Summary: ARMv6m supports dmb etc fench instructions but not ldrex/strex etc. So for some atomic load/store, LLVM should inline instructions instead of lowering to __sync_ calls.
Reviewers: rengolin, efriedma, t.p.northover, jmolloy
Subscribers: efriedma, aemerson, llvm-commits
Differential Revision: https://reviews.llvm.org/D26120
llvm-svn: 285969
This recommits r281323, which was backed out for two reasons. One, a selfhost failure, and two, it apparently caused Chromium failures. Actually, the latter was a red herring. The log has expired from the former, but I suspect that was a red herring too (actually caused by another problematic patch of mine). Therefore reapplying, and will watch the bots like a hawk.
For the common pattern (CMPZ (AND x, #bitmask), #0), we can do some more efficient instruction selection if the bitmask is one consecutive sequence of set bits (32 - clz(bm) - ctz(bm) == popcount(bm)).
1) If the bitmask touches the LSB, then we can remove all the upper bits and set the flags by doing one LSLS.
2) If the bitmask touches the MSB, then we can remove all the lower bits and set the flags with one LSRS.
3) If the bitmask has popcount == 1 (only one set bit), we can shift that bit into the sign bit with one LSLS and change the condition query from NE/EQ to MI/PL (we could also implement this by shifting into the carry bit and branching on BCC/BCS).
4) Otherwise, we can emit a sequence of LSLS+LSRS to remove the upper and lower zero bits of the mask.
1-3 require only one 16-bit instruction and can elide the CMP. 4 requires two 16-bit instructions but can elide the CMP and doesn't require materializing a complex immediate, so is also a win.
llvm-svn: 285893
[Reapplying r284580 and r285917 with fix and testing to ensure emitted jump tables for Thumb-1 have 4-byte alignment]
The TBB and TBH instructions in Thumb-2 allow jump tables to be compressed into sequences of bytes or shorts respectively. These instructions do not exist in Thumb-1, however it is possible to synthesize them out of a sequence of other instructions.
It turns out this sequence is so short that it's almost never a lose for performance and is ALWAYS a significant win for code size.
TBB example:
Before: lsls r0, r0, #2 After: add r0, pc
adr r1, .LJTI0_0 ldrb r0, [r0, #6]
ldr r0, [r0, r1] lsls r0, r0, #1
mov pc, r0 add pc, r0
=> No change in prologue code size or dynamic instruction count. Jump table shrunk by a factor of 4.
The only case that can increase dynamic instruction count is the TBH case:
Before: lsls r0, r4, #2 After: lsls r4, r4, #1
adr r1, .LJTI0_0 add r4, pc
ldr r0, [r0, r1] ldrh r4, [r4, #6]
mov pc, r0 lsls r4, r4, #1
add pc, r4
=> 1 more instruction in prologue. Jump table shrunk by a factor of 2.
So there is an argument that this should be disabled when optimizing for performance (and a TBH needs to be generated). I'm not so sure about that in practice, because on small cores with Thumb-1 performance is often tied to code size. But I'm willing to turn it off when optimizing for performance if people want (also note that TBHs are fairly rare in practice!)
llvm-svn: 285690
Generate the slowest possible codepath for noopt CodeGen. Even trying to be
clever with the negated jump can cause out-of-range jumps. Use a wide branch
instead. Although the code is modelled simplistically, the later optimizations
would recombine the branching into `cbz` if possible. This re-enables the
previous optimization as well as hopefully gives us working code in all cases.
Addresses PR30356!
llvm-svn: 285649
The Windows ARM target expects the compiler to emit a division-by-zero check.
The check would use the form of:
cmp r?, #0
cbz .Ltrap
b .Lbody
.Lbody:
...
.Ltrap:
udf #249 @ __brkdiv0
This works great most of the time. However, if the body of the function is
greater than 127 bytes, the branch target limitation of cbz becomes an issue.
This occurs in the unoptimized code generation cases sometimes (like in
compiler-rt).
Since this is a matter of correctness, possibly pay a small penalty instead. We
now form this slightly differently:
cbnz .Lbody
udf #249 @ __brkdiv0
.Lbody:
...
The positive case is through the branch instead of being the next instruction.
However, because of the basic block layout, the negated branch is going to be
a short distance always (2 bytes away, after the inserted __brkdiv0).
The new t__brkdiv0 instruction is required to explicitly mark the instruction as
a terminator as the generic UDF instruction is not a terminator.
Addresses PR30532!
llvm-svn: 285312
UMAAL is a DSP instruction and it is not available on thumbv7m
(Cortex-M3) and thumbv6m (Cortex-M0+1) targets. Also fix wrong
CHECK prefix in longMAC.ll test.
Patch by Vadzim Dambrouski.
Differential Revision: https://reviews.llvm.org/D25890
llvm-svn: 285278
It would be a very nice invariant to rely on, but unfortunately it doesn't
necessarily hold (and the causes of mis-sorted reglists appear to be quite
varied) so to be robust the frame lowering code can't assume that the first
register in the list is also the first one that actually gets pushed.
Should fix an issue where we were turning something like:
push {r8, r4, r7, lr}
sub sp, #24
into nonsense like:
push {r2, r3, r4, r5, r6, r7, r8, r4, r7, lr}
llvm-svn: 285232
Summary:
The original heuristic to break critical edge during machine sink is relatively conservertive: when there is only one instruction sinkable to the critical edge, it is likely that the machine sink pass will not break the critical edge. This leads to many speculative instructions executed at runtime. However, with profile info, we could model the splitting benefits: if the critical edge has 50% taken rate, it would always be beneficial to split the critical edge to avoid the speculated runtime instructions. This patch uses profile to guide critical edge splitting in machine sink pass.
