementedValue - True if this should use the
|
|
|
|
// post-incremented version of this IV, not the preincremented version.
|
|
|
|
// This can only be set in special cases, such as the terminating setcc
|
_test:
li r2, 0
LBB_test_1: ; no_exit.2
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
addi r3, r3, 4
cmpwi cr0, r2, 701
blt cr0, LBB_test_1 ; no_exit.2
LBB_test_2: ; loopexit.2.loopexit
addi r2, r2, 1
stw r2, 0(r4)
blr
[zion ~/llvm]$ cat > ~/xx
Uses of IV's outside of the loop should use hte post-incremented version
of the IV, not the preincremented version. This helps many loops (e.g. in sixtrack)
which used to generate code like this (this is the code from the
dont-hoist-simple-loop-constants.ll testcase):
_test:
li r2, 0 **** IV starts at 0
LBB_test_1: ; no_exit.2
or r5, r2, r2 **** Copy for loop exit
li r2, 0
stw r2, 0(r3)
addi r3, r3, 4
addi r2, r5, 1
addi r6, r5, 2 **** IV+2
cmpwi cr0, r6, 701
blt cr0, LBB_test_1 ; no_exit.2
LBB_test_2: ; loopexit.2.loopexit
addi r2, r5, 2 **** IV+2
stw r2, 0(r4)
blr
And now generated code like this:
_test:
li r2, 1 *** IV starts at 1
LBB_test_1: ; no_exit.2
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
addi r3, r3, 4
cmpwi cr0, r2, 701 *** IV.postinc + 0
blt cr0, LBB_test_1
LBB_test_2: ; loopexit.2.loopexit
stw r2, 0(r4) *** IV.postinc + 0
blr
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23313 91177308-0d34-0410-b5e6-96231b3b80d8
2005-09-12 06:04:47 +00:00
|
|
|
// instruction for a loop and uses outside the loop that are dominated by
|
|
|
|
// the loop.
|
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-08 05:28:22 +00:00
|
|
|
bool isUseOfPostIncrementedValue;
|
2005-08-08 22:56:21 +00:00
|
|
|
|
|
|
|
BasedUser(IVStrideUse &IVSU)
|
|
|
|
: Base(IVSU.Offset), Inst(IVSU.User),
|
|
|
|
OperandValToReplace(IVSU.OperandValToReplace),
|
|
|
|
Imm(SCEVUnknown::getIntegerSCEV(0, Base->getType())), EmittedBase(0),
|
|
|
|
isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
|
2005-07-30 00:15:07 +00:00
|
|
|
|
2005-08-04 20:03:32 +00:00
|
|
|
// Once we rewrite the code to insert the new IVs we want, update the
|
|
|
|
// operands of Inst to use the new expression 'NewBase', with 'Imm' added
|
|
|
|
// to it.
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
|
When splitting critical edges, make sure not to leave the new block in the
middle of the loop. This turns a critical loop in gzip into this:
.LBB_test_1: ; loopentry
or r27, r28, r28
add r28, r3, r27
lhz r28, 3(r28)
add r26, r4, r27
lhz r26, 3(r26)
cmpw cr0, r28, r26
bne .LBB_test_8 ; loopentry.loopexit_crit_edge
.LBB_test_2: ; shortcirc_next.0
add r28, r3, r27
lhz r28, 5(r28)
add r26, r4, r27
lhz r26, 5(r26)
cmpw cr0, r28, r26
bne .LBB_test_7 ; shortcirc_next.0.loopexit_crit_edge
.LBB_test_3: ; shortcirc_next.1
add r28, r3, r27
lhz r28, 7(r28)
add r26, r4, r27
lhz r26, 7(r26)
cmpw cr0, r28, r26
bne .LBB_test_6 ; shortcirc_next.1.loopexit_crit_edge
.LBB_test_4: ; shortcirc_next.2
add r28, r3, r27
lhz r26, 9(r28)
add r28, r4, r27
lhz r25, 9(r28)
addi r28, r27, 8
cmpw cr7, r26, r25
mfcr r26, 1
rlwinm r26, r26, 31, 31, 31
add r25, r8, r27
cmpw cr7, r25, r7
mfcr r25, 1
rlwinm r25, r25, 29, 31, 31
and. r26, r26, r25
bne .LBB_test_1 ; loopentry
instead of this:
.LBB_test_1: ; loopentry
or r27, r28, r28
add r28, r3, r27
lhz r28, 3(r28)
add r26, r4, r27
lhz r26, 3(r26)
cmpw cr0, r28, r26
beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2: ; loopentry.loopexit_crit_edge
add r2, r30, r27
add r8, r29, r27
b .LBB_test_9 ; loopexit
.LBB_test_3: ; shortcirc_next.0
add r28, r3, r27
lhz r28, 5(r28)
add r26, r4, r27
lhz r26, 5(r26)
cmpw cr0, r28, r26
beq .LBB_test_5 ; shortcirc_next.1
.LBB_test_4: ; shortcirc_next.0.loopexit_crit_edge
add r2, r11, r27
add r8, r12, r27
b .LBB_test_9 ; loopexit
.LBB_test_5: ; shortcirc_next.1
add r28, r3, r27
lhz r28, 7(r28)
add r26, r4, r27
lhz r26, 7(r26)
cmpw cr0, r28, r26
beq .LBB_test_7 ; shortcirc_next.2
.LBB_test_6: ; shortcirc_next.1.loopexit_crit_edge
add r2, r9, r27
add r8, r10, r27
b .LBB_test_9 ; loopexit
.LBB_test_7: ; shortcirc_next.2
add r28, r3, r27
lhz r26, 9(r28)
add r28, r4, r27
lhz r25, 9(r28)
addi r28, r27, 8
cmpw cr7, r26, r25
mfcr r26, 1
rlwinm r26, r26, 31, 31, 31
add r25, r8, r27
cmpw cr7, r25, r7
mfcr r25, 1
rlwinm r25, r25, 29, 31, 31
and. r26, r26, r25
bne .LBB_test_1 ; loopentry
Next up, improve the code for the loop.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22769 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-12 22:22:17 +00:00
|
|
|
SCEVExpander &Rewriter, Loop *L,
|
|
|
|
Pass *P);
|
2005-07-30 00:15:07 +00:00
|
|
|
|
2005-08-08 22:56:21 +00:00
|
|
|
// Sort by the Base field.
|
|
|
|
bool operator<(const BasedUser &BU) const { return Base < BU.Base; }
|
2005-07-30 00:15:07 +00:00
|
|
|
|
|
|
|
void dump() const;
|
|
|
|
};
|
|
|
|
}
|
|
|
|
|
|
|
|
void BasedUser::dump() const {
|
2005-08-08 22:56:21 +00:00
|
|
|
std::cerr << " Base=" << *Base;
|
2005-07-30 00:15:07 +00:00
|
|
|
std::cerr << " Imm=" << *Imm;
|
|
|
|
if (EmittedBase)
|
|
|
|
std::cerr << " EB=" << *EmittedBase;
|
|
|
|
|
|
|
|
std::cerr << " Inst: " << *Inst;
|
|
|
|
}
|
|
|
|
|
2005-08-04 20:03:32 +00:00
|
|
|
// Once we rewrite the code to insert the new IVs we want, update the
|
|
|
|
// operands of Inst to use the new expression 'NewBase', with 'Imm' added
|
|
|
|
// to it.
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
|
Fix a FIXME: if we are inserting code for a PHI argument, split the critical
edge so that the code is not always executed for both operands. This
prevents LSR from inserting code into loops whose exit blocks contain
PHI uses of IV expressions (which are outside of loops). On gzip, for
example, we turn this ugly code:
.LBB_test_1: ; loopentry
add r27, r3, r28
lhz r27, 3(r27)
add r26, r4, r28
lhz r26, 3(r26)
add r25, r30, r28 ;; Only live if exiting the loop
add r24, r29, r28 ;; Only live if exiting the loop
cmpw cr0, r27, r26
bne .LBB_test_5 ; loopexit
into this:
.LBB_test_1: ; loopentry
or r27, r28, r28
add r28, r3, r27
lhz r28, 3(r28)
add r26, r4, r27
lhz r26, 3(r26)
cmpw cr0, r28, r26
beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2: ; loopentry.loopexit_crit_edge
add r2, r30, r27
add r8, r29, r27
b .LBB_test_9 ; loopexit
.LBB_test_2: ; shortcirc_next.0
...
blt .LBB_test_1
into this:
.LBB_test_1: ; loopentry
or r27, r28, r28
add r28, r3, r27
lhz r28, 3(r28)
add r26, r4, r27
lhz r26, 3(r26)
cmpw cr0, r28, r26
beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2: ; loopentry.loopexit_crit_edge
add r2, r30, r27
add r8, r29, r27
b .LBB_t_3: ; shortcirc_next.0
.LBB_test_3: ; shortcirc_next.0
...
blt .LBB_test_1
Next step: get the block out of the loop so that the loop is all
fall-throughs again.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22766 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-12 22:06:11 +00:00
|
|
|
SCEVExpander &Rewriter,
|
When splitting critical edges, make sure not to leave the new block in the
middle of the loop. This turns a critical loop in gzip into this:
.LBB_test_1: ; loopentry
or r27, r28, r28
add r28, r3, r27
lhz r28, 3(r28)
add r26, r4, r27
lhz r26, 3(r26)
cmpw cr0, r28, r26
bne .LBB_test_8 ; loopentry.loopexit_crit_edge
.LBB_test_2: ; shortcirc_next.0
add r28, r3, r27
lhz r28, 5(r28)
add r26, r4, r27
lhz r26, 5(r26)
cmpw cr0, r28, r26
bne .LBB_test_7 ; shortcirc_next.0.loopexit_crit_edge
.LBB_test_3: ; shortcirc_next.1
add r28, r3, r27
lhz r28, 7(r28)
add r26, r4, r27
lhz r26, 7(r26)
cmpw cr0, r28, r26
bne .LBB_test_6 ; shortcirc_next.1.loopexit_crit_edge
.LBB_test_4: ; shortcirc_next.2
add r28, r3, r27
lhz r26, 9(r28)
add r28, r4, r27
lhz r25, 9(r28)
addi r28, r27, 8
cmpw cr7, r26, r25
mfcr r26, 1
rlwinm r26, r26, 31, 31, 31
add r25, r8, r27
cmpw cr7, r25, r7
mfcr r25, 1
rlwinm r25, r25, 29, 31, 31
and. r26, r26, r25
bne .LBB_test_1 ; loopentry
instead of this:
.LBB_test_1: ; loopentry
or r27, r28, r28
add r28, r3, r27
lhz r28, 3(r28)
add r26, r4, r27
lhz r26, 3(r26)
cmpw cr0, r28, r26
beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2: ; loopentry.loopexit_crit_edge
add r2, r30, r27
add r8, r29, r27
b .LBB_test_9 ; loopexit
.LBB_test_3: ; shortcirc_next.0
add r28, r3, r27
lhz r28, 5(r28)
add r26, r4, r27
lhz r26, 5(r26)
cmpw cr0, r28, r26
beq .LBB_test_5 ; shortcirc_next.1
.LBB_test_4: ; shortcirc_next.0.loopexit_crit_edge
add r2, r11, r27
add r8, r12, r27
b .LBB_test_9 ; loopexit
.LBB_test_5: ; shortcirc_next.1
add r28, r3, r27
lhz r28, 7(r28)
add r26, r4, r27
lhz r26, 7(r26)
cmpw cr0, r28, r26
beq .LBB_test_7 ; shortcirc_next.2
.LBB_test_6: ; shortcirc_next.1.loopexit_crit_edge
add r2, r9, r27
add r8, r10, r27
b .LBB_test_9 ; loopexit
.LBB_test_7: ; shortcirc_next.2
add r28, r3, r27
lhz r26, 9(r28)
add r28, r4, r27
lhz r25, 9(r28)
addi r28, r27, 8
cmpw cr7, r26, r25
mfcr r26, 1
rlwinm r26, r26, 31, 31, 31
add r25, r8, r27
cmpw cr7, r25, r7
mfcr r25, 1
rlwinm r25, r25, 29, 31, 31
and. r26, r26, r25
bne .LBB_test_1 ; loopentry
Next up, improve the code for the loop.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22769 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-12 22:22:17 +00:00
|
|
|
Loop *L, Pass *P) {
|
2005-08-04 20:03:32 +00:00
|
|
|
if (!isa<PHINode>(Inst)) {
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
SCEVHandle NewValSCEV = SCEVAddExpr::get(NewBase, Imm);
|
2005-08-04 20:03:32 +00:00
|
|
|
Value *NewVal = Rewriter.expandCodeFor(NewValSCEV, Inst,
|
|
|
|
OperandValToReplace->getType());
|
|
|
|
// Replace the use of the operand Value with the new Phi we just created.
