2004-07-23 17:49:16 +00:00
|
|
|
//===-- LiveInterval.cpp - Live Interval Representation -------------------===//
|
|
|
|
//
|
|
|
|
// The LLVM Compiler Infrastructure
|
|
|
|
//
|
|
|
|
// This file was developed by the LLVM research group and is distributed under
|
|
|
|
// the University of Illinois Open Source License. See LICENSE.TXT for details.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
//
|
|
|
|
// This file implements the LiveRange and LiveInterval classes. Given some
|
|
|
|
// numbering of each the machine instructions an interval [i, j) is said to be a
|
|
|
|
// live interval for register v if there is no instruction with number j' > j
|
|
|
|
// such that v is live at j' abd there is no instruction with number i' < i such
|
|
|
|
// that v is live at i'. In this implementation intervals can have holes,
|
|
|
|
// i.e. an interval might look like [1,20), [50,65), [1000,1001). Each
|
|
|
|
// individual range is represented as an instance of LiveRange, and the whole
|
|
|
|
// interval is represented as an instance of LiveInterval.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
#include "LiveInterval.h"
|
2004-09-01 22:55:40 +00:00
|
|
|
#include "llvm/ADT/STLExtras.h"
|
2004-09-28 02:38:58 +00:00
|
|
|
#include <algorithm>
|
2004-07-24 02:52:23 +00:00
|
|
|
#include <iostream>
|
|
|
|
#include <map>
|
2004-07-23 17:49:16 +00:00
|
|
|
using namespace llvm;
|
|
|
|
|
|
|
|
// An example for liveAt():
|
|
|
|
//
|
2004-07-23 18:40:00 +00:00
|
|
|
// this = [1,4), liveAt(0) will return false. The instruction defining this
|
|
|
|
// spans slots [0,3]. The interval belongs to an spilled definition of the
|
|
|
|
// variable it represents. This is because slot 1 is used (def slot) and spans
|
|
|
|
// up to slot 3 (store slot).
|
2004-07-23 17:49:16 +00:00
|
|
|
//
|
2004-07-23 18:13:24 +00:00
|
|
|
bool LiveInterval::liveAt(unsigned I) const {
|
|
|
|
Ranges::const_iterator r = std::upper_bound(ranges.begin(), ranges.end(), I);
|
|
|
|
|
2004-07-23 17:49:16 +00:00
|
|
|
if (r == ranges.begin())
|
|
|
|
return false;
|
|
|
|
|
|
|
|
--r;
|
2004-07-23 18:40:00 +00:00
|
|
|
return r->contains(I);
|
2004-07-23 17:49:16 +00:00
|
|
|
}
|
|
|
|
|
2004-11-18 03:47:34 +00:00
|
|
|
// overlaps - Return true if the intersection of the two live intervals is
|
|
|
|
// not empty.
|
|
|
|
//
|
2004-07-23 17:49:16 +00:00
|
|
|
// An example for overlaps():
|
|
|
|
//
|
|
|
|
// 0: A = ...
|
|
|
|
// 4: B = ...
|
|
|
|
// 8: C = A + B ;; last use of A
|
|
|
|
//
|
|
|
|
// The live intervals should look like:
|
|
|
|
//
|
|
|
|
// A = [3, 11)
|
|
|
|
// B = [7, x)
|
|
|
|
// C = [11, y)
|
|
|
|
//
|
|
|
|
// A->overlaps(C) should return false since we want to be able to join
|
|
|
|
// A and C.
|
2004-11-18 03:47:34 +00:00
|
|
|
//
|
|
|
|
bool LiveInterval::overlapsFrom(const LiveInterval& other,
|
|
|
|
const_iterator StartPos) const {
|
|
|
|
const_iterator i = begin();
|
|
|
|
const_iterator ie = end();
|
|
|
|
const_iterator j = StartPos;
|
|
|
|
const_iterator je = other.end();
|
|
|
|
|
|
|
|
assert((StartPos->start <= i->start || StartPos == other.begin()) &&
|
2004-11-18 04:02:11 +00:00
|
|
|
StartPos != other.end() && "Bogus start position hint!");
|
This patch makes use of the infrastructure implemented before to safely and
aggressively coallesce live ranges even if they overlap. Consider this LLVM
code for example:
int %test(int %X) {
%Y = mul int %X, 1 ;; Codegens to Y = X
%Z = add int %X, %Y
ret int %Z
}
The mul is just there to get a copy into the code stream. This produces
this machine code:
(0x869e5a8, LLVM BB @0x869b9a0):
%reg1024 = mov <fi#-2>, 1, %NOREG, 0 ;; "X"
