MemorySSAUpdater::getPreviousDefRecursive is a recursive algorithm, for
each block, it computes the previous definition for each predecessor,
then takes those definitions and combines them. But currently it doesn't
remember results which it already computed; this means it can visit the
same block multiple times, which adds up to exponential time overall.
To fix this, this patch adds a cache. If we computed the result for a
block already, we don't need to visit it again because we'll come up
with the same result. Well, unless we RAUW a MemoryPHI; in that case,
the TrackingVH will be updated automatically.
This matches the original source paper for this algorithm.
The testcase isn't really a test for the bug, but it adds coverage for
the case where tryRemoveTrivialPhi erases an existing PHI node. (It's
hard to write a good regression test for a performance issue.)
Differential Revision: https://reviews.llvm.org/D44715
llvm-svn: 328577
Summary:
Re-lands r328386 and r328443, reverting r328482.
Incorporates fixes from @mstorsjo in D44876 (thanks!) so that small
parameters in i8 and i16 do not end up in the SysV register parameters
(EDI, ESI, etc).
I added tests for how we receive small parameters, since that is the
important part. It's always safe to store more bytes than will be read,
but the assumptions you make when loading them are what really matter.
I also tested this by self-hosting clang and it passed tests on win64.
Reviewers: mstorsjo, hans
Subscribers: hiraditya, mstorsjo, llvm-commits
Differential Revision: https://reviews.llvm.org/D44900
llvm-svn: 328570
Give the bit count instructions their own scheduler classes instead of forcing them into existing classes.
These were mostly overridden anyway, but I had to add in costs from Agner for silvermont and znver1 and the Fam16h SoG for btver2 (Jaguar).
Differential Revision: https://reviews.llvm.org/D44879
llvm-svn: 328566
Summary:
r327219 added wrappers to std::sort which randomly shuffle the container before sorting.
This will help in uncovering non-determinism caused due to undefined sorting
order of objects having the same key.
To make use of that infrastructure we need to invoke llvm::sort instead of std::sort.
Note: This patch is one of a series of patches to replace *all* std::sort to llvm::sort.
Refer the comments section in D44363 for a list of all the required patches.
Reviewers: dblaikie, RKSimon, robertlytton
Reviewed By: robertlytton
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D44875
llvm-svn: 328564
Legalize and emit code for quad-precision floating point operation xscvdpqp
and add option to guard the quad precision operation support.
Differential Revision: https://reviews.llvm.org/D44746
llvm-svn: 328558
A new function getOpcodeForSpill should now be the only place to get
the opcode for a given spilled register.
Differential Revision: https://reviews.llvm.org/D43086
llvm-svn: 328556
First, we change the heuristic that is used to ignore the recurrent
node-sets in the node ordering. In certain cases it's not important
to focus on the recurrent node-sets. Instead, the algorithm begins
by considering all the instructions in the node ordering step.
Second, a minor change to the bottom up traversal, which needs to
consider loop carried dependences (modeled as anti dependences).
Previously, these instructions were skipped, which caused problems
because the instruction ends up having both predecessors and
sucessors in the schedule.
Third, consider anti-dependences as a tie breaker when choosing
between instructions in the node ordering. We want to make sure
that the source of the anti-dependence does not end up with both
predecesssors and sucessors in the final node ordering.
Patch by Brendon Cahoon.
llvm-svn: 328554
Summary:
llvm-objdump now disassembles unrecognised opcodes as data, using
the .long directive. We treat unrecognised opcodes as being 32 bit
values, so move along 4 bytes rather than the single byte which
previously resulted in a cascade of bogus disassembly following an
unrecognised opcode.
While no solution can always disassemble code that contains
embedded data correctly this provides a significant improvement.
The disassembler will now cope with an arbitrary length section
as it no longer truncates it to a multiple of 4 bytes, and will
use the .byte directive for trailing bytes.
Subscribers: arsenm, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D44685
llvm-svn: 328553
The pipeliner must add a loop carried dependence between two memory
operations if the base register is not an affine (linear) exression.
The current implementation doesn't check how the base register is
defined, which allows non-affine expressions, and then the pipeliner
does not add a loop carried dependence when one is needed.
This patch adds code to isLoopCarriedOrder that checks if the base
register of the memory operations is defined by a phi, and the loop
definition for the phi is a constant increment value. This is a very
simple check for a linear expression.
Patch by Brendon Cahoon.
llvm-svn: 328550
The pipeliner is not adding a dependence edge for a loop carried
dependence, and ends up scheduling a load from iteration n prior
to an aliased store in iteration n-1.
The code that adds the loop carried dependences in the pipeliner
doesn't check if the memory objects for loads and stores are
"identified" (i.e., distinct) objects. If they are not, then the
code that adds the dependences needs to be conservative. The
objects can be used to check dependences only when they are
distinct objects.
The code that checks for loop carried dependences has been updated
to classify loads and stores that are not identified as "unknown"
values. A store with an "unknown" value can potentially create
a loop carried dependence with any pending load.