The performance impact on speccpu2006 on Intel sandybridge machines:
spec/2006/fp/C++/444.namd 25.3 +0.26%
spec/2006/fp/C++/447.dealII 45.96 -0.10%
spec/2006/fp/C++/450.soplex 41.97 +1.49%
spec/2006/fp/C++/453.povray 36.83 -0.96%
spec/2006/fp/C/433.milc 23.81 +0.32%
spec/2006/fp/C/470.lbm 41.17 +0.34%
spec/2006/fp/C/482.sphinx3 48.13 +0.69%
spec/2006/int/C++/471.omnetpp 22.45 +3.25%
spec/2006/int/C++/473.astar 21.35 -2.06%
spec/2006/int/C++/483.xalancbmk 36.02 -2.39%
spec/2006/int/C/400.perlbench 33.7 -0.17%
spec/2006/int/C/401.bzip2 22.9 +0.52%
spec/2006/int/C/403.gcc 32.42 -0.54%
spec/2006/int/C/429.mcf 39.59 +0.19%
spec/2006/int/C/445.gobmk 26.98 -0.00%
spec/2006/int/C/456.hmmer 24.52 -0.18%
spec/2006/int/C/458.sjeng 28.26 +0.02%
spec/2006/int/C/462.libquantum 55.44 +3.74%
spec/2006/int/C/464.h264ref 46.67 -0.39%
geometric mean +0.20%
Manually checked 473 and 471 to verify the diff is in the noise range.
Reviewers: rengolin, davidxl
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D24818
llvm-svn: 284757
This code crashed on funclet-style EH instructions such as catchpad,
catchswitch, and cleanuppad. Just treat all EH pad instructions
equivalently and avoid merging the globals they reference through any
use.
llvm-svn: 284633
Use mask and negate for legalization of i1 source type with SIGN_EXTEND_INREG.
With the mask, this should be no worse than 2 shifts. The mask can be eliminated
in some cases, so that should be better than 2 shifts.
This change exposed some missing folds related to negation:
https://reviews.llvm.org/rL284239https://reviews.llvm.org/rL284395
There may be others, so please let me know if you see any regressions.
Differential Revision: https://reviews.llvm.org/D25485
llvm-svn: 284611
The TBB and TBH instructions in Thumb-2 allow jump tables to be compressed into sequences of bytes or shorts respectively. These instructions do not exist in Thumb-1, however it is possible to synthesize them out of a sequence of other instructions.
It turns out this sequence is so short that it's almost never a lose for performance and is ALWAYS a significant win for code size.
TBB example:
Before: lsls r0, r0, #2 After: add r0, pc
adr r1, .LJTI0_0 ldrb r0, [r0, #6]
ldr r0, [r0, r1] lsls r0, r0, #1
mov pc, r0 add pc, r0
=> No change in prologue code size or dynamic instruction count. Jump table shrunk by a factor of 4.
The only case that can increase dynamic instruction count is the TBH case:
Before: lsls r0, r4, #2 After: lsls r4, r4, #1
adr r1, .LJTI0_0 add r4, pc
ldr r0, [r0, r1] ldrh r4, [r4, #6]
mov pc, r0 lsls r4, r4, #1
add pc, r4
=> 1 more instruction in prologue. Jump table shrunk by a factor of 2.
So there is an argument that this should be disabled when optimizing for performance (and a TBH needs to be generated). I'm not so sure about that in practice, because on small cores with Thumb-1 performance is often tied to code size. But I'm willing to turn it off when optimizing for performance if people want (also note that TBHs are fairly rare in practice!)
llvm-svn: 284580
This renames the function for checking FP function attribute values and also
adds more build attribute tests (which are in separate files because build
attributes are set per file).
Differential Revision: https://reviews.llvm.org/D25625
llvm-svn: 284571
Summary:
The original heuristic to break critical edge during machine sink is relatively conservertive: when there is only one instruction sinkable to the critical edge, it is likely that the machine sink pass will not break the critical edge. This leads to many speculative instructions executed at runtime. However, with profile info, we could model the splitting benefits: if the critical edge has 50% taken rate, it would always be beneficial to split the critical edge to avoid the speculated runtime instructions. This patch uses profile to guide critical edge splitting in machine sink pass.
The performance impact on speccpu2006 on Intel sandybridge machines:
spec/2006/fp/C++/444.namd 25.3 +0.26%
spec/2006/fp/C++/447.dealII 45.96 -0.10%
spec/2006/fp/C++/450.soplex 41.97 +1.49%
spec/2006/fp/C++/453.povray 36.83 -0.96%
spec/2006/fp/C/433.milc 23.81 +0.32%
spec/2006/fp/C/470.lbm 41.17 +0.34%
spec/2006/fp/C/482.sphinx3 48.13 +0.69%
spec/2006/int/C++/471.omnetpp 22.45 +3.25%
spec/2006/int/C++/473.astar 21.35 -2.06%
spec/2006/int/C++/483.xalancbmk 36.02 -2.39%
spec/2006/int/C/400.perlbench 33.7 -0.17%
spec/2006/int/C/401.bzip2 22.9 +0.52%
spec/2006/int/C/403.gcc 32.42 -0.54%
spec/2006/int/C/429.mcf 39.59 +0.19%
spec/2006/int/C/445.gobmk 26.98 -0.00%
spec/2006/int/C/456.hmmer 24.52 -0.18%
spec/2006/int/C/458.sjeng 28.26 +0.02%
spec/2006/int/C/462.libquantum 55.44 +3.74%
spec/2006/int/C/464.h264ref 46.67 -0.39%
geometric mean +0.20%
Manually checked 473 and 471 to verify the diff is in the noise range.