%reg1025 = mov %reg1024 ;; "Y" (subsumed by X)
%reg1026 = add %reg1024, %reg1025
%EAX = mov %reg1026
ret
Note that the life times of reg1024 and reg1025 overlap, even though they
contain the same value. This results in this machine code:
test:
mov %EAX, DWORD PTR [%ESP + 4]
mov %ECX, %EAX
add %EAX, %ECX
ret
Another, worse case involves loops and PHI nodes. Consider this trivial loop:
testcase:
int %test2(int %X) {
entry:
br label %Loop
Loop:
%Y = phi int [%X, %entry], [%Z, %Loop]
%Z = add int %Y, 1
%cond = seteq int %Z, 100
br bool %cond, label %Out, label %Loop
Out:
ret int %Z
}
Because of interactions between the PHI elimination pass and the register
allocator, this got compiled to this code:
test2:
mov %ECX, DWORD PTR [%ESP + 4]
.LBBtest2_1:
*** mov %EAX, %ECX
inc %EAX
cmp %EAX, 100
*** mov %ECX, %EAX
jne .LBBtest2_1
ret
Or on powerpc, this code:
_test2:
mflr r0
stw r0, 8(r1)
stwu r1, -60(r1)
.LBB_test2_1:
addi r2, r3, 1
cmpwi cr0, r2, 100
*** or r3, r2, r2
bne cr0, .LBB_test2_1
*** or r3, r2, r2
lwz r0, 68(r1)
mtlr r0
addi r1, r1, 60
blr 0
With this improvement in place, we now generate this code for these two
testcases, which is what we want:
test:
mov %EAX, DWORD PTR [%ESP + 4]
add %EAX, %EAX
ret
test2:
mov %EAX, DWORD PTR [%ESP + 4]
.LBBtest2_1:
inc %EAX
cmp %EAX, 100
jne .LBBtest2_1 # Loop
ret
Or on PPC:
_test2:
mflr r0
stw r0, 8(r1)
stwu r1, -60(r1)
.LBB_test2_1:
addi r3, r3, 1
cmpwi cr0, r3, 100
bne cr0, .LBB_test2_1
lwz r0, 68(r1)
mtlr r0
addi r1, r1, 60
blr 0
Static numbers for spill code loads/stores/reg-reg copies (smaller is better):
em3d: before: 47/25/26 after: 44/22/24
164.gzip: before: 433/245/310 after: 403/231/278
175.vpr: before: 3721/2189/1581 after: 4144/2081/1423
176.gcc: before: 26195/8866/9235 after: 25942/8082/8275
186.crafty: before: 4295/2587/3079 after: 4119/2519/2916
252.eon: before: 12754/7585/5803 after: 12508/7425/5643
256.bzip2: before: 463/226/315 after: 482:241/309
Runtime perf number samples on X86:
gzip: before: 41.09 after: 39.86
bzip2: runtime: before: 56.71s after: 57.07s
gcc: before: 6.16 after: 6.12
eon: before: 2.03s after: 2.00s
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@15194 91177308-0d34-0410-b5e6-96231b3b80d8
2004-07-25 07:11:19 +00:00
|
|
|
|
2004-07-23 17:49:16 +00:00
|
|
|
if (i->start < j->start) {
|
2004-07-23 18:40:00 +00:00
|
|
|
i = std::upper_bound(i, ie, j->start);
|
2004-07-23 17:49:16 +00:00
|
|
|
if (i != ranges.begin()) --i;
|
2004-07-23 18:40:00 +00:00
|
|
|
} else if (j->start < i->start) {
|
2004-12-04 01:22:09 +00:00
|
|
|
++StartPos;
|
|
|
|
if (StartPos != other.end() && StartPos->start <= i->start) {
|
|
|
|
assert(StartPos < other.end() && i < end());
|
2004-11-18 04:02:11 +00:00
|
|
|
j = std::upper_bound(j, je, i->start);
|
|
|
|
if (j != other.ranges.begin()) --j;
|
|
|
|
}
|
2004-07-23 18:40:00 +00:00
|
|
|
} else {
|
|
|
|
return true;
|
2004-07-23 17:49:16 +00:00
|
|
|
}
|
|
|
|
|
2004-11-18 05:28:21 +00:00
|
|
|
if (j == je) return false;
|
|
|
|
|
|
|
|
while (i != ie) {
|
2004-07-23 17:49:16 +00:00
|
|
|
if (i->start > j->start) {
|
2004-07-24 11:44:15 +00:00
|
|
|
std::swap(i, j);
|
|
|
|
std::swap(ie, je);
|
2004-07-23 17:49:16 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
if (i->end > j->start)
|
|
|
|
return true;
|
|
|
|
++i;
|
|
|
|
}
|
|
|
|
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2004-07-24 02:52:23 +00:00
|
|
|
/// joinable - Two intervals are joinable if the either don't overlap at all
|
|
|
|
/// or if the destination of the copy is a single assignment value, and it
|
|
|
|
/// only overlaps with one value in the source interval.
