Splitting critical edges when one of the source edges is an indirectbr
is hard in general (because it requires changing the memory the indirectbr
reads). But if a block only has a single indirectbr predecessor (which is
the common case), we can simulate splitting that edge by splitting
the destination block, and retargeting the *direct* branches.
This is motivated by the use of computed gotos in python 2.7: PyEval_EvalFrame()
ends up using an indirect branch with ~100 successors, and passing a constant to
each of those. Since MachineSink can't break indirect critical edges on demand
(and doing this in MIR doesn't look feasible), this causes us to emit about ~100
defs of registers containing constants, which we in the predecessor block, where
only one of those constants is used in each successor. So, at each computed goto,
we needlessly spill about a 100 constants to stack. The end result is that a
clang-compiled python interpreter can be about ~2.5x slower on a simple python
reduction loop than a gcc-compiled interpreter.
Differential Revision: https://reviews.llvm.org/D29916
llvm-svn: 296416
The transform in question claims to be doing:
// fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
...starting in PerformADDCombineWithOperands(), but it wasn't actually checking for a setcc node
for the sext/zext patterns.
This is exactly the opposite of a transform I'd like to add to DAGCombiner's foldSelectOfConstants(),
so I was seeing infinite loops with my draft of a patch applied.
The changes in select_const.ll look positive (less instructions). The change in arm-and-tst-peephole.ll
is unrelated. We're changing the input IR in that test to preserve the intent of the test, but that's
not affected by this code change.
Differential Revision:
https://reviews.llvm.org/D30355
llvm-svn: 296389
This pattern is essentially a i16 load from p+1 address:
%p1.i16 = bitcast i8* %p to i16*
%p2.i8 = getelementptr i8, i8* %p, i64 2
%v1 = load i16, i16* %p1.i16
%v2.i8 = load i8, i8* %p2.i8
%v2 = zext i8 %v2.i8 to i16
%v1.shl = shl i16 %v1, 8
%res = or i16 %v1.shl, %v2
Current implementation would identify %v1 load as the first byte load and would mistakenly emit a i16 load from %p1.i16 address. This patch adds a check that the first byte is loaded from a non-zero offset of the first load address. This way this address can be used as the base address for the combined value. Otherwise just give up combining.
llvm-svn: 296336
Recommiting after fixup of 32-bit aliasing sign offset bug in DAGCombiner.
* Simplify Consecutive Merge Store Candidate Search
Now that address aliasing is much less conservative, push through
simplified store merging search and chain alias analysis which only
checks for parallel stores through the chain subgraph. This is cleaner
as the separation of non-interfering loads/stores from the
store-merging logic.
When merging stores search up the chain through a single load, and
finds all possible stores by looking down from through a load and a
TokenFactor to all stores visited.
This improves the quality of the output SelectionDAG and the output
Codegen (save perhaps for some ARM cases where we correctly constructs
wider loads, but then promotes them to float operations which appear
but requires more expensive constant generation).
Some minor peephole optimizations to deal with improved SubDAG shapes (listed below)
Additional Minor Changes:
1. Finishes removing unused AliasLoad code
2. Unifies the chain aggregation in the merged stores across code
paths
3. Re-add the Store node to the worklist after calling
SimplifyDemandedBits.
4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is
arbitrary, but seems sufficient to not cause regressions in
tests.
5. Remove Chain dependencies of Memory operations on CopyfromReg
nodes as these are captured by data dependence
6. Forward loads-store values through tokenfactors containing
{CopyToReg,CopyFromReg} Values.
7. Peephole to convert buildvector of extract_vector_elt to
extract_subvector if possible (see
CodeGen/AArch64/store-merge.ll)
8. Store merging for the ARM target is restricted to 32-bit as
some in some contexts invalid 64-bit operations are being
generated. This can be removed once appropriate checks are
added.
This finishes the change Matt Arsenault started in r246307 and
jyknight's original patch.
