In general, one should always complete CFG modifications first, update
CFG-based analyses, like Dominatores and LoopInfo, then generate
instruction sequences.
LoopVectorizer was creating a new loop, calling SCEVExpander to
generate checks, then updating LoopInfo. I just changed the order.
llvm-svn: 186241
Fixes a 35% degradation compared to unvectorized code in
MiBench/automotive-susan and an equally serious regression on a private
image processing benchmark.
radar://14351991
llvm-svn: 186188
against a constant."
This reverts commit r186107. It didn't handle wrapping arithmetic in the
loop correctly and thus caused the following C program to count from
0 to UINT64_MAX instead of from 0 to 255 as intended:
#include <stdio.h>
int main() {
unsigned char first = 0, last = 255;
do { printf("%d\n", first); } while (first++ != last);
}
Full test case and instructions to reproduce with just the -indvars pass
sent to the original review thread rather than to r186107's commit.
llvm-svn: 186152
Before we could vectorize PHINodes scanning successors was a good way of finding candidates. Now we can vectorize the phinodes which is simpler.
llvm-svn: 186139
Patch by Michele Scandale!
Adds a special handling of the case where, during the loop exit
condition rewriting, the exit value is a constant of bitwidth lower
than the type of the induction variable: instead of introducing a
trunc operation in order to match correctly the operand types, it
allows to convert the constant value to an equivalent constant,
depending on the initial value of the induction variable and the trip
count, in order have an equivalent comparison between the induction
variable and the new constant.
llvm-svn: 186107
We can vectorize them because in the case where we wrap in the address space the
unvectorized code would have had to access a pointer value of zero which is
undefined behavior in address space zero according to the LLVM IR semantics.
(Thank you Duncan, for pointing this out to me).
Fixes PR16592.
llvm-svn: 186088
predecessors of the two blocks it is attempting to merge supply the
same incoming values to any phi in the successor block. This change
allows merging in the case where there is one or more incoming values
that are undef. The undef values are rewritten to match the non-undef
value that flows from the other edge. Patch by Mark Lacey.
llvm-svn: 186069
Without the changes introduced into this patch, if TRE saw any allocas at all,
TRE would not perform TRE *or* mark callsites with the tail marker.
Because TRE runs after mem2reg, this inadequacy is not a death sentence. But
given a callsite A without escaping alloca argument, A may not be able to have
the tail marker placed on it due to a separate callsite B having a write-back
parameter passed in via an argument with the nocapture attribute.
Assume that B is the only other callsite besides A and B only has nocapture
escaping alloca arguments (*NOTE* B may have other arguments that are not passed
allocas). In this case not marking A with the tail marker is unnecessarily
conservative since:
1. By assumption A has no escaping alloca arguments itself so it can not
access the caller's stack via its arguments.
2. Since all of B's escaping alloca arguments are passed as parameters with
the nocapture attribute, we know that B does not stash said escaping
allocas in a manner that outlives B itself and thus could be accessed
indirectly by A.
With the changes introduced by this patch:
1. If we see any escaping allocas passed as a capturing argument, we do
nothing and bail early.
2. If we do not see any escaping allocas passed as captured arguments but we
do see escaping allocas passed as nocapture arguments:
i. We do not perform TRE to avoid PR962 since the code generator produces
significantly worse code for the dynamic allocas that would be created
by the TRE algorithm.
ii. If we do not return twice, mark call sites without escaping allocas
with the tail marker. *NOTE* This excludes functions with escaping
nocapture allocas.
3. If we do not see any escaping allocas at all (whether captured or not):
i. If we do not have usage of setjmp, mark all callsites with the tail
marker.
ii. If there are no dynamic/variable sized allocas in the function,
attempt to perform TRE on all callsites in the function.
Based off of a patch by Nick Lewycky.
rdar://14324281.
llvm-svn: 186057
(add nsw x, (and x, y)) isn't a power of two if x is zero, it's zero
(add nsw x, (xor x, y)) isn't a power of two if y has bits set that aren't set in x
llvm-svn: 185954
The following transforms are valid if -C is a power of 2:
(icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
(icmp ult (xor X, C), -C) -> (icmp uge X, C)
These are nice, they get rid of the xor.
llvm-svn: 185915
Back in r179493 we determined that two transforms collided with each
other. The fix back then was to reorder the transforms so that the
preferred transform would give it a try and then we would try the
secondary transform. However, it was noted that the best approach would
canonicalize one transform into the other, removing the collision and
allowing us to optimize IR given to us in that form.
llvm-svn: 185808
This is a complete re-write if the bottom-up vectorization class.