|
|
|
|
bool LiveInterval::joinable(const LiveInterval &other, unsigned CopyIdx) const {
|
This patch makes use of the infrastructure implemented before to safely and
aggressively coallesce live ranges even if they overlap. Consider this LLVM
code for example:
int %test(int %X) {
%Y = mul int %X, 1 ;; Codegens to Y = X
%Z = add int %X, %Y
ret int %Z
}
The mul is just there to get a copy into the code stream. This produces
this machine code:
(0x869e5a8, LLVM BB @0x869b9a0):
%reg1024 = mov <fi#-2>, 1, %NOREG, 0 ;; "X"
%reg1025 = mov %reg1024 ;; "Y" (subsumed by X)
%reg1026 = add %reg1024, %reg1025
%EAX = mov %reg1026
ret
Note that the life times of reg1024 and reg1025 overlap, even though they
contain the same value. This results in this machine code:
test:
mov %EAX, DWORD PTR [%ESP + 4]
mov %ECX, %EAX
add %EAX, %ECX
ret
Another, worse case involves loops and PHI nodes. Consider this trivial loop:
testcase:
int %test2(int %X) {
entry:
br label %Loop
Loop:
%Y = phi int [%X, %entry], [%Z, %Loop]
%Z = add int %Y, 1
%cond = seteq int %Z, 100
br bool %cond, label %Out, label %Loop
Out:
ret int %Z
}
Because of interactions between the PHI elimination pass and the register
allocator, this got compiled to this code:
test2:
mov %ECX, DWORD PTR [%ESP + 4]
.LBBtest2_1:
*** mov %EAX, %ECX
inc %EAX
cmp %EAX, 100
*** mov %ECX, %EAX
jne .LBBtest2_1
ret
Or on powerpc, this code:
_test2:
mflr r0
stw r0, 8(r1)
stwu r1, -60(r1)
.LBB_test2_1:
addi r2, r3, 1
cmpwi cr0, r2, 100
*** or r3, r2, r2
bne cr0, .LBB_test2_1
*** or r3, r2, r2
lwz r0, 68(r1)
mtlr r0
addi r1, r1, 60
blr 0
With this improvement in place, we now generate this code for these two
testcases, which is what we want:
test:
mov %EAX, DWORD PTR [%ESP + 4]
add %EAX, %EAX
ret
test2:
mov %EAX, DWORD PTR [%ESP + 4]
.LBBtest2_1:
inc %EAX
cmp %EAX, 100
jne .LBBtest2_1 # Loop
ret
Or on PPC:
_test2:
mflr r0
stw r0, 8(r1)
stwu r1, -60(r1)
.LBB_test2_1:
addi r3, r3, 1
cmpwi cr0, r3, 100
bne cr0, .LBB_test2_1
lwz r0, 68(r1)
mtlr r0
addi r1, r1, 60
blr 0
Static numbers for spill code loads/stores/reg-reg copies (smaller is better):
em3d: before: 47/25/26 after: 44/22/24
164.gzip: before: 433/245/310 after: 403/231/278
175.vpr: before: 3721/2189/1581 after: 4144/2081/1423
176.gcc: before: 26195/8866/9235 after: 25942/8082/8275
186.crafty: before: 4295/2587/3079 after: 4119/2519/2916
252.eon: before: 12754/7585/5803 after: 12508/7425/5643
256.bzip2: before: 463/226/315 after: 482:241/309
Runtime perf number samples on X86:
gzip: before: 41.09 after: 39.86
bzip2: runtime: before: 56.71s after: 57.07s
gcc: before: 6.16 after: 6.12
eon: before: 2.03s after: 2.00s
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@15194 91177308-0d34-0410-b5e6-96231b3b80d8
2004-07-25 07:11:19 +00:00
|
|
|
const LiveRange *SourceLR = other.getLiveRangeContaining(CopyIdx-1);
|
|
|
|
const LiveRange *DestLR = getLiveRangeContaining(CopyIdx);
|
|
|
|
assert(SourceLR && DestLR && "Not joining due to a copy?");
|
|
|
|
unsigned OtherValIdx = SourceLR->ValId;
|
|
|
|
unsigned ThisValIdx = DestLR->ValId;
|
|
|
|
|
|
|
|
Ranges::const_iterator i = ranges.begin();
|
|
|
|
Ranges::const_iterator ie = ranges.end();
|
|
|
|
Ranges::const_iterator j = other.ranges.begin();
|
|
|
|
Ranges::const_iterator je = other.ranges.end();
|
|
|
|
|
|
|
|
if (i->start < j->start) {
|
|
|
|
i = std::upper_bound(i, ie, j->start);
|
|
|
|
if (i != ranges.begin()) --i;
|
|
|
|
} else if (j->start < i->start) {
|
|
|
|
j = std::upper_bound(j, je, i->start);
|
|
|
|
if (j != other.ranges.begin()) --j;
|
|
|
|
}
|
|
|
|
|
|
|
|
while (i != ie && j != je) {
|
|
|
|
if (i->start == j->start) {
|
|
|
|
// If this is not the allowed value merge, we cannot join.