Many tests required some changes as memory operations are now
reorderable, improving load-store forwarding. One test in
particular is worth noting:
CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store
forwarding converts a load-store pair into a parallel store and
a memory-realized bitcast of the same value. However, because we
lose the sharing of the explicit and implicit store values we
must create another local store. A similar transformation
happens before SelectionDAG as well.
Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle
llvm-svn: 296252
Splitting critical edges when one of the source edges is an indirectbr
is hard in general (because it requires changing the memory the indirectbr
reads). But if a block only has a single indirectbr predecessor (which is
the common case), we can simulate splitting that edge by splitting
the destination block, and retargeting the *direct* branches.
This is motivated by the use of computed gotos in python 2.7: PyEval_EvalFrame()
ends up using an indirect branch with ~100 successors, and passing a constant to
each of those. Since MachineSink can't break indirect critical edges on demand
(and doing this in MIR doesn't look feasible), this causes us to emit about ~100
defs of registers containing constants, which we in the predecessor block, where
only one of those constants is used in each successor. So, at each computed goto,
we needlessly spill about a 100 constants to stack. The end result is that a
clang-compiled python interpreter can be about ~2.5x slower on a simple python
reduction loop than a gcc-compiled interpreter.
Differential Revision: https://reviews.llvm.org/D29916
llvm-svn: 296149
The motivation for filling out these select-of-constants cases goes back to D24480,
where we discussed removing an IR fold from add(zext) --> select. And that goes back to:
https://reviews.llvm.org/rL75531https://reviews.llvm.org/rL159230
The idea is that we should always canonicalize patterns like this to a select-of-constants
in IR because that's the smallest IR and the best for value tracking. Note that we currently
do the opposite in some cases (like the cases in *this* patch). Ie, the proposed folds in
this patch already exist in InstCombine today:
https://github.com/llvm-mirror/llvm/blob/master/lib/Transforms/InstCombine/InstCombineSelect.cpp#L1151
As this patch shows, most targets generate better machine code for simple ext/add/not ops
rather than a select of constants. So the follow-up steps to make this less of a patchwork
of special-case folds and missing IR canonicalization:
1. Have DAGCombiner convert any select of constants into ext/add/not ops.
2 Have InstCombine canonicalize in the other direction (create more selects).
Differential Revision: https://reviews.llvm.org/D30180
llvm-svn: 296137
clang will generate IR like this for input using packed bitfields;
very simple semantically, but it's a bit tricky to actually
generate good code.
llvm-svn: 296080
Splitting critical edges when one of the source edges is an indirectbr
is hard in general (because it requires changing the memory the indirectbr
reads). But if a block only has a single indirectbr predecessor (which is
the common case), we can simulate splitting that edge by splitting
the destination block, and retargeting the *direct* branches.
This is motivated by the use of computed gotos in python 2.7: PyEval_EvalFrame()
ends up using an indirect branch with ~100 successors, and passing a constant to
each of those. Since MachineSink can't break indirect critical edges on demand
(and doing this in MIR doesn't look feasible), this causes us to emit about ~100
defs of registers containing constants, which we in the predecessor block, where
only one of those constants is used in each successor. So, at each computed goto,
we needlessly spill about a 100 constants to stack. The end result is that a
clang-compiled python interpreter can be about ~2.5x slower on a simple python
reduction loop than a gcc-compiled interpreter.
Differential Revision: https://reviews.llvm.org/D29916
llvm-svn: 296060
FastISel wasn't checking the isFPOnlySP subtarget feature before emitting
double-precision operations, so it got completely invalid CodeGen for doubles
on Cortex-M4F.
The normal ISel testing wasn't spectacular either so I added a second RUN line
to improve that while I was in the area.
llvm-svn: 296031
Introduce a common ValueHandler for call returns and formal arguments, and
inherit two different versions for handling the differences (at the moment the
only difference is the way physical registers are marked as used).
llvm-svn: 295973
Add support for lowering calls with parameters than can fit into regs. Use the
same ValueHandler that we used for function returns, but rename it to match its
new, extended purpose.
llvm-svn: 295971
The ARMConstantIslandPass didn't have support for handling accesses to
constant island objects through ARM::t2LDRBpci instructions. This adds
support for that.