Reviewers: rengolin, davidxl
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D24818
llvm-svn: 284545
Summary:
The original heuristic to break critical edge during machine sink is relatively conservertive: when there is only one instruction sinkable to the critical edge, it is likely that the machine sink pass will not break the critical edge. This leads to many speculative instructions executed at runtime. However, with profile info, we could model the splitting benefits: if the critical edge has 50% taken rate, it would always be beneficial to split the critical edge to avoid the speculated runtime instructions. This patch uses profile to guide critical edge splitting in machine sink pass.
The performance impact on speccpu2006 on Intel sandybridge machines:
spec/2006/fp/C++/444.namd 25.3 +0.26%
spec/2006/fp/C++/447.dealII 45.96 -0.10%
spec/2006/fp/C++/450.soplex 41.97 +1.49%
spec/2006/fp/C++/453.povray 36.83 -0.96%
spec/2006/fp/C/433.milc 23.81 +0.32%
spec/2006/fp/C/470.lbm 41.17 +0.34%
spec/2006/fp/C/482.sphinx3 48.13 +0.69%
spec/2006/int/C++/471.omnetpp 22.45 +3.25%
spec/2006/int/C++/473.astar 21.35 -2.06%
spec/2006/int/C++/483.xalancbmk 36.02 -2.39%
spec/2006/int/C/400.perlbench 33.7 -0.17%
spec/2006/int/C/401.bzip2 22.9 +0.52%
spec/2006/int/C/403.gcc 32.42 -0.54%
spec/2006/int/C/429.mcf 39.59 +0.19%
spec/2006/int/C/445.gobmk 26.98 -0.00%
spec/2006/int/C/456.hmmer 24.52 -0.18%
spec/2006/int/C/458.sjeng 28.26 +0.02%
spec/2006/int/C/462.libquantum 55.44 +3.74%
spec/2006/int/C/464.h264ref 46.67 -0.39%
geometric mean +0.20%
Manually checked 473 and 471 to verify the diff is in the noise range.
Reviewers: rengolin, davidxl
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D24818
llvm-svn: 284541
The custom lowering is pretty straightforward: basically, just AND
together the two halves of a <4 x i32> compare.
Differential Revision: https://reviews.llvm.org/D25713
llvm-svn: 284536
This patch assigns cost of the scaling used in addressing for Cortex-R52.
On Cortex-R52 a negated register offset takes longer than a non-negated
register offset, in a register-offset addressing mode.
Differential Revision: http://reviews.llvm.org/D25670
Reviewer: jmolloy
llvm-svn: 284460
This patch adds simplified support for tail calls on ARM with XRay instrumentation.
Known issue: compiled with generic flags: `-O3 -g -fxray-instrument -Wall
-std=c++14 -ffunction-sections -fdata-sections` (this list doesn't include my
specific flags like --target=armv7-linux-gnueabihf etc.), the following program
#include <cstdio>
#include <cassert>
#include <xray/xray_interface.h>
[[clang::xray_always_instrument]] void __attribute__ ((noinline)) fC() {
std::printf("In fC()\n");
}
[[clang::xray_always_instrument]] void __attribute__ ((noinline)) fB() {
std::printf("In fB()\n");
fC();
}
[[clang::xray_always_instrument]] void __attribute__ ((noinline)) fA() {
std::printf("In fA()\n");
fB();
}
// Avoid infinite recursion in case the logging function is instrumented (so calls logging
// function again).
[[clang::xray_never_instrument]] void simplyPrint(int32_t functionId, XRayEntryType xret)
{
printf("XRay: functionId=%d type=%d.\n", int(functionId), int(xret));
}
int main(int argc, char* argv[]) {
__xray_set_handler(simplyPrint);
printf("Patching...\n");
__xray_patch();
fA();
printf("Unpatching...\n");
__xray_unpatch();
fA();
return 0;
}
gives the following output:
Patching...
XRay: functionId=3 type=0.
In fA()
XRay: functionId=3 type=1.
XRay: functionId=2 type=0.
In fB()
XRay: functionId=2 type=1.
XRay: functionId=1 type=0.
XRay: functionId=1 type=1.
In fC()
Unpatching...
In fA()
In fB()
In fC()
So for function fC() the exit sled seems to be called too much before function
exit: before printing In fC().
Debugging shows that the above happens because printf from fC is also called as
a tail call. So first the exit sled of fC is executed, and only then printf is
jumped into. So it seems we can't do anything about this with the current
approach (i.e. within the simplification described in
https://reviews.llvm.org/D23988 ).
Differential Revision: https://reviews.llvm.org/D25030
llvm-svn: 284456
SelectionDAG::getConstantPool will automatically determine an appropriate alignment if one is not specified. It does this by querying the type's preferred alignment. This can end up creating quite a lot of padding when the preferred alignment for vectors is 128.