Patch by Brendon Cahoon.
llvm-svn: 328547
The phi renaming code in the pipeliner uses the wrong value when
rewriting phi uses, which results in an undefined value. In this
case, the original phi is no longer needed due to the order of
instruction in the pipelined loop. The pipeliner was assuming, in
this case, the the phi loop definition should be used to
rewrite the uses. However, the pipeliner needs to check to make
sure that the loop definition has already been scheduled. If not,
then the phi initial value needs to be used instead.
Patch by Brendon Cahoon.
llvm-svn: 328545
The pipeliner was generating too many phis in the epilog blocks, which
caused incorrect code generation when rewriting an instruction that uses
the phi.
In this case, there 3 prolog and epilog stages. An existing phi was
scheduled at stage 1. When generating the code for the 2nd epilog an
extra new phi was generated.
To fix this, we need to update the code that calculates the maximum
number of phis that can be generated, which is based upon the current
prolog stage and the stage of the original phi. In this case, when the
prolog stage is 1 and the original phi stage is 1, the maximum number
of phis to generate is 2.
Patch by Brendon Cahoon.
llvm-svn: 328543
The patch contains severals changes needed to pipeline an example
that was transformed so that a Phi with a subreg is converted to
copies.
The pipeliner wasn't working for a couple of reasons.
- The RecMII was 3 instead of 2 due to the extra copies.
- Copy instructions contained a latency of 1.
- The node order algorithm was not choosing the best "bottom"
node, which caused an instruction to be scheduled that had a
predecessor and successor already scheduled.
- Updated the Hexagon Machine Scheduler to check if the node is
latency bound when adding the cost for a 0-latency dependence.
The RecMII was 3 because the computation looks at the number of
nodes in the recurrence. The extra copy is an extra node but
it shouldn't increase the latency. The new RecMII computation
looks at the latency of the instructions in the recurrence. We
changed the latency of the dependence of a copy to 0. The latency
computation for the copy also checks the use of the copy (similar
to a reg_sequence).
The node order algorithm was not choosing the last instruction
in the recurrence for a bottom up traversal. This was when the
last instruction is a copy. A check was added when choosing the
instruction to check for NodeNum if the maxASAP is the same. This
means that the scheduler will not end up with another node in
the recurrence that has both a predecessor and successor already
scheduled.
The cost computation in Hexagon Machine Scheduler adds cost when
an instruction can be packetized with a zero-latency instruction.
We should only do this if the schedule is latency bound.
Patch by Brendon Cahoon.
llvm-svn: 328542
The pipeliner is asserting because the serialization step that
occurs at the end is deleting an instruction. The assert
occurs later on because there is a use without a definition.
The problem occurs when an instruction defines a value used
by a REQ_SEQUENCE and that value is used by a COPY instruction.
The latencies between these instructions are zero, so they are
put in to the same packet. The serialization code is unable to
handle this correctly, and ends up putting the REG_SEQUENCE
before its definition.
There is special code in the serialization step that attempts
to handle zero-cost instructions (phis, copy, reg_sequence)
differently than regular instructions. Unfortunately, this means
the order does not come out correct.
This patch simplifies the code by changing the seperate steps for
handling zero-cost and regular instructions. Only phis are
handled separate now, since they should occurs first. Then, this
patch adds checks to make use the MoveUse is set to the smallest
value if there are multiple uses in a cycle.
Patch by Brendon Cahoon.
llvm-svn: 328540
This change brings performance of zlib up by 10%. The example below is from a
hot loop in longest_match() from zlib.
do.body:
%cur_match.addr.0 = phi i32 [ %cur_match, %entry ], [ %2, %do.cond ]
%idx.ext = zext i32 %cur_match.addr.0 to i64
%add.ptr = getelementptr inbounds i8, i8* %win, i64 %idx.ext
%add.ptr2 = getelementptr inbounds i8, i8* %add.ptr, i64 %idx.ext1
%add.ptr3 = getelementptr inbounds i8, i8* %add.ptr2, i64 -1
In this example %idx.ext1 is a loop invariant. It will be moved above the use of
loop induction variable %idx.ext such that it can be hoisted out of the loop by
LICM. The operands that have dependences carried by the loop will be sinked down
in the GEP chain. This patch will produce the following output:
do.body:
%cur_match.addr.0 = phi i32 [ %cur_match, %entry ], [ %2, %do.cond ]
%idx.ext = zext i32 %cur_match.addr.0 to i64
%add.ptr = getelementptr inbounds i8, i8* %win, i64 %idx.ext1
%add.ptr2 = getelementptr inbounds i8, i8* %add.ptr, i64 -1
%add.ptr3 = getelementptr inbounds i8, i8* %add.ptr2, i64 %idx.ext
llvm-svn: 328539
The pipeliner changes dependences between base+offset instructions
(loads and stores) so that the instructions have more flexibility
to be scheduled with respect to each other. This occurs when the
pipeliner is able to compute that the instructions will not alias
if their order is changed. The prevous code enforced the alias
property by checking if the base register is the same, and that the
offset values are either both positive or negative.