|
|
|
|
Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
|
|
|
|
DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
// PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
|
2005-08-10 00:35:32 +00:00
|
|
|
// expression into each operand block that uses it. Note that PHI nodes can
|
|
|
|
// have multiple entries for the same predecessor. We use a map to make sure
|
|
|
|
// that a PHI node only has a single Value* for each predecessor (which also
|
|
|
|
// prevents us from inserting duplicate code in some blocks).
|
|
|
|
std::map<BasicBlock*, Value*> InsertedCode;
|
2005-08-04 20:03:32 +00:00
|
|
|
PHINode *PN = cast<PHINode>(Inst);
|
|
|
|
for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
|
|
|
|
if (PN->getIncomingValue(i) == OperandValToReplace) {
|
Fix a FIXME: if we are inserting code for a PHI argument, split the critical
edge so that the code is not always executed for both operands. This
prevents LSR from inserting code into loops whose exit blocks contain
PHI uses of IV expressions (which are outside of loops). On gzip, for
example, we turn this ugly code:
.LBB_test_1: ; loopentry
add r27, r3, r28
lhz r27, 3(r27)
add r26, r4, r28
lhz r26, 3(r26)
add r25, r30, r28 ;; Only live if exiting the loop
add r24, r29, r28 ;; Only live if exiting the loop
cmpw cr0, r27, r26
bne .LBB_test_5 ; loopexit
into this:
.LBB_test_1: ; loopentry
or r27, r28, r28
add r28, r3, r27
lhz r28, 3(r28)
add r26, r4, r27
lhz r26, 3(r26)
cmpw cr0, r28, r26
beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2: ; loopentry.loopexit_crit_edge
add r2, r30, r27
add r8, r29, r27
b .LBB_test_9 ; loopexit
.LBB_test_2: ; shortcirc_next.0
...
blt .LBB_test_1
into this:
.LBB_test_1: ; loopentry
or r27, r28, r28
add r28, r3, r27
lhz r28, 3(r28)
add r26, r4, r27
lhz r26, 3(r26)
cmpw cr0, r28, r26
beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2: ; loopentry.loopexit_crit_edge
add r2, r30, r27
add r8, r29, r27
b .LBB_t_3: ; shortcirc_next.0
.LBB_test_3: ; shortcirc_next.0
...
blt .LBB_test_1
Next step: get the block out of the loop so that the loop is all
fall-throughs again.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22766 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-12 22:06:11 +00:00
|
|
|
// If this is a critical edge, split the edge so that we do not insert the
|
2005-09-13 02:09:55 +00:00
|
|
|
// code on all predecessor/successor paths. We do this unless this is the
|
|
|
|
// canonical backedge for this loop, as this can make some inserted code
|
|
|
|
// be in an illegal position.
|
2005-10-03 00:31:52 +00:00
|
|
|
BasicBlock *PHIPred = PN->getIncomingBlock(i);
|
|
|
|
if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
|
|
|
|
(PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
|
|
|
|
|
2005-08-17 06:35:16 +00:00
|
|
|
// First step, split the critical edge.
|
2005-10-03 00:31:52 +00:00
|
|
|
SplitCriticalEdge(PHIPred, PN->getParent(), P);
|
When splitting critical edges, make sure not to leave the new block in the
middle of the loop. This turns a critical loop in gzip into this:
.LBB_test_1: ; loopentry
or r27, r28, r28
add r28, r3, r27
lhz r28, 3(r28)
add r26, r4, r27
lhz r26, 3(r26)
cmpw cr0, r28, r26
bne .LBB_test_8 ; loopentry.loopexit_crit_edge
.LBB_test_2: ; shortcirc_next.0
add r28, r3, r27
lhz r28, 5(r28)
add r26, r4, r27
lhz r26, 5(r26)
cmpw cr0, r28, r26
bne .LBB_test_7 ; shortcirc_next.0.loopexit_crit_edge
.LBB_test_3: ; shortcirc_next.1
add r28, r3, r27
lhz r28, 7(r28)
add r26, r4, r27
lhz r26, 7(r26)
cmpw cr0, r28, r26
bne .LBB_test_6 ; shortcirc_next.1.loopexit_crit_edge
.LBB_test_4: ; shortcirc_next.2
add r28, r3, r27
lhz r26, 9(r28)
add r28, r4, r27
lhz r25, 9(r28)
addi r28, r27, 8
cmpw cr7, r26, r25
mfcr r26, 1
rlwinm r26, r26, 31, 31, 31
add r25, r8, r27
cmpw cr7, r25, r7
mfcr r25, 1
rlwinm r25, r25, 29, 31, 31
and. r26, r26, r25
bne .LBB_test_1 ; loopentry
instead of this:
.LBB_test_1: ; loopentry
or r27, r28, r28
add r28, r3, r27
lhz r28, 3(r28)
add r26, r4, r27
lhz r26, 3(r26)
cmpw cr0, r28, r26
beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2: ; loopentry.loopexit_crit_edge
add r2, r30, r27
add r8, r29, r27
b .LBB_test_9 ; loopexit
.LBB_test_3: ; shortcirc_next.0
add r28, r3, r27
lhz r28, 5(r28)
add r26, r4, r27
lhz r26, 5(r26)
cmpw cr0, r28, r26
beq .LBB_test_5 ; shortcirc_next.1
.LBB_test_4: ; shortcirc_next.0.loopexit_crit_edge
add r2, r11, r27
add r8, r12, r27
b .LBB_test_9 ; loopexit
.LBB_test_5: ; shortcirc_next.1
add r28, r3, r27
lhz r28, 7(r28)
add r26, r4, r27
lhz r26, 7(r26)
cmpw cr0, r28, r26
beq .LBB_test_7 ; shortcirc_next.2
.LBB_test_6: ; shortcirc_next.1.loopexit_crit_edge
add r2, r9, r27
add r8, r10, r27
b .LBB_test_9 ; loopexit
.LBB_test_7: ; shortcirc_next.2
add r28, r3, r27
lhz r26, 9(r28)
add r28, r4, r27
lhz r25, 9(r28)
addi r28, r27, 8
cmpw cr7, r26, r25
mfcr r26, 1
rlwinm r26, r26, 31, 31, 31
add r25, r8, r27
cmpw cr7, r25, r7
mfcr r25, 1
rlwinm r25, r25, 29, 31, 31
and. r26, r26, r25
bne .LBB_test_1 ; loopentry
Next up, improve the code for the loop.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22769 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-12 22:22:17 +00:00
|
|
|
|
2005-08-17 06:35:16 +00:00
|
|
|
// Next step: move the basic block. In particular, if the PHI node
|
|
|
|
// is outside of the loop, and PredTI is in the loop, we want to
|
|
|
|
// move the block to be immediately before the PHI block, not
|
|
|
|
// immediately after PredTI.
|
2005-10-03 00:31:52 +00:00
|
|
|
if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
|
2005-08-17 06:35:16 +00:00
|
|
|
BasicBlock *NewBB = PN->getIncomingBlock(i);
|
|
|
|
NewBB->moveBefore(PN->getParent());
|
Fix a FIXME: if we are inserting code for a PHI argument, split the critical
edge so that the code is not always executed for both operands. This
prevents LSR from inserting code into loops whose exit blocks contain
PHI uses of IV expressions (which are outside of loops). On gzip, for
example, we turn this ugly code:
.LBB_test_1: ; loopentry
add r27, r3, r28
lhz r27, 3(r27)
add r26, r4, r28
lhz r26, 3(r26)
add r25, r30, r28 ;; Only live if exiting the loop
add r24, r29, r28 ;; Only live if exiting the loop
cmpw cr0, r27, r26
bne .LBB_test_5 ; loopexit
into this:
.LBB_test_1: ; loopentry
or r27, r28, r28
add r28, r3, r27
lhz r28, 3(r28)
add r26, r4, r27
lhz r26, 3(r26)
cmpw cr0, r28, r26
beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2: ; loopentry.loopexit_crit_edge
add r2, r30, r27
add r8, r29, r27
b .LBB_test_9 ; loopexit
.LBB_test_2: ; shortcirc_next.0
...
blt .LBB_test_1
into this:
.LBB_test_1: ; loopentry
or r27, r28, r28
add r28, r3, r27
lhz r28, 3(r28)
add r26, r4, r27
lhz r26, 3(r26)
cmpw cr0, r28, r26
beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2: ; loopentry.loopexit_crit_edge
add r2, r30, r27
add r8, r29, r27
b .LBB_t_3: ; shortcirc_next.0
.LBB_test_3: ; shortcirc_next.0
...
blt .LBB_test_1
Next step: get the block out of the loop so that the loop is all
fall-throughs again.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22766 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-12 22:06:11 +00:00
|
|
|
}
|
|
|
|
}
|
2005-08-04 20:03:32 +00:00
|
|
|
|
2005-08-10 00:35:32 +00:00
|
|
|
Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
|
|
|
|
if (!Code) {
|
|
|
|
// Insert the code into the end of the predecessor block.
|
|
|
|
BasicBlock::iterator InsertPt =PN->getIncomingBlock(i)->getTerminator();
|
2005-08-04 20:03:32 +00:00
|
|
|
|
2005-08-10 00:35:32 +00:00
|
|
|
SCEVHandle NewValSCEV = SCEVAddExpr::get(NewBase, Imm);
|
|
|
|
Code = Rewriter.expandCodeFor(NewValSCEV, InsertPt,
|
|
|
|
OperandValToReplace->getType());
|
|
|
|
}
|
2005-08-04 20:03:32 +00:00
|
|
|
|
|
|
|
// Replace the use of the operand Value with the new Phi we just created.
|
2005-08-10 00:35:32 +00:00
|
|
|
PN->setIncomingValue(i, Code);
|
2005-08-04 20:03:32 +00:00
|
|
|
Rewriter.clear();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2005-07-30 00:15:07 +00:00
|
|
|
/// isTargetConstant - Return true if the following can be referenced by the
|
|
|
|
/// immediate field of a target instruction.
|
|
|
|
static bool isTargetConstant(const SCEVHandle &V) {
|
2005-07-30 18:33:25 +00:00
|
|
|
|
2005-07-30 00:15:07 +00:00
|
|
|
// FIXME: Look at the target to decide if &GV is a legal constant immediate.
|
2005-08-08 06:25:50 +00:00
|
|
|
if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
|
|
|
|
// PPC allows a sign-extended 16-bit immediate field.
|
|
|
|
if ((int64_t)SC->getValue()->getRawValue() > -(1 << 16) &&
|
|
|
|
(int64_t)SC->getValue()->getRawValue() < (1 << 16)-1)
|
|
|
|
return true;
|
|
|
|
return false;
|
|
|
|
}
|
2005-07-30 18:33:25 +00:00
|
|
|
|
2005-07-30 00:15:07 +00:00
|
|
|
return false; // ENABLE this for x86
|
2005-07-30 18:33:25 +00:00
|
|
|
|
2005-07-30 00:15:07 +00:00
|
|
|
if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
|
|
|
|
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
|
|
|
|
if (CE->getOpcode() == Instruction::Cast)
|
|
|
|
if (isa<GlobalValue>(CE->getOperand(0)))
|
|
|
|
// FIXME: should check to see that the dest is uintptr_t!