In optimize-for-size mode, it makes sense to instead query the ABI type alignment which is often smaller and causes less padding.
llvm-svn: 284381
Retrying after upstream changes.
Simplify Consecutive Merge Store Candidate Search
Now that address aliasing is much less conservative, push through
simplified store merging search 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.
Whem 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 generally the output CodeGen (with some
exceptions).
Additional Minor Changes:
1. Finishes removing unused AliasLoad code
2. Unifies the 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 seemed sufficient to not cause regressions in
tests.
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. Some tests relying on the order were changed to use
volatile memory operations
Noteworthy tests:
CodeGen/AArch64/argument-blocks.ll -
It's not entirely clear what the test_varargs_stackalign test is
supposed to be asserting, but the new code looks right.
CodeGen/AArch64/arm64-memset-inline.lli -
CodeGen/AArch64/arm64-stur.ll -
CodeGen/ARM/memset-inline.ll -
The backend now generates *worse* code due to store merging
succeeding, as we do do a 16-byte constant-zero store efficiently.
CodeGen/AArch64/merge-store.ll -
Improved, but there still seems to be an extraneous vector insert
from an element to itself?
CodeGen/PowerPC/ppc64-align-long-double.ll -
Worse code emitted in this case, due to the improved store->load
forwarding.
CodeGen/X86/dag-merge-fast-accesses.ll -
CodeGen/X86/MergeConsecutiveStores.ll -
CodeGen/X86/stores-merging.ll -
CodeGen/Mips/load-store-left-right.ll -
Restored correct merging of non-aligned stores
CodeGen/AMDGPU/promote-alloca-stored-pointer-value.ll -
Improved. Correctly merges buffer_store_dword calls
CodeGen/AMDGPU/si-triv-disjoint-mem-access.ll -
Improved. Sidesteps loading a stored value and
merges two stores
CodeGen/X86/pr18023.ll -
This test has been removed, as it was asserting incorrect
behavior. Non-volatile stores *CAN* be moved past volatile loads,
and now are.
CodeGen/X86/vector-idiv.ll -
CodeGen/X86/vector-lzcnt-128.ll -
It's basically impossible to tell what these tests are actually
testing. But, looks like the code got better due to the memory
operations being recognized as non-aliasing.
CodeGen/X86/win32-eh.ll -
Both loads of the securitycookie are now merged.
CodeGen/AMDGPU/vgpr-spill-emergency-stack-slot-compute.ll -
This test appears to work but no longer exhibits the spill behavior.
Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle
Subscribers: wdng, nhaehnle, nemanjai, arsenm, weimingz, niravd, RKSimon, aemerson, qcolombet, dsanders, resistor, tstellarAMD, t.p.northover, spatel
Differential Revision: https://reviews.llvm.org/D14834
llvm-svn: 284151
This patch assigns cost of the scaling used in addressing.
On many ARM cores, a negated register offset takes longer than a
non-negated register offset, in a register-offset addressing mode.
For instance:
LDR R0, [R1, R2 LSL #2]
LDR R0, [R1, -R2 LSL #2]
Above, (1) takes less cycles than (2).
By assigning appropriate scaling factor cost, we enable the LLVM
to make the right trade-offs in the optimization and code-selection phase.
Differential Revision: http://reviews.llvm.org/D24857
Reviewers: jmolloy, rengolin
llvm-svn: 284127
- Use storage class C_STAT for 'PrivateLinkage' The storage class for
PrivateLinkage should equal to the Internal Linkage.
- Set 'PrivateGlobalPrefix' from "L" to ".L" for MM_WinCOFF (includes
x86_64) MM_WinCOFF has empty GlobalPrefix '\0' so PrivateGlobalPrefix
"L" may conflict to the normal symbol name starting with 'L'.
Based on a patch by Han Sangjin! Manually updated test cases.
llvm-svn: 284096
This combiner breaks debug experience and should not be run when optimizations are disabled.
For example:
int main() {
int j = 0;
j += 2;
if (j == 2)
return 0;
return 5;
}
When debugging this code compiled in /O0, it should be valid to break at line "j+=2;" and edit the value of j. It should change the return value of the function.
Differential Revision: https://reviews.llvm.org/D19268
llvm-svn: 284014
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
llvm-svn: 283934
Currently, the Int_eh_sjlj_dispatchsetup intrinsic is marked as
clobbering all registers, including floating-point registers that may
not be present on the target. This is technically true, as we could get
linked against code that does use the FP registers, but that will not
actually work, as the soft-float code cannot save and restore the FP
registers. SjLj exception handling can only work correctly if either all
or none of the code is built for a target with FP registers. Therefore,
we can assume that, when Int_eh_sjlj_dispatchsetup is compiled for a
soft-float target, it is only going to be linked against other
soft-float code, and so only clobbers the general-purpose registers.
This allows us to check that no non-savable registers are clobbered when
generating the prologue/epilogue.
Differential Revision: https://reviews.llvm.org/D25180
llvm-svn: 283866
This reverts commit r283842.
test/CodeGen/X86/tail-dup-repeat.ll causes and llc crash with our
internal testing. I'll share a link with you.
llvm-svn: 283857
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
llvm-svn: 283842
The instructions VLDM/VSTM can only access word-aligned memory
locations and produce alignment fault if the condition is not met.