Before this commit we scanned the instruction tree 3 times. First in search of merge points for the trees. Second, for estimating the cost. And finally for vectorization.
There was a lot of code duplication and adding the DCE exposed bugs. The new design is simpler and DCE was a part of the design.
In this implementation we build the tree once. After that we estimate the cost by scanning the different entries in the constructed tree (in any order). The vectorization phase also works on the built tree.
llvm-svn: 185774
This transform was originally added in r185257 but later removed in
r185415. The original transform would create instructions speculatively
and then discard them if the speculation was proved incorrect. This has
been replaced with a scheme that splits the transform into two parts:
preflight and fold. While we preflight, we build up fold actions that
inform the folding stage on how to act.
llvm-svn: 185667
This allows us to create switches even if instcombine has munged two of the
incombing compares into one and some bit twiddling. This was motivated by enum
compares that are common in clang.
llvm-svn: 185632
This implies annotating it as nounwind and its arguments as nocapture. To be
conservative, we do not annotate the arguments with noalias since some platforms
do not have restrict on the declaration for gettimeofday.
llvm-svn: 185502
I'm reverting this commit because:
1. As discussed during review, it needs to be rewritten (to avoid creating and
then deleting instructions).
2. This is causing optimizer crashes. Specifically, I'm seeing things like
this:
While deleting: i1 %
Use still stuck around after Def is destroyed: <badref> = select i1 <badref>, i32 0, i32 1
opt: /src/llvm-trunk/lib/IR/Value.cpp:79: virtual llvm::Value::~Value(): Assertion `use_empty() && "Uses remain when a value is destroyed!"' failed.
I'd guess that these will go away once we're no longer creating/deleting
instructions here, but just in case, I'm adding a regression test.
Because the code is bring rewritten, I've just XFAIL'd the original regression test. Original commit message:
InstCombine: Be more agressive optimizing 'udiv' instrs with 'select' denoms
Real world code sometimes has the denominator of a 'udiv' be a
'select'. LLVM can handle such cases but only when the 'select'
operands are symmetric in structure (both select operands are a constant
power of two or a left shift, etc.). This falls apart if we are dealt a
'udiv' where the code is not symetric or if the select operands lead us
to more select instructions.
Instead, we should treat the LHS and each select operand as a distinct
divide operation and try to optimize them independently. If we can
to simplify each operation, then we can replace the 'udiv' with, say, a
'lshr' that has a new select with a bunch of new operands for the
select.
llvm-svn: 185415
Math functions are mark as readonly because they read the floating point
rounding mode. Because we don't vectorize loops that would contain function
calls that set the rounding mode it is safe to ignore this memory read.
llvm-svn: 185299
Changing the sign when comparing the base pointer would introduce all
sorts of unexpected things like:
%gep.i = getelementptr inbounds [1 x i8]* %a, i32 0, i32 0
%gep2.i = getelementptr inbounds [1 x i8]* %b, i32 0, i32 0
%cmp.i = icmp ult i8* %gep.i, %gep2.i
%cmp.i1 = icmp ult [1 x i8]* %a, %b
%cmp = icmp ne i1 %cmp.i, %cmp.i1
ret i1 %cmp
into:
%cmp.i = icmp slt [1 x i8]* %a, %b
%cmp.i1 = icmp ult [1 x i8]* %a, %b
%cmp = xor i1 %cmp.i, %cmp.i1
ret i1 %cmp
By preserving the original sign, we now get:
ret i1 false
This fixes PR16483.
llvm-svn: 185259
Real world code sometimes has the denominator of a 'udiv' be a
'select'. LLVM can handle such cases but only when the 'select'
operands are symmetric in structure (both select operands are a constant
power of two or a left shift, etc.). This falls apart if we are dealt a
'udiv' where the code is not symetric or if the select operands lead us
to more select instructions.
Instead, we should treat the LHS and each select operand as a distinct
divide operation and try to optimize them independently. If we can
to simplify each operation, then we can replace the 'udiv' with, say, a
'lshr' that has a new select with a bunch of new operands for the
select.
llvm-svn: 185257
We may, after other optimizations, find ourselves with IR that looks
like:
%shl = shl i32 1, %y
%cmp = icmp ult i32 %shl, 32
Instead, we should just compare the shift count:
%cmp = icmp ult i32 %y, 5
llvm-svn: 185242
To support this we have to insert 'extractelement' instructions to pick the right lane.