|
|
|
|
return true;
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2005-08-08 22:32:34 +00:00
|
|
|
/// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
|
|
|
|
/// loop varying to the Imm operand.
|
|
|
|
static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
|
|
|
|
Loop *L) {
|
|
|
|
if (Val->isLoopInvariant(L)) return; // Nothing to do.
|
|
|
|
|
|
|
|
if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
|
|
|
|
std::vector<SCEVHandle> NewOps;
|
|
|
|
NewOps.reserve(SAE->getNumOperands());
|
|
|
|
|
|
|
|
for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
|
|
|
|
if (!SAE->getOperand(i)->isLoopInvariant(L)) {
|
|
|
|
// If this is a loop-variant expression, it must stay in the immediate
|
|
|
|
// field of the expression.
|
|
|
|
Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
|
|
|
|
} else {
|
|
|
|
NewOps.push_back(SAE->getOperand(i));
|
|
|
|
}
|
|
|
|
|
|
|
|
if (NewOps.empty())
|
|
|
|
Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
|
|
|
|
else
|
|
|
|
Val = SCEVAddExpr::get(NewOps);
|
|
|
|
} else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
|
|
|
|
// Try to pull immediates out of the start value of nested addrec's.
|
|
|
|
SCEVHandle Start = SARE->getStart();
|
|
|
|
MoveLoopVariantsToImediateField(Start, Imm, L);
|
|
|
|
|
|
|
|
std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
|
|
|
|
Ops[0] = Start;
|
|
|
|
Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
|
|
|
|
} else {
|
|
|
|
// Otherwise, all of Val is variant, move the whole thing over.
|
|
|
|
Imm = SCEVAddExpr::get(Imm, Val);
|
|
|
|
Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2005-08-04 22:34:05 +00:00
|
|
|
/// MoveImmediateValues - Look at Val, and pull out any additions of constants
|
2005-07-30 00:15:07 +00:00
|
|
|
/// that can fit into the immediate field of instructions in the target.
|
2005-08-04 22:34:05 +00:00
|
|
|
/// Accumulate these immediate values into the Imm value.
|
|
|
|
static void MoveImmediateValues(SCEVHandle &Val, SCEVHandle &Imm,
|
|
|
|
bool isAddress, Loop *L) {
|
Teach loop-reduce to see into nested loops, to pull out immediate values
pushed down by SCEV.
In a nested loop case, this allows us to emit this:
lis r3, ha16(L_A$non_lazy_ptr)
lwz r3, lo16(L_A$non_lazy_ptr)(r3)
add r2, r2, r3
li r3, 1
.LBB_foo_2: ; no_exit.1
lfd f0, 8(r2) ;; Uses offset of 8 instead of 0
stfd f0, 0(r2)
addi r4, r3, 1
addi r2, r2, 8
cmpwi cr0, r3, 100
or r3, r4, r4
bne .LBB_foo_2 ; no_exit.1
instead of this:
lis r3, ha16(L_A$non_lazy_ptr)
lwz r3, lo16(L_A$non_lazy_ptr)(r3)
add r2, r2, r3
addi r3, r3, 8
li r4, 1
.LBB_foo_2: ; no_exit.1
lfd f0, 0(r3)
stfd f0, 0(r2)
addi r5, r4, 1
addi r2, r2, 8
addi r3, r3, 8
cmpwi cr0, r4, 100
or r4, r5, r5
bne .LBB_foo_2 ; no_exit.1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22639 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-03 23:44:42 +00:00
|
|
|
if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
|
2005-08-04 22:34:05 +00:00
|
|
|
std::vector<SCEVHandle> NewOps;
|
|
|
|
NewOps.reserve(SAE->getNumOperands());
|
|
|
|
|
|
|
|
for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
|
2005-08-04 19:08:16 +00:00
|
|
|
if (isAddress && isTargetConstant(SAE->getOperand(i))) {
|
|
|
|
Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
|
|
|
|
} else if (!SAE->getOperand(i)->isLoopInvariant(L)) {
|
|
|
|
// If this is a loop-variant expression, it must stay in the immediate
|
|
|
|
// field of the expression.
|
|
|
|
Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
|
2005-08-04 22:34:05 +00:00
|
|
|
} else {
|
|
|
|
NewOps.push_back(SAE->getOperand(i));
|
2005-07-30 00:15:07 +00:00
|
|
|
}
|
2005-08-04 22:34:05 +00:00
|
|
|
|
|
|
|
if (NewOps.empty())
|
|
|
|
Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
|
|
|
|
else
|
|
|
|
Val = SCEVAddExpr::get(NewOps);
|
|
|
|
return;
|
Teach loop-reduce to see into nested loops, to pull out immediate values
pushed down by SCEV.
In a nested loop case, this allows us to emit this:
lis r3, ha16(L_A$non_lazy_ptr)
lwz r3, lo16(L_A$non_lazy_ptr)(r3)
add r2, r2, r3
li r3, 1
.LBB_foo_2: ; no_exit.1
lfd f0, 8(r2) ;; Uses offset of 8 instead of 0
stfd f0, 0(r2)
addi r4, r3, 1
addi r2, r2, 8
cmpwi cr0, r3, 100
or r3, r4, r4
bne .LBB_foo_2 ; no_exit.1
instead of this:
lis r3, ha16(L_A$non_lazy_ptr)
lwz r3, lo16(L_A$non_lazy_ptr)(r3)
add r2, r2, r3
addi r3, r3, 8
li r4, 1
.LBB_foo_2: ; no_exit.1
lfd f0, 0(r3)
stfd f0, 0(r2)
addi r5, r4, 1
addi r2, r2, 8
addi r3, r3, 8
cmpwi cr0, r4, 100
or r4, r5, r5
bne .LBB_foo_2 ; no_exit.1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22639 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-03 23:44:42 +00:00
|
|
|
} else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
|
|
|
|
// Try to pull immediates out of the start value of nested addrec's.
|
2005-08-04 22:34:05 +00:00
|
|
|
SCEVHandle Start = SARE->getStart();
|
|
|
|
MoveImmediateValues(Start, Imm, isAddress, L);
|
|
|
|
|
|
|
|
if (Start != SARE->getStart()) {
|
|
|
|
std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
|
|
|
|
Ops[0] = Start;
|
|
|
|
Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
|
|
|
|
}
|
|
|
|
return;
|
2005-07-30 00:15:07 +00:00
|
|
|
}
|
|
|
|
|
2005-08-04 22:34:05 +00:00
|
|
|
// Loop-variant expressions must stay in the immediate field of the
|
|
|
|
// expression.
|
|
|
|
if ((isAddress && isTargetConstant(Val)) ||
|
|
|
|
!Val->isLoopInvariant(L)) {
|
|
|
|
Imm = SCEVAddExpr::get(Imm, Val);
|
|
|
|
Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
|
|
|
|
return;
|
2005-08-04 19:26:19 +00:00
|
|
|
}
|
2005-08-04 22:34:05 +00:00
|
|
|
|
|
|
|
// Otherwise, no immediates to move.
|
2005-07-30 00:15:07 +00:00
|
|
|
}
|
|
|
|
|
Recursively scan scev expressions for common subexpressions. This allows us
to handle nested loops much better, for example, by being able to tell that
these two expressions:
{( 8 + ( 16 * ( 1 + %Tmp11 + %Tmp12)) + %c_),+,( 16 * %Tmp 12)}<loopentry.1>
{(( 16 * ( 1 + %Tmp11 + %Tmp12)) + %c_),+,( 16 * %Tmp12)}<loopentry.1>
Have the following common part that can be shared:
{(( 16 * ( 1 + %Tmp11 + %Tmp12)) + %c_),+,( 16 * %Tmp12)}<loopentry.1>
This allows us to codegen an important inner loop in 168.wupwise as:
.LBB_foo_4: ; no_exit.1
lfd f2, 16(r9)
fmul f3, f0, f2
fmul f2, f1, f2
fadd f4, f3, f2
stfd f4, 8(r9)
fsub f2, f3, f2
stfd f2, 16(r9)
addi r8, r8, 1
addi r9, r9, 16
cmpw cr0, r8, r4
ble .LBB_foo_4 ; no_exit.1
instead of:
.LBB_foo_3: ; no_exit.1
lfdx f2, r6, r9
add r10, r6, r9
lfd f3, 8(r10)
fmul f4, f1, f2
fmadd f4, f0, f3, f4
stfd f4, 8(r10)
fmul f3, f1, f3
fmsub f2, f0, f2, f3
stfdx f2, r6, r9
addi r9, r9, 16
addi r8, r8, 1
cmpw cr0, r8, r4
ble .LBB_foo_3 ; no_exit.1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22781 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-13 07:27:18 +00:00
|
|
|
|
|
|
|
/// IncrementAddExprUses - Decompose the specified expression into its added
|
|
|
|
/// subexpressions, and increment SubExpressionUseCounts for each of these
|
|
|
|
/// decomposed parts.
|
|
|
|
static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
|
|
|
|
SCEVHandle Expr) {
|
|
|
|
if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
|
|
|
|
for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
|
|
|
|
SeparateSubExprs(SubExprs, AE->getOperand(j));
|
|
|
|
} else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
|
|
|
|
SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Expr->getType());
|
|
|
|
if (SARE->getOperand(0) == Zero) {
|
|
|
|
SubExprs.push_back(Expr);
|
|
|
|
} else {
|
|
|
|
// Compute the addrec with zero as its base.
|
|
|
|
std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
|
|
|
|
Ops[0] = Zero; // Start with zero base.
|
|
|
|
SubExprs.push_back(SCEVAddRecExpr::get(Ops, SARE->getLoop()));
|
|
|
|
|
|
|
|
|
|
|
|
SeparateSubExprs(SubExprs, SARE->getOperand(0));
|
|
|
|
}
|
|
|
|
} else if (!isa<SCEVConstant>(Expr) ||
|
|
|
|
!cast<SCEVConstant>(Expr)->getValue()->isNullValue()) {
|
|
|
|
// Do not add zero.
|
|
|
|
SubExprs.push_back(Expr);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
/// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
|
|
|
|
/// removing any common subexpressions from it. Anything truly common is
|
|
|
|
/// removed, accumulated, and returned. This looks for things like (a+b+c) and
|
|
|
|
/// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
|
|
|
|
static SCEVHandle
|
|
|
|
RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses) {
|
|
|
|
unsigned NumUses = Uses.size();
|
|
|
|
|
|
|
|
// Only one use? Use its base, regardless of what it is!
|
|
|
|
SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Uses[0].Base->getType());
|
|
|
|
SCEVHandle Result = Zero;
|
|
|
|
if (NumUses == 1) {
|
|
|
|
std::swap(Result, Uses[0].Base);
|
|
|
|
return Result;
|
|
|
|
}
|
|
|
|
|
|
|
|
// To find common subexpressions, count how many of Uses use each expression.
|
|
|
|
// If any subexpressions are used Uses.size() times, they are common.
|
|
|
|
std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
|
|
|
|
|
Recursively scan scev expressions for common subexpressions. This allows us
to handle nested loops much better, for example, by being able to tell that
these two expressions:
{( 8 + ( 16 * ( 1 + %Tmp11 + %Tmp12)) + %c_),+,( 16 * %Tmp 12)}<loopentry.1>
{(( 16 * ( 1 + %Tmp11 + %Tmp12)) + %c_),+,( 16 * %Tmp12)}<loopentry.1>
Have the following common part that can be shared:
{(( 16 * ( 1 + %Tmp11 + %Tmp12)) + %c_),+,( 16 * %Tmp12)}<loopentry.1>
This allows us to codegen an important inner loop in 168.wupwise as:
.LBB_foo_4: ; no_exit.1
lfd f2, 16(r9)
fmul f3, f0, f2
fmul f2, f1, f2
fadd f4, f3, f2
stfd f4, 8(r9)
fsub f2, f3, f2
stfd f2, 16(r9)
addi r8, r8, 1
addi r9, r9, 16
cmpw cr0, r8, r4
ble .LBB_foo_4 ; no_exit.1
instead of:
.LBB_foo_3: ; no_exit.1
lfdx f2, r6, r9
add r10, r6, r9
lfd f3, 8(r10)
fmul f4, f1, f2
fmadd f4, f0, f3, f4
stfd f4, 8(r10)
fmul f3, f1, f3
fmsub f2, f0, f2, f3
stfdx f2, r6, r9
addi r9, r9, 16
addi r8, r8, 1
cmpw cr0, r8, r4
ble .LBB_foo_3 ; no_exit.1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22781 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-13 07:27:18 +00:00
|
|
|
std::vector<SCEVHandle> SubExprs;
|
|
|
|
for (unsigned i = 0; i != NumUses; ++i) {
|
|
|
|
// If the base is zero (which is common), return zero now, there are no
|
|
|
|
// CSEs we can find.
