Summary:
Reserving registers x1-7 is used to support CONFIG_ARM64_LSE_ATOMICS in Linux kernel. This change adds support for reserving registers x1 through x7.
Reviewers: javed.absar, phosek, srhines, nickdesaulniers, efriedma
Reviewed By: nickdesaulniers, efriedma
Subscribers: niravd, jfb, manojgupta, nickdesaulniers, jyknight, efriedma, kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D48580
llvm-svn: 341706
Summary:
I added a few ARM64 memset codegen tests in r341406 and r341493, and annotated
where the generated code was bad. This patch fixes the majority of the issues by
requesting that a 2xi64 vector be used for memset of 32 bytes and above.
The patch leaves the former request for f128 unchanged, despite f128
materialization being suboptimal: doing otherwise runs into other asserts in
isel and makes this patch too broad.
This patch hides the issue that was present in bzero_40_stack and bzero_72_stack
because the code now generates in a better order which doesn't have the store
offset issue. I'm not aware of that issue appearing elsewhere at the moment.
<rdar://problem/44157755>
Reviewers: t.p.northover, MatzeB, javed.absar
Subscribers: eraman, kristof.beyls, chrib, dexonsmith, llvm-commits
Differential Revision: https://reviews.llvm.org/D51706
llvm-svn: 341558
I'm looking at some codegen optimization in this area and want to make sure I understand the current codegen and don't regress it. This patch further expands the tests (which I already expanded in r341406) to capture more of the current code generation when it comes to stack-based small non-zero memset on arm64. This patch annotates some potential fixes.
llvm-svn: 341493
This was proposed as an IR transform in D49306, but it was not clearly justifiable as a canonicalization.
Here, we only do the transform when the target tells us that sqrt can be lowered with inline code.
This is the basic case. Some potential enhancements are in the TODO comments:
1. Generalize the transform for other exponents (allow more than 2 sqrt calcs if that's really cheaper).
2. If we have less fast-math-flags, generate code to avoid -0.0 and/or INF.
3. Allow the transform when optimizing/minimizing size (might require a target hook to get that right).
Note that by default, x86 converts single-precision sqrt calcs into sqrt reciprocal estimate with
refinement. That codegen is controlled by CPU attributes and can be manually overridden. We have plenty
of test coverage for that already, so I didn't bother to include extra testing for that here. AArch uses
its full-precision ops in all cases (not sure if that's the intended behavior or not, but that should
also be covered by existing tests).
Differential Revision: https://reviews.llvm.org/D51630
llvm-svn: 341481
Reland r341269. Use std::stable_sort when sorting constant condidates.
Reverting commit, r341365:
Revert r341269: [Constant Hoisting] Hoisting Constant GEP Expressions
One of the tests is failing 50% of the time when expensive checks are
enabled. Not sure how deep the problem is so just reverting while the
author can investigate so that the bots stop repeatedly failing and
blaming things incorrectly. Will respond with details on the original
commit.
Original commit, r341269:
[Constant Hoisting] Hoisting Constant GEP Expressions
Leverage existing logic in constant hoisting pass to transform constant GEP
expressions sharing the same base global variable. Multi-dimensional GEPs are
rewritten into single-dimensional GEPs.
https://reviews.llvm.org/D51396
Differential Revision: https://reviews.llvm.org/D51654
llvm-svn: 341417
I'm looking at some codegen optimization in this area and want to make sure I understand the current codegen and don't regress it. This patch simply expands the two existing tests to capture more of the current code generation when it comes to heap-based and stack-based small memset on arm64. The tested code is already pretty good, notably when it comes to using STP, FP stores, FP immediate generation, and folding one of the stores into a stack spill when possible. The uses of STUR could be improved, and some more pairing could occur. Straying from bzero patterns currently yield suboptimal code, and I expect a variety of small changes could make things way better.
llvm-svn: 341406
The runtime pseudo relocations can't handle the AArch64 format PC
relative addressing in adrp+add/ldr pairs. By using stubs, the potentially
dllimported addresses can be touched up by the runtime pseudo relocation
framework.
Differential Revision: https://reviews.llvm.org/D51452
llvm-svn: 341401
One of the tests is failing 50% of the time when expensive checks are
enabled. Not sure how deep the problem is so just reverting while the
author can investigate so that the bots stop repeatedly failing and
blaming things incorrectly. Will respond with details on the original
commit.
llvm-svn: 341365
For instructions that spill/fill to and from multiple frame-indices
in a single instruction, hasStoreToStackSlot and hasLoadFromStackSlot
should return an array of accesses, rather than just the first encounter
of such an access.
This better describes FI accesses for AArch64 (paired) LDP/STP
instructions.
Reviewers: t.p.northover, gberry, thegameg, rengolin, javed.absar, MatzeB
Reviewed By: MatzeB
Differential Revision: https://reviews.llvm.org/D51537
llvm-svn: 341301
Summary:
A follow-up for D49266 / rL337166 + D49497 / rL338044.
This is still the same pattern to check for the [lack of]
signed truncation, but in this case the constants and the predicate
are negated.
https://rise4fun.com/Alive/BDVhttps://rise4fun.com/Alive/n7Z
Reviewers: spatel, craig.topper, RKSimon, javed.absar, efriedma, dmgreen
Reviewed By: spatel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D51532
llvm-svn: 341287
Leverage existing logic in constant hoisting pass to transform constant GEP
expressions sharing the same base global variable. Multi-dimensional GEPs are
rewritten into single-dimensional GEPs.
Differential Revision: https://reviews.llvm.org/D51396
llvm-svn: 341269
Summary:
This is a continuation of https://reviews.llvm.org/D49727
Below the original text, current changes in the comments:
Currently, in line with GCC, when specifying reserved registers like sp or pc on an inline asm() clobber list, we don't always preserve the original value across the statement. And in general, overwriting reserved registers can have surprising results.
For example:
extern int bar(int[]);
int foo(int i) {
int a[i]; // VLA
asm volatile(
"mov r7, #1"
:
:
: "r7"
);
return 1 + bar(a);
}
Compiled for thumb, this gives:
$ clang --target=arm-arm-none-eabi -march=armv7a -c test.c -o - -S -O1 -mthumb
...
foo:
.fnstart
@ %bb.0: @ %entry
.save {r4, r5, r6, r7, lr}
push {r4, r5, r6, r7, lr}
.setfp r7, sp, #12
add r7, sp, #12
.pad #4
sub sp, #4
movs r1, #7
add.w r0, r1, r0, lsl #2
bic r0, r0, #7
sub.w r0, sp, r0
mov sp, r0
@APP
mov.w r7, #1
@NO_APP
bl bar
adds r0, #1
sub.w r4, r7, #12
mov sp, r4
pop {r4, r5, r6, r7, pc}
...
r7 is used as the frame pointer for thumb targets, and this function needs to restore the SP from the FP because of the variable-length stack allocation a. r7 is clobbered by the inline assembly (and r7 is included in the clobber list), but LLVM does not preserve the value of the frame pointer across the assembly block.
This type of behavior is similar to GCC's and has been discussed on the bugtracker: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=11807 . No consensus seemed to have been reached on the way forward. Clang behavior has briefly been discussed on the CFE mailing (starting here: http://lists.llvm.org/pipermail/cfe-dev/2018-July/058392.html). I've opted for following Eli Friedman's advice to print warnings when there are reserved registers on the clobber list so as not to diverge from GCC behavior for now.
The patch uses MachineRegisterInfo's target-specific knowledge of reserved registers, just before we convert the inline asm string in the AsmPrinter.
If we find a reserved register, we print a warning:
repro.c:6:7: warning: inline asm clobber list contains reserved registers: R7 [-Winline-asm]
"mov r7, #1"
^
Reviewers: efriedma, olista01, javed.absar
Reviewed By: efriedma
Subscribers: eraman, kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D51165
llvm-svn: 341062
Providing that the load is known to be 4 byte aligned, we can optimise a
ldr(adr address) to just ldr address.
Differential Revision: https://reviews.llvm.org/D51030
llvm-svn: 341058
Hacker's Delight 10-17: when C is constant,
the result of X % C == 0 can be computed more cheaply
without actually calculating the remainder.
The motivation is discussed here:
https://bugs.llvm.org/show_bug.cgi?id=35479.
Patch by: hermord (Dmytro Shynkevych)!
For https://reviews.llvm.org/D50222
llvm-svn: 341047
Summary:
Global variables that are external and zero initialized are
supposed to be merged with global variables in the bss section
rather than the data section.
Reviewers: efriedma, rengolin, t.p.northover, javed.absar, asl, john.brawn, pcc
Reviewed By: efriedma
Subscribers: dmgreen, llvm-commits
Differential Revision: https://reviews.llvm.org/D51379
llvm-svn: 341008
https://reviews.llvm.org/D51197
Currently, IRTranslator (and GISel) seems to be arbitrarily picking
which overflow intrinsics get mapped into opcodes which either have a
carry as an input or not.
For intrinsics such as Intrinsic::uadd_with_overflow, translate it to an
opcode (G_UADDO) which doesn't have any carry inputs (similar to LLVM
IR).
This patch adds 4 missing opcodes for completeness - G_UADDO, G_USUBO,
G_SSUBE and G_SADDE.
llvm-svn: 340865
This solves the motivating case from:
https://bugs.llvm.org/show_bug.cgi?id=38527
If we are legalizing an FP vector op that maps to 1 of the LLVM intrinsics that mimic libm calls,
but we're going to end up with scalar libcalls for that vector type anyway, then we should unroll
the vector op into scalars before widening. This avoids libcalls because we've lost the knowledge
that some of the scalar elements are undef.
Differential Revision: https://reviews.llvm.org/D50791
llvm-svn: 340469
This adds the plumbing for the Tiny code model for the AArch64 backend. This,
instead of loading addresses through the normal ADRP;ADD pair used in the Small
model, uses a single ADR. The 21 bit range of an ADR means that the code and
its statically defined symbols need to be within 1MB of each other.
This makes it mostly interesting for embedded applications where we want to fit
as much as we can in as small a space as possible.
Differential Revision: https://reviews.llvm.org/D49673
llvm-svn: 340397
This reverts commit 7debc334e6421bb5251ef8f18e97166dfc7dd787.
