When deciding if it is safe to optimize a conditional branch to a CBZ or
CBNZ the offsets of the BasicBlocks from the start of the function are
estimated. For inline assembly the generic getInlineAsmLength() function is
used to get a worst case estimate of the inline assembly by multiplying the
number of instructions by the max instruction size of 4 bytes. This
unfortunately doesn't take into account the generation of Thumb implicit IT
instructions. In edge cases such as when all the instructions in the block
are 4-bytes in size and there is an implicit IT then the size is
underestimated. This can cause an out of range CBZ or CBNZ to be generated.
The patch takes a conservative approach and assumes that every instruction
in the inline assembly block may have an implicit IT.
Fixes pr31805
Differential Revision: https://reviews.llvm.org/D52834
llvm-svn: 343960
Correctly check for relocations in the constant to promote. And don't
allow promoting a constant multiple times.
This partially fixes https://bugs.llvm.org//show_bug.cgi?id=32780 ;
it's not a complete fix because we also need to prevent
ARMConstantIslands from cloning the constant.
(-arm-promote-constant is currently off by default, and it stays off
with this patch. I'll look into turning it on again when all the known
issues are fixed.)
Differential Revision: https://reviews.llvm.org/D51472
llvm-svn: 343361
This mostly affects IR generated by non-clang frontends because clang
generally sets the alignment of globals explicitly.
Fixes https://bugs.llvm.org//show_bug.cgi?id=32394 .
(-arm-promote-constant is currently off by default, and it stays off
with this patch. I'll look into turning it on again when all the known
issues are fixed.)
Differential Revision: https://reviews.llvm.org/D51469
llvm-svn: 343359
The NoMovt feature prevents the use of MOVW/MOVT
instructions on Cortex-M23 for performance reasons.
These instructions are required for execute only code
so NoMovt should be disabled when that option is enabled.
Differential Revision: https://reviews.llvm.org/D52551
llvm-svn: 343302
It was the case when calling MO::dump(), but MI::dump() was still
depending on hasComplexRegisterTies().
The MIR output is not affected.
llvm-svn: 343107
When calculating whether a value can safely overflow for use by an
icmp, we weren't checking that the value couldn't wrap around. To do
this we need the icmp to be using a constant, as well as the incoming
add or sub.
bugzilla report: https://bugs.llvm.org/show_bug.cgi?id=39060
Differential Revision: https://reviews.llvm.org/D52463
llvm-svn: 343092
Adding NonNull as attributes to returned pointers has the unfortunate side
effect of disabling tail calls. This patch ignores the NonNull attribute when
we decide whether to tail merge, in the same way that we ignore the NoAlias
attribute, as it has no affect on the call sequence.
Differential Revision: https://reviews.llvm.org/D52238
llvm-svn: 343091
This broke Chromium's Android build (https://crbug.com/889390) and the
polly-aosp buildbot
(http://lab.llvm.org:8011/builders/aosp-O3-polly-before-vectorizer-unprofitable).
> Originally committed in rL342210 but was reverted in rL342260 because
> it was causing issues in vectorized code, because I had forgotten to
> ensure that we're operating on scalar values.
>
> Original commit message:
>
> On failing to find sequences that can be converted into dual macs,
> try to find sequential 16-bit loads that are used by muls which we
> can then use smultb, smulbt, smultt with a wide load.
>
> Differential Revision: https://reviews.llvm.org/D51983
llvm-svn: 343082
A sequence of VMUL and VADD instructions always give the same or better
performance than a fused VMLA instruction on the Cortex-M4 and Cortex-M33.
Executing the VMUL and VADD back-to-back requires the same cycles, but
having separate instructions allows scheduling to avoid the hazard between
these 2 instructions.
Differential Revision: https://reviews.llvm.org/D52289
llvm-svn: 342874
- The load store optimizer is currently merging multiple loads/stores into VLDM/VSTM with more than 16 doubleword registers
- This is an UNPREDICTABLE instruction and shouldn't be done
- It looks like the Limit for how many registers included in a merge got dropped at some point so I am reintroducing it in this patch
- This fixes https://bugs.llvm.org/show_bug.cgi?id=38389
Differential Revision: https://reviews.llvm.org/D52085
llvm-svn: 342872
Originally committed in rL342210 but was reverted in rL342260 because
it was causing issues in vectorized code, because I had forgotten to
ensure that we're operating on scalar values.