The compiler currently generates VLDM/VSTM for v2f64 load/store
regardless the alignment of the memory access. Instead, if a v2f64
load/store is not word-aligned, the compiler should generate
VLD1/VST1. For each non double-word-aligned VLD1/VST1, a VREV
instruction should be generated when targeting Big Endian.
Differential Revision: https://reviews.llvm.org/D25281
llvm-svn: 283763
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Differential revision: https://reviews.llvm.org/D18226
llvm-svn: 283619
The code used llvm basic block predecessors to decided where to insert phi
nodes. Instruction selection can and will liberally insert new machine basic
block predecessors. There is not a guaranteed one-to-one mapping from pred.
llvm basic blocks and machine basic blocks.
Therefore the current approach does not work as it assumes we can mark
predecessor machine basic block as needing a copy, and needs to know the set of
all predecessor machine basic blocks to decide when to insert phis.
Instead of computing the swifterror vregs as we select instructions, propagate
them at the end of instruction selection when the MBB CFG is complete.
When an instruction needs a swifterror vreg and we don't know the value yet,
generate a new vreg and remember this "upward exposed" use, and reconcile this
at the end of instruction selection.
This will only happen if the target supports promoting swifterror parameters to
registers and the swifterror attribute is used.
rdar://28300923
llvm-svn: 283617
Reapplying r283383 after revert in r283442. The additional fix
is a getting rid of a stray space in a function name, in the
refactoring part of the commit.
This avoids falling back to calling out to the GCC rem functions
(__moddi3, __umoddi3) when targeting Windows.
The __rt_div functions have flipped the two arguments compared
to the __aeabi_divmod functions. To match MSVC, we emit a
check for division by zero before actually calling the library
function (even if the library function itself also might do
the same check).
Not all calls to __rt_div functions for division are currently
merged with calls to the same function with the same parameters
for the remainder. This is more wasteful than a div + mls as before,
but avoids calls to __moddi3.
Differential Revision: https://reviews.llvm.org/D25332
llvm-svn: 283550
This reverts commit r283383 because it broke some of the bots:
undefined reference to ` __aeabi_uldivmod'
It affected (at least) clang-cmake-armv7-a15-selfhost,
clang-cmake-armv7-a15-selfhost and clang-native-arm-lnt.
llvm-svn: 283442
Global variables are GlobalValues, so they have explicit alignment. Querying
DataLayout for the alignment was incorrect.
Testcase added.
llvm-svn: 283423
We can work around a shortcoming of FileCheck by using {{\[}} to match a square
bracket before a [[ sequence.
Thanks to Eli Friedman for the heads up!
llvm-svn: 283422
This avoids falling back to calling out to the GCC rem functions
(__moddi3, __umoddi3) when targeting Windows.
The __rt_div functions have flipped the two arguments compared
to the __aeabi_divmod functions. To match MSVC, we emit a
check for division by zero before actually calling the library
function (even if the library function itself also might do
the same check).
Not all calls to __rt_div functions for division are currently
merged with calls to the same function with the same parameters
for the remainder. This is more wasteful than a div + mls as before,
but avoids calls to __moddi3.
Differential Revision: https://reviews.llvm.org/D24076
llvm-svn: 283383
This is not a valid encoding - these instructions cannot do PC-relative addressing.
The underlying problem here is of whitelist in ARMISelDAGToDAG that unwraps ARMISD::Wrappers during addressing-mode selection. This didn't realise TargetConstantPool was actually possible, so didn't handle it.
llvm-svn: 283323
This reverts commit 062ace9764953e9769142c1099281a345f9b6bdc.
Issue with loop info and block removal revealed by polly.
I have a fix for this issue already in another patch, I'll re-roll this
together with that fix, and a test case.
llvm-svn: 283292
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well.
Differential revision: https://reviews.llvm.org/D18226
llvm-svn: 283274
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
llvm-svn: 283164
library call to __aeabi_uidivmod. This is an improved implementation of
r280808, see also D24133, that got reverted because isel was stuck in a loop.
That was caused by the optimisation incorrectly triggering on i64 ints, which
shouldn't happen because there is no 64bit hwdiv support; that put isel's type
legalization and this optimisation in a loop. A native ARM compiler and testing
now shows that this is fixed.
Patch mostly by Pablo Barrio.
Differential Revision: https://reviews.llvm.org/D25077
llvm-svn: 283098
This addresses PR26055 LiveDebugValues is very slow.
Contrary to the old LiveDebugVariables pass LiveDebugValues currently
doesn't look at the lexical scopes before inserting a DBG_VALUE
intrinsic. This means that we often propagate DBG_VALUEs much further
down than necessary. This is especially noticeable in large C++
functions with many inlined method calls that all use the same
"this"-pointer.
For example, in the following code it makes no sense to propagate the
inlined variable a from the first inlined call to f() into any of the
subsequent basic blocks, because the variable will always be out of
scope:
void sink(int a);
void __attribute((always_inline)) f(int a) { sink(a); }
void foo(int i) {
f(i);
if (i)
f(i);
f(i);
}
This patch reuses the LexicalScopes infrastructure we have for
LiveDebugVariables to take this into account.