|
|
|
|
if (i->ValId != ThisValIdx || j->ValId != OtherValIdx)
|
2004-07-25 07:47:25 +00:00
|
|
|
return false;
|
This patch makes use of the infrastructure implemented before to safely and
aggressively coallesce live ranges even if they overlap. Consider this LLVM
code for example:
int %test(int %X) {
%Y = mul int %X, 1 ;; Codegens to Y = X
%Z = add int %X, %Y
ret int %Z
}
The mul is just there to get a copy into the code stream. This produces
this machine code:
(0x869e5a8, LLVM BB @0x869b9a0):
%reg1024 = mov <fi#-2>, 1, %NOREG, 0 ;; "X"
%reg1025 = mov %reg1024 ;; "Y" (subsumed by X)
%reg1026 = add %reg1024, %reg1025
%EAX = mov %reg1026
ret
Note that the life times of reg1024 and reg1025 overlap, even though they
contain the same value. This results in this machine code:
test:
mov %EAX, DWORD PTR [%ESP + 4]
mov %ECX, %EAX
add %EAX, %ECX
ret
Another, worse case involves loops and PHI nodes. Consider this trivial loop:
testcase:
int %test2(int %X) {
entry:
br label %Loop
Loop:
%Y = phi int [%X, %entry], [%Z, %Loop]
%Z = add int %Y, 1
%cond = seteq int %Z, 100
br bool %cond, label %Out, label %Loop
Out:
ret int %Z
}
Because of interactions between the PHI elimination pass and the register
allocator, this got compiled to this code:
test2:
mov %ECX, DWORD PTR [%ESP + 4]
.LBBtest2_1:
*** mov %EAX, %ECX
inc %EAX
cmp %EAX, 100
*** mov %ECX, %EAX
jne .LBBtest2_1
ret
Or on powerpc, this code:
_test2:
mflr r0
stw r0, 8(r1)
stwu r1, -60(r1)
.LBB_test2_1:
addi r2, r3, 1
cmpwi cr0, r2, 100
*** or r3, r2, r2
bne cr0, .LBB_test2_1
*** or r3, r2, r2
lwz r0, 68(r1)
mtlr r0
addi r1, r1, 60
blr 0
With this improvement in place, we now generate this code for these two
testcases, which is what we want:
test:
mov %EAX, DWORD PTR [%ESP + 4]
add %EAX, %EAX
ret
test2:
mov %EAX, DWORD PTR [%ESP + 4]
.LBBtest2_1:
inc %EAX
cmp %EAX, 100
jne .LBBtest2_1 # Loop
ret
Or on PPC:
_test2:
mflr r0
stw r0, 8(r1)
stwu r1, -60(r1)
.LBB_test2_1:
addi r3, r3, 1
cmpwi cr0, r3, 100
bne cr0, .LBB_test2_1
lwz r0, 68(r1)
mtlr r0
addi r1, r1, 60
blr 0
Static numbers for spill code loads/stores/reg-reg copies (smaller is better):
em3d: before: 47/25/26 after: 44/22/24
164.gzip: before: 433/245/310 after: 403/231/278
175.vpr: before: 3721/2189/1581 after: 4144/2081/1423
176.gcc: before: 26195/8866/9235 after: 25942/8082/8275
186.crafty: before: 4295/2587/3079 after: 4119/2519/2916
252.eon: before: 12754/7585/5803 after: 12508/7425/5643
256.bzip2: before: 463/226/315 after: 482:241/309
Runtime perf number samples on X86:
gzip: before: 41.09 after: 39.86
bzip2: runtime: before: 56.71s after: 57.07s
gcc: before: 6.16 after: 6.12
eon: before: 2.03s after: 2.00s
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@15194 91177308-0d34-0410-b5e6-96231b3b80d8
2004-07-25 07:11:19 +00:00
|
|
|
} else if (i->start < j->start) {
|
|
|
|
if (i->end > j->start) {
|
|
|
|
if (i->ValId != ThisValIdx || j->ValId != OtherValIdx)
|
2004-07-25 07:47:25 +00:00
|
|
|
return false;
|
This patch makes use of the infrastructure implemented before to safely and
aggressively coallesce live ranges even if they overlap. Consider this LLVM
code for example:
int %test(int %X) {
%Y = mul int %X, 1 ;; Codegens to Y = X
%Z = add int %X, %Y
ret int %Z
}
The mul is just there to get a copy into the code stream. This produces
this machine code:
(0x869e5a8, LLVM BB @0x869b9a0):
%reg1024 = mov <fi#-2>, 1, %NOREG, 0 ;; "X"
%reg1025 = mov %reg1024 ;; "Y" (subsumed by X)
%reg1026 = add %reg1024, %reg1025
%EAX = mov %reg1026
ret
Note that the life times of reg1024 and reg1025 overlap, even though they
contain the same value. This results in this machine code:
test:
mov %EAX, DWORD PTR [%ESP + 4]
mov %ECX, %EAX
add %EAX, %ECX
ret
Another, worse case involves loops and PHI nodes. Consider this trivial loop:
