it was a use, def, or both. This allows us to be less pessimistic in our
analysis of them. In practice, this doesn't make a big difference, but it
doesn't hurt either.
llvm-svn: 16632
and delete them if they turn out to be dead. This is a useful little hack
that even speeds up some programs. For example, it speeds up Ptrdist/ks
from 17.53s to 15.59s, and 188.ammp from 149s to 146s.
This also speeds up llc :)
llvm-svn: 16630
generated code over the simple spiller. The new local spiller generates
substantially better code than the simple one in some cases, by reusing
values that are loaded out of stack slots and kept available in registers.
This primarily helps programs that are spilling a lot, and there is still
stuff that can be done to improve it. This patch makes the local spiller
the default, as it's only a tiny bit slower than the simple spiller (it
increases the runtime of llc by < 1%).
Here are some numbers with speedups.
Program #reuse old(s) new(s) Speedup
Povray: 3452, 16.87 -> 15.93 (5.5%)
177.mesa: 2176, 2.77 -> 2.76 (0%)
179.art: 35, 28.43 -> 28.01 (1.5%)
183.equake: 55, 61.44 -> 61.41 (0%)
188.ammp: 869, 174 -> 149 (15%)
164.gzip: 43, 40.73 -> 40.71 (0%)
175.vpr: 351, 18.54 -> 17.34 (6.5%)
176.gcc: 2471, 5.01 -> 4.92 (1.8%)
181.mcf 42, 79.30 -> 75.20 (5.2%)
186.crafty: 484, 29.73 -> 30.04 (-1%)
197.parser: 251, 10.47 -> 10.67 (-1%)
252.eon: 1501, 1.98 -> 1.75 (12%)
253.perlbm: 1183, 14.83 -> 14.42 (2.8%)
254.gap: 825, 7.46 -> 7.29 (2.3%)
255.vortex: 285, 10.51 -> 10.27 (2.3%)
256.bzip2: 63, 55.70 -> 55.20 (0.9%)
300.twolf: 830, 21.63 -> 22.00 (-1%)
PtrDist/ks 14, 32.75 -> 17.53 (46.5%)
Olden/tsp 46, 8.71 -> 8.24 (5.4%)
Free/distray 70, 1.09 -> 0.99 (9.2%)
llvm-svn: 16629
two spillers produce perfectly identical code (at least on povray and eon),
but the simple spiller is substantially faster than the local spiller. Once
the local spiller is improved, we can switch back.
Switching cuts 5.2% off of the llc time for povray (about 1.3s).
llvm-svn: 16608
use a simple vector. This speeds up -spiller=simple from taking 22s to taking
.1s on povray (debug build). This change does not modify the generated code.
llvm-svn: 16607
Move include/Config and include/Support into include/llvm/Config,
include/llvm/ADT and include/llvm/Support. From here on out, all LLVM
public header files must be under include/llvm/.
llvm-svn: 16137
lists. Instead of scanning the vector backwards, scan it forward and
swap each element we want to erase. Then at the end erase all removed
intervals at once. This doesn't save much: 0.08s out of 4s when
compiling 176.gcc.
llvm-svn: 16136
Regression.CodeGen.Generic.2004-04-09-SameValueCoalescing.llx and the
code size problem.
This bug prevented us from doing most register coallesces.
llvm-svn: 16031
with emitting .xwords when not on an 8-byte boundary (.xword 0 is not the
same as 8 .byte 0's). Because we do not know when or when we are not aligned,
just emit bytes like the old V9 asmprinter did.
llvm-svn: 16006
Add support for targets that must spill certain physregs at certain locations.
Patch contributed by Nate Begeman, slightly hacked by me.
llvm-svn: 15701
These side-effects seem to make a difference when using llc -march=sparcv9
in Release mode (i.e., with -DNDEBUG); when they are left out, lots of
instructions just get dropped on the floor, because they never end up
in the schedule.
llvm-svn: 15339
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
llvm-svn: 15194
same as the PHI use. This is not correct as the PHI use value is different
depending on which branch is taken. This fixes espresso with aggressive
coallescing, and perhaps others.
llvm-svn: 15189
us back to taking about 10.5s on gcc, instead of taking 15.6s! The net result
is that my big patches have hand no significant effect on compile time or code
quality. heh.
llvm-svn: 15156
Interval. This generalizes the isDefinedOnce mechanism that we used before
to help us coallesce ranges that overlap. As part of this, every logical
range with a different value is assigned a different number in the interval.
For example, for code that looks like this:
0 X = ...
4 X += ...
...
N = X
We now generate a live interval that contains two ranges: [2,6:0),[6,?:1)
reflecting the fact that there are two different values in the range at
different positions in the code.
Currently we are not using this information at all, so this just slows down
liveintervals. In the future, this will change.
Note that this change also substantially refactors the joinIntervalsInMachineBB
method to merge the cases for virt-virt and phys-virt joining into a single
case, adds comments, and makes the code a bit easier to follow.
llvm-svn: 15154
* Fix comment typeo
* add dump() methods
* add a few new methods like getLiveRangeContaining, removeRange & joinable
(which is currently the same as overlaps)
* Remove the unused operator==
Bigger change:
* In LiveInterval, instead of using a boolean isDefinedOnce to keep track of
if there are > 1 definitions in a particular interval, keep a counter,
NumValues to keep track of exactly how many there are.