The effect on compile time and memory consumption is quite noticeable:
I tested a benchmark that is a large C++ source with an enormous
amount of inlined "this"-pointers that would previously eat >24GiB
(most of them for DBG_VALUE intrinsics) and whose compile time was
dominated by LiveDebugValues. With this patch applied the memory
consumption is 1GiB and 1.7% of the time is spent in LiveDebugValues.
https://reviews.llvm.org/D24994
Thanks to Daniel Berlin and Keith Walker for reviewing!
llvm-svn: 282611
Simplify Consecutive Merge Store Candidate Search
Now that address aliasing is much less conservative, push through
simplified store merging search 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.
Whem 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 generally the output CodeGen (with some
exceptions).
Additional Minor Changes:
1. Finishes removing unused AliasLoad code
2. Unifies the 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 seemed sufficient to not cause regressions in
tests.
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. Some tests relying on the order were changed to use
volatile memory operations
Noteworthy tests:
CodeGen/AArch64/argument-blocks.ll -
It's not entirely clear what the test_varargs_stackalign test is
supposed to be asserting, but the new code looks right.
CodeGen/AArch64/arm64-memset-inline.lli -
CodeGen/AArch64/arm64-stur.ll -
CodeGen/ARM/memset-inline.ll -
The backend now generates *worse* code due to store merging
succeeding, as we do do a 16-byte constant-zero store efficiently.
CodeGen/AArch64/merge-store.ll -
Improved, but there still seems to be an extraneous vector insert
from an element to itself?
CodeGen/PowerPC/ppc64-align-long-double.ll -
Worse code emitted in this case, due to the improved store->load
forwarding.
CodeGen/X86/dag-merge-fast-accesses.ll -
CodeGen/X86/MergeConsecutiveStores.ll -
CodeGen/X86/stores-merging.ll -
CodeGen/Mips/load-store-left-right.ll -
Restored correct merging of non-aligned stores
CodeGen/AMDGPU/promote-alloca-stored-pointer-value.ll -
Improved. Correctly merges buffer_store_dword calls
CodeGen/AMDGPU/si-triv-disjoint-mem-access.ll -
Improved. Sidesteps loading a stored value and merges two stores
CodeGen/X86/pr18023.ll -
This test has been removed, as it was asserting incorrect
behavior. Non-volatile stores *CAN* be moved past volatile loads,
and now are.
CodeGen/X86/vector-idiv.ll -
CodeGen/X86/vector-lzcnt-128.ll -
It's basically impossible to tell what these tests are actually
testing. But, looks like the code got better due to the memory
operations being recognized as non-aliasing.
CodeGen/X86/win32-eh.ll -
Both loads of the securitycookie are now merged.
CodeGen/AMDGPU/vgpr-spill-emergency-stack-slot-compute.ll -
This test appears to work but no longer exhibits the spill
behavior.
Reviewers: arsenm, hfinkel, tstellarAMD, nhaehnle, jyknight
Subscribers: wdng, nhaehnle, nemanjai, arsenm, weimingz, niravd, RKSimon, aemerson, qcolombet, resistor, tstellarAMD, t.p.northover, spatel
Differential Revision: https://reviews.llvm.org/D14834
llvm-svn: 282600
Variables are sometimes missing their debug location information in
blocks in which the variables should be available. This would occur
when one or more predecessor blocks had not yet been visited by the
routine which propagated the information from predecessor blocks.
This is addressed by only considering predecessor blocks which have
already been visited.
The solution to this problem was suggested by Daniel Berlin on the
LLVM developer mailing list.
Differential Revision: https://reviews.llvm.org/D24927
llvm-svn: 282506
If a constant is unamed_addr and is only used within one function, we can save
on the code size and runtime cost of an indirection by changing the global's storage
to inside the constant pool. For example, instead of:
ldr r0, .CPI0
bl printf
bx lr
.CPI0: &format_string
format_string: .asciz "hello, world!\n"
We can emit:
adr r0, .CPI0
bl printf
bx lr
.CPI0: .asciz "hello, world!\n"
This can cause significant code size savings when many small strings are used in one
function (4 bytes per string).
This recommit contains fixes for a nasty bug related to fast-isel fallback - because
fast-isel doesn't know about this optimization, if it runs and emits references to
a string that we inline (because fast-isel fell back to SDAG) we will end up
with an inlined string and also an out-of-line string, and we won't emit the
out-of-line string, causing backend failures.
It also contains fixes for emitting .text relocations which made the sanitizer
bots unhappy.
llvm-svn: 282387
If a constant is unamed_addr and is only used within one function, we can save
on the code size and runtime cost of an indirection by changing the global's storage
to inside the constant pool. For example, instead of:
ldr r0, .CPI0
bl printf
bx lr
.CPI0: &format_string
format_string: .asciz "hello, world!\n"
We can emit:
adr r0, .CPI0
bl printf
bx lr
.CPI0: .asciz "hello, world!\n"
This can cause significant code size savings when many small strings are used in one
function (4 bytes per string).
This recommit contains fixes for a nasty bug related to fast-isel fallback - because
fast-isel doesn't know about this optimization, if it runs and emits references to
a string that we inline (because fast-isel fell back to SDAG) we will end up
with an inlined string and also an out-of-line string, and we won't emit the
out-of-line string, causing backend failures.
It also contains fixes for emitting .text relocations which made the sanitizer
bots unhappy.
llvm-svn: 282241
ldm and stm instructions always require 4-byte alignment on the pointer, but we
weren't checking this before trying to reduce code-size by replacing a
post-indexed load/store with them. Unfortunately, we were also dropping this
incormation in DAG ISel too, but that's easy enough to fix.
llvm-svn: 281893
This is a port of XRay to ARM 32-bit, without Thumb support yet. The XRay instrumentation support is moving up to AsmPrinter.
