If a block has N predecessors, then the current algorithm will try to
sink common code to this block N times (whenever we visit a
predecessor). Every attempt to sink the common code includes going
through all predecessors, so the complexity of the algorithm becomes
O(N^2).
With this patch we try to sink common code only when we visit the block
itself. With this, the complexity goes down to O(N).
As a side effect, the moment the code is sunk is slightly different than
before (the order of simplifications has been changed), that's why I had
to adjust two tests (note that neither of the tests is supposed to test
SimplifyCFG):
* test/CodeGen/AArch64/arm64-jumptable.ll - changes in this test mimic
the changes that previous implementation of SimplifyCFG would do.
* test/CodeGen/ARM/avoid-cpsr-rmw.ll - in this test I disabled common
code sinking by a command line flag.
llvm-svn: 321236
[Recommitting now an unrelated assertion in SROA is sorted out]
The new version has several advantages:
1) IMSHO it's more readable and neater
2) It handles loads and stores properly
3) It can handle any number of incoming blocks rather than just two. I'll be taking advantage of this in a followup patch.
With this change we can now finally sink load-modify-store idioms such as:
if (a)
return *b += 3;
else
return *b += 4;
=>
%z = load i32, i32* %y
%.sink = select i1 %a, i32 5, i32 7
%b = add i32 %z, %.sink
store i32 %b, i32* %y
ret i32 %b
When this works for switches it'll be even more powerful.
Round 4. This time we should handle all instructions correctly, and not replace any operands that need to be constant with variables.
This was really hard to determine safely, so the helper function should be put into the Instruction API. I'll do that as a followup.
llvm-svn: 279460
The new version has several advantages:
1) IMSHO it's more readable and neater
2) It handles loads and stores properly
3) It can handle any number of incoming blocks rather than just two. I'll be taking advantage of this in a followup patch.
With this change we can now finally sink load-modify-store idioms such as:
if (a)
return *b += 3;
else
return *b += 4;
=>
%z = load i32, i32* %y
%.sink = select i1 %a, i32 5, i32 7
%b = add i32 %z, %.sink
store i32 %b, i32* %y
ret i32 %b
When this works for switches it'll be even more powerful.
Round 4. This time we should handle all instructions correctly, and not replace any operands that need to be constant with variables.
This was really hard to determine safely, so the helper function should be put into the Instruction API. I'll do that as a followup.
llvm-svn: 279443
The new version has several advantages:
1) IMSHO it's more readable and neater
2) It handles loads and stores properly
3) It can handle any number of incoming blocks rather than just two. I'll be taking advantage of this in a followup patch.
With this change we can now finally sink load-modify-store idioms such as:
if (a)
return *b += 3;
else
return *b += 4;
=>
%z = load i32, i32* %y
%.sink = select i1 %a, i32 5, i32 7
%b = add i32 %z, %.sink
store i32 %b, i32* %y
ret i32 %b
When this works for switches it'll be even more powerful.
llvm-svn: 279229
This reverts commit r278660.
It causes downstream assertion failure in InstCombine on shuffle
instructions. Comes up in __mm_swizzle_epi32.
llvm-svn: 278672
The new version has several advantages:
1) IMSHO it's more readable and neater
2) It handles loads and stores properly
3) It can handle any number of incoming blocks rather than just two. I'll be taking advantage of this in a followup patch.
With this change we can now finally sink load-modify-store idioms such as:
if (a)
return *b += 3;
else
return *b += 4;
=>
%z = load i32, i32* %y
%.sink = select i1 %a, i32 5, i32 7
%b = add i32 %z, %.sink
store i32 %b, i32* %y
ret i32 %b
When this works for switches it'll be even more powerful.
llvm-svn: 278660
Reapply r242500 now that the swift schedmodel includes LDRLIT.
This is mostly done to disable the PostRAScheduler which optimizes for
instruction latencies which isn't a good fit for out-of-order
architectures. This also allows to leave out the itinerary table in
swift in favor of the SchedModel ones.
This change leads to performance improvements/regressions by as much as
10% in some benchmarks, in fact we loose 0.4% performance over the
llvm-testsuite for reasons that appear to be unknown or out of the
compilers control. rdar://20803802 documents the investigation of
these effects.
While it is probably a good idea to perform the same switch for the
other ARM out-of-order CPUs, I limited this change to swift as I cannot
perform the benchmark verification on the other CPUs.
Differential Revision: http://reviews.llvm.org/D10513
llvm-svn: 242588
This is mostly done to disable the PostRAScheduler which optimizes for
instruction latencies which isn't a good fit for out-of-order
architectures. This also allows to leave out the itinerary table in
swift in favor of the SchedModel ones.