I missed updating legalizer-info-validation.mir as I had assertions
turned off in my build and that specific test requires asserts. Fixed it
now.
llvm-svn: 340197
- Generate pointer authentication instructions
- The functions instrumented depend on function attribtues:
all (all functions instrumentent)
non-leaf (only those that spill LR)
none
- Function epilogues sign the LR before spilling to the stack and authenticate
the LR once restored
- If the target is v8.3a or greater than can use the combined authenticate and
return instruction
Differential revision: https://reviews.llvm.org/D49793
llvm-svn: 340018
https://reviews.llvm.org/D50401
Add opcodes for llvm.intrinsic.trunc, round, and update the IRTranslator
for the same.
Reviewed by: dsanders.
llvm-svn: 339977
There is no way in the universe, that doing a full-width division in
software will be faster than doing overflowing multiplication in
software in the first place, especially given that this same full-width
multiplication needs to be done anyway.
This patch replaces the previous implementation with a direct lowering
into an overflowing multiplication algorithm based on half-width
operations.
Correctness of the algorithm was verified by exhaustively checking the
output of this algorithm for overflowing multiplication of 16 bit
integers against an obviously correct widening multiplication. Baring
any oversights introduced by porting the algorithm to DAG, confidence in
correctness of this algorithm is extremely high.
Following table shows the change in both t = runtime and s = space. The
change is expressed as a multiplier of original, so anything under 1 is
“better” and anything above 1 is worse.
+-------+-----------+-----------+-------------+-------------+
| Arch | u64*u64 t | u64*u64 s | u128*u128 t | u128*u128 s |
+-------+-----------+-----------+-------------+-------------+
| X64 | - | - | ~0.5 | ~0.64 |
| i686 | ~0.5 | ~0.6666 | ~0.05 | ~0.9 |
| armv7 | - | ~0.75 | - | ~1.4 |
+-------+-----------+-----------+-------------+-------------+
Performance numbers have been collected by running overflowing
multiplication in a loop under `perf` on two x86_64 (one Intel Haswell,
other AMD Ryzen) based machines. Size numbers have been collected by
looking at the size of function containing an overflowing multiply in
a loop.
All in all, it can be seen that both performance and size has improved
except in the case of armv7 where code size has regressed for 128-bit
multiply. u128*u128 overflowing multiply on 32-bit platforms seem to
benefit from this change a lot, taking only 5% of the time compared to
original algorithm to calculate the same thing.
The final benefit of this change is that LLVM is now capable of lowering
the overflowing unsigned multiply for integers of any bit-width as long
as the target is capable of lowering regular multiplication for the same
bit-width. Previously, 128-bit overflowing multiply was the widest
possible.
Patch by Simonas Kazlauskas!
Differential Revision: https://reviews.llvm.org/D50310
llvm-svn: 339922
These correspond to the x86 tests added with rL339790 / rL339791, but I widened
the non-fsin tests to v3f32 to show the problem because AArch supports v2f32 ops.
llvm-svn: 339793
Similar to rL337966 - if the DAGCombiner's rotate matching was
working as expected, I don't think we'd see any test diffs here.
AArch only goes right, and PPC only goes left.
x86 has both, so no diffs there.
Differential Revision: https://reviews.llvm.org/D50091
llvm-svn: 339359
Summary:
Currently, in line with GCC, when specifying reserved registers like sp or pc on an inline asm() clobber list, we don't always preserve the original value across the statement. And in general, overwriting reserved registers can have surprising results.
For example:
```
extern int bar(int[]);
int foo(int i) {
int a[i]; // VLA
asm volatile(
"mov r7, #1"
:
:
: "r7"
);
return 1 + bar(a);
}
```
Compiled for thumb, this gives:
```
$ clang --target=arm-arm-none-eabi -march=armv7a -c test.c -o - -S -O1 -mthumb
...
foo:
.fnstart
@ %bb.0: @ %entry
.save {r4, r5, r6, r7, lr}
push {r4, r5, r6, r7, lr}
.setfp r7, sp, #12
add r7, sp, #12
.pad #4
sub sp, #4
movs r1, #7
add.w r0, r1, r0, lsl #2
bic r0, r0, #7
sub.w r0, sp, r0
mov sp, r0
@APP
mov.w r7, #1
@NO_APP
bl bar
adds r0, #1
sub.w r4, r7, #12
mov sp, r4
pop {r4, r5, r6, r7, pc}
...
```
r7 is used as the frame pointer for thumb targets, and this function needs to restore the SP from the FP because of the variable-length stack allocation a. r7 is clobbered by the inline assembly (and r7 is included in the clobber list), but LLVM does not preserve the value of the frame pointer across the assembly block.
This type of behavior is similar to GCC's and has been discussed on the bugtracker: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=11807 . No consensus seemed to have been reached on the way forward. Clang behavior has briefly been discussed on the CFE mailing (starting here: http://lists.llvm.org/pipermail/cfe-dev/2018-July/058392.html). I've opted for following Eli Friedman's advice to print warnings when there are reserved registers on the clobber list so as not to diverge from GCC behavior for now.
The patch uses MachineRegisterInfo's target-specific knowledge of reserved registers, just before we convert the inline asm string in the AsmPrinter.
If we find a reserved register, we print a warning:
```
repro.c:6:7: warning: inline asm clobber list contains reserved registers: R7 [-Winline-asm]
"mov r7, #1"
^
```
Reviewers: eli.friedman, olista01, javed.absar, efriedma
Reviewed By: efriedma
Subscribers: efriedma, eraman, kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D49727
llvm-svn: 339257
Summary:
Ensure that NormalizedBuildVector returns a BUILD_VECTOR with operands of the
same type. This fixes an assertion failure in VerifySDNode.
Reviewers: SjoerdMeijer, t.p.northover, javed.absar
Reviewed By: SjoerdMeijer
Subscribers: kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D50202
llvm-svn: 339013
In expansion of FCOPYSIGN, the shift node is missing when the two
operands of FCOPYSIGN are of the same size. We should always generate
shift node (if the required shift bit is not zero) to put the sign
bit into the right position, regardless of the size of underlying
types.
Differential Revision: https://reviews.llvm.org/D49973
llvm-svn: 338665
The bug is visible in the constant-folded x86 tests. We can't use the
negated shift amount when the type is not power-of-2:
https://rise4fun.com/Alive/US1r
...so in that case, use the regular lowering that includes a select
to guard against a shift-by-bitwidth. This path is improved by only
calculating the modulo shift amount once now.
Also, improve the rotate (with power-of-2 size) lowering to use
a negate rather than subtract from bitwidth. This improves the
codegen whether we have a rotate instruction or not (although
we can still see that we're not matching to a legal rotate in
all cases).
llvm-svn: 338592
Previously we were just visiting the blocks in the function in IR order, which
is rather arbitrary. Therefore we wouldn't always visit defs before uses, but
the translation code relies on this assumption in some places.
Only codegen change seen in tests is an elision of a redundant copy.
Fixes PR38396
llvm-svn: 338476
Also refactors some existing code to materialize addresses for the large code
model so it can be shared between G_GLOBAL_VALUE and G_BLOCK_ADDR.
This implements PR36390.
Differential Revision: https://reviews.llvm.org/D49903
llvm-svn: 338337
This is exchanging a sub-of-1 with add-of-minus-1:
https://rise4fun.com/Alive/plKAH
This is another step towards improving select-of-constants codegen (see D48970).
x86 is the motivating target, and those diffs all appear to be wins. PPC and AArch64 look neutral.
I've limited this to early combining (!LegalOperations) in case a target wants to reverse it, but
I think canonicalizing to 'add' is more likely to produce further transforms because we have more
folds for 'add'.
Differential Revision: https://reviews.llvm.org/D49924
llvm-svn: 338317
This teaches the outliner to save LR to a register rather than the stack when
possible. This allows us to avoid bumping the stack in outlined functions in
some cases. By doing this, in a later patch, we can teach the outliner to do
something like this:
f1:
...
bl OUTLINED_FUNCTION
...
f2:
...
move LR's contents to a register
bl OUTLINED_FUNCTION
move the register's contents back
instead of falling back to saving LR in both cases.
llvm-svn: 338278
Previously, I thought this was a Windows failure. Then I realized it failed on
every bot that used the verifier. This makes it use the verifier always, and
adds that pass to the pipeline checks so that it's consistent across all bots.
llvm-svn: 338272
Summary:
Attempt to extract a shrl from a udiv or a shl from a mul if this allows a rotate to be formed. This targets cases where the input to a rotate pattern was a mul or udiv by a constant and InstCombine merged one of the shifts with the op.
Patch by: sameconrad (Sam Conrad)
Reviewers: RKSimon, craig.topper, spatel, lebedev.ri, javed.absar
Reviewed By: lebedev.ri
Subscribers: efriedma, kparzysz, llvm-commits
Differential Revision: https://reviews.llvm.org/D47681
llvm-svn: 338270
It seems like the pass pipeline on Windows is slightly different than on Linux
and macOS. As a result, the arm64-opt-remarks-lazy-bfi test has been failing.
This switches a CHECK-NEXT to a CHECK-DAG to try and get this running properly
again.
It'd be nice to switch it back to a CHECK-NEXT if possible, but the CHECK-NEXT
lines following the line we care about (the optimization remark emitter)
do a pretty good job of enforcing the ordering we want.
Hopefully this works, since I don't have a Windows machine. ;)
Example failure: http://lab.llvm.org:8011/builders/llvm-clang-x86_64-expensive-checks-win/builds/11295
llvm-svn: 338267
Fixed the ASAN failure from before in r338148, so recommiting.
This patch enables the MachineOutliner by default in AArch64 under -Oz.
The MachineOutliner offers around a 4.5% improvement on the current -Oz code
size improvements.
We have done work into improving the debuggability of outlined code, so that
users of -Oz won't be surprised by the optimization. We have also been executing
the LLVM test suite and common external tests such as the SPEC suites
continuously with no issue. The outliner has a low compile-time overhead of
roughly 1%. At this point, the outliner would be a really good addition to the
-Oz pass pipeline!
llvm-svn: 338160
The tests with a constant sub operand were added with rL338143,
but the potential transform doesn't have that requirement, so
adding more tests with variable operands.
llvm-svn: 338150
This patch enables the MachineOutliner by default in AArch64 under -Oz.
The MachineOutliner offers around a 4.5% improvement on the current -Oz code
size improvements.
We have done work into improving the debuggability of outlined code, so that
users of -Oz won't be surprised by the optimization. We have also been executing
the LLVM test suite and common external tests such as the SPEC suites
continuously with no issue. The outliner has a low compile-time overhead of
roughly 1%. At this point, the outliner would be a really good addition to the
-Oz pass pipeline!
llvm-svn: 338133
This is a follow-up suggested in D48970.