This is one of 3 commits to different repositories of XRay ARM port. The other 2 are:
https://reviews.llvm.org/D23932 (Clang test)
https://reviews.llvm.org/D23933 (compiler-rt)
Differential Revision: https://reviews.llvm.org/D23931
llvm-svn: 281878
(and the same for SREM)
This was causing buildbot failures earlier (time outs in the LNT suite).
However, we haven't been able to reproduce this and are suspecting this
was caused by another (reverted) patch.
llvm-svn: 281719
If a constant is unamed_addr and is only used within one function, we can save
on the code size and runtime cost of an indirection by changing the global's storage
to inside the constant pool. For example, instead of:
ldr r0, .CPI0
bl printf
bx lr
.CPI0: &format_string
format_string: .asciz "hello, world!\n"
We can emit:
adr r0, .CPI0
bl printf
bx lr
.CPI0: .asciz "hello, world!\n"
This can cause significant code size savings when many small strings are used in one
function (4 bytes per string).
This recommit contains fixes for a nasty bug related to fast-isel fallback - because
fast-isel doesn't know about this optimization, if it runs and emits references to
a string that we inline (because fast-isel fell back to SDAG) we will end up
with an inlined string and also an out-of-line string, and we won't emit the
out-of-line string, causing backend failures.
It also contains fixes for emitting .text relocations which made the sanitizer
bots unhappy.
llvm-svn: 281715
If a constant is unamed_addr and is only used within one function, we can save
on the code size and runtime cost of an indirection by changing the global's storage
to inside the constant pool. For example, instead of:
ldr r0, .CPI0
bl printf
bx lr
.CPI0: &format_string
format_string: .asciz "hello, world!\n"
We can emit:
adr r0, .CPI0
bl printf
bx lr
.CPI0: .asciz "hello, world!\n"
This can cause significant code size savings when many small strings are used in one
function (4 bytes per string).
This recommit contains fixes for a nasty bug related to fast-isel fallback - because
fast-isel doesn't know about this optimization, if it runs and emits references to
a string that we inline (because fast-isel fell back to SDAG) we will end up
with an inlined string and also an out-of-line string, and we won't emit the
out-of-line string, causing backend failures.
llvm-svn: 281604
If a constant is unamed_addr and is only used within one function, we can save
on the code size and runtime cost of an indirection by changing the global's storage
to inside the constant pool. For example, instead of:
ldr r0, .CPI0
bl printf
bx lr
.CPI0: &format_string
format_string: .asciz "hello, world!\n"
We can emit:
adr r0, .CPI0
bl printf
bx lr
.CPI0: .asciz "hello, world!\n"
This can cause significant code size savings when many small strings are used in one
function (4 bytes per string).
llvm-svn: 281484
Before, only Thumb functions were marked as ".code 16". These
".code x" directives are effective until the next directive of its
kind is encountered. Therefore, in code with interleaved ARM and
Thumb functions, it was possible to declare a function as ARM and
end up with a Thumb function after assembly. A test has been added.
An existing test has also been fixed to take this change into
account.
Reviewers: aschwaighofer, t.p.northover, jmolloy, rengolin
Subscribers: aemerson, rengolin, llvm-commits
Differential Revision: https://reviews.llvm.org/D24337
llvm-svn: 281324
For the common pattern (CMPZ (AND x, #bitmask), #0), we can do some more efficient instruction selection if the bitmask is one consecutive sequence of set bits (32 - clz(bm) - ctz(bm) == popcount(bm)).
1) If the bitmask touches the LSB, then we can remove all the upper bits and set the flags by doing one LSLS.
2) If the bitmask touches the MSB, then we can remove all the lower bits and set the flags with one LSRS.
3) If the bitmask has popcount == 1 (only one set bit), we can shift that bit into the sign bit with one LSLS and change the condition query from NE/EQ to MI/PL (we could also implement this by shifting into the carry bit and branching on BCC/BCS).
4) Otherwise, we can emit a sequence of LSLS+LSRS to remove the upper and lower zero bits of the mask.
1-3 require only one 16-bit instruction and can elide the CMP. 4 requires two 16-bit instructions but can elide the CMP and doesn't require materializing a complex immediate, so is also a win.
llvm-svn: 281323
If a constant is unamed_addr and is only used within one function, we can save
on the code size and runtime cost of an indirection by changing the global's storage
to inside the constant pool. For example, instead of:
ldr r0, .CPI0
bl printf
bx lr
.CPI0: &format_string
format_string: .asciz "hello, world!\n"
We can emit:
adr r0, .CPI0
bl printf
bx lr
.CPI0: .asciz "hello, world!\n"
This can cause significant code size savings when many small strings are used in one
function (4 bytes per string).
llvm-svn: 281314
This patch reverses the edge from DIGlobalVariable to GlobalVariable.
This will allow us to more easily preserve debug info metadata when
manipulating global variables.
Fixes PR30362. A program for upgrading test cases is attached to that
bug.
Differential Revision: http://reviews.llvm.org/D20147
llvm-svn: 281284
For the common pattern (CMPZ (AND x, #bitmask), #0), we can do some more efficient instruction selection if the bitmask is one consecutive sequence of set bits (32 - clz(bm) - ctz(bm) == popcount(bm)).