We had this functionality before but I removed it when we moved to the multi-block design because it was too complicated.
llvm-svn: 185230
- lit tests verify that each line of input LLVM IR gets a !dbg node and a
corresponding entry of metadata that contains the line number
- unit tests verify that DebugIR works as advertised in the interface
- refactored some useful IR generation functionality from the MCJIT unit tests
so it can be reused
llvm-svn: 185212
No functionality change.
It should suffice to check the type of a debug info metadata, instead of
calling Verify. For cases where we know the type of a DI metadata, use
assert.
Also update testing cases to make them conform to the format of DI classes.
llvm-svn: 185135
When we store values for reversed induction stores we must not store the
reversed value in the vectorized value map. Another instruction might use this
value.
This fixes 3 test cases of PR16455.
llvm-svn: 185051
The Builtin attribute is an attribute that can be placed on function call site that signal that even though a function is declared as being a builtin,
rdar://problem/13727199
llvm-svn: 185049
When a 1-element vector alloca is promoted, a store instruction can often be
rewritten without converting the value to a scalar and using an insertelement
instruction to stuff it into the new alloca. This patch just adds a check
to skip that conversion when it is unnecessary. This turns out to be really
important for some ARM Neon operations where <1 x i64> is used to get around
the fact that i64 is not a legal type.
llvm-svn: 184870
This should hopefully have fixed the stage2/stage3 miscompare on the dragonegg
testers.
"LoopVectorize: Use the dependence test utility class
We now no longer need alias analysis - the cases that alias analysis would
handle are now handled as accesses with a large dependence distance.
We can now vectorize loops with simple constant dependence distances.
for (i = 8; i < 256; ++i) {
a[i] = a[i+4] * a[i+8];
}
for (i = 8; i < 256; ++i) {
a[i] = a[i-4] * a[i-8];
}
We would be able to vectorize about 200 more loops (in many cases the cost model
instructs us no to) in the test suite now. Results on x86-64 are a wash.
I have seen one degradation in ammp. Interestingly, the function in which we
now vectorize a loop is never executed so we probably see some instruction
cache effects. There is a 2% improvement in h264ref. There is one or the other
TSCV loop kernel that speeds up.
radar://13681598"
llvm-svn: 184724
We now no longer need alias analysis - the cases that alias analysis would
handle are now handled as accesses with a large dependence distance.
We can now vectorize loops with simple constant dependence distances.
for (i = 8; i < 256; ++i) {
a[i] = a[i+4] * a[i+8];
}
for (i = 8; i < 256; ++i) {
a[i] = a[i-4] * a[i-8];
}
We would be able to vectorize about 200 more loops (in many cases the cost model
instructs us no to) in the test suite now. Results on x86-64 are a wash.
I have seen one degradation in ammp. Interestingly, the function in which we
now vectorize a loop is never executed so we probably see some instruction
cache effects. There is a 2% improvement in h264ref. There is one or the other
TSCV loop kernel that speeds up.
radar://13681598
llvm-svn: 184685
Untill now we detected the vectorizable tree and evaluated the cost of the
entire tree. With this patch we can decide to trim-out branches of the tree
that are not profitable to vectorizer.
Also, increase the max depth from 6 to 12. In the worse possible case where all
of the code is made of diamond-shaped graph this can bring the cost to 2**10,
but diamonds are not very common.
llvm-svn: 184681
Rewrote the SLP-vectorization as a whole-function vectorization pass. It is now able to vectorize chains across multiple basic blocks.
It still does not vectorize PHIs, but this should be easy to do now that we scan the entire function.
I removed the support for extracting values from trees.
We are now able to vectorize more programs, but there are some serious regressions in many workloads (such as flops-6 and mandel-2).
llvm-svn: 184647
We collect gather sequences when we vectorize basic blocks. Gather sequences are excellent
hints for vectorization of other basic blocks.
llvm-svn: 184444
Prior to this change, the considered addressing modes may be invalid since the
maximum and minimum offsets were not taking into account.
This was causing an assertion failure.
The added test case exercices that behavior.
<rdar://problem/14199725> Assertion failed: (CurScaleCost >= 0 && "Legal
addressing mode has an illegal cost!")
llvm-svn: 184341
The type <3 x i8> is a common in graphics and we want to be able to vectorize it.
This changes accelerates bullet by 12% and 471_omnetpp by 5%.
llvm-svn: 184317
This pass was assuming that if hasAddressTaken() returns false for a
function, the function's only uses are call sites. That's not true
because there can be references by BlockAddresses too.
Fix the pass to handle this case. Fix
BlockAddress::replaceUsesOfWithOnConstant() to allow a function's type
to be changed by RAUW'ing the function with a bitcast of the recreated
function.