Alive proofs:
https://rise4fun.com/Alive/sII
We can eliminate an instruction in the usual select-of-constants
to bit hack transform by adjusting the add/sub with constant.
This is always a win.
There are more transforms that are likely wins, but they may need
target hooks in case some targets do not benefit.
This is another step towards making up for canonicalizing to
select-of-constants in rL331486.
llvm-svn: 338132
When fusing instructions A and B, we must add all predecessors of B as
predecessors of A to avoid instructions getting scheduling in between.
There is a special case involving ExitSU: Every other node must be
scheduled before it by design and we don't need to make this explicit in
the graph, however when fusing with a different node we need to schedule
every othere node before the fused node too and we need to make this
explicit now: This patch adds a dependency from the fused node to all
roots in the graph.
Differential Revision: https://reviews.llvm.org/D49830
llvm-svn: 338046
Summary:
A follow-up for D49266 / rL337166.
At least one of these cases is more canonical,
so we really do have to handle it.
https://godbolt.org/g/pkzP3Xhttps://rise4fun.com/Alive/pQyhZZ
We won't get to these cases with I1 being -1,
as that will be constant-folded to true or false.
I'm also not sure we actually hit the 'ule' case,
but i think the worst think that could happen is that being dead code.
Reviewers: spatel, craig.topper, RKSimon, javed.absar, efriedma
Reviewed By: spatel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D49497
llvm-svn: 338044
This reverts commit r337951.
While that kind of shared constant generally works fine in a MinGW
setting, it broke some cases of inline assembly that worked before:
$ cat const-asm.c
int MULH(int a, int b) {
int rt, dummy;
__asm__ (
"imull %3"
:"=d"(rt), "=a"(dummy)
:"a"(a), "rm"(b)
);
return rt;
}
int func(int a) {
return MULH(a, 1);
}
$ clang -target x86_64-win32-gnu -c const-asm.c -O2
const-asm.c:4:9: error: invalid variant '00000001'
"imull %3"
^
<inline asm>:1:15: note: instantiated into assembly here
imull __real@00000001(%rip)
^
A similar error is produced for i686 as well. The same test with a
target of x86_64-win32-msvc or i686-win32-msvc works fine.
llvm-svn: 338018
If the DAGCombiner's rotate matching was working as expected,
I don't think we'd see any test diffs here.
This sidesteps the issue of custom lowering for rotates raised in PR38243:
https://bugs.llvm.org/show_bug.cgi?id=38243
...by only dealing with legal operations.
llvm-svn: 337966
GNU binutils tools have no problems with this kind of shared constants,
provided that we actually hook it up completely in AsmPrinter and
produce a global symbol.
This effectively reverts SVN r335918 by hooking the rest of it up
properly.
This feature was implemented originally in SVN r213006, with no reason
for why it can't be used for MinGW other than the fact that GCC doesn't
do it while MSVC does.
Differential Revision: https://reviews.llvm.org/D49646
llvm-svn: 337951
In SVN r334523, the first half of comdat constant pool handling was
hoisted from X86WindowsTargetObjectFile (which despite the name only
was used for msvc targets) into the arch independent
TargetLoweringObjectFileCOFF, but the other half of the handling was
left behind in X86AsmPrinter::GetCPISymbol.
With only half of the handling in place, inconsistent comdat
sections/symbols are created, causing issues with both GNU binutils
(avoided for X86 in SVN r335918) and with the MS linker, which
would complain like this:
fatal error LNK1143: invalid or corrupt file: no symbol for COMDAT section 0x4
Differential Revision: https://reviews.llvm.org/D49644
llvm-svn: 337950
This matches the structure used on X86 and ARM. This requires
a little bit of duplication of the parts that are equal in both
AArch64 COFF variants though.
Before SVN r335286, these classes didn't add anything that MCAsmInfoCOFF
didn't, but now they do.
This makes AArch64 match X86 in how comdat is used for float constants
for MinGW.
Differential Revision: https://reviews.llvm.org/D49637
llvm-svn: 337755
As discussed here:
http://lists.llvm.org/pipermail/llvm-dev/2018-May/123292.htmlhttp://lists.llvm.org/pipermail/llvm-dev/2018-July/124400.html
We want to add rotate intrinsics because the IR expansion of that pattern is 4+ instructions,
and we can lose pieces of the pattern before it gets to the backend. Generalizing the operation
by allowing 2 different input values (plus the 3rd shift/rotate amount) gives us a "funnel shift"
operation which may also be a single hardware instruction.
Initially, I thought we needed to define new DAG nodes for these ops, and I spent time working
on that (much larger patch), but then I concluded that we don't need it. At least as a first
step, we have all of the backend support necessary to match these ops...because it was required.
And shepherding these through the IR optimizer is the primary concern, so the IR intrinsics are
likely all that we'll ever need.
There was also a question about converting the intrinsics to the existing ROTL/ROTR DAG nodes
(along with improving the oversized shift documentation). Again, I don't think that's strictly
necessary (as the test results here prove). That can be an efficiency improvement as a small
follow-up patch.
So all we're left with is documentation, definition of the IR intrinsics, and DAG builder support.
Differential Revision: https://reviews.llvm.org/D49242
llvm-svn: 337221
Summary:
[[ https://bugs.llvm.org/show_bug.cgi?id=38149 | PR38149 ]]
As discussed in https://reviews.llvm.org/D49179#1158957 and later,
the IR for 'check for [no] signed truncation' pattern can be improved:
https://rise4fun.com/Alive/gBf
^ that pattern will be produced by Implicit Integer Truncation sanitizer,
https://reviews.llvm.org/D48958https://bugs.llvm.org/show_bug.cgi?id=21530
in signed case, therefore it is probably a good idea to improve it.
But the IR-optimal patter does not lower efficiently, so we want to undo it..
This handles the simple pattern.
There is a second pattern with predicate and constants inverted.
NOTE: we do not check uses here. we always do the transform.
Reviewers: spatel, craig.topper, RKSimon, javed.absar
Reviewed By: spatel
Subscribers: kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D49266
llvm-svn: 337166
This is almost the same as an existing IR canonicalization in instcombine,
so I'm assuming this is a good early generic DAG combine too.
The motivation comes from reduced bit-hacking for select-of-constants in IR
after rL331486. We want to restore that functionality in the DAG as noted in
the commit comments for that change and the llvm-dev discussion here:
http://lists.llvm.org/pipermail/llvm-dev/2018-July/124433.html
The PPC and AArch tests show that those targets are already doing something
similar. x86 will be neutral in the minimal case and generally better when
this pattern is extended with other ops as shown in the signbit-shift.ll tests.
Note the asymmetry: we don't include the (extend (ifneg X)) transform because
it already exists in SimplifySelectCC(), and that is verified in the later
unchanged tests in the signbit-shift.ll files. Without the 'not' op, the
general transform to use a shift is always a win because that's a single
instruction.
Alive proofs:
https://rise4fun.com/Alive/ysli
Name: if pos, get -1
%c = icmp sgt i16 %x, -1
%r = sext i1 %c to i16
=>
%n = xor i16 %x, -1
%r = ashr i16 %n, 15
Name: if pos, get 1
%c = icmp sgt i16 %x, -1
%r = zext i1 %c to i16
=>
%n = xor i16 %x, -1
%r = lshr i16 %n, 15
Differential Revision: https://reviews.llvm.org/D48970
llvm-svn: 337130
See D49247, D49266
I'm only adding the sane negative tests, and not
adding the one-use tests yet. Also, not adding
negative tests for the second pattern with inverted operands yet,
since it's handling will be added in later differential.
llvm-svn: 337014
As suggested by @efriedma on D49262 - changed the extractelement to a store to prevent SimplifyDemandedVectorElts from simplifying the build vectors - this keeps the immediate generation which was the point of the tests.
llvm-svn: 336981
See https://reviews.llvm.org/D47106 for details.
Reviewed By: probinson
Differential Revision: https://reviews.llvm.org/D47171
This commit drops that patch's changes to:
llvm/test/CodeGen/NVPTX/f16x2-instructions.ll
llvm/test/CodeGen/NVPTX/param-load-store.ll
For some reason, the dos line endings there prevent me from commiting
via the monorepo. A follow-up commit (not via the monorepo) will
finish the patch.
llvm-svn: 336843
Summary:
This patch adds support for the atomicrmw instructions and the strong
cmpxchg instruction to the IRTranslator.
I've left out weak cmpxchg because LangRef.rst isn't entirely clear on what
difference it makes to the backend. As far as I can tell from the code, it
only matters to AtomicExpandPass which is run at the LLVM-IR level.
Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, javed.absar
Reviewed By: qcolombet
Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits
Differential Revision: https://reviews.llvm.org/D40092
llvm-svn: 336589
D48278
Allow to reduce redundant shift masks.
For example:
x1 = x & 0xAB00
x2 = (x >> 8) & 0xAB
can be reduced to:
x1 = x & 0xAB00
x2 = x1 >> 8
It only allows folding when the masks and shift values are constants.
llvm-svn: 336426
r336120 resulted in falling back to SelectionDAG more often due to the G_STORE
MMOs not matching the vreg size. This fixes that by explicitly any-extending the
value.
llvm-svn: 336209
We currently don't any-extend vararg parameters before storing them to the stack
locations on Darwin. However, SelectionDAG however does this, and so user code
is in the wild which inadvertently relies on this extension. This can manifest
in cases where the value stored is (int)0, but the actual parameter is interpreted
by va_arg as a pointer, and so not extending to 64 bits causes the callee to
load additional undefined bits.
llvm-svn: 336120
This adds functionality to the outliner that allows targets to
specify certain functions that should be outlined from by default.
If a target supports default outlining, then it specifies that in
its TargetOptions. In the case that it does, and the user hasn't
specified that they *never* want to outline, the outliner will
be added to the pass pipeline and will run on those default functions.
This is a preliminary patch for turning the outliner on by default
under -Oz for AArch64.
https://reviews.llvm.org/D48776
llvm-svn: 336040
This is a recommit of r335887, which was erroneously committed earlier.
To enable the MachineOutliner by default on AArch64, we need to be able to
disable the MachineOutliner and also provide an option to "always" enable the
outliner.
This adds that capability. It allows the user to still use the old
-enable-machine-outliner option, which defaults to "always". This is building
up to allowing the user to specify "always" versus the target default
outlining behaviour.
https://reviews.llvm.org/D48682
llvm-svn: 335986
This is a recommit of r335879.