|
|
|
|
if (Uses[i].Base == Zero) return Zero;
|
|
|
|
|
|
|
|
// Split the expression into subexprs.
|
|
|
|
SeparateSubExprs(SubExprs, Uses[i].Base);
|
|
|
|
// Add one to SubExpressionUseCounts for each subexpr present.
|
|
|
|
for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
|
|
|
|
SubExpressionUseCounts[SubExprs[j]]++;
|
|
|
|
SubExprs.clear();
|
|
|
|
}
|
|
|
|
|
|
|
|
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
// Now that we know how many times each is used, build Result.
|
|
|
|
for (std::map<SCEVHandle, unsigned>::iterator I =
|
|
|
|
SubExpressionUseCounts.begin(), E = SubExpressionUseCounts.end();
|
|
|
|
I != E; )
|
|
|
|
if (I->second == NumUses) { // Found CSE!
|
|
|
|
Result = SCEVAddExpr::get(Result, I->first);
|
|
|
|
++I;
|
|
|
|
} else {
|
|
|
|
// Remove non-cse's from SubExpressionUseCounts.
|
|
|
|
SubExpressionUseCounts.erase(I++);
|
|
|
|
}
|
|
|
|
|
|
|
|
// If we found no CSE's, return now.
|
|
|
|
if (Result == Zero) return Result;
|
|
|
|
|
|
|
|
// Otherwise, remove all of the CSE's we found from each of the base values.
|
Recursively scan scev expressions for common subexpressions. This allows us
to handle nested loops much better, for example, by being able to tell that
these two expressions:
{( 8 + ( 16 * ( 1 + %Tmp11 + %Tmp12)) + %c_),+,( 16 * %Tmp 12)}<loopentry.1>
{(( 16 * ( 1 + %Tmp11 + %Tmp12)) + %c_),+,( 16 * %Tmp12)}<loopentry.1>
Have the following common part that can be shared:
{(( 16 * ( 1 + %Tmp11 + %Tmp12)) + %c_),+,( 16 * %Tmp12)}<loopentry.1>
This allows us to codegen an important inner loop in 168.wupwise as:
.LBB_foo_4: ; no_exit.1
lfd f2, 16(r9)
fmul f3, f0, f2
fmul f2, f1, f2
fadd f4, f3, f2
stfd f4, 8(r9)
fsub f2, f3, f2
stfd f2, 16(r9)
addi r8, r8, 1
addi r9, r9, 16
cmpw cr0, r8, r4
ble .LBB_foo_4 ; no_exit.1
instead of:
.LBB_foo_3: ; no_exit.1
lfdx f2, r6, r9
add r10, r6, r9
lfd f3, 8(r10)
fmul f4, f1, f2
fmadd f4, f0, f3, f4
stfd f4, 8(r10)
fmul f3, f1, f3
fmsub f2, f0, f2, f3
stfdx f2, r6, r9
addi r9, r9, 16
addi r8, r8, 1
cmpw cr0, r8, r4
ble .LBB_foo_3 ; no_exit.1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22781 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-13 07:27:18 +00:00
|
|
|
for (unsigned i = 0; i != NumUses; ++i) {
|
|
|
|
// Split the expression into subexprs.
|
|
|
|
SeparateSubExprs(SubExprs, Uses[i].Base);
|
|
|
|
|
|
|
|
// Remove any common subexpressions.
|
|
|
|
for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
|
|
|
|
if (SubExpressionUseCounts.count(SubExprs[j])) {
|
|
|
|
SubExprs.erase(SubExprs.begin()+j);
|
|
|
|
--j; --e;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Finally, the non-shared expressions together.
|
|
|
|
if (SubExprs.empty())
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
Uses[i].Base = Zero;
|
Recursively scan scev expressions for common subexpressions. This allows us
to handle nested loops much better, for example, by being able to tell that
these two expressions:
{( 8 + ( 16 * ( 1 + %Tmp11 + %Tmp12)) + %c_),+,( 16 * %Tmp 12)}<loopentry.1>
{(( 16 * ( 1 + %Tmp11 + %Tmp12)) + %c_),+,( 16 * %Tmp12)}<loopentry.1>
Have the following common part that can be shared:
{(( 16 * ( 1 + %Tmp11 + %Tmp12)) + %c_),+,( 16 * %Tmp12)}<loopentry.1>
This allows us to codegen an important inner loop in 168.wupwise as:
.LBB_foo_4: ; no_exit.1
lfd f2, 16(r9)
fmul f3, f0, f2
fmul f2, f1, f2
fadd f4, f3, f2
stfd f4, 8(r9)
fsub f2, f3, f2
stfd f2, 16(r9)
addi r8, r8, 1
addi r9, r9, 16
cmpw cr0, r8, r4
ble .LBB_foo_4 ; no_exit.1
instead of:
.LBB_foo_3: ; no_exit.1
lfdx f2, r6, r9
add r10, r6, r9
lfd f3, 8(r10)
fmul f4, f1, f2
fmadd f4, f0, f3, f4
stfd f4, 8(r10)
fmul f3, f1, f3
fmsub f2, f0, f2, f3
stfdx f2, r6, r9
addi r9, r9, 16
addi r8, r8, 1
cmpw cr0, r8, r4
ble .LBB_foo_3 ; no_exit.1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22781 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-13 07:27:18 +00:00
|
|
|
else
|
|
|
|
Uses[i].Base = SCEVAddExpr::get(SubExprs);
|
Ooops, don't forget to clear this. The real inner loop is now:
.LBB_foo_3: ; no_exit.1
lfd f2, 0(r9)
lfd f3, 8(r9)
fmul f4, f1, f2
fmadd f4, f0, f3, f4
stfd f4, 8(r9)
fmul f3, f1, f3
fmsub f2, f0, f2, f3
stfd f2, 0(r9)
addi r9, r9, 16
addi r8, r8, 1
cmpw cr0, r8, r4
ble .LBB_foo_3 ; no_exit.1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22782 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-13 07:42:01 +00:00
|
|
|
SubExprs.clear();
|
Recursively scan scev expressions for common subexpressions. This allows us
to handle nested loops much better, for example, by being able to tell that
these two expressions:
{( 8 + ( 16 * ( 1 + %Tmp11 + %Tmp12)) + %c_),+,( 16 * %Tmp 12)}<loopentry.1>
{(( 16 * ( 1 + %Tmp11 + %Tmp12)) + %c_),+,( 16 * %Tmp12)}<loopentry.1>
Have the following common part that can be shared:
{(( 16 * ( 1 + %Tmp11 + %Tmp12)) + %c_),+,( 16 * %Tmp12)}<loopentry.1>
This allows us to codegen an important inner loop in 168.wupwise as:
.LBB_foo_4: ; no_exit.1
lfd f2, 16(r9)
fmul f3, f0, f2
fmul f2, f1, f2
fadd f4, f3, f2
stfd f4, 8(r9)
fsub f2, f3, f2
stfd f2, 16(r9)
addi r8, r8, 1
addi r9, r9, 16
cmpw cr0, r8, r4
ble .LBB_foo_4 ; no_exit.1
instead of:
.LBB_foo_3: ; no_exit.1
lfdx f2, r6, r9
add r10, r6, r9
lfd f3, 8(r10)
fmul f4, f1, f2
fmadd f4, f0, f3, f4
stfd f4, 8(r10)
fmul f3, f1, f3
fmsub f2, f0, f2, f3
stfdx f2, r6, r9
addi r9, r9, 16
addi r8, r8, 1
cmpw cr0, r8, r4
ble .LBB_foo_3 ; no_exit.1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22781 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-13 07:27:18 +00:00
|
|
|
}
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
|
|
|
|
return Result;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2005-07-30 00:15:07 +00:00
|
|
|
/// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
|
|
|
|
/// stride of IV. All of the users may have different starting values, and this
|
|
|
|
/// may not be the only stride (we know it is if isOnlyStride is true).
|
Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
For code like this:
void foo(float *a, float *b, int n, int stride_a, int stride_b) {
int i;
for (i=0; i<n; i++)
a[i*stride_a] = b[i*stride_b];
}
we now emit:
.LBB_foo2_2: ; no_exit
lfs f0, 0(r4)
stfs f0, 0(r3)
addi r7, r7, 1
add r4, r2, r4
add r3, r6, r3
cmpw cr0, r7, r5
blt .LBB_foo2_2 ; no_exit
instead of:
.LBB_foo_2: ; no_exit
mullw r8, r2, r7 ;; multiply!
slwi r8, r8, 2
lfsx f0, r4, r8
mullw r8, r2, r6 ;; multiply!
slwi r8, r8, 2
stfsx f0, r3, r8
addi r2, r2, 1
cmpw cr0, r2, r5
blt .LBB_foo_2 ; no_exit
loops with variable strides occur pretty often. For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.
Now we can allow indvars to turn functions written like this:
void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
int i, ai = 0, bi = 0;
for (i=0; i<n; i++)
{
a[ai] = b[bi];
ai += stride_a;
bi += stride_b;
}
}
into code like the above for better analysis. With this patch, they generate
identical code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22740 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-10 00:45:21 +00:00
|
|
|
void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
|
2005-08-03 22:21:05 +00:00
|
|
|
IVUsersOfOneStride &Uses,
|
|
|
|
Loop *L,
|
2005-07-30 00:15:07 +00:00
|
|
|
bool isOnlyStride) {
|
|
|
|
// Transform our list of users and offsets to a bit more complex table. In
|
2005-08-08 22:56:21 +00:00
|
|
|
// this new vector, each 'BasedUser' contains 'Base' the base of the
|
|
|
|
// strided accessas well as the old information from Uses. We progressively
|
|
|
|
// move information from the Base field to the Imm field, until we eventually
|
|
|
|
// have the full access expression to rewrite the use.
|
|
|
|
std::vector<BasedUser> UsersToProcess;
|
2005-07-30 00:15:07 +00:00
|
|
|
UsersToProcess.reserve(Uses.Users.size());
|
2005-08-08 22:56:21 +00:00
|
|
|
for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
|
|
|
|
UsersToProcess.push_back(Uses.Users[i]);
|
|
|
|
|
|
|
|
// Move any loop invariant operands from the offset field to the immediate
|
|
|
|
// field of the use, so that we don't try to use something before it is
|
|
|
|
// computed.
|
|
|
|
MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
|
|
|
|
UsersToProcess.back().Imm, L);
|
|
|
|
assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
|
2005-08-04 22:34:05 +00:00
|
|
|
"Base value is not loop invariant!");
|
2005-03-06 21:58:22 +00:00
|
|
|
}
|
2005-08-08 22:32:34 +00:00
|
|
|
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
// We now have a whole bunch of uses of like-strided induction variables, but
|
|
|
|
// they might all have different bases. We want to emit one PHI node for this
|
|
|
|
// stride which we fold as many common expressions (between the IVs) into as
|
|
|
|
// possible. Start by identifying the common expressions in the base values
|
|
|
|
// for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
|
|
|
|
// "A+B"), emit it to the preheader, then remove the expression from the
|
|
|
|
// UsersToProcess base values.