Patch by Mark Seaborn.
llvm-svn: 183933
Instead of a custom implementation of replaceAllUsesWith, we just call
replaceAllUsesWith and recreate llvm.used and llvm.compiler-used.
This change is particularity interesting because it makes llvm see
through what clang is doing with static used functions in extern "C"
contexts. With this change, running clang -O2 in
extern "C" {
__attribute__((used)) static void foo() {}
}
produces
@llvm.used = appending global [1 x i8*] [i8* bitcast (void ()* @foo to
i8*)], section "llvm.metadata"
define internal void @foo() #0 {
entry:
ret void
}
llvm-svn: 183756
r183584 tries to derive some info from the code *AFTER* a call and apply
these derived info to the code *BEFORE* the call, which is not always safe
as the call in question may never return, and in this case, the derived
info is invalid.
Thank Duncan for pointing out this potential bug.
rdar://14073661
llvm-svn: 183606
The MemCpyOpt pass is capable of optimizing:
callee(&S); copy N bytes from S to D.
into:
callee(&D);
subject to some legality constraints.
Assertion is triggered when the compiler tries to evalute "sizeof(typeof(D))",
while D is an opaque-typed, 'sret' formal argument of function being compiled.
i.e. the signature of the func being compiled is something like this:
T caller(...,%opaque* noalias nocapture sret %D, ...)
The fix is that when come across such situation, instead of calling some
utility functions to get the size of D's type (which will crash), we simply
assume D has at least N bytes as implified by the copy-instruction.
rdar://14073661
llvm-svn: 183584
IndVarSimplify is willing to move divide instructions outside of their
loop bodies if they are invariant of the loop. However, it may not be
safe to expand them if we do not know if they can trap.
Instead, check to see if it is not safe to expand the instruction and
skip the expansion.
This fixes PR16041.
Testcase by Rafael Ávila de Espíndola.
llvm-svn: 183239
The problem this time seems to be a thinko. We were assuming that in the CFG
A
| \
| B
| /
C
speculating the basic block B would cause only the phi value for the B->C edge
to be speculated. That is not true, the phi's are semantically in the edges, so
if the A->B->C path is taken, any code needed for A->C is not executed and we
have to consider it too when deciding to speculate B.
llvm-svn: 183226
PR16069 is an interesting case where an incoming value to a PHI is a
trap value while also being a 'ConstantExpr'.
We do not consider this case when performing the 'HoistThenElseCodeToIf'
optimization.
Instead, make our modifications more conservative if we detect that we
cannot transform the PHI to a select.
llvm-svn: 183152
index greater than the size of the vector is invalid. The shuffle may be
shrinking the size of the vector. Fixes a crash!
Also drop the maximum recursion depth of the safety check for this
optimization to five.
llvm-svn: 183080
Fixes rdar:14036816, PR16130.
There is an opportunity to compute precise trip counts for 'or'
expressions and multi-exit loops.
rdar:14038809: Optimize trip count computation for multi-exit loops.
To do this we need to record the fact that ExitLimit assumes NSW. When
it does not we can safely assume that the loop trip count is the
minimum ExitLimt across all subexpressions and loop exits.
llvm-svn: 183060
We check that instructions in the loop don't have outside users (except if
they are reduction values). Unfortunately, we skipped this check for
if-convertable PHIs.
Fixes PR16184.
llvm-svn: 183035
Namely, check if the target allows to fold more that one register in the
addressing mode and if yes, adjust the cost accordingly.
Prior to this commit, reg1 + scale * reg2 accesses were artificially preferred
to reg1 + reg2 accesses. Indeed, the cost model wrongly assumed that reg1 + reg2
needs a temporary register for the computation, whereas it was correctly
estimated for reg1 + scale * reg2.
<rdar://problem/13973908>
llvm-svn: 183021
- llvm.loop.parallel metadata has been renamed to llvm.loop to be more generic
by making the root of additional loop metadata.
- Loop::isAnnotatedParallel now looks for llvm.loop and associated
llvm.mem.parallel_loop_access
- document llvm.loop and update llvm.mem.parallel_loop_access
- add support for llvm.vectorizer.width and llvm.vectorizer.unroll
- document llvm.vectorizer.* metadata
- add utility class LoopVectorizerHints for getting/setting loop metadata
- use llvm.vectorizer.width=1 to indicate already vectorized instead of
already_vectorized
- update existing tests that used llvm.loop.parallel and
llvm.vectorizer.already_vectorized
Reviewed by: Nadav Rotem
llvm-svn: 182802