We shouldn't add the outliner when compiling at -O0 even if
-enable-machine-outliner is passed in. This makes sure that we
don't add it in this case.
This also removes -O0 from the outliner DWARF test.
llvm-svn: 335930
This reverts commit 9c7c10e4073a0bc6a759ce5cd33afbac74930091.
It relies on r335872 since that introduces the machine outliner
flags test. I meant to commit D48683 in that commit, but got mixed
up and committed D48682 instead. So, I'm reverting this and
r335872, since D48682 hasn't made it through review yet.
llvm-svn: 335882
We shouldn't add the outliner when compiling at -O0 even if
-enable-machine-outliner is passed in. This makes sure that we
don't add it in this case.
This also updates machine-outliner-flags to reflect the change
and improves the comment describing what that test does.
llvm-svn: 335879
To enable the MachineOutliner by default on AArch64, we need to be able to
disable the MachineOutliner and also provide an option to "always" enable the
outliner.
This adds that capability. It allows the user to still use the old
-enable-machine-outliner option, which defaults to "always". This is building
up to allowing the user to specify "always" versus the target-default
outlining behaviour.
llvm-svn: 335872
Now that we have the ability to legalize based on MMO's. Add support for
legalizing based on AtomicOrdering and use it to correct the legalization
of the atomic instructions.
Also extend all() to be a variadic template as this ruleset now requires
3 and 4 argument versions.
llvm-svn: 335767
As noted in the D44909 review, the transform from (fptosi+sitofp) to ftrunc
can produce -0.0 where the original code does not:
#include <stdio.h>
int main(int argc) {
float x;
x = -0.8 * argc;
printf("%f\n", (float)((int)x));
return 0;
}
$ clang -O0 -mavx fp.c ; ./a.out
0.000000
$ clang -O1 -mavx fp.c ; ./a.out
-0.000000
Ideally, we'd use IR/node flags to predicate the transform, but the IR parser
doesn't currently allow fast-math-flags on the cast instructions. So for now,
just use the function attribute that corresponds to clang's "-fno-signed-zeros"
option.
Differential Revision: https://reviews.llvm.org/D48085
llvm-svn: 335761
It isn't safe to outline sequences of instructions where x16/x17/nzcv live
across the sequence.
This teaches the outliner to check whether or not a specific canidate has
x16/x17/nzcv live across it and discard the candidate in the case that that is
true.
https://bugs.llvm.org/show_bug.cgi?id=37573https://reviews.llvm.org/D47655
llvm-svn: 335758
This patch adds a custom trunc store lowering for v4i8 vector types.
Since there is not v.4b register, the v4i8 is promoted to v4i16 (v.4h)
and default action for v4i8 is to extract each element and issue 4
byte stores.
A better strategy would be to extended the promoted v4i16 to v8i16
(with undef elements) and extract and store the word lane which
represents the v4i8 subvectores. The construction:
define void @foo(<4 x i16> %x, i8* nocapture %p) {
%0 = trunc <4 x i16> %x to <4 x i8>
%1 = bitcast i8* %p to <4 x i8>*
store <4 x i8> %0, <4 x i8>* %1, align 4, !tbaa !2
ret void
}
Can be optimized from:
umov w8, v0.h[3]
umov w9, v0.h[2]
umov w10, v0.h[1]
umov w11, v0.h[0]
strb w8, [x0, #3]
strb w9, [x0, #2]
strb w10, [x0, #1]
strb w11, [x0]
ret
To:
xtn v0.8b, v0.8h
str s0, [x0]
ret
The patch also adjust the memory cost for autovectorization, so the C
code:
void foo (const int *src, int width, unsigned char *dst)
{
for (int i = 0; i < width; i++)
*dst++ = *src++;
}
can be vectorized to:
.LBB0_4: // %vector.body
// =>This Inner Loop Header: Depth=1
ldr q0, [x0], #16
subs x12, x12, #4 // =4
xtn v0.4h, v0.4s
xtn v0.8b, v0.8h
st1 { v0.s }[0], [x2], #4
b.ne .LBB0_4
Instead of byte operations.
llvm-svn: 335735
This removes debug locations from ConstantSDNode and ConstantSDFPNode.
When this kind of node is materialized we no longer create a line table
entry which jumps back to the constant's first point of use. This makes
single-stepping behavior smoother, and it matches the model used by IR,
where Constants have no locations. See this thread for more context:
http://lists.llvm.org/pipermail/llvm-dev/2018-June/124164.html
I'd like to handle constant BuildVectorSDNodes and to try to eliminate
passing SDLocs to SelectionDAG::getConstant*() in follow-up commits.
Differential Revision: https://reviews.llvm.org/D48468
llvm-svn: 335497
This reverts commit d8f57105010cc7e78026e511d5def873fc91e0e7.
Original Commit:
Author: Haicheng Wu <haicheng@codeaurora.org>
Date: Sun Feb 18 13:51:33 2018 +0000
[AArch64] Coalesce Copy Zero during instruction selection
Add special case for copy of zero to avoid a double copy.
Differential Revision: https://reviews.llvm.org/D36104
Author's intention is to remove a BB that has one mov instruction. In
order to do that, d8f571050 pessmizes MachineSinking by introducing a
copy, such that mov instruction is NOT moved to the BB. Optimization
downstream gets rid of the BB with only mov instruction. This works well
if we have only one fall through branch as there is only one "extra"
mov instruction.
If we have multiple fall throughs, we will have a lot of redundant movs.
In such a case, it's better to have this BB which has one mov instruction.
This is causing degradation in jpeg, fft and other codebases. I believe
if we want to remove a BB with only one branch instruction, we should not
pessimize Machine Sinking at all, and find some other solution.
llvm-svn: 335251
Summary: This patch originated from D46562 and is a proper subset, with some issues addressed.
Reviewers: spatel, hfinkel, wristow, arsenm, javed.absar
Reviewed By: spatel
Subscribers: wdng, nhaehnle
Differential Revision: https://reviews.llvm.org/D47909
llvm-svn: 334996
Summary: This patch originated from D47388 and is a proper subset of the originating changes, containing only the fmf optimization guard extensions.
Reviewers: spatel, hfinkel, wristow, arsenm, javed.absar, rampitec, nhaehnle, nemanjai
Reviewed By: rampitec, nhaehnle
Subscribers: tpr, nemanjai, wdng
Differential Revision: https://reviews.llvm.org/D47918
llvm-svn: 334876
This is r334750 (which was reverted in r334754) with a fix for an
uninitialized variable that was caught by msan.
Original commit message:
> If a copy bundle happens to involve overlapping registers, we can end
> up with emitting the copies in an order that ends up clobbering some
> of the subregisters. Since instructions in the copy bundle
> semantically happen at the same time, this is incorrect and we need to
> make sure we order the copies such that this doesn't happen.
llvm-svn: 334756
If a copy bundle happens to involve overlapping registers, we can end
up with emitting the copies in an order that ends up clobbering some
of the subregisters. Since instructions in the copy bundle
semantically happen at the same time, this is incorrect and we need to
make sure we order the copies such that this doesn't happen.
Differential Revision: https://reviews.llvm.org/D48154
llvm-svn: 334750
We're constant folding here, so we shouldn't check uses. This matches
the IR optimizer behavior.
The x86 test shows the expected win. The AArch64 test shows something
else. This only seems to happen if the "generic" AArch64 CPU model is
used by MachineCombiner, so I'll file a bug report to follow-up.
llvm-svn: 334608
Register x20 is a callee-saved register which may be used for other
purposes in certain contexts, for example to hold special variables
within the kernel. This change adds support for reserving this register
both to frontend and backend to make this register usable for these
purposes.
Differential Revision: https://reviews.llvm.org/D46552
llvm-svn: 334531
Apparently, MachineInstr class definition as well as pretty much all of
the machine passes assume that the only kind of MachineInstr's operands
that is variadic for variadic opcodes is explicit non-definitions.
In particular, this assumption is made by MachineInstr::defs(), uses(),
and explicit_uses() methods, as well as by MachineCSE pass.
The assumption is incorrect judging from at least TableGen backend
implementation, that recognizes variable_ops in OutOperandList, and the
very existence of G_UNMERGE_VALUES generic opcode, or ARM load multiple
instructions, all of which have variadic defs.
In particular, MachineCSE pass breaks MIR with CSE'able G_UNMERGE_VALUES
instructions in it.
This commit implements MachineInstr::getNumExplicitDefs() similar to
pre-existing MachineInstr::getNumExplicitOperands(), fixes
MachineInstr::defs(), uses(), and explicit_uses(), and fixes MachineCSE
pass.
As the issue addressed seems to affect only machine passes that could be
ran mid-GlobalISel pipeline at the moment, the other passes aren't fixed
by this commit, like MachineLICM: that could be done on per-pass basis
when (if ever) they get adopted for GlobalISel.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D45640
llvm-svn: 334520
Summary:
In D47428, i propose to choose the `~(-(1 << nbits))` as the canonical form of low-bit-mask formation.
As it is seen from these tests, there is a reason for that.
AArch64 currently better handles `~(-(1 << nbits))`, but not the more traditional `(1 << nbits) - 1` (sic!).
The other way around for X86.
It would be much better to canonicalize.
It would seem that there is too much tests, but this is most of all the auto-generated possible variants
of C code that one would expect for BZHI to be formed, and then manually cleaned up a bit.
So this should be pretty representable, which somewhat good coverage...
Related links:
https://bugs.llvm.org/show_bug.cgi?id=36419https://bugs.llvm.org/show_bug.cgi?id=37603https://bugs.llvm.org/show_bug.cgi?id=37610https://rise4fun.com/Alive/idM
Reviewers: javed.absar, craig.topper, RKSimon, spatel
Reviewed By: RKSimon
Subscribers: kristof.beyls, llvm-commits, RKSimon, craig.topper, spatel
Differential Revision: https://reviews.llvm.org/D47452
llvm-svn: 334124
If no alignment is set, the abi/preferred alignment of structs will be
used which may be higher than required. This can lead to extra padding
and in the end an increase in data size.
Differential Revision: https://reviews.llvm.org/D47633
llvm-svn: 334099
Start by emitting remarks for very basic unsupported cases such as
irreducible CFGs and EHFunclets. The end goal is to be able to cover all
the cases where we give up with an explanation.
llvm-svn: 333972
Before we were relying on the any extend of the s1 to s32, but
for AAPCS we need to zero-extend it to at least s8.