testcase:
int %test2(int %X) {
entry:
br label %Loop
Loop:
%Y = phi int [%X, %entry], [%Z, %Loop]
%Z = add int %Y, 1
%cond = seteq int %Z, 100
br bool %cond, label %Out, label %Loop
Out:
ret int %Z
}
Because of interactions between the PHI elimination pass and the register
allocator, this got compiled to this code:
test2:
mov %ECX, DWORD PTR [%ESP + 4]
.LBBtest2_1:
*** mov %EAX, %ECX
inc %EAX
cmp %EAX, 100
*** mov %ECX, %EAX
jne .LBBtest2_1
ret
Or on powerpc, this code:
_test2:
mflr r0
stw r0, 8(r1)
stwu r1, -60(r1)
.LBB_test2_1:
addi r2, r3, 1
cmpwi cr0, r2, 100
*** or r3, r2, r2
bne cr0, .LBB_test2_1
*** or r3, r2, r2
lwz r0, 68(r1)
mtlr r0
addi r1, r1, 60
blr 0
With this improvement in place, we now generate this code for these two
testcases, which is what we want:
test:
mov %EAX, DWORD PTR [%ESP + 4]
add %EAX, %EAX
ret
test2:
mov %EAX, DWORD PTR [%ESP + 4]
.LBBtest2_1:
inc %EAX
cmp %EAX, 100
jne .LBBtest2_1 # Loop
ret
Or on PPC:
_test2:
mflr r0
stw r0, 8(r1)
stwu r1, -60(r1)
.LBB_test2_1:
addi r3, r3, 1
cmpwi cr0, r3, 100
bne cr0, .LBB_test2_1
lwz r0, 68(r1)
mtlr r0
addi r1, r1, 60
blr 0
Static numbers for spill code loads/stores/reg-reg copies (smaller is better):
em3d: before: 47/25/26 after: 44/22/24
164.gzip: before: 433/245/310 after: 403/231/278
175.vpr: before: 3721/2189/1581 after: 4144/2081/1423
176.gcc: before: 26195/8866/9235 after: 25942/8082/8275
186.crafty: before: 4295/2587/3079 after: 4119/2519/2916
252.eon: before: 12754/7585/5803 after: 12508/7425/5643
256.bzip2: before: 463/226/315 after: 482:241/309
Runtime perf number samples on X86:
gzip: before: 41.09 after: 39.86
bzip2: runtime: before: 56.71s after: 57.07s
gcc: before: 6.16 after: 6.12
eon: before: 2.03s after: 2.00s
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@15194 91177308-0d34-0410-b5e6-96231b3b80d8
2004-07-25 07:11:19 +00:00
|
|
|
}
|
|
|
|
} else {
|
|
|
|
if (j->end > i->start) {
|
|
|
|
if (i->ValId != ThisValIdx || j->ValId != OtherValIdx)
|
2004-07-25 07:47:25 +00:00
|
|
|
return false;
|
This patch makes use of the infrastructure implemented before to safely and
aggressively coallesce live ranges even if they overlap. Consider this LLVM
code for example:
int %test(int %X) {
%Y = mul int %X, 1 ;; Codegens to Y = X
%Z = add int %X, %Y
ret int %Z
}
The mul is just there to get a copy into the code stream. This produces
this machine code:
(0x869e5a8, LLVM BB @0x869b9a0):
%reg1024 = mov <fi#-2>, 1, %NOREG, 0 ;; "X"
%reg1025 = mov %reg1024 ;; "Y" (subsumed by X)
%reg1026 = add %reg1024, %reg1025
%EAX = mov %reg1026
ret
Note that the life times of reg1024 and reg1025 overlap, even though they
contain the same value. This results in this machine code:
test:
mov %EAX, DWORD PTR [%ESP + 4]
mov %ECX, %EAX
add %EAX, %ECX
ret
Another, worse case involves loops and PHI nodes. Consider this trivial loop:
testcase:
int %test2(int %X) {
entry:
br label %Loop
Loop:
%Y = phi int [%X, %entry], [%Z, %Loop]
%Z = add int %Y, 1
%cond = seteq int %Z, 100
br bool %cond, label %Out, label %Loop
Out:
ret int %Z
}
Because of interactions between the PHI elimination pass and the register
allocator, this got compiled to this code:
test2:
mov %ECX, DWORD PTR [%ESP + 4]
.LBBtest2_1:
*** mov %EAX, %ECX
inc %EAX
cmp %EAX, 100
*** mov %ECX, %EAX
jne .LBBtest2_1
ret
Or on powerpc, this code:
_test2:
mflr r0
stw r0, 8(r1)
stwu r1, -60(r1)
.LBB_test2_1:
addi r2, r3, 1
cmpwi cr0, r2, 100
*** or r3, r2, r2
bne cr0, .LBB_test2_1
*** or r3, r2, r2
lwz r0, 68(r1)
mtlr r0
addi r1, r1, 60
blr 0
With this improvement in place, we now generate this code for these two
testcases, which is what we want:
test:
mov %EAX, DWORD PTR [%ESP + 4]
add %EAX, %EAX
ret
test2:
mov %EAX, DWORD PTR [%ESP + 4]
.LBBtest2_1:
inc %EAX
cmp %EAX, 100
jne .LBBtest2_1 # Loop