This fixes PR31997.
llvm-svn: 295964
This patch adds missing sched classes for Thumb2 instructions.
This has been missing so far, and as a consequence, machine
scheduler models for individual sub-targets have tended to
be larger than they needed to be. These patches should help
write schedulers better and faster in the future
for ARM sub-targets.
Reviewer: Diana Picus
Differential Revision: https://reviews.llvm.org/D29953
llvm-svn: 295811
Summary:
This file was missed in the commit for Cortex-M23 and Cortex-M33
support. See https://reviews.llvm.org/D29073?id=85814 .
Reviewers: rengolin, javed.absar, samparker
Reviewed By: samparker
Subscribers: llvm-commits, aemerson
Differential Revision: https://reviews.llvm.org/D30162
llvm-svn: 295655
Start using the Subtarget to make decisions about what's legal. In particular,
we only mark floating point operations as legal if we have VFP2, which is
something we should've done from the very start.
llvm-svn: 295439
Resubmit -r295314 with PowerPC and AMDGPU tests updated.
Support {a|s}ext, {a|z|s}ext load nodes as a part of load combine patters.
Reviewed By: filcab
Differential Revision: https://reviews.llvm.org/D29591
llvm-svn: 295336
Support {a|s}ext, {a|z|s}ext load nodes as a part of load combine patters.
Reviewed By: filcab
Differential Revision: https://reviews.llvm.org/D29591
llvm-svn: 295314
Since they're only used for passing around double precision floating point
values into the general purpose registers, we'll lower them to VMOVDRR and
VMOVRRD.
llvm-svn: 295310
For now we just mark them as legal all the time and let the other passes bail
out if they can't handle it. In the future, we'll want to move more of the
brains into the legalizer.
llvm-svn: 295300
For the hard float calling convention, we just use the D registers.
For the soft-fp calling convention, we use the R registers and move values
to/from the D registers by means of G_SEQUENCE/G_EXTRACT. While doing so, we
make sure to honor the endianness of the target, since the CCAssignFn doesn't do
that for us.
For pure soft float targets, we still bail out because we don't support the
libcalls yet.
llvm-svn: 295295
Lay out trellis-shaped CFGs optimally.
A trellis of the shape below:
A B
|\ /|
| \ / |
| X |
| / \ |
|/ \|
C D
would be laid out A; B->C ; D by the current layout algorithm. Now we identify
trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an
increasing number of predecessors. A trellis is a a group of 2 or more
predecessor blocks that all have the same successors.
because of this we can tail duplicate to extend existing trellises.
As an example consider the following CFG:
B D F H
/ \ / \ / \ / \
A---C---E---G---Ret
Where A,C,E,G are all small (Currently 2 instructions).
The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret.