|
|
|
|
SCEVHandle CommonExprs = RemoveCommonExpressionsFromUseBases(UsersToProcess);
|
|
|
|
|
2005-08-08 22:32:34 +00:00
|
|
|
// Next, figure out what we can represent in the immediate fields of
|
|
|
|
// instructions. If we can represent anything there, move it to the imm
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
// fields of the BasedUsers. We do this so that it increases the commonality
|
|
|
|
// of the remaining uses.
|
2005-08-08 22:32:34 +00:00
|
|
|
for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
|
2005-08-16 00:38:11 +00:00
|
|
|
// If the user is not in the current loop, this means it is using the exit
|
|
|
|
// value of the IV. Do not put anything in the base, make sure it's all in
|
|
|
|
// the immediate field to allow as much factoring as possible.
|
|
|
|
if (!L->contains(UsersToProcess[i].Inst->getParent())) {
|
2005-08-17 21:22:41 +00:00
|
|
|
UsersToProcess[i].Imm = SCEVAddExpr::get(UsersToProcess[i].Imm,
|
|
|
|
UsersToProcess[i].Base);
|
|
|
|
UsersToProcess[i].Base =
|
|
|
|
SCEVUnknown::getIntegerSCEV(0, UsersToProcess[i].Base->getType());
|
2005-08-16 00:38:11 +00:00
|
|
|
} else {
|
|
|
|
|
|
|
|
// Addressing modes can be folded into loads and stores. Be careful that
|
|
|
|
// the store is through the expression, not of the expression though.
|
|
|
|
bool isAddress = isa<LoadInst>(UsersToProcess[i].Inst);
|
|
|
|
if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
|
|
|
|
if (SI->getOperand(1) == UsersToProcess[i].OperandValToReplace)
|
|
|
|
isAddress = true;
|
|
|
|
|
|
|
|
MoveImmediateValues(UsersToProcess[i].Base, UsersToProcess[i].Imm,
|
|
|
|
isAddress, L);
|
|
|
|
}
|
2005-08-08 22:32:34 +00:00
|
|
|
}
|
|
|
|
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
// Now that we know what we need to do, insert the PHI node itself.
|
|
|
|
//
|
|
|
|
DEBUG(std::cerr << "INSERTING IV of STRIDE " << *Stride << " and BASE "
|
|
|
|
<< *CommonExprs << " :\n");
|
|
|
|
|
|
|
|
SCEVExpander Rewriter(*SE, *LI);
|
|
|
|
SCEVExpander PreheaderRewriter(*SE, *LI);
|
|
|
|
|
|
|
|
BasicBlock *Preheader = L->getLoopPreheader();
|
|
|
|
Instruction *PreInsertPt = Preheader->getTerminator();
|
|
|
|
Instruction *PhiInsertBefore = L->getHeader()->begin();
|
|
|
|
|
2005-09-12 17:11:27 +00:00
|
|
|
BasicBlock *LatchBlock = L->getLoopLatch();
|
2005-08-08 22:32:34 +00:00
|
|
|
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
// Create a new Phi for this base, and stick it in the loop header.
|
|
|
|
const Type *ReplacedTy = CommonExprs->getType();
|
|
|
|
PHINode *NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
|
|
|
|
++NumInserted;
|
2005-08-08 22:32:34 +00:00
|
|
|
|
Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
For code like this:
void foo(float *a, float *b, int n, int stride_a, int stride_b) {
int i;
for (i=0; i<n; i++)
a[i*stride_a] = b[i*stride_b];
}
we now emit:
.LBB_foo2_2: ; no_exit
lfs f0, 0(r4)
stfs f0, 0(r3)
addi r7, r7, 1
add r4, r2, r4
add r3, r6, r3
cmpw cr0, r7, r5
blt .LBB_foo2_2 ; no_exit
instead of:
.LBB_foo_2: ; no_exit
mullw r8, r2, r7 ;; multiply!
slwi r8, r8, 2
lfsx f0, r4, r8
mullw r8, r2, r6 ;; multiply!
slwi r8, r8, 2
stfsx f0, r3, r8
addi r2, r2, 1
cmpw cr0, r2, r5
blt .LBB_foo_2 ; no_exit
loops with variable strides occur pretty often. For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.
Now we can allow indvars to turn functions written like this:
void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
int i, ai = 0, bi = 0;
for (i=0; i<n; i++)
{
a[ai] = b[bi];
ai += stride_a;
bi += stride_b;
}
}
into code like the above for better analysis. With this patch, they generate
identical code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22740 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-10 00:45:21 +00:00
|
|
|
// Insert the stride into the preheader.
|
|
|
|
Value *StrideV = PreheaderRewriter.expandCodeFor(Stride, PreInsertPt,
|
|
|
|
ReplacedTy);
|
|
|
|
if (!isa<ConstantInt>(StrideV)) ++NumVariable;
|
|
|
|
|
|
|
|
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
// Emit the initial base value into the loop preheader, and add it to the
|
|
|
|
// Phi node.
|
|
|
|
Value *PHIBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt,
|
|
|
|
ReplacedTy);
|
|
|
|
NewPHI->addIncoming(PHIBaseV, Preheader);
|
2005-08-03 23:30:08 +00:00
|
|
|
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
// Emit the increment of the base value before the terminator of the loop
|
|
|
|
// latch block, and add it to the Phi node.
|
|
|
|
SCEVHandle IncExp = SCEVAddExpr::get(SCEVUnknown::get(NewPHI),
|
Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
For code like this:
void foo(float *a, float *b, int n, int stride_a, int stride_b) {
int i;
for (i=0; i<n; i++)
a[i*stride_a] = b[i*stride_b];
}
we now emit:
.LBB_foo2_2: ; no_exit
lfs f0, 0(r4)
stfs f0, 0(r3)
addi r7, r7, 1
add r4, r2, r4
add r3, r6, r3
cmpw cr0, r7, r5
blt .LBB_foo2_2 ; no_exit
instead of:
.LBB_foo_2: ; no_exit
mullw r8, r2, r7 ;; multiply!
slwi r8, r8, 2
lfsx f0, r4, r8
mullw r8, r2, r6 ;; multiply!
slwi r8, r8, 2
stfsx f0, r3, r8
addi r2, r2, 1
cmpw cr0, r2, r5
blt .LBB_foo_2 ; no_exit
loops with variable strides occur pretty often. For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.
Now we can allow indvars to turn functions written like this:
void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
int i, ai = 0, bi = 0;
for (i=0; i<n; i++)
{
a[ai] = b[bi];
ai += stride_a;
bi += stride_b;
}
}
into code like the above for better analysis. With this patch, they generate
identical code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22740 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-10 00:45:21 +00:00
|
|
|
SCEVUnknown::get(StrideV));
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
|
|
|
|
Value *IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator(),
|
|
|
|
ReplacedTy);
|
|
|
|
IncV->setName(NewPHI->getName()+".inc");
|
|
|
|
NewPHI->addIncoming(IncV, LatchBlock);
|
|
|
|
|
Move from Stage 0 to Stage 1.
Only emit one PHI node for IV uses with identical bases and strides (after
moving foldable immediates to the load/store instruction).
This implements LoopStrengthReduce/dont_insert_redundant_ops.ll, allowing
us to generate this PPC code for test1:
or r30, r3, r3
.LBB_test1_1: ; Loop
li r2, 0
stw r2, 0(r30)
stw r2, 4(r30)
bl L_pred$stub
addi r30, r30, 8
cmplwi cr0, r3, 0
bne .LBB_test1_1 ; Loop
instead of this code:
or r30, r3, r3
or r29, r3, r3
.LBB_test1_1: ; Loop
li r2, 0
stw r2, 0(r29)
stw r2, 4(r30)
bl L_pred$stub
addi r30, r30, 8 ;; Two iv's with step of 8
addi r29, r29, 8
cmplwi cr0, r3, 0
bne .LBB_test1_1 ; Loop
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22635 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-03 22:51:21 +00:00
|
|
|
// Sort by the base value, so that all IVs with identical bases are next to
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
// each other.
|
Move from Stage 0 to Stage 1.
Only emit one PHI node for IV uses with identical bases and strides (after
moving foldable immediates to the load/store instruction).
This implements LoopStrengthReduce/dont_insert_redundant_ops.ll, allowing
us to generate this PPC code for test1:
or r30, r3, r3
.LBB_test1_1: ; Loop
li r2, 0
stw r2, 0(r30)
stw r2, 4(r30)
bl L_pred$stub
addi r30, r30, 8
cmplwi cr0, r3, 0
bne .LBB_test1_1 ; Loop
instead of this code:
or r30, r3, r3
or r29, r3, r3
.LBB_test1_1: ; Loop
li r2, 0
stw r2, 0(r29)
stw r2, 4(r30)
bl L_pred$stub
addi r30, r30, 8 ;; Two iv's with step of 8
addi r29, r29, 8
cmplwi cr0, r3, 0
bne .LBB_test1_1 ; Loop
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22635 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-03 22:51:21 +00:00
|
|
|
std::sort(UsersToProcess.begin(), UsersToProcess.end());
|
2005-07-30 00:15:07 +00:00
|
|
|
while (!UsersToProcess.empty()) {
|
2005-08-08 22:56:21 +00:00
|
|
|
SCEVHandle Base = UsersToProcess.front().Base;
|
2005-08-03 23:30:08 +00:00
|
|
|
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
DEBUG(std::cerr << " INSERTING code for BASE = " << *Base << ":\n");
|
2005-08-03 23:30:08 +00:00
|
|
|
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
// Emit the code for Base into the preheader.
|
2005-08-08 05:47:49 +00:00
|
|
|
Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt,
|
|
|
|
ReplacedTy);
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
|
|
|
|
// If BaseV is a constant other than 0, make sure that it gets inserted into
|
|
|
|
// the preheader, instead of being forward substituted into the uses. We do
|
|
|
|
// this by forcing a noop cast to be inserted into the preheader in this
|
|
|
|
// case.
|
|
|
|
if (Constant *C = dyn_cast<Constant>(BaseV))
|
implement Transforms/LoopStrengthReduce/dont-hoist-simple-loop-constants.ll.
We used to emit this code for it:
_test:
li r2, 1 ;; Value tying up a register for the whole loop
li r5, 0
LBB_test_1: ; no_exit.2
or r6, r5, r5
li r5, 0
stw r5, 0(r3)
addi r5, r6, 1
addi r3, r3, 4
add r7, r2, r5 ;; should be addi r7, r5, 1
cmpwi cr0, r7, 701
blt cr0, LBB_test_1 ; no_exit.2
LBB_test_2: ; loopexit.2.loopexit
addi r2, r6, 2
stw r2, 0(r4)
blr
now we emit this:
_test:
li r2, 0
LBB_test_1: ; no_exit.2
or r5, r2, r2
li r2, 0
stw r2, 0(r3)
addi r3, r3, 4
addi r2, r5, 1
addi r6, r5, 2 ;; whoa, fold those adds!
cmpwi cr0, r6, 701
blt cr0, LBB_test_1 ; no_exit.2
LBB_test_2: ; loopexit.2.loopexit
addi r2, r5, 2
stw r2, 0(r4)
blr
more improvement coming.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23306 91177308-0d34-0410-b5e6-96231b3b80d8
2005-09-10 01:18:45 +00:00
|
|
|
if (!C->isNullValue() && !isTargetConstant(Base)) {
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
// We want this constant emitted into the preheader!
|
|
|
|
BaseV = new CastInst(BaseV, BaseV->getType(), "preheaderinsert",
|
|
|
|
PreInsertPt);
|
|
|
|
}
|
|
|
|
|
2005-07-30 00:15:07 +00:00
|
|
|
// Emit the code to add the immediate offset to the Phi value, just before
|
Move from Stage 0 to Stage 1.
Only emit one PHI node for IV uses with identical bases and strides (after
moving foldable immediates to the load/store instruction).