Fixes PR36719
Differential Revision: https://reviews.llvm.org/D47425
llvm-svn: 333747
This is to make it clear what kind of bugs the LegalizerInfo::verifier
is able to catch and test its output
Reviewers: aemerson, qcolombet
Reviewed By: aemerson
Differential Revision: https://reviews.llvm.org/D46338
llvm-svn: 333597
As suggested in https://bugs.llvm.org/show_bug.cgi?id=32384#c1, this change
makes the inlining of `memset()` and `memcpy()` more aggressive when
compiling for speed. The tuning remains the same when optimizing for size.
Patch by: Sebastian Pop <s.pop@samsung.com>
Evandro Menezes <e.menezes@samsung.com>
Differential revision: https://reviews.llvm.org/D45098
llvm-svn: 333429
The existing code has three different ways to try to lower a 64-bit
immediate to the sequence ORR+MOVK. The result is messy: it misses
some possible sequences, and the order of the checks means we sometimes
emit two MOVKs when we only need one.
Instead, just use a simple loop to try all possible two-instruction
ORR+MOVK sequences.
Differential Revision: https://reviews.llvm.org/D47176
llvm-svn: 333218
When we're outlining a sequence that ends in a call, we can save up to
three instructions in the outlined function by turning the call into
a tail-call. I refer to this as thunk outlining because the resulting
outlined function looks like a thunk; suggestions welcome for a better
name.
In addition to making the outlined function shorter, thunk outlining
allows outlining calls which would otherwise be illegal to outline:
we don't need to save/restore LR, so we don't need to prove anything
about the stack access patterns of the callee.
To make this work effectively, I also added
MachineOutlinerInstrType::LegalTerminator to the generic MachineOutliner
code; this allows treating an arbitrary instruction as a terminator in
the suffix tree.
Differential Revision: https://reviews.llvm.org/D47173
llvm-svn: 333015
This is the FP sibling of D43141 with the corresponding IR change in rL327212.
We can't propagate undef here because if a variable operand is a NaN, these
binops must propagate NaN. Neither global nor node-level fast-math makes a
difference. If we have 'nnan', I think later folds can turn the NaN into undef.
The tests in X86/fp-undef.ll are meant to be the definitive verification for
these folds - everything reduces identically now.
The other test changes are collateral damage. They may need to be altered to
preserve their intent.
Differential Revision: https://reviews.llvm.org/D47026
llvm-svn: 332920
Summary:
This **appears** to be the last missing piece for the masked merge pattern handling in the backend.
This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]].
[[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly.
Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`)
Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated.
Differential Revision: https://reviews.llvm.org/D46528
llvm-svn: 332904
Counting the number of instructions is both unintuitive and inaccurate.
On AArch64, this only affects the generated remarks and certain rare
pseudo-instructions, but it will have a bigger impact on other targets.
Differential Revision: https://reviews.llvm.org/D46921
llvm-svn: 332685
We need to clean up the DAG floating-point undef logic.
This process is similar to how we handled integer undef
logic in D43141.
And as we did there, I'm trying to reduce the patch by
changing tests that would probably become meaningless
once we correct FP undef folding.
Follow-up to:
https://reviews.llvm.org/rL332534
...because that change wasn't enough.
llvm-svn: 332636
We need to clean up the DAG floating-point undef logic.
This process is similar to how we handled integer undef
logic in D43141.
And as we did there, I'm trying to reduce the patch by
changing tests that would probably become meaningless
once we correct FP undef folding.
llvm-svn: 332534
This breaks the code which saves and restores LR, so we can't outline
without doing something more complicated for stack adjustment.
Found by inspection; we get lucky in most cases because getMemOpInfo
only handles STRWpost, not any other pre/post-increment forms. But it
hits a couple of artificial testcases in the tree.
Differential Revision: https://reviews.llvm.org/D46920
llvm-svn: 332529
The cost computation assumes we do this correctly, but the actual
lowering was wrong.
Differential Revision: https://reviews.llvm.org/D46923
llvm-svn: 332514
Keep loads and stores together (target defines how many loads
and stores to gang up), such that it will help in pairing
and vectorization.
Differential Revision https://reviews.llvm.org/D46477
llvm-svn: 332482
We currently handle all aggregates by creating one large LLT, and letting the
legalizer deal with splitting them up. However using this approach means that
we can't support big endian code correctly.
This patch changes the way that the IRTranslator deals with aggregate values,
by splitting them up into their constituent element values. To do this, parts
of the translator need to be modified to deal with multiple VRegs for a single
Value.
A new Value to VReg mapper is introduced to help keep compile time under
control, currently there is no measurable impact on CTMark despite the extra
code being generated in some cases.
Patch is based on the original work of Tim Northover.
Differential Revision: https://reviews.llvm.org/D46018
llvm-svn: 332449
This patch re-introduces the "S" inline assembler constraint. This matches
an absolute symbolic address or a label reference. The primary use case is
asm("adrp %0, %1\n\t"
"add %0, %0, :lo12:%1" : "=r"(addr) : "S"(&var));
I say re-introduces as it seems like "S" was implemented in the original
AArch64 backend, but it looks like it wasn't carried forward to the merged
backend. The original implementation had A and L modifiers that could be
used to print ":lo12:" to the string. It looks like gcc doesn't use these
and :lo12: is expected to be written in the inline assembly string so I've
not implemented A and L. Clang already supports the S modifier.
Fixes PR37180
Differential Revision: https://reviews.llvm.org/D46745
llvm-svn: 332444
It doesn't matter much this late in the pipeline, but one place that
does check for it is the function alignment code.
Differential Revision: https://reviews.llvm.org/D46373
llvm-svn: 332415
When storing the 0th lane of a vector, use a simpler and usually more
efficient scalar store instead. In this case, also using the unscaled
offset.
Differential revision: https://reviews.llvm.org/D46762
llvm-svn: 332394
The test case added in r332265 had incomplete livein information which
was caught by the EXPENSIVE_CHECKS bot. Fix the livein information and
add -verify-machineinstrs to the test case.
llvm-svn: 332367
This is a simple hack based on what's proposed in D37686, but we can extend it if needed in follow-ups.
It gets us most of the FMF functionality that we want without adding any state bits to the flags. It
also intentionally leaves out non-FMF flags (nsw, etc) to minimize the patch.
It should provide a superset of the functionality from D46563 - the extra tests show propagation and
codegen diffs for fcmp, vecreduce, and FP libcalls.
The PPC log2() test shows the limits of this most basic approach - we only applied 'afn' to the last
node created for the call. AFAIK, there aren't any libcall optimizations based on the flags currently,
so that shouldn't make any difference.
Differential Revision: https://reviews.llvm.org/D46854
llvm-svn: 332358
Summary:
The BranchFolding pass is currently missing opportunities to hoist
common code if the hoisted-to block contains a single conditional branch
that has register uses. This occurs somewhat frequently on AArch64 with
CBZ/TBZ opcodes.
This change also eliminates some code differences when debug info is
present since the presence of e.g. DBG_VALUE instructions in the
hoisted-to block can enable hoisting that wouldn't have occurred without
them.
Reviewers: MatzeB, rnk, kparzysz, twoh, aprantl, javed.absar
Subscribers: kristof.beyls, JDevlieghere, mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D46324
llvm-svn: 332265
When storing the 0th lane of a vector, use a simpler and usually more efficient scalar store instead.
Differential revision: https://reviews.llvm.org/D46655
llvm-svn: 332251
This teaches tryToFoldExtOfLoad to set the right location on a
newly-created extload. With that in place, the logic for performing a
certain ([s|z]ext (load ...)) combine becomes identical for sexts and
zexts, and we can get rid of one copy of the logic.
The test case churn is due to dependencies on IROrders inherited from
the wrong SDLoc.
Part of: llvm.org/PR37262
Differential Revision: https://reviews.llvm.org/D46158
llvm-svn: 332118
Summary:
performPostLD1Combine in AArch64ISelLowering looks for vector
insert_vector_elt of a loaded value which it can optimize into a single
LD1LANE instruction. The code checking for the pattern was not checking
if the lane index was a constant which could cause two problems:
- an assert when lowering the LD1LANE ISD node since it assumes an
constant operand
- an assert in isel if the lane index value depends on the
post-incremented base register
Both of these issues are avoided by simply checking that the lane index
is a constant.
Fixes bug 35822.
Reviewers: t.p.northover, javed.absar
Subscribers: rengolin, kristof.beyls, mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D46591
llvm-svn: 332103
The second source operand of G_SHL, G_ASHR, and G_LSHR must preserve its
value as a (small) unsigned integer, therefore its incorrect to widen it
in any way but by zero extending it.
G_SHL was using G_ANYEXT and G_ASHR - G_SEXT (which is correct for their
destination and first source operands, but not the "number of bits to
shift" operand).
Generally, shifts aren't as similar to regular binary operations as it
might seem, for instance, they aren't commutative nor associative and
the second source operand usually requires a special treatment.
Reviewers: bogner, javed.absar, aivchenk, rovka
Reviewed By: bogner
Subscribers: igorb, kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D46413
llvm-svn: 331926
Reverting this to see if the clang-cmake-aarch64-global-isel and
clang-cmake-aarch64-quick bots are failing because of this commit.
We know it wasn't r331819.
llvm-svn: 331846
In order to set breakpoints on labels and list source code around
labels, we need collect debug information for labels, i.e., label
name, the function label belong, line number in the file, and the
address label located. In order to keep these information in LLVM
IR and to allow backend to generate debug information correctly.
We create a new kind of metadata for labels, DILabel. The format
of DILabel is
!DILabel(scope: !1, name: "foo", file: !2, line: 3)
We hope to keep debug information as much as possible even the
code is optimized. So, we create a new kind of intrinsic for label
metadata to avoid the metadata is eliminated with basic block.
The intrinsic will keep existing if we keep it from optimized out.
The format of the intrinsic is
llvm.dbg.label(metadata !1)
It has only one argument, that is the DILabel metadata. The
intrinsic will follow the label immediately. Backend could get the
label metadata through the intrinsic's parameter.
We also create DIBuilder API for labels to be used by Frontend.
Frontend could use createLabel() to allocate DILabel objects, and use
insertLabel() to insert llvm.dbg.label intrinsic in LLVM IR.