This patch improves the alias check by using the API
areMemAccessesTriviallyDisjoint instead. This enables more cases,
especially if the offset is a negative value. The pipeliner uses
the function by creating a new instruction with the offset used
in the next iteration.
Patch by Brendon Cahoon.
llvm-svn: 328538
A schedule may require that a phi from the original loop is used in
multiple iterations in the scheduled loop. When this occurs, we generate
multiple phis in the pipelined loop to save the value across iterations.
When we generate the new phis and update the register names in the
pipelined loop, the pipeliner attempts to reuse a previously generated
phi, when possible. The calculation for the name of the new phi needs
to account for the version/iteration of the original phi. Also, in the
epilog, the code only needs to check backwards for a previous iteration
until reaching the first prolog block.
Patch by Brendon Cahoon.
llvm-svn: 328537
The code in orderDepdences that looks at the order dependences between
instructions was processing all the successor and predecessor order
dependences. However, we really only want to check for an order dependence
for instructions scheduled in the same cycle.
Also, fixed how the pipeliner handles output dependences. An output
dependence is also a potential loop carried dependence. The pipeliner
didn't handle this case properly so an invalid schedule could be created
that allowed an output dependence to be scheduled in the next iteration
at the same cycle.
Patch by Brendon Cahoon.
llvm-svn: 328516
When the definition of a phi is used by a phi in the next iteration,
the pipeliner was assuming that the definition is processed first.
Because of the assumption, an incorrect phi name was used. This patch
has a check to see if the phi definition has been processed already.
Patch by Brendon Cahoon.
llvm-svn: 328510
The software pipeliner attempts to delete dead instructions after
generating the pipelined loop. The code looks for uses of each
instruction. Physical registers should be treated differently because
the use chains do not exist. The code that checks for dead
instructions should assume that definitions of physical registers
are used if the operand doesn't contain the dead flag.
Patch by Brendon Cahoon.
llvm-svn: 328509
The pipeliner needs to be conservative when updating the memoperands
of instructions in the epilog. Previously, the pipeliner was changing
the offset of the memoperand based upon the scheduling stage. However,
that is incorrect when control flow branches around the kernel code.
The bug enabled a load and store to the same stack offset to be swapped.
This patch fixes the bug by updating the size of the memoperands to be
UINT_MAX. This conservative value means that dependences will be created
between other loads and stores.
Patch by Brendon Cahoon.
llvm-svn: 328508
This replaces a large chunk of code that was looking for compound
patterns that include these sub-patterns. Existing tests ensure that
all of the previous examples are still folded as expected.
We still need to loosen the FMF check.
llvm-svn: 328502
Implement TTI interface for targets to indicate that the LSR should give
priority to post-incrementing addressing modes.
Combination of patches by Sebastian Pop and Brendon Cahoon.
Differential Revision: https://reviews.llvm.org/D44758
llvm-svn: 328490
Current logic of loop SCEV invalidation in Loop Unroller implicitly relies on
fact that exit count of outer loops cannot rely on exiting blocks of
inner loops, which is true in current implementation of backedge taken count
calculation but is wrong in general. As result, when we only forget the loop that
we have just unrolled, we may still have cached data for its outer loops (in particular,
exit counts) which keeps references on blocks of inner loop that could have been
changed or even deleted.
The attached test demonstrates a situaton when after unrolling of innermost loop
the outermost loop contains a dangling pointer on non-existant block. The problem
shows up when we apply patch https://reviews.llvm.org/D44677 that makes SCEV
smarter about exit count calculation. I am not sure if the bug exists without this patch,
it appears that now it is accidentally correct just because in practice exact backedge
taken count for outer loops with complex control flow inside is never calculated.
But when SCEV learns to do so, this problem shows up.
This patch replaces existing logic of SCEV loop invalidation with a correct one, which
happens to be invalidation of outermost loop (which also leads to invalidation of all
loops inside of it). It is the only way to ensure that no outer loop keeps dangling pointers
on removed blocks, or just outdated information that has changed after unrolling.
Differential Revision: https://reviews.llvm.org/D44818
Reviewed By: samparker
llvm-svn: 328483
This broke Chromium (see crbug.com/825748). It looks like mstorsjo's follow-up
patch at D44876 fixes this, but let's revert back to green for now until that's
ready to land.
(Also reverts r328443.)
> Both GCC and MSVC only look at the low byte of a boolean when it is
> passed.
llvm-svn: 328482
CanBeMin is currently used which will report true for any unknown
values, but often a check is performed outside the loop which covers
this situation:
for (int i = 0; i < N; ++i)
...
if (N > 0)
for (int i = 0; i < N; ++i)
...
So I've add 'LoopGuardedAgainstMin' which reports whether N is
greater than the minimum value which then allows loop with a variable
loop count to be optimised. I've also moved the increasing bound
checking into its own function and replaced SumCanReachMax is another
isLoopEntryGuardedByCond function.
llvm-svn: 328480
Paul Robinson