This implements LoopStrengthReduce/dont_insert_redundant_ops.ll, allowing
us to generate this PPC code for test1:
or r30, r3, r3
.LBB_test1_1: ; Loop
li r2, 0
stw r2, 0(r30)
stw r2, 4(r30)
bl L_pred$stub
addi r30, r30, 8
cmplwi cr0, r3, 0
bne .LBB_test1_1 ; Loop
instead of this code:
or r30, r3, r3
or r29, r3, r3
.LBB_test1_1: ; Loop
li r2, 0
stw r2, 0(r29)
stw r2, 4(r30)
bl L_pred$stub
addi r30, r30, 8 ;; Two iv's with step of 8
addi r29, r29, 8
cmplwi cr0, r3, 0
bne .LBB_test1_1 ; Loop
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22635 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-03 22:51:21 +00:00
|
|
|
// the instructions that we identified as using this stride and base.
|
2005-08-08 22:56:21 +00:00
|
|
|
while (!UsersToProcess.empty() && UsersToProcess.front().Base == Base) {
|
|
|
|
BasedUser &User = UsersToProcess.front();
|
Move from Stage 0 to Stage 1.
Only emit one PHI node for IV uses with identical bases and strides (after
moving foldable immediates to the load/store instruction).
This implements LoopStrengthReduce/dont_insert_redundant_ops.ll, allowing
us to generate this PPC code for test1:
or r30, r3, r3
.LBB_test1_1: ; Loop
li r2, 0
stw r2, 0(r30)
stw r2, 4(r30)
bl L_pred$stub
addi r30, r30, 8
cmplwi cr0, r3, 0
bne .LBB_test1_1 ; Loop
instead of this code:
or r30, r3, r3
or r29, r3, r3
.LBB_test1_1: ; Loop
li r2, 0
stw r2, 0(r29)
stw r2, 4(r30)
bl L_pred$stub
addi r30, r30, 8 ;; Two iv's with step of 8
addi r29, r29, 8
cmplwi cr0, r3, 0
bne .LBB_test1_1 ; Loop
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22635 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-03 22:51:21 +00:00
|
|
|
|
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-08 05:28:22 +00:00
|
|
|
// If this instruction wants to use the post-incremented value, move it
|
|
|
|
// after the post-inc and use its value instead of the PHI.
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
Value *RewriteOp = NewPHI;
|
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-08 05:28:22 +00:00
|
|
|
if (User.isUseOfPostIncrementedValue) {
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
RewriteOp = IncV;
|
_test:
li r2, 0
LBB_test_1: ; no_exit.2
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
addi r3, r3, 4
cmpwi cr0, r2, 701
blt cr0, LBB_test_1 ; no_exit.2
LBB_test_2: ; loopexit.2.loopexit
addi r2, r2, 1
stw r2, 0(r4)
blr
[zion ~/llvm]$ cat > ~/xx
Uses of IV's outside of the loop should use hte post-incremented version
of the IV, not the preincremented version. This helps many loops (e.g. in sixtrack)
which used to generate code like this (this is the code from the
dont-hoist-simple-loop-constants.ll testcase):
_test:
li r2, 0 **** IV starts at 0
LBB_test_1: ; no_exit.2
or r5, r2, r2 **** Copy for loop exit
li r2, 0
stw r2, 0(r3)
addi r3, r3, 4
addi r2, r5, 1
addi r6, r5, 2 **** IV+2
cmpwi cr0, r6, 701
blt cr0, LBB_test_1 ; no_exit.2
LBB_test_2: ; loopexit.2.loopexit
addi r2, r5, 2 **** IV+2
stw r2, 0(r4)
blr
And now generated code like this:
_test:
li r2, 1 *** IV starts at 1
LBB_test_1: ; no_exit.2
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
addi r3, r3, 4
cmpwi cr0, r2, 701 *** IV.postinc + 0
blt cr0, LBB_test_1
LBB_test_2: ; loopexit.2.loopexit
stw r2, 0(r4) *** IV.postinc + 0
blr
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23313 91177308-0d34-0410-b5e6-96231b3b80d8
2005-09-12 06:04:47 +00:00
|
|
|
|
|
|
|
// If this user is in the loop, make sure it is the last thing in the
|
|
|
|
// loop to ensure it is dominated by the increment.
|
|
|
|
if (L->contains(User.Inst->getParent()))
|
|
|
|
User.Inst->moveBefore(LatchBlock->getTerminator());
|
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-08 05:28:22 +00:00
|
|
|
}
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
SCEVHandle RewriteExpr = SCEVUnknown::get(RewriteOp);
|
|
|
|
|
|
|
|
// Clear the SCEVExpander's expression map so that we are guaranteed
|
|
|
|
// to have the code emitted where we expect it.
|
|
|
|
Rewriter.clear();
|
|
|
|
|
|
|
|
// Now that we know what we need to do, insert code before User for the
|
|
|
|
// immediate and any loop-variant expressions.
|
|
|
|
if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isNullValue())
|
|
|
|
// Add BaseV to the PHI value if needed.
|
|
|
|
RewriteExpr = SCEVAddExpr::get(RewriteExpr, SCEVUnknown::get(BaseV));
|
|
|
|
|
When splitting critical edges, make sure not to leave the new block in the
middle of the loop. This turns a critical loop in gzip into this:
.LBB_test_1: ; loopentry
or r27, r28, r28
add r28, r3, r27
lhz r28, 3(r28)
add r26, r4, r27
lhz r26, 3(r26)
cmpw cr0, r28, r26
bne .LBB_test_8 ; loopentry.loopexit_crit_edge
.LBB_test_2: ; shortcirc_next.0
add r28, r3, r27
lhz r28, 5(r28)
add r26, r4, r27
lhz r26, 5(r26)
cmpw cr0, r28, r26
bne .LBB_test_7 ; shortcirc_next.0.loopexit_crit_edge
.LBB_test_3: ; shortcirc_next.1
add r28, r3, r27
lhz r28, 7(r28)
add r26, r4, r27
lhz r26, 7(r26)
cmpw cr0, r28, r26
bne .LBB_test_6 ; shortcirc_next.1.loopexit_crit_edge
.LBB_test_4: ; shortcirc_next.2
add r28, r3, r27
lhz r26, 9(r28)
add r28, r4, r27
lhz r25, 9(r28)
addi r28, r27, 8
cmpw cr7, r26, r25
mfcr r26, 1
rlwinm r26, r26, 31, 31, 31
add r25, r8, r27
cmpw cr7, r25, r7
mfcr r25, 1
rlwinm r25, r25, 29, 31, 31
and. r26, r26, r25
bne .LBB_test_1 ; loopentry
instead of this:
.LBB_test_1: ; loopentry
or r27, r28, r28
add r28, r3, r27
lhz r28, 3(r28)
add r26, r4, r27
lhz r26, 3(r26)
cmpw cr0, r28, r26
beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2: ; loopentry.loopexit_crit_edge
add r2, r30, r27
add r8, r29, r27
b .LBB_test_9 ; loopexit
.LBB_test_3: ; shortcirc_next.0
add r28, r3, r27
lhz r28, 5(r28)
add r26, r4, r27
lhz r26, 5(r26)
cmpw cr0, r28, r26
beq .LBB_test_5 ; shortcirc_next.1
.LBB_test_4: ; shortcirc_next.0.loopexit_crit_edge
add r2, r11, r27
add r8, r12, r27
b .LBB_test_9 ; loopexit
.LBB_test_5: ; shortcirc_next.1
add r28, r3, r27
lhz r28, 7(r28)
add r26, r4, r27
lhz r26, 7(r26)
cmpw cr0, r28, r26
beq .LBB_test_7 ; shortcirc_next.2
.LBB_test_6: ; shortcirc_next.1.loopexit_crit_edge
add r2, r9, r27
add r8, r10, r27
b .LBB_test_9 ; loopexit
.LBB_test_7: ; shortcirc_next.2
add r28, r3, r27
lhz r26, 9(r28)
add r28, r4, r27
lhz r25, 9(r28)
addi r28, r27, 8
cmpw cr7, r26, r25
mfcr r26, 1
rlwinm r26, r26, 31, 31, 31
add r25, r8, r27
cmpw cr7, r25, r7
mfcr r25, 1
rlwinm r25, r25, 29, 31, 31
and. r26, r26, r25
bne .LBB_test_1 ; loopentry
Next up, improve the code for the loop.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22769 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-12 22:22:17 +00:00
|
|
|
User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter, L, this);
|
Move from Stage 0 to Stage 1.
Only emit one PHI node for IV uses with identical bases and strides (after
moving foldable immediates to the load/store instruction).
This implements LoopStrengthReduce/dont_insert_redundant_ops.ll, allowing
us to generate this PPC code for test1:
or r30, r3, r3
.LBB_test1_1: ; Loop
li r2, 0
stw r2, 0(r30)
stw r2, 4(r30)
bl L_pred$stub
addi r30, r30, 8
cmplwi cr0, r3, 0
bne .LBB_test1_1 ; Loop
instead of this code:
or r30, r3, r3
or r29, r3, r3
.LBB_test1_1: ; Loop
li r2, 0
stw r2, 0(r29)
stw r2, 4(r30)
bl L_pred$stub
addi r30, r30, 8 ;; Two iv's with step of 8
addi r29, r29, 8
cmplwi cr0, r3, 0
bne .LBB_test1_1 ; Loop
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22635 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-03 22:51:21 +00:00
|
|
|
|
|
|
|
// Mark old value we replaced as possibly dead, so that it is elminated
|
|
|
|
// if we just replaced the last use of that value.
|
2005-08-04 20:03:32 +00:00
|
|
|
DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
|
Move from Stage 0 to Stage 1.
Only emit one PHI node for IV uses with identical bases and strides (after
moving foldable immediates to the load/store instruction).
This implements LoopStrengthReduce/dont_insert_redundant_ops.ll, allowing
us to generate this PPC code for test1:
or r30, r3, r3
.LBB_test1_1: ; Loop
li r2, 0
stw r2, 0(r30)
stw r2, 4(r30)
bl L_pred$stub
addi r30, r30, 8
cmplwi cr0, r3, 0
bne .LBB_test1_1 ; Loop
instead of this code:
or r30, r3, r3
or r29, r3, r3
.LBB_test1_1: ; Loop
li r2, 0
stw r2, 0(r29)
stw r2, 4(r30)
bl L_pred$stub
addi r30, r30, 8 ;; Two iv's with step of 8
addi r29, r29, 8
cmplwi cr0, r3, 0
bne .LBB_test1_1 ; Loop
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22635 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-03 22:51:21 +00:00
|
|
|
|
|
|
|
UsersToProcess.erase(UsersToProcess.begin());
|
|
|
|
++NumReduced;
|
|
|
|
}
|
2005-07-30 00:15:07 +00:00
|
|
|
// TODO: Next, find out which base index is the most common, pull it out.
|
|
|
|
}
|
|
|
|
|
|
|
|
// IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
|
|
|
|
// different starting values, into different PHIs.
|
2004-10-18 21:08:22 +00:00
|
|
|
}
|
|
|
|
|
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-08 05:28:22 +00:00
|
|
|
// OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
|
|
|
|
// uses in the loop, look to see if we can eliminate some, in favor of using
|
|
|
|
// common indvars for the different uses.
|
|
|
|
void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
|
|
|
|
// TODO: implement optzns here.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
// Finally, get the terminating condition for the loop if possible. If we
|
|
|
|
// can, we want to change it to use a post-incremented version of its
|
|
|
|
// induction variable, to allow coallescing the live ranges for the IV into
|
|
|
|
// one register value.
|
|
|
|
PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
|
|
|
|
BasicBlock *Preheader = L->getLoopPreheader();
|
|
|
|
BasicBlock *LatchBlock =
|
|
|
|
SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
|
|
|
|
BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
|
|
|
|
if (!TermBr || TermBr->isUnconditional() ||
|
|
|
|
!isa<SetCondInst>(TermBr->getCondition()))
|
|
|
|
return;
|
|
|
|
SetCondInst *Cond = cast<SetCondInst>(TermBr->getCondition());
|
|
|
|
|
|
|
|
// Search IVUsesByStride to find Cond's IVUse if there is one.
|
|
|
|
IVStrideUse *CondUse = 0;
|
Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
For code like this:
void foo(float *a, float *b, int n, int stride_a, int stride_b) {
int i;
for (i=0; i<n; i++)
a[i*stride_a] = b[i*stride_b];
}
we now emit:
.LBB_foo2_2: ; no_exit
lfs f0, 0(r4)
stfs f0, 0(r3)
addi r7, r7, 1
add r4, r2, r4
add r3, r6, r3
cmpw cr0, r7, r5
blt .LBB_foo2_2 ; no_exit
instead of:
.LBB_foo_2: ; no_exit
mullw r8, r2, r7 ;; multiply!
slwi r8, r8, 2
lfsx f0, r4, r8
mullw r8, r2, r6 ;; multiply!
slwi r8, r8, 2
stfsx f0, r3, r8
addi r2, r2, 1
cmpw cr0, r2, r5
blt .LBB_foo_2 ; no_exit
loops with variable strides occur pretty often. For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.
