The code changes here are hopefully straightforward:
1. Use MachineInstruction flags to decide if FP ops can be reassociated
(use both "reassoc" and "nsz" to be consistent with IR transforms;
we probably don't need "nsz", but that's a safer interpretation of
the FMF).
2. Check that both nodes allow reassociation to change instructions.
This is a stronger requirement than we've usually implemented in
IR/DAG, but this is needed to solve the motivating bug (see below),
and it seems unlikely to impede optimization at this late stage.
3. Intersect/propagate MachineIR flags to enable further reassociation
in MachineCombiner.
We managed to make MachineCombiner flexible enough that no changes are
needed to that pass itself. So this patch should only affect x86
(assuming no other targets have implemented the hooks using MachineIR
flags yet).
The motivating example in PR43609 is another case of fast-math transforms
interacting badly with special FP ops created during lowering:
https://bugs.llvm.org/show_bug.cgi?id=43609
The special fadd ops used for converting int to FP assume that they will
not be altered, so those are created without FMF.
However, the MachineCombiner pass was being enabled for FP ops using the
global/function-level TargetOption for "UnsafeFPMath". We managed to run
instruction/node-level FMF all the way down to MachineIR sometime in the
last 1-2 years though, so we can do better now.
The test diffs require some explanation:
1. llvm/test/CodeGen/X86/fmf-flags.ll - no target option for unsafe math was
specified here, so MachineCombiner kicks in where it did not previously;
to make it behave consistently, we need to specify a CPU schedule model,
so use the default model, and there are no code diffs.
2. llvm/test/CodeGen/X86/machine-combiner.ll - replace the target option for
unsafe math with the equivalent IR-level flags, and there are no code diffs;
we can't remove the NaN/nsz options because those are still used to drive
x86 fmin/fmax codegen (special SDAG opcodes).
3. llvm/test/CodeGen/X86/pow.ll - similar to #1
4. llvm/test/CodeGen/X86/sqrt-fastmath.ll - similar to #1, but MachineCombiner
does some reassociation of the estimate sequence ops; presumably these are
perf wins based on latency/throughput (and we get some reduction of move
instructions too); I'm not sure how it affects numerical accuracy, but the
test reflects reality better now because we would expect MachineCombiner to
be enabled if the IR was generated via something like "-ffast-math" with clang.
5. llvm/test/CodeGen/X86/vec_int_to_fp.ll - this is the test added to model PR43609;
the fadds are not reassociated now, so we should get the expected results.
6. llvm/test/CodeGen/X86/vector-reduce-fadd-fast.ll - similar to #1
7. llvm/test/CodeGen/X86/vector-reduce-fmul-fast.ll - similar to #1
Differential Revision: https://reviews.llvm.org/D74851
In D41587, @mssimpso discovered that the order of some patterns for
AArch64 was sub-optimal. I thought a bit about how we could avoid that
case in the future. I do not think there is a need for evaluating all
patterns for now. But this patch adds an extra (expensive) check, that
evaluates the latencies of all patterns, and ensures that the latency
saved decreases for subsequent patterns.
This catches the sub-optimal order fixed in D41587, but I am not
entirely happy with the check, as it only applies to sub-optimal
patterns seen while building with EXPENSIVE_CHECKS on. It did not
discover any other sub-optimal pattern ordering.
Reviewers: Gerolf, spatel, mssimpso
Reviewed By: Gerolf, mssimpso
Differential Revision: https://reviews.llvm.org/D41766
llvm-svn: 323873
As part of the unification of the debug format and the MIR format, print
MBB references as '%bb.5'.
The MIR printer prints the IR name of a MBB only for block definitions.
* find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)->getNumber\(\)/" << printMBBReference(*\1)/g'
* find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#" << ([a-zA-Z0-9_]+)\.getNumber\(\)/" << printMBBReference(\1)/g'
* find . \( -name "*.txt" -o -name "*.s" -o -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E 's/BB#([0-9]+)/%bb.\1/g'
* grep -nr 'BB#' and fix
Differential Revision: https://reviews.llvm.org/D40422
llvm-svn: 319665
This version of the patch fixes an off-by-one error causing PR34596. We
do not need to use std::next(BlockIter) when calling updateDepths, as
BlockIter already points to the next element.
Original commit message:
> For large basic blocks with lots of combinable instructions, the
> MachineTraceMetrics computations in MachineCombiner can dominate the compile
> time, as computing the trace information is quadratic in the number of
> instructions in a BB and it's relevant successors/predecessors.
> In most cases, knowing the instruction depth should be enough to make
> combination decisions. As we already iterate over all instructions in a basic
> block, the instruction depth can be computed incrementally. This reduces the
> cost of machine-combine drastically in cases where lots of instructions
> are combined. The major drawback is that AFAIK, computing the critical path
> length cannot be done incrementally. Therefore we only compute
> instruction depths incrementally, for basic blocks with more
> instructions than inc_threshold. The -machine-combiner-inc-threshold
> option can be used to set the threshold and allows for easier
> experimenting and checking if using incremental updates for all basic
> blocks has any impact on the performance.
