of this, and use it to our advantage (bwahahah). This allows us to eliminate another
60 instructions from smg2000 on PPC (probably significantly more on X86). A common
old-new diff looks like this:
stw r2, 3304(r1)
- lwz r2, 3192(r1)
stw r2, 3300(r1)
- lwz r2, 3192(r1)
stw r2, 3296(r1)
- lwz r2, 3192(r1)
stw r2, 3200(r1)
- lwz r2, 3192(r1)
stw r2, 3196(r1)
- lwz r2, 3192(r1)
+ or r2, r2, r2
stw r2, 3188(r1)
and
- lwz r31, 604(r1)
- lwz r13, 604(r1)
- lwz r14, 604(r1)
- lwz r15, 604(r1)
- lwz r16, 604(r1)
- lwz r30, 604(r1)
+ or r31, r30, r30
+ or r13, r30, r30
+ or r14, r30, r30
+ or r15, r30, r30
+ or r16, r30, r30
+ or r30, r30, r30
Removal of the R = R copies is coming next...
llvm-svn: 25919
this code:
store [stack slot #0], R10
= add R14, [stack slot #0]
The spiller didn't know that the store made the value of [stackslot#0] available
in R10 *IF* the store came from a copy instruction with the store folded into it.
This patch teaches VirtRegMap to look at these stores and recognize the values
they make available. In one case Evan provided, this code:
divsd %XMM0, %XMM1
movsd %XMM1, QWORD PTR [%ESP + 40]
1) movsd QWORD PTR [%ESP + 48], %XMM1
2) movsd %XMM1, QWORD PTR [%ESP + 48]
addsd %XMM1, %XMM0
3) movsd QWORD PTR [%ESP + 48], %XMM1
movsd QWORD PTR [%ESP + 4], %XMM0
turns into:
divsd %XMM0, %XMM1
movsd %XMM1, QWORD PTR [%ESP + 40]
addsd %XMM1, %XMM0
3) movsd QWORD PTR [%ESP + 48], %XMM1
movsd QWORD PTR [%ESP + 4], %XMM0
In this case, instruction #2 was removed because of the value made
available by #1, and inst #1 was later deleted because it is now
never used before the stack slot is redefined by #3.
This occurs here and there in a lot of code with high spilling, on PPC
most of the removed loads/stores are LSU-reject-causing loads, which is
nice.
On X86, things are much better (because it spills more), where we nuke
about 1% of the instructions from SMG2000 and several hundred from eon.
More improvements to come...
llvm-svn: 25917
previous copy elisions and we discover we need to reload a register, make
sure to use the regclass of the original register for the reload, not the
class of the current register. This avoid using 16-bit loads to reload 32-bit
values.
llvm-svn: 23645
store r12 -> [ss#2]
R3 = load [ss#1]
use R3
R3 = load [ss#2]
R4 = load [ss#1]
and turn it into this code:
store R12 -> [ss#2]
R3 = load [ss#1]
use R3
R3 = R12
R4 = R3 <- oops!
The problem was that promoting R3 = load[ss#2] to a copy missed the fact that
the instruction invalidated R3 at that point.
llvm-svn: 23638
when possible, avoiding the load (and avoiding the copy if the value is already
in the right register).
This patch came about when I noticed code like the following being generated:
store R17 -> [SS1]
...blah...
R4 = load [SS1]
This was causing an LSU reject on the G5. This problem was due to the register
allocator folding spill code into a reg-reg copy (producing the load), which
prevented the spiller from being able to rewrite the load into a copy, despite
the fact that the value was already available in a register. In the case
above, we now rip out the R4 load and replace it with a R4 = R17 copy.
This speeds up several programs on X86 (which spills a lot :) ), e.g.
smg2k from 22.39->20.60s, povray from 12.93->12.66s, 168.wupwise from
68.54->53.83s (!), 197.parser from 7.33->6.62s (!), etc. This may have a larger
impact in some cases on the G5 (by avoiding LSU rejects), though it probably
won't trigger as often (less spilling in general).
Targets that implement folding of loads/stores into copies should implement
the isLoadFromStackSlot hook to get this.
llvm-svn: 23388
are, simplify logic, and cause things to not be nested as deeply. This also
uses MRI->areAliases instead of an explicit loop.
No functionality change, just code cleanup.
llvm-svn: 23296
The problem occurred when trying to reload this instruction:
MOV32mr %reg2326, 8, %reg2297, 4, %reg2295
The value of reg2326 was available in EBX, so it was reused from there, instead
of reloading it into EDX.
The value of reg2297 was available in EDX, so it was reused from there, instead
of reloading it into EDI.
The value of reg2295 was not available, so we tried reloading it into EBX, its
assigned register. However, we checked and saw that we already reloaded
something into EBX, so we chose what reg2326 was assigned to (EDX) and reloaded
into that register instead.
Unfortunately EDX had already been used by reg2297, so reloading into EDX
clobbered the value used by the reg2326 operand, breaking the program.
The fix for this is to check that the newly picked register is ok. In this
case we now find that EDX is already used and try using EDI, which succeeds.
llvm-svn: 17006
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
allocator.
The implementation is completely rewritten and now employs several
optimizations not exercised before. For example for 164.gzip we have
997 loads and 699 stores vs the 1221 loads and 880 stores we have
before.
llvm-svn: 11798
Chris Lattner