Now we can allow indvars to turn functions written like this:
void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
int i, ai = 0, bi = 0;
for (i=0; i<n; i++)
{
a[ai] = b[bi];
ai += stride_a;
bi += stride_b;
}
}
into code like the above for better analysis. With this patch, they generate
identical code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22740 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-10 00:45:21 +00:00
|
|
|
const SCEVHandle *CondStride = 0;
|
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-08 05:28:22 +00:00
|
|
|
|
Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
For code like this:
void foo(float *a, float *b, int n, int stride_a, int stride_b) {
int i;
for (i=0; i<n; i++)
a[i*stride_a] = b[i*stride_b];
}
we now emit:
.LBB_foo2_2: ; no_exit
lfs f0, 0(r4)
stfs f0, 0(r3)
addi r7, r7, 1
add r4, r2, r4
add r3, r6, r3
cmpw cr0, r7, r5
blt .LBB_foo2_2 ; no_exit
instead of:
.LBB_foo_2: ; no_exit
mullw r8, r2, r7 ;; multiply!
slwi r8, r8, 2
lfsx f0, r4, r8
mullw r8, r2, r6 ;; multiply!
slwi r8, r8, 2
stfsx f0, r3, r8
addi r2, r2, 1
cmpw cr0, r2, r5
blt .LBB_foo_2 ; no_exit
loops with variable strides occur pretty often. For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.
Now we can allow indvars to turn functions written like this:
void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
int i, ai = 0, bi = 0;
for (i=0; i<n; i++)
{
a[ai] = b[bi];
ai += stride_a;
bi += stride_b;
}
}
into code like the above for better analysis. With this patch, they generate
identical code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22740 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-10 00:45:21 +00:00
|
|
|
for (std::map<SCEVHandle, IVUsersOfOneStride>::iterator
|
|
|
|
I = IVUsesByStride.begin(), E = IVUsesByStride.end();
|
|
|
|
I != E && !CondUse; ++I)
|
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-08 05:28:22 +00:00
|
|
|
for (std::vector<IVStrideUse>::iterator UI = I->second.Users.begin(),
|
|
|
|
E = I->second.Users.end(); UI != E; ++UI)
|
|
|
|
if (UI->User == Cond) {
|
|
|
|
CondUse = &*UI;
|
Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
For code like this:
void foo(float *a, float *b, int n, int stride_a, int stride_b) {
int i;
for (i=0; i<n; i++)
a[i*stride_a] = b[i*stride_b];
}
we now emit:
.LBB_foo2_2: ; no_exit
lfs f0, 0(r4)
stfs f0, 0(r3)
addi r7, r7, 1
add r4, r2, r4
add r3, r6, r3
cmpw cr0, r7, r5
blt .LBB_foo2_2 ; no_exit
instead of:
.LBB_foo_2: ; no_exit
mullw r8, r2, r7 ;; multiply!
slwi r8, r8, 2
lfsx f0, r4, r8
mullw r8, r2, r6 ;; multiply!
slwi r8, r8, 2
stfsx f0, r3, r8
addi r2, r2, 1
cmpw cr0, r2, r5
blt .LBB_foo_2 ; no_exit
loops with variable strides occur pretty often. For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.
Now we can allow indvars to turn functions written like this:
void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
int i, ai = 0, bi = 0;
for (i=0; i<n; i++)
{
a[ai] = b[bi];
ai += stride_a;
bi += stride_b;
}
}
into code like the above for better analysis. With this patch, they generate
identical code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22740 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-10 00:45:21 +00:00
|
|
|
CondStride = &I->first;
|
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-08 05:28:22 +00:00
|
|
|
// NOTE: we could handle setcc instructions with multiple uses here, but
|
|
|
|
// InstCombine does it as well for simple uses, it's not clear that it
|
|
|
|
// occurs enough in real life to handle.
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
if (!CondUse) return; // setcc doesn't use the IV.
|
|
|
|
|
|
|
|
// setcc stride is complex, don't mess with users.
|
Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
For code like this:
void foo(float *a, float *b, int n, int stride_a, int stride_b) {
int i;
for (i=0; i<n; i++)
a[i*stride_a] = b[i*stride_b];
}
we now emit:
.LBB_foo2_2: ; no_exit
lfs f0, 0(r4)
stfs f0, 0(r3)
addi r7, r7, 1
add r4, r2, r4
add r3, r6, r3
cmpw cr0, r7, r5
blt .LBB_foo2_2 ; no_exit
instead of:
.LBB_foo_2: ; no_exit
mullw r8, r2, r7 ;; multiply!
slwi r8, r8, 2
lfsx f0, r4, r8
mullw r8, r2, r6 ;; multiply!
slwi r8, r8, 2
stfsx f0, r3, r8
addi r2, r2, 1
cmpw cr0, r2, r5
blt .LBB_foo_2 ; no_exit
loops with variable strides occur pretty often. For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.
Now we can allow indvars to turn functions written like this:
void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
int i, ai = 0, bi = 0;
for (i=0; i<n; i++)
{
a[ai] = b[bi];
ai += stride_a;
bi += stride_b;
}
}
into code like the above for better analysis. With this patch, they generate
identical code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22740 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-10 00:45:21 +00:00
|
|
|
// FIXME: Evaluate whether this is a good idea or not.
|
|
|
|
if (!isa<SCEVConstant>(*CondStride)) return;
|
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-08 05:28:22 +00:00
|
|
|
|
|
|
|
// It's possible for the setcc instruction to be anywhere in the loop, and
|
|
|
|
// possible for it to have multiple users. If it is not immediately before
|
|
|
|
// the latch block branch, move it.
|
|
|
|
if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
|
|
|
|
if (Cond->hasOneUse()) { // Condition has a single use, just move it.
|
|
|
|
Cond->moveBefore(TermBr);
|
|
|
|
} else {
|
|
|
|
// Otherwise, clone the terminating condition and insert into the loopend.
|
|
|
|
Cond = cast<SetCondInst>(Cond->clone());
|
|
|
|
Cond->setName(L->getHeader()->getName() + ".termcond");
|
|
|
|
LatchBlock->getInstList().insert(TermBr, Cond);
|
|
|
|
|
|
|
|
// Clone the IVUse, as the old use still exists!
|
Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
For code like this:
void foo(float *a, float *b, int n, int stride_a, int stride_b) {
int i;
for (i=0; i<n; i++)
a[i*stride_a] = b[i*stride_b];
}
we now emit:
.LBB_foo2_2: ; no_exit
lfs f0, 0(r4)
stfs f0, 0(r3)
addi r7, r7, 1
add r4, r2, r4
add r3, r6, r3
cmpw cr0, r7, r5
blt .LBB_foo2_2 ; no_exit
instead of:
.LBB_foo_2: ; no_exit
mullw r8, r2, r7 ;; multiply!
slwi r8, r8, 2
lfsx f0, r4, r8
mullw r8, r2, r6 ;; multiply!
slwi r8, r8, 2
stfsx f0, r3, r8
addi r2, r2, 1
cmpw cr0, r2, r5
blt .LBB_foo_2 ; no_exit
loops with variable strides occur pretty often. For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.
Now we can allow indvars to turn functions written like this:
void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
int i, ai = 0, bi = 0;
for (i=0; i<n; i++)
{
a[ai] = b[bi];
ai += stride_a;
bi += stride_b;
}
}
into code like the above for better analysis. With this patch, they generate
identical code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22740 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-10 00:45:21 +00:00
|
|
|
IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
|
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-08 05:28:22 +00:00
|
|
|
CondUse->OperandValToReplace);
|
Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
For code like this:
void foo(float *a, float *b, int n, int stride_a, int stride_b) {
int i;
for (i=0; i<n; i++)
a[i*stride_a] = b[i*stride_b];
}
we now emit:
.LBB_foo2_2: ; no_exit
lfs f0, 0(r4)
stfs f0, 0(r3)
addi r7, r7, 1
add r4, r2, r4
add r3, r6, r3
cmpw cr0, r7, r5
blt .LBB_foo2_2 ; no_exit
instead of:
.LBB_foo_2: ; no_exit
mullw r8, r2, r7 ;; multiply!
slwi r8, r8, 2
lfsx f0, r4, r8
mullw r8, r2, r6 ;; multiply!
slwi r8, r8, 2
stfsx f0, r3, r8
addi r2, r2, 1
cmpw cr0, r2, r5
blt .LBB_foo_2 ; no_exit
loops with variable strides occur pretty often. For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.
Now we can allow indvars to turn functions written like this:
void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
int i, ai = 0, bi = 0;
for (i=0; i<n; i++)
{
a[ai] = b[bi];
ai += stride_a;
bi += stride_b;
}
}
into code like the above for better analysis. With this patch, they generate
identical code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22740 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-10 00:45:21 +00:00
|
|
|
CondUse = &IVUsesByStride[*CondStride].Users.back();
|
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-08 05:28:22 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// If we get to here, we know that we can transform the setcc instruction to
|
|
|
|
// use the post-incremented version of the IV, allowing us to coallesce the
|
|
|
|
// live ranges for the IV correctly.
|
Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
For code like this:
void foo(float *a, float *b, int n, int stride_a, int stride_b) {
int i;
for (i=0; i<n; i++)
a[i*stride_a] = b[i*stride_b];
}
we now emit:
.LBB_foo2_2: ; no_exit
lfs f0, 0(r4)
stfs f0, 0(r3)
addi r7, r7, 1
add r4, r2, r4
add r3, r6, r3
cmpw cr0, r7, r5
blt .LBB_foo2_2 ; no_exit
instead of:
.LBB_foo_2: ; no_exit
mullw r8, r2, r7 ;; multiply!
slwi r8, r8, 2
lfsx f0, r4, r8
mullw r8, r2, r6 ;; multiply!
slwi r8, r8, 2
stfsx f0, r3, r8
addi r2, r2, 1
cmpw cr0, r2, r5
blt .LBB_foo_2 ; no_exit
loops with variable strides occur pretty often. For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.