>
> Reviewers: sanjoy, Gerolf, MatzeB, efriedma, fhahn
>
> Reviewed By: fhahn
>
> Subscribers: kiranchandramohan, javed.absar, efriedma, llvm-commits
>
> Differential Revision: https://reviews.llvm.org/D36619
llvm-svn: 313751
This caused PR34596.
> [MachineCombiner] Update instruction depths incrementally for large BBs.
>
> Summary:
> For large basic blocks with lots of combinable instructions, the
> MachineTraceMetrics computations in MachineCombiner can dominate the compile
> time, as computing the trace information is quadratic in the number of
> instructions in a BB and it's relevant successors/predecessors.
>
> In most cases, knowing the instruction depth should be enough to make
> combination decisions. As we already iterate over all instructions in a basic
> block, the instruction depth can be computed incrementally. This reduces the
> cost of machine-combine drastically in cases where lots of instructions
> are combined. The major drawback is that AFAIK, computing the critical path
> length cannot be done incrementally. Therefore we only compute
> instruction depths incrementally, for basic blocks with more
> instructions than inc_threshold. The -machine-combiner-inc-threshold
> option can be used to set the threshold and allows for easier
> experimenting and checking if using incremental updates for all basic
> blocks has any impact on the performance.
>
> Reviewers: sanjoy, Gerolf, MatzeB, efriedma, fhahn
>
> Reviewed By: fhahn
>
> Subscribers: kiranchandramohan, javed.absar, efriedma, llvm-commits
>
> Differential Revision: https://reviews.llvm.org/D36619
llvm-svn: 313213
Summary:
For large basic blocks with lots of combinable instructions, the
MachineTraceMetrics computations in MachineCombiner can dominate the compile
time, as computing the trace information is quadratic in the number of
instructions in a BB and it's relevant successors/predecessors.
In most cases, knowing the instruction depth should be enough to make
combination decisions. As we already iterate over all instructions in a basic
block, the instruction depth can be computed incrementally. This reduces the
cost of machine-combine drastically in cases where lots of instructions
are combined. The major drawback is that AFAIK, computing the critical path
length cannot be done incrementally. Therefore we only compute
instruction depths incrementally, for basic blocks with more
instructions than inc_threshold. The -machine-combiner-inc-threshold
option can be used to set the threshold and allows for easier
experimenting and checking if using incremental updates for all basic
blocks has any impact on the performance.
Reviewers: sanjoy, Gerolf, MatzeB, efriedma, fhahn
Reviewed By: fhahn
Subscribers: kiranchandramohan, javed.absar, efriedma, llvm-commits
Differential Revision: https://reviews.llvm.org/D36619
llvm-svn: 312719
This is one of the problems noted in PR25016:
https://llvm.org/bugs/show_bug.cgi?id=25016
and:
http://lists.llvm.org/pipermail/llvm-dev/2015-October/090998.html
The spilling problem is independent and not addressed by this patch.
The MachineCombiner was doing reassociations that don't improve or even worsen the critical path.
This is caused by inclusion of the "slack" factor when calculating the critical path of the original
code sequence. If we don't add that, then we have a more conservative cost comparison of the old code
sequence vs. a new sequence. The more liberal calculation must be preserved, however, for the AArch64
MULADD patterns because benchmark regressions were observed without that.
The two failing test cases now have identical asm that does what we want:
a + b + c + d ---> (a + b) + (c + d)
Differential Revision: http://reviews.llvm.org/D13417
llvm-svn: 252616
Extend the reassociation optimization of http://reviews.llvm.org/rL240361 (D10460)
to SSE scalar FP SP adds in addition to AVX scalar FP SP adds.
With the 'switch' in place, we can trivially add other opcodes and test cases in
future patches.
Differential Revision: http://reviews.llvm.org/D10975
llvm-svn: 241515
Currently ( D10321, http://reviews.llvm.org/rL239486 ), we can use the machine combiner pass
to reassociate the following sequence to reduce the critical path:
A = ? op ?
B = A op X
C = B op Y
-->
A = ? op ?
B = X op Y
C = A op B
'op' is currently limited to x86 AVX scalar FP adds (with fast-math on), but in theory, it could
be any associative math/logic op (see TODO in code comment).
This patch generalizes the pattern match to ignore the instruction that defines 'A'. So instead of
a sequence of 3 adds, we now only need to find 2 dependent adds and decide if it's worth
reassociating them.
This generalization has a compile-time cost because we can now match more instruction sequences
and we rely more heavily on the machine combiner to discard sequences where reassociation doesn't
improve the critical path.
For example, in the new test case:
A = M div N
B = A add X
C = B add Y
We'll match 2 reassociation patterns, but this transform doesn't reduce the critical path:
A = M div N
B = A add Y
C = B add X
We need the combiner to reject that pattern but select this:
A = M div N
B = X add Y
C = B add A
Differential Revision: http://reviews.llvm.org/D10460
llvm-svn: 240361