Differential Revision: https://reviews.llvm.org/D45024
Patch by Hsiangkai Wang.
llvm-svn: 331841
Refactoring LegalizerHelper::widenScalar member function reducing its
size by approximately a factor of 2 and (hopefuly) making it more
straightforward and regular by introducing widenScalarSrc and
widenScalarDst helper methods.
The new widenScalar* methods mutate the instructions in place instead
of recreating them from scratch and removing the originals. The
compile time implications of this were measured on sqlite3
amalgamation, targeting AArch64 in -O0:
LegalizerHelper::widenScalar: > 25% faster
Legalizer::runOnMachineFunction: ~ 4.0 - 4.5% faster
Also adding MachineOperand::setCImm and refactoring out
MachineIRBuilder::recordInsertion methods to make the change possible.
Reviewers: aditya_nandakumar, bogner, javed.absar, t.p.northover, ab, dsanders, arsenm
Reviewed By: aditya_nandakumar
Subscribers: wdng, rovka, kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D46414
llvm-svn: 331819
As Roman Tereshin pointed out in https://reviews.llvm.org/D45541, the
-global-isel option is redundant when -run-pass is given. -global-isel sets up
the GlobalISel passes in the pass manager but -run-pass skips that entirely and
configures it's own pipeline.
llvm-svn: 331603
Summary:
Previously, a extending load was represented at (G_*EXT (G_LOAD x)).
This had a few drawbacks:
* G_LOAD had to be legal for all sizes you could extend from, even if
registers didn't naturally hold those sizes.
* All sizes you could extend from had to be allocatable just in case the
extend went missing (e.g. by optimization).
* At minimum, G_*EXT and G_TRUNC had to be legal for these sizes. As we
improve optimization of extends and truncates, this legality requirement
would spread without considerable care w.r.t when certain combines were
permitted.
* The SelectionDAG importer required some ugly and fragile pattern
rewriting to translate patterns into this style.
This patch changes the representation to:
* (G_[SZ]EXTLOAD x)
* (G_LOAD x) any-extends when MMO.getSize() * 8 < ResultTy.getSizeInBits()
which resolves these issues by allowing targets to work entirely in their
native register sizes, and by having a more direct translation from
SelectionDAG patterns.
Each extending load can be lowered by the legalizer into separate extends
and loads, however a target that supports s1 will need the any-extending
load to extend to at least s8 since LLVM does not represent memory accesses
smaller than 8 bit. The legalizer can widenScalar G_LOAD into an
any-extending load but sign/zero-extending loads need help from something
else like a combiner pass. A follow-up patch that adds combiner helpers for
for this will follow.
The new representation requires that the MMO correctly reflect the memory
access so this has been corrected in a couple tests. I've also moved the
extending loads to their own tests since they are (mostly) separate opcodes
now. Additionally, the re-write appears to have invalidated two tests from
select-with-no-legality-check.mir since the matcher table no longer contains
loads that result in s1's and they aren't legal in AArch64 anymore.
Depends on D45540
Reviewers: ab, aditya_nandakumar, bogner, rtereshin, volkan, rovka, javed.absar
Reviewed By: rtereshin
Subscribers: javed.absar, llvm-commits, kristof.beyls
Differential Revision: https://reviews.llvm.org/D45541
llvm-svn: 331601
Summary:
Split off form D46031.
It seems we don't want to transform the pattern if the `xor`'s are actually `not`'s.
In vector case, this breaks `andnpd` / `vandnps` patterns.
That being said, we may want to re-visit this `not` handling, maybe in D46073.
Reviewers: spatel, craig.topper, javed.absar
Reviewed By: spatel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D46492
llvm-svn: 331595
Summary:
When checking if an instruction stores to a given frame index, check
that the instruction can write to memory before looking at the memory
operands list to avoid e.g. DBG_VALUE instructions that reference a
frame index preventing a load from that index from being hoisted.
Reviewers: dblaikie, MatzeB, qcolombet, reames, javed.absar
Subscribers: mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D46284
llvm-svn: 331549
This patch adds a custom lowering for ISD::MULH{S,U} used on divide by
constant optimization (DAGCombiner::BuildSDIV and DAGCombiner::BuildUDIV).
New patterns for smull and umull are added, so AArch64ISD::{S,U}MULL
can be correctly lowered to smull2 and umull2.
Reviewed By: SjoerdMeijer
Differential Revision: https://reviews.llvm.org/D46009
llvm-svn: 331522
This adds a some more tests, and adds some notes to tests which are using
a suboptimal lowering.
The constants with suboptimal lowerings seem to be relatively rare in
practice, but it might be a fun project to work on improvements.
llvm-svn: 331304
Setting the right SDLoc on a newly-created zextload fixes a line table
bug which resulted in non-linear stepping behavior.
Several backend tests contained CHECK lines which relied on the IROrder
inherited from the wrong SDLoc. This patch breaks that dependence where
feasbile and regenerates test cases where not.
In some cases, changing a node's IROrder may alter register allocation
and spill behavior. This can affect performance. I have chosen not to
prevent this by applying a "known good" IROrder to SDLocs, as this may
hide a more general bug in the scheduler, or cause regressions on other
test inputs.
rdar://33755881, Part of: llvm.org/PR37262
Differential Revision: https://reviews.llvm.org/D45995
llvm-svn: 331300
There are two separate fixes here:
* The lowering code for non-extending loads should report UnableToLegalize instead of emitting the same instruction.
* The target should not be requesting lowering of non-extending loads.
llvm-svn: 331201
Summary:
Previously, a extending load was represented at (G_*EXT (G_LOAD x)).
This had a few drawbacks:
* G_LOAD had to be legal for all sizes you could extend from, even if
registers didn't naturally hold those sizes.
* All sizes you could extend from had to be allocatable just in case the
extend went missing (e.g. by optimization).
* At minimum, G_*EXT and G_TRUNC had to be legal for these sizes. As we
improve optimization of extends and truncates, this legality requirement
would spread without considerable care w.r.t when certain combines were
permitted.
* The SelectionDAG importer required some ugly and fragile pattern
rewriting to translate patterns into this style.
This patch begins changing the representation to:
* (G_[SZ]EXTLOAD x)
* (G_LOAD x) any-extends when MMO.getSize() * 8 < ResultTy.getSizeInBits()
which resolves these issues by allowing targets to work entirely in their
native register sizes, and by having a more direct translation from
SelectionDAG patterns.
This patch introduces the new generic instructions and new variation on
G_LOAD and adds lowering for them to convert back to the existing
representations.
Depends on D45466
Reviewers: ab, aditya_nandakumar, bogner, rtereshin, volkan, rovka, aemerson, javed.absar
Reviewed By: aemerson
Subscribers: aemerson, kristof.beyls, javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D45540
llvm-svn: 331115
This commit makes it so that if you outline a def of some register, then the
call instruction created by the outliner actually reflects that the register
is defined by the call. It also makes it so that outlined functions don't
have the TracksLiveness property.
Outlined calls shouldn't break liveness assumptions that someone might make.
This also un-XFAILs the noredzone test, and updates the calls test.
llvm-svn: 331095
Extend the live-in check for all aliased registers so that we can
allow sinking Copy instructions when only implicit def is in successor's
live-in.
llvm-svn: 331072
Summary:
Currently only the memory size is supported but others can be added as
needed.
narrowScalar for G_LOAD and G_STORE now correctly update the
MachineMemOperand and will refuse to legalize atomics since those need more
careful expansions to maintain atomicity.
Reviewers: ab, aditya_nandakumar, bogner, rtereshin, aemerson, javed.absar
Reviewed By: aemerson
Subscribers: aemerson, rovka, kristof.beyls, javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D45466
llvm-svn: 331071
Put the first ldp at the end, so that the load-store optimizer can run
and merge the ldp and the add into a post-index ldp.
This didn't work in case no frame was needed and resulted in code size
regressions.
llvm-svn: 331044
This adds IR intrinsics for the AArch64 dot-product instructions introduced in
v8.2-A.
Differential revisioon: https://reviews.llvm.org/D46107
llvm-svn: 331036
The program might have unusual expectations for functions; for example,
the Linux kernel's build system warns if it finds references from .text
to .init.data.
I'm not sure this is something we actually want to make any guarantees
about (there isn't any explicit rule that would disallow outlining
in this case), but we might want to be conservative anyway.
Differential Revision: https://reviews.llvm.org/D46091
llvm-svn: 331007
Summary:
Use the FP for scavenged spill slot accesses to prevent corruption of
the callee-save region when the SP is re-aligned.
Based on problem and patch reported by @paulwalker-arm
This is an alternative to solution proposed in D45770
Reviewers: t.p.northover, paulwalker-arm, thegameg, javed.absar
Subscribers: qcolombet, mcrosier, paulwalker-arm, kristof.beyls, rengolin, javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D46063
llvm-svn: 330976
Debug var, expr and loc were only supported for non-fixed stack objects.
This patch adds the following fields to the "fixedStack:" entries, and
renames the ones from "stack:" to:
* debug-info-variable
* debug-info-expression
* debug-info-location
Differential Revision: https://reviews.llvm.org/D46032
llvm-svn: 330859
Before, the outliner would grab ADRPs that used LR/W30. This patch fixes
that by checking for explicit uses of those registers before the special-casing
for ADRPs.
This also adds a test that ensures that those sorts of ADRPs won't be outlined.
llvm-svn: 330783
This patch aims to provide correct dwarf unwind information in function
epilogue for X86.
It consists of two parts. The first part inserts CFI instructions that set
appropriate cfa offset and cfa register in emitEpilogue() in
X86FrameLowering. This part is X86 specific.
The second part is platform independent and ensures that:
* CFI instructions do not affect code generation (they are not counted as
instructions when tail duplicating or tail merging)
* Unwind information remains correct when a function is modified by
different passes. This is done in a late pass by analyzing information
about cfa offset and cfa register in BBs and inserting additional CFI
directives where necessary.
Added CFIInstrInserter pass:
* analyzes each basic block to determine cfa offset and register are valid
at its entry and exit
* verifies that outgoing cfa offset and register of predecessor blocks match
incoming values of their successors
* inserts additional CFI directives at basic block beginning to correct the
rule for calculating CFA
Having CFI instructions in function epilogue can cause incorrect CFA
calculation rule for some basic blocks. This can happen if, due to basic
block reordering, or the existence of multiple epilogue blocks, some of the
blocks have wrong cfa offset and register values set by the epilogue block
above them.