The current code will copy C into B, E into D and G into F and yield the layout
A,C,B(C),E,D(E),F(G),G,H,ret
define void @straight_test(i32 %tag) {
entry:
br label %test1
test1: ; A
%tagbit1 = and i32 %tag, 1
%tagbit1eq0 = icmp eq i32 %tagbit1, 0
br i1 %tagbit1eq0, label %test2, label %optional1
optional1: ; B
call void @a()
br label %test2
test2: ; C
%tagbit2 = and i32 %tag, 2
%tagbit2eq0 = icmp eq i32 %tagbit2, 0
br i1 %tagbit2eq0, label %test3, label %optional2
optional2: ; D
call void @b()
br label %test3
test3: ; E
%tagbit3 = and i32 %tag, 4
%tagbit3eq0 = icmp eq i32 %tagbit3, 0
br i1 %tagbit3eq0, label %test4, label %optional3
optional3: ; F
call void @c()
br label %test4
test4: ; G
%tagbit4 = and i32 %tag, 8
%tagbit4eq0 = icmp eq i32 %tagbit4, 0
br i1 %tagbit4eq0, label %exit, label %optional4
optional4: ; H
call void @d()
br label %exit
exit:
ret void
}
here is the layout after D27742:
straight_test: # @straight_test
; ... Prologue elided
; BB#0: # %entry ; A (merged with test1)
; ... More prologue elided
mr 30, 3
andi. 3, 30, 1
bc 12, 1, .LBB0_2
; BB#1: # %test2 ; C
rlwinm. 3, 30, 0, 30, 30
beq 0, .LBB0_3
b .LBB0_4
.LBB0_2: # %optional1 ; B (copy of C)
bl a
nop
rlwinm. 3, 30, 0, 30, 30
bne 0, .LBB0_4
.LBB0_3: # %test3 ; E
rlwinm. 3, 30, 0, 29, 29
beq 0, .LBB0_5
b .LBB0_6
.LBB0_4: # %optional2 ; D (copy of E)
bl b
nop
rlwinm. 3, 30, 0, 29, 29
bne 0, .LBB0_6
.LBB0_5: # %test4 ; G
rlwinm. 3, 30, 0, 28, 28
beq 0, .LBB0_8
b .LBB0_7
.LBB0_6: # %optional3 ; F (copy of G)
bl c
nop
rlwinm. 3, 30, 0, 28, 28
beq 0, .LBB0_8
.LBB0_7: # %optional4 ; H
bl d
nop
.LBB0_8: # %exit ; Ret
ld 30, 96(1) # 8-byte Folded Reload
addi 1, 1, 112
ld 0, 16(1)
mtlr 0
blr
The tail-duplication has produced some benefit, but it has also produced a
trellis which is not laid out optimally. With this patch, we improve the layouts
of such trellises, and decrease the cost calculation for tail-duplication
accordingly.
This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have
back edges, which is a negative, but it has a bigger compensating
positive, which is that it handles the case where there are long strings
of skipped blocks much better than the original layout. Both layouts
handle runs of executed blocks equally well. Branch prediction also
improves if there is any correlation between subsequent optional blocks.
Here is the resulting concrete layout:
straight_test: # @straight_test
; BB#0: # %entry ; A (merged with test1)
mr 30, 3
andi. 3, 30, 1
bc 12, 1, .LBB0_4
; BB#1: # %test2 ; C
rlwinm. 3, 30, 0, 30, 30
bne 0, .LBB0_5
.LBB0_2: # %test3 ; E
rlwinm. 3, 30, 0, 29, 29
bne 0, .LBB0_6
.LBB0_3: # %test4 ; G
rlwinm. 3, 30, 0, 28, 28
bne 0, .LBB0_7
b .LBB0_8
.LBB0_4: # %optional1 ; B (Copy of C)
bl a
nop
rlwinm. 3, 30, 0, 30, 30
beq 0, .LBB0_2
.LBB0_5: # %optional2 ; D (Copy of E)
bl b
nop
rlwinm. 3, 30, 0, 29, 29
beq 0, .LBB0_3
.LBB0_6: # %optional3 ; F (Copy of G)
bl c
nop
rlwinm. 3, 30, 0, 28, 28
beq 0, .LBB0_8
.LBB0_7: # %optional4 ; H
bl d
nop
.LBB0_8: # %exit
Differential Revision: https://reviews.llvm.org/D28522
llvm-svn: 295223
Summary:
Blocks ending in unreachable are typically cold because they end the
program or throw an exception, so merging them with other identical
blocks is usually profitable because it reduces the size of cold code.
MachineBlockPlacement generally does not arrange to fall through to such
blocks, so commoning these blocks will not introduce additional
unconditional branches.
Reviewers: hans, iteratee, haicheng
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D29153
llvm-svn: 295105