This change leads to performance improvements/regressions by as much as
10% in some benchmarks, in fact we loose 0.4% performance over the
llvm-testsuite for reasons that appear to be unknown or out of the
compilers control. rdar://20803802 documents the investigation of
these effects.
While it is probably a good idea to perform the same switch for the
other ARM out-of-order CPUs, I limited this change to swift as I cannot
perform the benchmark verification on the other CPUs.
Differential Revision: http://reviews.llvm.org/D10513
llvm-svn: 242500
Essentially the same as the GEP change in r230786.
A similar migration script can be used to update test cases, though a few more
test case improvements/changes were required this time around: (r229269-r229278)
import fileinput
import sys
import re
pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)")
for line in sys.stdin:
sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line))
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7649
llvm-svn: 230794
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
llvm-svn: 230786
The previous change to local live range allocation also suppressed
eviction of local ranges. In rare cases, this could result in more
expensive register choices. This commit actually revives a feature
that I added long ago: check if live ranges can be reassigned before
eviction. But now it only happens in rare cases of evicting a local
live range because another local live range wants a cheaper register.
The benefit is improved code size for some benchmarks on x86 and armv7.
I measured no significant compile time increase and performance
changes are noise.
llvm-svn: 187140
Also avoid locals evicting locals just because they want a cheaper register.
Problem: MI Sched knows exactly how many registers we have and assumes
they can be colored. In cases where we have large blocks, usually from
unrolled loops, greedy coloring fails. This is a source of
"regressions" from the MI Scheduler on x86. I noticed this issue on
x86 where we have long chains of two-address defs in the same live
range. It's easy to see this in matrix multiplication benchmarks like
IRSmk and even the unit test misched-matmul.ll.
A fundamental difference between the LLVM register allocator and
conventional graph coloring is that in our model a live range can't
discover its neighbors, it can only verify its neighbors. That's why
we initially went for greedy coloring and added eviction to deal with
the hard cases. However, for singly defined and two-address live
ranges, we can optimally color without visiting neighbors simply by
processing the live ranges in instruction order.
Other beneficial side effects:
It is much easier to understand and debug regalloc for large blocks
when the live ranges are allocated in order. Yes, global allocation is
still very confusing, but it's nice to be able to comprehend what
happened locally.
Heuristics could be added to bias register assignment based on
instruction locality (think late register pairing, banks...).
Intuituvely this will make some test cases that are on the threshold
of register pressure more stable.
llvm-svn: 187139
This update was done with the following bash script:
find test/CodeGen -name "*.ll" | \
while read NAME; do
echo "$NAME"
if ! grep -q "^; *RUN: *llc.*debug" $NAME; then
TEMP=`mktemp -t temp`
cp $NAME $TEMP
sed -n "s/^define [^@]*@\([A-Za-z0-9_]*\)(.*$/\1/p" < $NAME | \
while read FUNC; do
sed -i '' "s/;\(.*\)\([A-Za-z0-9_-]*\):\( *\)$FUNC: *\$/;\1\2-LABEL:\3$FUNC:/g" $TEMP
done
sed -i '' "s/;\(.*\)-LABEL-LABEL:/;\1-LABEL:/" $TEMP
sed -i '' "s/;\(.*\)-NEXT-LABEL:/;\1-NEXT:/" $TEMP
sed -i '' "s/;\(.*\)-NOT-LABEL:/;\1-NOT:/" $TEMP
sed -i '' "s/;\(.*\)-DAG-LABEL:/;\1-DAG:/" $TEMP
mv $TEMP $NAME
fi
done
llvm-svn: 186280
The Thumb2SizeReduction pass avoids false CPSR dependencies, except it
still aggressively creates tMOVi8 instructions because they are so
common.
Avoid creating false CPSR dependencies even for tMOVi8 instructions when
the the CPSR flags are known to have high latency. This allows integer
computation to overlap floating point computations.
Also process blocks in a reverse post-order and propagate high-latency
flags to successors.
<rdar://problem/13468102>
llvm-svn: 178773
To not over constrain the scheduler for ARM in thumb mode, some optimizations for code size reduction, specific to ARM thumb, are blocked when they add a dependency (like write after read dependency).
Disables this check when code size is the priority, i.e., code is compiled with -Oz.
llvm-svn: 170462
The tied source operand of tMUL is the second source operand, not the
first like every other two-address thumb instruction. Special case it
in the size reduction pass to make sure we create the tMUL instruction
properly.
llvm-svn: 151315
Michael Zolotukhin