CFIInstrInserter is currently run only on X86, but can be used by any target
that implements support for adding CFI instructions in epilogue.
Patch by Violeta Vukobrat.
Differential Revision: https://reviews.llvm.org/D42848
llvm-svn: 330706
As we're becoming stricter w/ respect to not allowing vregs having LLTs
and regclasses assigned both mid-globalisel pipeline, the number of
extra copies grows, some of which separate G_UNMERGE's from their
corresponding G_MERGE's, becoming a performance concern.
It's worth mentioning that we're already looking through copies while
combining legalization artifacts for every kind of artifact but
G_UNMERGE.
Reviewed By: aditya_nandakumar
Reviewers: ab, t.p.northover, volkan, javed.absar
Subscribers: rovka, kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D45644
llvm-svn: 330660
Summary:
This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]].
[[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly.
Previously, `andl`+`andn`/`andps`+`andnps` / `bic`/`bsl` would be generated. (see `@out`)
Now, they would no longer be generated (see `@in`).
So we need to make sure that they are still generated.
If the mask is constant, we do nothing. InstCombine should have unfolded it.
Else, i use `hasAndNot()` TLI hook.
For now, only handle scalars.
https://rise4fun.com/Alive/bO6
----
I *really* don't like the code i wrote in `DAGCombiner::unfoldMaskedMerge()`.
It is super fragile. Is there something like IR Pattern Matchers for this?
Reviewers: spatel, craig.topper, RKSimon, javed.absar
Reviewed By: spatel
Subscribers: andreadb, courbet, kristof.beyls, javed.absar, rengolin, nemanjai, llvm-commits
Differential Revision: https://reviews.llvm.org/D45733
llvm-svn: 330646
Summary:
This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]].
[[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly.
Previously, `andl`+`andn`/`andps`+`andnps` / `bic`/`bsl` would be generated. (see `@out`)
Now, they would no longer be generated (see `@in`).
I'm guessing `llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp` should be able to unfold this.
Reviewers: spatel, craig.topper, RKSimon, javed.absar
Reviewed By: spatel
Subscribers: nemanjai, rengolin, javed.absar, kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D45563
llvm-svn: 330645
This reland includes a check to prevent the DAG combiner from folding an
offset that is smaller than the existing one. This can cause oscillations
between two possible DAGs, which was the cause of the hang and later assertion
failure observed on the lnt-ctmark-aarch64-O3-flto bot.
http://green.lab.llvm.org/green/job/lnt-ctmark-aarch64-O3-flto/2024/
Original commit message:
> This is a code size win in code that takes offseted addresses
> frequently, such as C++ constructors that typically need to compute
> an offseted address of a vtable. This reduces the size of Chromium
> for Android's .text section by 108KB.
Differential Revision: https://reviews.llvm.org/D45199
llvm-svn: 330630
In certain cases, the compiler might try to merge __stack_chk_guard with
another global variable. (Or someone could theoretically define
__stack_chk_guard as an alias.) In that case, make sure we don't crash.
Differential Revision: https://reviews.llvm.org/D45746
llvm-svn: 330495
First off, this is more correct than having the B. Second off, this was making
a bot upset. This fixes that.
Update the test to include -verify-machineinstrs as well to prevent stuff like
this slipping by non debug/assert builds in the future.
llvm-svn: 330459
This was originally committed at rL328921 and reverted at rL329920 to
investigate failures in Chrome. This time I've added to the ReleaseNotes
to warn users of the potential of exposing UB and let me repeat that
here for more exposure:
Optimization of floating-point casts is improved. This may cause surprising
results for code that is relying on undefined behavior. Code sanitizers can
be used to detect affected patterns such as this:
int main() {
float x = 4294967296.0f;
x = (float)((int)x);
printf("junk in the ftrunc: %f\n", x);
return 0;
}
$ clang -O1 ftrunc.c -fsanitize=undefined ; ./a.out
ftrunc.c:5:15: runtime error: 4.29497e+09 is outside the range of
representable values of type 'int'
junk in the ftrunc: 0.000000
Original commit message:
fptosi / fptoui round towards zero, and that's the same behavior as ISD::FTRUNC,
so replace a pair of casts with the equivalent node. We don't have to account for
special cases (NaN, INF) because out-of-range casts are undefined.
Differential Revision: https://reviews.llvm.org/D44909
llvm-svn: 330437
This is a code size win in code that takes offseted addresses
frequently, such as C++ constructors that typically need to compute
an offseted address of a vtable. This reduces the size of Chromium
for Android's .text section by 108KB.
Differential Revision: https://reviews.llvm.org/D45199
llvm-svn: 329956
AFI->setRedZone(false) was put in the wrong place before, and so it only fired
on functions that didn't have stack frames. This moves that to the top of
emitPrologue to make sure that every function without a redzone has it set
correctly.
This also adds a function representing one of the early exit cases (GHC calling
convention) to the MachineOutliner noredzone test to ensure that we can outline
from functions like these, where we never use a redzone.
llvm-svn: 329922
This change is exposing UB in source code - as was warned/predicted. :)
See D44909 for discussion. Reverting while we figure out how to fix things.
llvm-svn: 329920
This is causing compilation timeouts on code with long sequences of
local values and calls (i.e. foo(1); foo(2); foo(3); ...). It turns out
that code coverage instrumentation is a great way to create sequences
like this, which how our users ran into the issue in practice.
Intel has a tool that detects these kinds of non-linear compile time
issues, and Andy Kaylor reported it as PR37010.
The current sinking code scans the whole basic block once per local
value sink, which happens before emitting each call. In theory, local
values should only be introduced to be used by instructions between the
current flush point and the last flush point, so we should only need to
scan those instructions.
llvm-svn: 329822
Summary:
When inserting MOVs to avoid Falkor HWPF collisions, the non-base
register operand of load instructions (e.g. a register offset) was not
being considered live, so it could potentially have been used as a
scratch register, clobbering the actual offset value.
Reviewers: mcrosier
Subscribers: rengolin, javed.absar, kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D45502
llvm-svn: 329761
In the presence of variable-sized stack objects, we always picked the
base pointer when resolving frame indices if it was available.
This makes us hit an assert where we can't reach the emergency spill
slot if it's too far away from the base pointer. Since on AArch64 we
decide to place the emergency spill slot at the top of the frame, it
makes more sense to use FP to access it.
The changes here don't affect only emergency spill slots but all the
frame indices. The goal here is to try to choose between FP, BP and SP
so that we minimize the offset and avoid scavenging, or worse, asserting
when trying to access a slot allocated by the scavenger.
Previously discussed here: https://reviews.llvm.org/D40876.
Differential Revision: https://reviews.llvm.org/D45358
llvm-svn: 329691
This is a code size win in code that takes offseted addresses
frequently, such as C++ constructors that typically need to compute
an offseted address of a vtable. It reduces the size of Chromium for
Android's .text section by 46KB, or 56KB with ThinLTO (which exposes
more opportunities to use a direct access rather than a GOT access).
Because the addend range is limited in COFF and Mach-O, this is
enabled for ELF only.
Differential Revision: https://reviews.llvm.org/D45199
llvm-svn: 329611
Summary:
Currently MachineLoopInfo is used in only two places:
1) for computing IsBasicBlockInsideInnermostLoop field of MCCodePaddingContext, and it is never used.
2) in emitBasicBlockLoopComments, which is called only if `isVerbose()` is true.
Despite that, we currently have a dependency on MachineLoopInfo, which makes
pass manager to compute it and MachineDominator Tree. This patch removes the
use (1) and makes the use (2) lazy, thus avoiding some redundant
recomputations.
Reviewers: opaparo, gadi.haber, rafael, craig.topper, zvi
Subscribers: rengolin, javed.absar, hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D44812
llvm-svn: 329542
In our real world application, we found the following optimization is missed in DAGCombiner
(zext (and/or/xor (shl/shr (load x), cst), cst)) -> (and/or/xor (shl/shr (zextload x), (zext cst)), (zext cst))
If the user of original zext is an add, it may enable further lea optimization on x86.
This patch add a new function CombineZExtLogicopShiftLoad to do this optimization.
Differential Revision: https://reviews.llvm.org/D44402
llvm-svn: 329516
Should fix UBSan bot by also checking there's no "uwtable" attribute
before skipping. Otherwise the unwind table will be useless since its
moves expect CSRs to actually be preserved.
A noreturn nounwind function can be expected to never return in any way, and by
never returning it will also never have to restore any callee-saved registers
for its caller. This makes it possible to skip spills of those registers during
function entry, saving some stack space and time in the process. This is rather
useful for embedded targets with limited stack space.
Should fix PR9970.
Patch mostly by myeisha (pmb).
llvm-svn: 329494
A noreturn nounwind function can be expected to never return in any way, and by
never returning it will also never have to restore any callee-saved registers
for its caller. This makes it possible to skip spills of those registers during
function entry, saving some stack space and time in the process. This is rather
useful for embedded targets with limited stack space.
Should fix PR9970.
Patch by myeisha (pmb).
llvm-svn: 329287
The implementation of shadow call stack on aarch64 is quite different to
the implementation on x86_64. Instead of reserving a segment register for
the shadow call stack, we reserve the platform register, x18. Any function
that spills lr to sp also spills it to the shadow call stack, a pointer to
which is stored in x18.
Differential Revision: https://reviews.llvm.org/D45239
llvm-svn: 329236
This patch adds a hasRedZone() function to AArch64MachineFunctionInfo. It
returns true if the function is known to use a redzone, false if it is known
to not use a redzone, and no value otherwise.
This removes the requirement to pass -mno-red-zone when outlining for AArch64.
https://reviews.llvm.org/D45189
llvm-svn: 329120
The linkage type on outlined functions was private before. This meant that if
you set a breakpoint in an outlined function, the debugger wouldn't be able to
give a sane name to the outlined function.
This commit changes the linkage type to internal and updates any tests that
relied on the prefixes on the names of outlined functions.
llvm-svn: 329116
fptosi / fptoui round towards zero, and that's the same behavior as ISD::FTRUNC,
so replace a pair of casts with the equivalent node. We don't have to account for
special cases (NaN, INF) because out-of-range casts are undefined.
Differential Revision: https://reviews.llvm.org/D44909
llvm-svn: 328921
The code has bugs dealing with -0.0.
Since D44550 introduced FABS pattern folding in InstCombine,
this patch removes the now-redundant code that causes
https://bugs.llvm.org/show_bug.cgi?id=36600.
Patch by Mikhail Dvoretckii!