* In LiveRange, add a new ValId element to indicate which of the numbered
values each LiveRange belongs to. We now no longer merge LiveRanges if
they are of differing value ID's even if they are neighbors.
llvm-svn: 15152
want to insert a new range into the middle of the vector, then delete ranges
one at a time next to the inserted one as they are merged.
Instead, if the inserted interval overlaps, just start merging. The only time
we insert into the middle of the vector is when we don't overlap at all. Also
delete blocks of live ranges if we overlap with many of them.
This patch speeds up joining by .7 seconds on a large testcase, but more
importantly gets all of the range adding code into addRangeFrom.
llvm-svn: 15141
will soon be renamed) into their own file. The new file should not emit
DEBUG output or have other side effects. The LiveInterval class also now
doesn't know whether its working on registers or some other thing.
In the future we will want to use the LiveInterval class and friends to do
stack packing. In addition to a code simplification, this will allow us to
do it more easily.
llvm-svn: 15134
Use an explicit LiveRange class to represent ranges instead of an std::pair.
This is a minor cleanup, but is really intended to make a future patch simpler
and less invasive.
Alkis, could you please take a look at LiveInterval::liveAt? I suspect that
you can add an operator<(unsigned) to LiveRange, allowing us to speed up the
upper_bound call by quite a bit (this would also apply to other callers of
upper/lower_bound). I would do it myself, but I still don't understand that
crazy liveAt function, despite the comment. :)
Basically I would like to see this:
LiveRange dummy(index, index+1);
Ranges::const_iterator r = std::upper_bound(ranges.begin(),
ranges.end(),
dummy);
Turn into:
Ranges::const_iterator r = std::upper_bound(ranges.begin(),
ranges.end(),
index);
llvm-svn: 15130
interfere. Because these intervals have a single definition, and one of them
is a copy instruction, they are always safe to merge even if their lifetimes
interfere. This slightly reduces the amount of spill code, for example on
252.eon, from:
12837 spiller - Number of loads added
7604 spiller - Number of stores added
5842 spiller - Number of register spills
18155 liveintervals - Number of identity moves eliminated after coalescing
to:
12754 spiller - Number of loads added
7585 spiller - Number of stores added
5803 spiller - Number of register spills
18262 liveintervals - Number of identity moves eliminated after coalescing
The much much bigger win would be to merge intervals with multiple definitions
(aka phi nodes) but this is not that day.
llvm-svn: 15124
compilation of gcc:
* Use vectors instead of lists for the intervals sets
* Use a heap for the unhandled set to keep intervals always sorted and
makes insertions back to the heap very fast (compared to scanning a
list)
llvm-svn: 15103
fortunately, they are easy to handle if we know about them. This patch fixes
some serious pessimization of code produced by the linscan register allocator.
llvm-svn: 15092
* vreg <-> vreg joining now works, enable it unconditionally when joining
is enabled (which is the default).
* Fix a serious pessimization of spill code where we were saying that a
spilled DEF operand was live into the subsequent instruction. This allows
for substantially better code when spilling starts to happen.
llvm-svn: 14993
order, causing the inactive list in the linearscan list to get unsorted, which
basically fuxored everything up severely.
These seems to fix the joiner, so with more testing I will enable it by default.
llvm-svn: 14992
Heavily refactor handleVirtualRegisterDef, adding comments and making it more
efficient. It is also much easier to follow and convince ones self that it is
correct :)
Add -debug output to the joine, showing the result of joining the intervals.
llvm-svn: 14989
basic block clear()'s all of the operands lists, including phis. This
caused removePredecessor to get confused later. Because of this, we just
nuke (without prejudice) PHI nodes in unreachable blocks.
llvm-svn: 14635
pass is required to paper over problems in the code generator (primarily
live variables and its clients) which doesn't really have any well defined
semantics for unreachable code.
The proper solution to this problem is to have instruction selectors not
select blocks that are unreachable. Until we have a instruction selection
framework available for use, however, we can't expect all instruction
selector writers to do this. Until then, this pass should be used.
llvm-svn: 14563
instructions. Instead, keep a map of instructions -> MCFI objects in the
already sparc-specific class MachineFunctionInfo. This will slow down the
sparc backend a bit, but it does not penalize the rest of LLVM!
llvm-svn: 14438
The vector may actually be empty if the register that we are marking as
recently used is not actually allocatable. This happens for physical registers
that are not allocatable, like the ST(x) registers on X86.
llvm-svn: 14195
broke obsequi and a lot of other things. It all boiled down to MBB being
overloaded in an inner scope and me confusing it with the one in the outer
scope. Ugh!
llvm-svn: 13517
MBBs start out as #-1. When a MBB is added to a MachineFunction, it
gets the next available unique MBB number. If it is removed from a
MachineFunction, it goes back to being #-1.
llvm-svn: 13514
in the basic block being processed. This fixes PhiElimination on kimwitu++
from taking 105s to taking a much more reasonable 0.6s (in a debug build).
llvm-svn: 13453
than before. Because this is the case, we can compute the first non-phi
instruction once when de-phi'ing a block. This shaves ~4s off of
phielimination of _Z7yyparsev in kimwitu++ from 109s -> 105s. There are
still much more important gains to come.
llvm-svn: 13452
when we see a read of a register. This is important in cases like:
AL = ...
AH = ...
= AX
The read of AX must make both the AL and AH defs live until the use.
llvm-svn: 13444
use MachineBasicBlocks. To do this, we traverse the Machine CFG instead of
the LLVM CFG, which is also *MUCH* more efficient by having fewer levels of
indirections and mappings.
llvm-svn: 13301
Reid Spencer