Now we can allow indvars to turn functions written like this:
void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
int i, ai = 0, bi = 0;
for (i=0; i<n; i++)
{
a[ai] = b[bi];
ai += stride_a;
bi += stride_b;
}
}
into code like the above for better analysis. With this patch, they generate
identical code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22740 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-10 00:45:21 +00:00
|
|
|
CondUse->Offset = SCEV::getMinusSCEV(CondUse->Offset, *CondStride);
|
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-08 05:28:22 +00:00
|
|
|
CondUse->isUseOfPostIncrementedValue = true;
|
|
|
|
}
|
2005-07-30 00:15:07 +00:00
|
|
|
|
2004-10-18 21:08:22 +00:00
|
|
|
void LoopStrengthReduce::runOnLoop(Loop *L) {
|
|
|
|
// First step, transform all loops nesting inside of this loop.
|
|
|
|
for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
|
|
|
|
runOnLoop(*I);
|
|
|
|
|
2005-07-30 00:15:07 +00:00
|
|
|
// Next, find all uses of induction variables in this loop, and catagorize
|
|
|
|
// them by stride. Start by finding all of the PHI nodes in the header for
|
|
|
|
// this loop. If they are induction variables, inspect their uses.
|
2005-08-04 17:40:30 +00:00
|
|
|
std::set<Instruction*> Processed; // Don't reprocess instructions.
|
2005-07-30 00:15:07 +00:00
|
|
|
for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
|
2005-08-04 17:40:30 +00:00
|
|
|
AddUsersIfInteresting(I, L, Processed);
|
2005-07-30 00:15:07 +00:00
|
|
|
|
|
|
|
// If we have nothing to do, return.
|
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-08 05:28:22 +00:00
|
|
|
if (IVUsesByStride.empty()) return;
|
|
|
|
|
|
|
|
// Optimize induction variables. Some indvar uses can be transformed to use
|
|
|
|
// strides that will be needed for other purposes. A common example of this
|
|
|
|
// is the exit test for the loop, which can often be rewritten to use the
|
|
|
|
// computation of some other indvar to decide when to terminate the loop.
|
|
|
|
OptimizeIndvars(L);
|
|
|
|
|
2005-07-30 00:15:07 +00:00
|
|
|
|
|
|
|
// FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
|
|
|
|
// doing computation in byte values, promote to 32-bit values if safe.
|
|
|
|
|
|
|
|
// FIXME: Attempt to reuse values across multiple IV's. In particular, we
|
|
|
|
// could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
|
|
|
|
// codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
|
|
|
|
// to be careful that IV's are all the same type. Only works for intptr_t
|
|
|
|
// indvars.
|
|
|
|
|
|
|
|
// If we only have one stride, we can more aggressively eliminate some things.
|
|
|
|
bool HasOneStride = IVUsesByStride.size() == 1;
|
2005-10-09 06:20:55 +00:00
|
|
|
|
Implement: LoopStrengthReduce/share_ivs.ll
Two changes:
* Only insert one PHI node for each stride. Other values are live in
values. This cannot introduce higher register pressure than the
previous approach, and can take advantage of reg+reg addressing modes.
* Factor common base values out of uses before moving values from the
base to the immediate fields. This improves codegen by starting the
stride-specific PHI node out at a common place for each IV use.
As an example, we used to generate this for a loop in swim:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfd f0, 0(r8)
stfd f0, 0(r3)
lfd f0, 0(r6)
stfd f0, 0(r7)
lfd f0, 0(r2)
stfd f0, 0(r5)
addi r9, r9, 1
addi r2, r2, 8
addi r5, r5, 8
addi r6, r6, 8
addi r7, r7, 8
addi r8, r8, 8
addi r3, r3, 8
cmpw cr0, r9, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
now we emit:
.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2: ; no_exit.7.i
lfdx f0, r8, r2
stfdx f0, r9, r2
lfdx f0, r5, r2
stfdx f0, r7, r2
lfdx f0, r3, r2
stfdx f0, r6, r2
addi r10, r10, 1
addi r2, r2, 8
cmpw cr0, r10, r4
bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1
As another more dramatic example, we used to emit this:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfd f0, 8(r21)
lfd f4, 8(r3)
lfd f5, 8(r27)
lfd f6, 8(r22)
lfd f7, 8(r5)
lfd f8, 8(r6)
lfd f9, 8(r30)
lfd f10, 8(r11)
lfd f11, 8(r12)
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfd f0, 8(r4)
lfd f0, 8(r25)
lfd f5, 8(r26)
lfd f6, 8(r23)
lfd f9, 8(r28)
lfd f10, 8(r10)
lfd f12, 8(r9)
lfd f13, 8(r29)
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfd f0, 8(r24)
lfd f0, 8(r8)
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfd f0, 8(r2)
addi r20, r20, 1
addi r2, r2, 8
addi r8, r8, 8
addi r10, r10, 8
addi r12, r12, 8
addi r6, r6, 8
addi r29, r29, 8
addi r28, r28, 8
addi r26, r26, 8
addi r25, r25, 8
addi r24, r24, 8
addi r5, r5, 8
addi r23, r23, 8
addi r22, r22, 8
addi r3, r3, 8
addi r9, r9, 8
addi r11, r11, 8
addi r30, r30, 8
addi r27, r27, 8
addi r21, r21, 8
addi r4, r4, 8
cmpw cr0, r20, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
we now emit:
.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2: ; no_exit.1.i19
lfdx f0, r21, r20
lfdx f4, r3, r20
lfdx f5, r27, r20
lfdx f6, r22, r20
lfdx f7, r5, r20
lfdx f8, r6, r20
lfdx f9, r30, r20
lfdx f10, r11, r20
lfdx f11, r12, r20
fsub f10, f10, f11
fadd f5, f4, f5
fmul f5, f5, f1
fadd f6, f6, f7
fadd f6, f6, f8
fadd f6, f6, f9
fmadd f0, f5, f6, f0
fnmsub f0, f10, f2, f0
stfdx f0, r4, r20
lfdx f0, r25, r20
lfdx f5, r26, r20
lfdx f6, r23, r20
lfdx f9, r28, r20
lfdx f10, r10, r20
lfdx f12, r9, r20
lfdx f13, r29, r20
fsub f11, f13, f11
fadd f4, f4, f5
fmul f4, f4, f1
fadd f5, f6, f9
fadd f5, f5, f10
fadd f5, f5, f12
fnmsub f0, f4, f5, f0
fnmsub f0, f11, f3, f0
stfdx f0, r24, r20
lfdx f0, r8, r20
fsub f4, f7, f8
fsub f5, f12, f10
fnmsub f0, f5, f2, f0
fnmsub f0, f4, f3, f0
stfdx f0, r2, r20
addi r19, r19, 1
addi r20, r20, 8
cmpw cr0, r19, r7
bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 00:18:09 +00:00
|
|
|
// Note: this processes each stride/type pair individually. All users passed
|
2005-10-09 06:20:55 +00:00
|
|
|
// into StrengthReduceStridedIVUsers have the same type AND stride. Also,
|
|
|
|
// node that we iterate over IVUsesByStride indirectly by using StrideOrder.
|
|
|
|
// This extra layer of indirection makes the ordering of strides deterministic
|
|
|
|
// - not dependent on map order.
|
|
|
|
for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
|
|
|
|
std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
|
|
|
|
IVUsesByStride.find(StrideOrder[Stride]);
|
|
|
|
assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
|
2005-07-30 00:15:07 +00:00
|
|
|
StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
|
2005-10-09 06:20:55 +00:00
|
|
|
}
|
2004-10-18 21:08:22 +00:00
|
|
|
|
|
|
|
// Clean up after ourselves
|
|
|
|
if (!DeadInsts.empty()) {
|
|
|
|
DeleteTriviallyDeadInstructions(DeadInsts);
|
|
|
|
|
2005-07-30 00:15:07 +00:00
|
|
|
BasicBlock::iterator I = L->getHeader()->begin();
|
|
|
|
PHINode *PN;
|
2005-08-02 02:44:31 +00:00
|
|
|
while ((PN = dyn_cast<PHINode>(I))) {
|
2005-08-02 00:41:11 +00:00
|
|
|
++I; // Preincrement iterator to avoid invalidating it when deleting PN.
|
|
|
|
|
Fix some 80 column violations.
Once we compute the evolution for a GEP, tell SE about it. This allows users
of the GEP to know it, if the users are not direct. This allows us to compile
this testcase:
void fbSolidFillmmx(int w, unsigned char *d) {
while (w >= 64) {
*(unsigned long long *) (d + 0) = 0;
*(unsigned long long *) (d + 8) = 0;
*(unsigned long long *) (d + 16) = 0;
*(unsigned long long *) (d + 24) = 0;
*(unsigned long long *) (d + 32) = 0;
*(unsigned long long *) (d + 40) = 0;
*(unsigned long long *) (d + 48) = 0;
*(unsigned long long *) (d + 56) = 0;
w -= 64;
d += 64;
}
}
into:
.LBB_fbSolidFillmmx_2: ; no_exit
li r2, 0
stw r2, 0(r4)
stw r2, 4(r4)
stw r2, 8(r4)
stw r2, 12(r4)
stw r2, 16(r4)
stw r2, 20(r4)
stw r2, 24(r4)
stw r2, 28(r4)
stw r2, 32(r4)
stw r2, 36(r4)
stw r2, 40(r4)
stw r2, 44(r4)
stw r2, 48(r4)
stw r2, 52(r4)
stw r2, 56(r4)
stw r2, 60(r4)
addi r4, r4, 64
addi r3, r3, -64
cmpwi cr0, r3, 63
bgt .LBB_fbSolidFillmmx_2 ; no_exit
instead of:
.LBB_fbSolidFillmmx_2: ; no_exit
li r11, 0
stw r11, 0(r4)
stw r11, 4(r4)
stwx r11, r10, r4
add r12, r10, r4
stw r11, 4(r12)
stwx r11, r9, r4
add r12, r9, r4
stw r11, 4(r12)
stwx r11, r8, r4
add r12, r8, r4
stw r11, 4(r12)
stwx r11, r7, r4
add r12, r7, r4
stw r11, 4(r12)
stwx r11, r6, r4
add r12, r6, r4
stw r11, 4(r12)
stwx r11, r5, r4
add r12, r5, r4
stw r11, 4(r12)
stwx r11, r2, r4
add r12, r2, r4
stw r11, 4(r12)
addi r4, r4, 64
addi r3, r3, -64
cmpwi cr0, r3, 63
bgt .LBB_fbSolidFillmmx_2 ; no_exit
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22737 91177308-0d34-0410-b5e6-96231b3b80d8
2005-08-09 23:39:36 +00:00
|
|
|
// At this point, we know that we have killed one or more GEP
|
|
|
|
// instructions. It is worth checking to see if the cann indvar is also
|
|
|
|
// dead, so that we can remove it as well. The requirements for the cann
|
|
|
|
// indvar to be considered dead are:
|
2005-07-30 00:15:07 +00:00
|
|
|
// 1. the cann indvar has one use
|
|
|
|
// 2. the use is an add instruction
|
|
|
|
// 3. the add has one use
|
|
|
|
// 4. the add is used by the cann indvar
|
|
|
|
// If all four cases above are true, then we can remove both the add and
|
|
|
|
// the cann indvar.
|
|
|
|
// FIXME: this needs to eliminate an induction variable even if it's being
|
|
|
|
// compared against some value to decide loop termination.
|
|
|
|
if (PN->hasOneUse()) {
|
|
|
|
BinaryOperator *BO = dyn_cast<BinaryOperator>(*(PN->use_begin()));
|
2005-08-02 02:52:02 +00:00
|
|
|
if (BO && BO->hasOneUse()) {
|
|
|
|
if (PN == *(BO->use_begin())) {
|
|
|
|
DeadInsts.insert(BO);
|
|
|
|
// Break the cycle, then delete the PHI.
|
|
|
|
PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
|
2005-08-03 21:36:09 +00:00
|
|
|
SE->deleteInstructionFromRecords(PN);
|
2005-08-02 02:52:02 +00:00
|
|
|
PN->eraseFromParent();
|
2004-10-18 21:08:22 +00:00
|
|
|
}
|
2005-08-02 02:52:02 +00:00
|
|
|
}
|
2005-07-30 00:15:07 +00:00
|
|
|
}
|
2004-10-18 21:08:22 +00:00
|
|
|
}
|
2005-07-30 00:15:07 +00:00
|
|
|
DeleteTriviallyDeadInstructions(DeadInsts);
|
2004-10-18 21:08:22 +00:00
|
|
|
}
|
2005-07-30 00:15:07 +00:00
|
|
|
|
2005-08-05 01:30:11 +00:00
|
|
|
CastedPointers.clear();
|
2005-07-30 00:15:07 +00:00
|
|
|
IVUsesByStride.clear();
|
2005-10-09 06:20:55 +00:00
|
|
|
StrideOrder.clear();
|
2005-07-30 00:15:07 +00:00
|
|
|
return;
|
2004-10-18 21:08:22 +00:00
|
|
|
}
|