Differential Revision: https://reviews.llvm.org/D44683
llvm-svn: 328872
MachineCopyPropagation::CopyPropagateBlock has a bunch of special
handling for COPY instructions. This handling assumes that COPY
instructions do not modify the source of the copy; this is wrong if
the COPY destination overlaps the source.
To fix the bug, check explicitly for this situation, and fall back to
the generic instruction handling.
This bug can't happen for most register classes because they don't
have this sort of overlap, but there are a few register classes
where this is possible. The testcase uses the AArch64 QQQQ register
class.
Differential Revision: https://reviews.llvm.org/D44911
llvm-svn: 328851
This commit simplifies the call outlining logic by removing references to the
Function associated with the callee. To do this, it requires that valid
callee save info is available to the outliner.
llvm-svn: 328719
If an ADRP appears with, say, a CPI operand, we shouldn't outline it.
This moves the check for unsafe operands so that it occurs before the special-case
for ADRPs. Also add a test for outlining ADRPs.
llvm-svn: 328674
%tmp = bitcast i32* %arg to i8*
%tmp1 = getelementptr inbounds i8, i8* %tmp, i32 0
- %tmp2 = load i8, i8* %tmp, align 1
+ %tmp2 = load i8, i8* %tmp1, align 1
This doesn't change the semantics of the tests but makes use of %tmp1 which was originally intended.
llvm-svn: 328642
On Hexagon "x = y" is a syntax used in most instructions, and is not
treated as a directive.
Differential Revision: https://reviews.llvm.org/D44256
llvm-svn: 328635
Loads and stores can only shift the offset register by the size of the value
being loaded, but currently the DAGCombiner will reduce the width of the load
if it's followed by a trunc making it impossible to later combine the shift.
Solve this by implementing shouldReduceLoadWidth for the AArch64 backend and
make it prevent the width reduction if this is what would happen, though do
allow it if reducing the load width will let us eliminate a later sign or zero
extend.
Differential Revision: https://reviews.llvm.org/D44794
llvm-svn: 328321
This was being masked because GISel is enabled by default for -O0 and
the abort was disabled. Modified test to explicitly enable abort.
llvm-svn: 328311
Summary:
This pass sinks COPY instructions into a successor block, if the COPY is not
used in the current block and the COPY is live-in to a single successor
(i.e., doesn't require the COPY to be duplicated). This avoids executing the
the copy on paths where their results aren't needed. This also exposes
additional opportunites for dead copy elimination and shrink wrapping.
These copies were either not handled by or are inserted after the MachineSink
pass. As an example of the former case, the MachineSink pass cannot sink
COPY instructions with allocatable source registers; for AArch64 these type
of copy instructions are frequently used to move function parameters (PhyReg)
into virtual registers in the entry block..
For the machine IR below, this pass will sink %w19 in the entry into its
successor (%bb.1) because %w19 is only live-in in %bb.1.
```
%bb.0:
%wzr = SUBSWri %w1, 1
%w19 = COPY %w0
Bcc 11, %bb.2
%bb.1:
Live Ins: %w19
BL @fun
%w0 = ADDWrr %w0, %w19
RET %w0
%bb.2:
%w0 = COPY %wzr
RET %w0
```
As we sink %w19 (CSR in AArch64) into %bb.1, the shrink-wrapping pass will be
able to see %bb.0 as a candidate.
With this change I observed 12% more shrink-wrapping candidate and 13% more dead copies deleted in spec2000/2006/2017 on AArch64.
Reviewers: qcolombet, MatzeB, thegameg, mcrosier, gberry, hfinkel, john.brawn, twoh, RKSimon, sebpop, kparzysz
Reviewed By: sebpop
Subscribers: evandro, sebpop, sfertile, aemerson, mgorny, javed.absar, kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D41463
llvm-svn: 328237
When outlining calls, the outliner needs to update CFI to ensure that, say,
exception handling works. This commit adds that functionality and adds a test
just for call outlining.
Call outlining stuff in machine-outliner.mir should be moved into
machine-outliner-calls.mir in a later commit.
llvm-svn: 327917
This extends the use of this attribute on ARM and AArch64 from
SVN r325900 (where it was only checked for fixed stack
allocations on ARM/AArch64, but for all stack allocations on X86).
This also adds a testcase for the existing use of disabling the
fixed stack probe with the attribute on ARM and AArch64.
Differential Revision: https://reviews.llvm.org/D44291
llvm-svn: 327897
At the point the outliner runs, KILLs don't impact anything, but they're still
considered unique instructions. This commit makes them invisible like
DebugValues so that they can still be outlined without impacting outlining
decisions.
llvm-svn: 327760
This is a follow up of the AArch64 FP16 intrinsics work;
the codegen tests had not been added yet.
Differential Revision: https://reviews.llvm.org/D44510
llvm-svn: 327624
Summary:
Local values are constants, global addresses, and stack addresses that
can't be folded into the instruction that uses them. For example, when
storing the address of a global variable into memory, we need to
materialize that address into a register.
FastISel doesn't want to materialize any given local value more than
once, so it generates all local value materialization code at
EmitStartPt, which always dominates the current insertion point. This
allows it to maintain a map of local value registers, and it knows that
the local value area will always dominate the current insertion point.
The downside is that local value instructions are always emitted without
a source location. This is done to prevent jumpy line tables, but it
means that the local value area will be considered part of the previous
statement. Consider this C code:
call1(); // line 1
++global; // line 2
++global; // line 3
call2(&global, &local); // line 4
Today we end up with assembly and line tables like this:
.loc 1 1
callq call1
leaq global(%rip), %rdi
leaq local(%rsp), %rsi
.loc 1 2
addq $1, global(%rip)
.loc 1 3
addq $1, global(%rip)
.loc 1 4
callq call2
The LEA instructions in the local value area have no source location and
are treated as being on line 1. Stepping through the code in a debugger
and correlating it with the assembly won't make much sense, because
these materializations are only required for line 4.
This is actually problematic for the VS debugger "set next statement"
feature, which effectively assumes that there are no registers live
across statement boundaries. By sinking the local value code into the
statement and fixing up the source location, we can make that feature
work. This was filed as https://bugs.llvm.org/show_bug.cgi?id=35975 and
https://crbug.com/793819.
This change is obviously not enough to make this feature work reliably
in all cases, but I felt that it was worth doing anyway because it
usually generates smaller, more comprehensible -O0 code. I measured a
0.12% regression in code generation time with LLC on the sqlite3
amalgamation, so I think this is worth doing.
There are some special cases worth calling out in the commit message:
1. local values materialized for phis
2. local values used by no-op casts
3. dead local value code
Local values can be materialized for phis, and this does not show up as
a vreg use in MachineRegisterInfo. In this case, if there are no other
uses, this patch sinks the value to the first terminator, EH label, or
the end of the BB if nothing else exists.
Local values may also be used by no-op casts, which adds the register to
the RegFixups table. Without reversing the RegFixups map direction, we
don't have enough information to sink these instructions.
Lastly, if the local value register has no other uses, we can delete it.
This comes up when fastisel tries two instruction selection approaches
and the first materializes the value but fails and the second succeeds
without using the local value.
Reviewers: aprantl, dblaikie, qcolombet, MatzeB, vsk, echristo
Subscribers: dotdash, chandlerc, hans, sdardis, amccarth, javed.absar, zturner, llvm-commits, hiraditya
Differential Revision: https://reviews.llvm.org/D43093
llvm-svn: 327581
Get rid of the "; mem:" suffix and use the one we use in MIR: ":: (load 2)".
rdar://38163529
Differential Revision: https://reviews.llvm.org/D42377
llvm-svn: 327580
Optionally allow the order of restoring the callee-saved registers in the
epilogue to be reversed.
The flag -reverse-csr-restore-seq generates the following code:
```
stp x26, x25, [sp, #-64]!
stp x24, x23, [sp, #16]
stp x22, x21, [sp, #32]
stp x20, x19, [sp, #48]
; [..]
ldp x24, x23, [sp, #16]
ldp x22, x21, [sp, #32]
ldp x20, x19, [sp, #48]
ldp x26, x25, [sp], #64
ret
```
Note how the CSRs are restored in the same order as they are saved.
One exception to this rule is the last `ldp`, which allows us to merge
the stack adjustment and the ldp into a post-index ldp. This is done by
first generating:
ldp x26, x27, [sp]
add sp, sp, #64
which gets merged by the arm64 load store optimizer into
ldp x26, x25, [sp], #64
The flag is disabled by default.
llvm-svn: 327569
r327100 made us stop producing vecreduce-propagate-sd-flags.s, but it's
still sticking around on some bots. This makes the bots unhappy.
I'll revert this tomorrow.
llvm-svn: 327199
The code to match and produce more x86 vector blends was enabled for all
architectures even though the transform may pessimize the code for other
architectures that do not provide a vector blend instruction.
Added an aarch64 testcase to check that a VZIP instruction is generated instead
of byte movs.
Differential Revision: https://reviews.llvm.org/D44118
llvm-svn: 327132
This patch is a fix for PR36642.
While legalizing long vector types, make sure the smaller types get the
flags of the wider type.
bugzilla link: https://bugs.llvm.org/show_bug.cgi?id=36642
Change-Id: I0c2829639f094c862c10a6b51b342d4c2563e1fa
llvm-svn: 327079
The attached testcase started failing after the patch to define
isExtractSubvectorCheap with the following pattern mismatch:
ISEL: Starting pattern match
Initial Opcode index to 85068
Match failed at index 85076
LLVM ERROR: Cannot select: t47: v8i16 = insert_subvector undef:v8i16, t43, Constant:i64<0>
The code generated from llvm/lib/Target/AArch64/AArch64InstrInfo.td
def : Pat<(insert_subvector undef, (v4i16 FPR64:$src), (i32 0)),
(INSERT_SUBREG (v8i16 (IMPLICIT_DEF)), FPR64:$src, dsub)>;
is in ninja/lib/Target/AArch64/AArch64GenDAGISel.inc
At the location of the error it is:
/* 85076*/ OPC_CheckChild2Type, MVT::i32,
And it failed to match the type of operand 2.
Adding another def-pat for i64 fixes the failed def-pat error:
def : Pat<(insert_subvector undef, (v4i16 FPR64:$src), (i64 0)),
(INSERT_SUBREG (v8i16 (IMPLICIT_DEF)), FPR64:$src, dsub)>;
llvm-svn: 326949
Nick Desaulniers