ret
Or on PPC:
_test2:
mflr r0
stw r0, 8(r1)
stwu r1, -60(r1)
.LBB_test2_1:
addi r3, r3, 1
cmpwi cr0, r3, 100
bne cr0, .LBB_test2_1
lwz r0, 68(r1)
mtlr r0
addi r1, r1, 60
blr 0
Static numbers for spill code loads/stores/reg-reg copies (smaller is better):
em3d: before: 47/25/26 after: 44/22/24
164.gzip: before: 433/245/310 after: 403/231/278
175.vpr: before: 3721/2189/1581 after: 4144/2081/1423
176.gcc: before: 26195/8866/9235 after: 25942/8082/8275
186.crafty: before: 4295/2587/3079 after: 4119/2519/2916
252.eon: before: 12754/7585/5803 after: 12508/7425/5643
256.bzip2: before: 463/226/315 after: 482:241/309
Runtime perf number samples on X86:
gzip: before: 41.09 after: 39.86
bzip2: runtime: before: 56.71s after: 57.07s
gcc: before: 6.16 after: 6.12
eon: before: 2.03s after: 2.00s
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@15194 91177308-0d34-0410-b5e6-96231b3b80d8
2004-07-25 07:11:19 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
if (i->end < j->end)
|
|
|
|
++i;
|
|
|
|
else
|
|
|
|
++j;
|
|
|
|
}
|
|
|
|
|
2004-07-25 07:47:25 +00:00
|
|
|
return true;
|
2004-07-24 02:52:23 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2004-07-23 19:38:44 +00:00
|
|
|
/// extendIntervalEndTo - This method is used when we want to extend the range
|
|
|
|
/// specified by I to end at the specified endpoint. To do this, we should
|
|
|
|
/// merge and eliminate all ranges that this will overlap with. The iterator is
|
|
|
|
/// not invalidated.
|
|
|
|
void LiveInterval::extendIntervalEndTo(Ranges::iterator I, unsigned NewEnd) {
|
|
|
|
assert(I != ranges.end() && "Not a valid interval!");
|
2004-07-24 02:52:23 +00:00
|
|
|
unsigned ValId = I->ValId;
|
2004-07-23 19:38:44 +00:00
|
|
|
|
|
|
|
// Search for the first interval that we can't merge with.
|
|
|
|
Ranges::iterator MergeTo = next(I);
|
2004-07-24 02:52:23 +00:00
|
|
|
for (; MergeTo != ranges.end() && NewEnd >= MergeTo->end; ++MergeTo) {
|
|
|
|
assert(MergeTo->ValId == ValId && "Cannot merge with differing values!");
|
|
|
|
}
|
2004-07-23 19:38:44 +00:00
|
|
|
|
|
|
|
// If NewEnd was in the middle of an interval, make sure to get its endpoint.
|
|
|
|
I->end = std::max(NewEnd, prior(MergeTo)->end);
|
|
|
|
|
|
|
|
// Erase any dead ranges
|
|
|
|
ranges.erase(next(I), MergeTo);
|
2004-07-23 17:49:16 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2004-07-23 19:38:44 +00:00
|
|
|
/// extendIntervalStartTo - This method is used when we want to extend the range
|
|
|
|
/// specified by I to start at the specified endpoint. To do this, we should
|
|
|
|
/// merge and eliminate all ranges that this will overlap with.
|
|
|
|
LiveInterval::Ranges::iterator
|
|
|
|
LiveInterval::extendIntervalStartTo(Ranges::iterator I, unsigned NewStart) {
|
|
|
|
assert(I != ranges.end() && "Not a valid interval!");
|
2004-07-24 02:52:23 +00:00
|
|
|
unsigned ValId = I->ValId;
|
2004-07-23 19:38:44 +00:00
|
|
|
|
|
|
|
// Search for the first interval that we can't merge with.
|
|
|
|
Ranges::iterator MergeTo = I;
|
|
|
|
do {
|
|
|
|
if (MergeTo == ranges.begin()) {
|
|
|
|
I->start = NewStart;
|
2004-07-24 02:52:23 +00:00
|
|
|
ranges.erase(MergeTo, I);
|
2004-07-23 19:38:44 +00:00
|
|
|
return I;
|
|
|
|
}
|
2004-07-24 02:52:23 +00:00
|
|
|
assert(MergeTo->ValId == ValId && "Cannot merge with differing values!");
|
2004-07-23 19:38:44 +00:00
|
|
|
--MergeTo;
|
|
|
|
} while (NewStart <= MergeTo->start);
|
|
|
|
|
|
|
|
// If we start in the middle of another interval, just delete a range and
|
|
|
|
// extend that interval.
|
2004-07-24 02:52:23 +00:00
|
|
|
if (MergeTo->end >= NewStart && MergeTo->ValId == ValId) {
|
2004-07-23 19:38:44 +00:00
|
|
|
MergeTo->end = I->end;
|
|
|
|
} else {
|
|
|
|
// Otherwise, extend the interval right after.