1) If the bitmask touches the LSB, then we can remove all the upper bits and set the flags by doing one LSLS.
2) If the bitmask touches the MSB, then we can remove all the lower bits and set the flags with one LSRS.
3) If the bitmask has popcount == 1 (only one set bit), we can shift that bit into the sign bit with one LSLS and change the condition query from NE/EQ to MI/PL (we could also implement this by shifting into the carry bit and branching on BCC/BCS).
4) Otherwise, we can emit a sequence of LSLS+LSRS to remove the upper and lower zero bits of the mask.
1-3 require only one 16-bit instruction and can elide the CMP. 4 requires two 16-bit instructions but can elide the CMP and doesn't require materializing a complex immediate, so is also a win.
llvm-svn: 281215
If a constant is unamed_addr and is only used within one function, we can save
on the code size and runtime cost of an indirection by changing the global's storage
to inside the constant pool. For example, instead of:
ldr r0, .CPI0
bl printf
bx lr
.CPI0: &format_string
format_string: .asciz "hello, world!\n"
We can emit:
adr r0, .CPI0
bl printf
bx lr
.CPI0: .asciz "hello, world!\n"
This can cause significant code size savings when many small strings are used in one
function (4 bytes per string).
llvm-svn: 281213
Summary:
This test was not testing the intrinsics. A function like this:
define %v4f32 @test_v4f32.floor(%v4f32 %a){
...
%1 = call %v4f32 @llvm.floor.v4f32(%v4f32 %a)
...
}
is transformed into the following assembly:
_test_v4f32.floor: @ @test_v4f32.floor
...
bl _floorf
...
In each function tested, there are two CHECK: one that checked
for the label and another one for the intrinsic that should be used
inside the function (in our case, "floor"). However, although the
first CHECK was matching the label, the second was not matching the
intrinsic, but the second "floor" in the same line as the label.
This is fixed by making the first CHECK match the entire line.
Reviewers: jmolloy, rengolin
Subscribers: rengolin, llvm-commits
Differential Revision: https://reviews.llvm.org/D24398
llvm-svn: 281211
The CMPZ #0 disappears during peepholing, leaving just a tADDi3, tADDi8 or t2ADDri. This avoids having to materialize the expensive negative constant in Thumb-1, and allows a shrinking from a 32-bit CMN to a 16-bit ADDS in Thumb-2.
llvm-svn: 281040
And associated commits, as they broke the Thumb bots.
This reverts commit r280935.
This reverts commit r280891.
This reverts commit r280888.
llvm-svn: 280967
Materializing something like "-3" can be done as 2 instructions:
MOV r0, #3
MVN r0, r0
This has a cost of 2, not 3. It looks like we were already trying to detect this pattern in TII::getIntImmCost(), but were taking the complement of the zero-extended value instead of the sign-extended value which is unlikely to ever produce a number < 256.
There were no tests failing after changing this... :/
llvm-svn: 280928
This reverts commit r280808.
It is possible that this change results in an infinite loop. This
is causing timeouts in some tests on ARM, and a Chromebook bot is
failing.
llvm-svn: 280918
This is a port of XRay to ARM 32-bit, without Thumb support yet. The XRay instrumentation support is moving up to AsmPrinter.
This is one of 3 commits to different repositories of XRay ARM port. The other 2 are:
1. https://reviews.llvm.org/D23932 (Clang test)
2. https://reviews.llvm.org/D23933 (compiler-rt)
Differential Revision: https://reviews.llvm.org/D23931
llvm-svn: 280888
Summary:
This saves a library call to __aeabi_uidivmod. However, the
processor must feature hardware division in order to benefit from
the transformation.
Reviewers: scott-0, jmolloy, compnerd, rengolin
Subscribers: t.p.northover, compnerd, aemerson, rengolin, samparker, llvm-commits
Differential Revision: https://reviews.llvm.org/D24133
llvm-svn: 280808
This is a Windows ARM specific issue. If the code path in the if conversion
ends up using a relocation which will form a IMAGE_REL_ARM_MOV32T, we end up
with a bundle to ensure that the mov.w/mov.t pair is not split up. This is
normally fine, however, if the branch is also predicated, then we end up trying
to predicate the bundle.
For now, report a bundle as being unpredicatable. Although this is false, this
would trigger a failure case previously anyways, so this is no worse. That is,
there should not be any code which would previously have been if converted and
predicated which would not be now.
Under certain circumstances, it may be possible to "predicate the bundle". This
would require scanning all bundle instructions, and ensure that the bundle
contains only predicatable instructions, and converting the bundle into an IT
block sequence. If the bundle is larger than the maximal IT block length (4
instructions), it would require materializing multiple IT blocks from the single
bundle.
llvm-svn: 280689
All of the builtins are designed to be invoked with ARM AAPCS CC even on ARM
AAPCS VFP CC hosts. Tweak the default initialisation to ARM AAPCS CC rather
than C CC for ARM/thumb targets.
The changes to the tests are necessary to ensure that the calling convention for
the lowered library calls are honoured. Furthermore, these adjustments cause
certain branch invocations to change to branch-and-link since the returned value
needs to be moved across registers (d0 -> r0, r1).
llvm-svn: 280683
types. This is the LLVM counterpart and it adds options that map onto FP
exceptions and denormal build attributes allowing better fp math library
selections.
Differential Revision: https://reviews.llvm.org/D24070
llvm-svn: 280246
Weiming Zhao