|
|
|
|
++MergeTo;
|
|
|
|
MergeTo->start = NewStart;
|
|
|
|
MergeTo->end = I->end;
|
2004-07-23 17:49:16 +00:00
|
|
|
}
|
2004-07-23 19:38:44 +00:00
|
|
|
|
|
|
|
ranges.erase(next(MergeTo), next(I));
|
|
|
|
return MergeTo;
|
2004-07-23 17:49:16 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
LiveInterval::Ranges::iterator
|
2004-07-23 19:38:44 +00:00
|
|
|
LiveInterval::addRangeFrom(LiveRange LR, Ranges::iterator From) {
|
|
|
|
unsigned Start = LR.start, End = LR.end;
|
|
|
|
Ranges::iterator it = std::upper_bound(From, ranges.end(), Start);
|
|
|
|
|
|
|
|
// If the inserted interval starts in the middle or right at the end of
|
|
|
|
// another interval, just extend that interval to contain the range of LR.
|
|
|
|
if (it != ranges.begin()) {
|
|
|
|
Ranges::iterator B = prior(it);
|
2004-07-24 02:52:23 +00:00
|
|
|
if (LR.ValId == B->ValId) {
|
|
|
|
if (B->start <= Start && B->end >= Start) {
|
|
|
|
extendIntervalEndTo(B, End);
|
|
|
|
return B;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// Check to make sure that we are not overlapping two live ranges with
|
|
|
|
// different ValId's.
|
|
|
|
assert(B->end <= Start &&
|
2004-11-16 06:52:35 +00:00
|
|
|
"Cannot overlap two LiveRanges with differing ValID's"
|
|
|
|
" (did you def the same reg twice in a MachineInstr?)");
|
2004-07-23 19:38:44 +00:00
|
|
|
}
|
2004-07-23 17:49:16 +00:00
|
|
|
}
|
2004-07-23 19:38:44 +00:00
|
|
|
|
|
|
|
// Otherwise, if this range ends in the middle of, or right next to, another
|
|
|
|
// interval, merge it into that interval.
|
2004-07-24 02:52:23 +00:00
|
|
|
if (it != ranges.end())
|
|
|
|
if (LR.ValId == it->ValId) {
|
|
|
|
if (it->start <= End) {
|
|
|
|
it = extendIntervalStartTo(it, Start);
|
|
|
|
|
|
|
|
// If LR is a complete superset of an interval, we may need to grow its
|
|
|
|
// endpoint as well.
|
|
|
|
if (End > it->end)
|
|
|
|
extendIntervalEndTo(it, End);
|
|
|
|
return it;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// Check to make sure that we are not overlapping two live ranges with
|
|
|
|
// different ValId's.
|
|
|
|
assert(it->start >= End &&
|
|
|
|
"Cannot overlap two LiveRanges with differing ValID's");
|
|
|
|
}
|
2004-07-23 19:38:44 +00:00
|
|
|
|
|
|
|
// Otherwise, this is just a new range that doesn't interact with anything.
|
|
|
|
// Insert it.
|
|
|
|
return ranges.insert(it, LR);
|
2004-07-23 17:49:16 +00:00
|
|
|
}
|
|
|
|
|
2004-07-24 02:52:23 +00:00
|
|
|
|
|
|
|
/// removeRange - Remove the specified range from this interval. Note that
|
|
|
|
/// the range must already be in this interval in its entirety.
|
|
|
|
void LiveInterval::removeRange(unsigned Start, unsigned End) {
|
|
|
|
// Find the LiveRange containing this span.
|
|
|
|
Ranges::iterator I = std::upper_bound(ranges.begin(), ranges.end(), Start);
|
|
|
|
assert(I != ranges.begin() && "Range is not in interval!");
|
|
|
|
--I;
|
|
|
|
assert(I->contains(Start) && I->contains(End-1) &&
|
|
|
|
"Range is not entirely in interval!");
|
|
|
|
|
|
|
|
// If the span we are removing is at the start of the LiveRange, adjust it.
|
|
|
|
if (I->start == Start) {
|
|
|
|
if (I->end == End)
|
|
|
|
ranges.erase(I); // Removed the whole LiveRange.
|
|
|
|
else
|
|
|
|
I->start = End;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Otherwise if the span we are removing is at the end of the LiveRange,
|
|
|
|
// adjust the other way.
|
|
|
|
if (I->end == End) {
|
2004-07-25 05:43:53 +00:00
|
|
|
I->end = Start;
|
2004-07-24 02:52:23 +00:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Otherwise, we are splitting the LiveRange into two pieces.
|
|
|
|
unsigned OldEnd = I->end;
|
|
|
|
I->end = Start; // Trim the old interval.
|
|
|
|
|
|
|
|
// Insert the new one.
|
|
|
|
ranges.insert(next(I), LiveRange(End, OldEnd, I->ValId));
|
|
|
|
}
|
|
|
|
|
|
|
|
/// getLiveRangeContaining - Return the live range that contains the
|
|
|
|
/// specified index, or null if there is none.
|
2004-07-25 06:23:01 +00:00
|
|
|
const LiveRange *LiveInterval::getLiveRangeContaining(unsigned Idx) const {
|
|
|
|
Ranges::const_iterator It = std::upper_bound(ranges.begin(),ranges.end(),Idx);
|
2004-07-24 02:52:23 +00:00
|
|
|
if (It != ranges.begin()) {
|
2004-07-25 06:23:01 +00:00
|
|
|
const LiveRange &LR = *prior(It);
|
2004-07-24 02:52:23 +00:00
|
|
|
if (LR.contains(Idx))
|
|
|
|
return &LR;
|
|
|
|
}
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
/// join - Join two live intervals (this, and other) together. This operation
|
|
|
|
/// is the result of a copy instruction in the source program, that occurs at
|
|
|
|
/// index 'CopyIdx' that copies from 'Other' to 'this'.
|
|
|
|
void LiveInterval::join(LiveInterval &Other, unsigned CopyIdx) {
|
2004-07-25 06:23:01 +00:00
|
|
|
const LiveRange *SourceLR = Other.getLiveRangeContaining(CopyIdx-1);
|
|
|
|
const LiveRange *DestLR = getLiveRangeContaining(CopyIdx);
|
2004-07-24 02:52:23 +00:00
|
|
|
assert(SourceLR && DestLR && "Not joining due to a copy?");
|
|
|
|
unsigned MergedSrcValIdx = SourceLR->ValId;
|
|
|
|
unsigned MergedDstValIdx = DestLR->ValId;
|
2004-07-23 17:49:16 +00:00
|
|
|
|
2004-07-24 03:41:50 +00:00
|
|
|
// Try to do the least amount of work possible. In particular, if there are
|
|
|
|
// more liverange chunks in the other set than there are in the 'this' set,
|
|
|
|
// swap sets to merge the fewest chunks in possible.
|
|
|
|
if (Other.ranges.size() > ranges.size()) {
|
|
|
|
std::swap(MergedSrcValIdx, MergedDstValIdx);
|
|
|
|
std::swap(ranges, Other.ranges);
|
|
|
|
std::swap(NumValues, Other.NumValues);
|
|
|
|
}
|
|
|
|
|
2004-07-23 19:38:44 +00:00
|
|
|
// Join the ranges of other into the ranges of this interval.
|
2004-07-24 02:52:23 +00:00
|
|
|
Ranges::iterator InsertPos = ranges.begin();
|
|
|
|
std::map<unsigned, unsigned> Dst2SrcIdxMap;
|
|
|
|
for (Ranges::iterator I = Other.ranges.begin(),
|
|
|
|
E = Other.ranges.end(); I != E; ++I) {
|
|
|
|
// Map the ValId in the other live range to the current live range.
|
|
|
|
if (I->ValId == MergedSrcValIdx)
|
|
|
|
I->ValId = MergedDstValIdx;
|
|
|
|
else {
|
|
|
|
unsigned &NV = Dst2SrcIdxMap[I->ValId];
|
|
|
|
if (NV == 0) NV = getNextValue();
|
|
|
|
I->ValId = NV;
|
|
|
|
}
|
2004-07-23 19:38:44 +00:00
|
|
|
|
2004-07-24 02:52:23 +00:00
|
|
|
InsertPos = addRangeFrom(*I, InsertPos);
|
|
|
|
}
|
|
|
|
|
|
|
|
weight += Other.weight;
|
2004-07-23 17:49:16 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
std::ostream& llvm::operator<<(std::ostream& os, const LiveRange &LR) {
|
2004-07-24 02:52:23 +00:00
|
|
|
return os << '[' << LR.start << ',' << LR.end << ':' << LR.ValId << ")";
|
|
|
|
}
|
|
|
|
|
|
|
|
void LiveRange::dump() const {
|
|
|
|
std::cerr << *this << "\n";
|
2004-07-23 17:49:16 +00:00
|
|
|
}
|
|
|
|
|
2004-07-24 02:52:23 +00:00
|
|
|
|
2004-07-23 17:49:16 +00:00
|
|
|
std::ostream& llvm::operator<<(std::ostream& os, const LiveInterval& li) {
|
|
|
|
os << "%reg" << li.reg << ',' << li.weight;
|
|
|
|
if (li.empty())
|
|
|
|
return os << "EMPTY";
|
|
|
|
|
|
|
|
os << " = ";
|
|
|
|
for (LiveInterval::Ranges::const_iterator i = li.ranges.begin(),
|
|
|
|
e = li.ranges.end(); i != e; ++i)
|
|
|
|
os << *i;
|
|
|
|
return os;
|
|
|
|
}
|
2004-07-24 02:52:23 +00:00
|
|
|
|
|
|
|
void LiveInterval::dump() const {
|
|
|
|
std::cerr << *this << "\n";
|
|
|
|
}
|