Original commit message:
On failing to find sequences that can be converted into dual macs,
try to find sequential 16-bit loads that are used by muls which we
can then use smultb, smulbt, smultt with a wide load.
Differential Revision: https://reviews.llvm.org/D51983
llvm-svn: 342870
Fixes the unwind information generated for floating-point registers.
Previously, all padding registers were assumed to be four bytes wide. Now, the
width of the register is used to specify the amount of padding.
Patch by Jackson Woodruff!
Differential revision: https://reviews.llvm.org/D51494
llvm-svn: 342545
Since Android API version 9 the Android libm has had the sincos functions, so
they should be recognised as libcalls and sincos optimisation should be applied.
Differential Revision: https://reviews.llvm.org/D52025
llvm-svn: 342471
This reverts r342395 as it caused error
> Argument value type does not match pointer operand type!
> %0 = atomicrmw volatile xchg i8* %_Value1, i32 1 monotonic, !dbg !25
> i8in function atomic_flag_test_and_set
> fatal error: error in backend: Broken function found, compilation aborted!
on bot http://green.lab.llvm.org/green/job/clang-stage1-configure-RA/
More details are available at https://reviews.llvm.org/D52080
llvm-svn: 342431
isSupportedValue explicitly checked and accepted many types of value,
primarily for debugging reasons. Remove most of these checks and do a
bit of refactoring now that the pass is more stable. This also enables
ZExts to be sources, but this has very little practical benefit at the
moment extend instructions will still be introduced.
Differential Revision: https://reviews.llvm.org/D52080
llvm-svn: 342395
We allow overflowing instructions if they're decreasing and only used
by an unsigned compare. Add the extra condition that the icmp cannot
be using a negative immediate.
Differential Revision: https://reviews.llvm.org/D52102
llvm-svn: 342392
On failing to find sequences that can be converted into dual macs,
try to find sequential 16-bit loads that are used by muls which we
can then use smultb, smulbt, smultt with a wide load.
Differential Revision: https://reviews.llvm.org/D51983
llvm-svn: 342210
We previously only allowed truncs as sinks, but now allow them as
sources too. We do this by checking that the result type is the
narrow type that we're trying to optimise for.
Differential Revision: https://reviews.llvm.org/D51978
llvm-svn: 342141
Part of FixConsts wrongly assumes either a 8- or 16-bit constant
which can result in the wrong constants being generated during
promotion.
Differential Revision: https://reviews.llvm.org/D52032
llvm-svn: 342140
The Technical Reference Manuals for these two CPUs state that branching
to an unaligned 32-bit instruction incurs an extra pipeline reload
penalty. That's bad.
This also enables the optimization at -Os since it costs on average one
byte per loop in return for 1 cycle per iteration, which is pretty good
going.
llvm-svn: 342127
SMLAD and SMLALD instructions also come in the form of SMLADX and
SMLALDX which perform an exchange on their second operand. To support
this, more of the loads in the MAC candidates are compared for
sequential access and a boolean value has been added to BinOpChain.
AddMACCandiate has been refactored into a small pattern matching
state machine to reduce the amount of duplicated code, but also to
enable the matching to be more flexible. CreateParallelMACPairs now
iterates through all the candidates to find parallel ones.
Differential Revision: https://reviews.llvm.org/D51424
llvm-svn: 342033
Search from i64 reducing phis, as well as i32, to allow the
generation of smlald instructions.
Differential Revision: https://reviews.llvm.org/D51101
llvm-svn: 341941
Because t2LDREX (& t2STREX) were marked as AddrModeNone, but did allow a
FrameIndex operand, rewriteT2FrameIndex asserted. This gives them a
proper addressing-mode and tells the rewriter about it so that encodable
offsets are exploited and others are rejected.
Should fix PR38828.
llvm-svn: 341642
SHF_ARM_PURECODE flag when being built with the -mexecute-only flag.
All code sections of an ELF must have the flag set for the final .text
section to be execute-only, otherwise the flag gets removed.
A HasData flag is added to MCSection to aid in the determination that
the section is empty. A virtual setTargetSectionFlags is added to
MCELFObjectTargetWriter to allow subclasses to set target specific
section flags to be added to sections which we then use in the ARM
backend to set SHF_ARM_PURECODE.
Patch by Ivan Lozano!
Reviewed By: echristo
Differential Revision: https://reviews.llvm.org/D48792
llvm-svn: 341593
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
The runtime pseudo relocations can't handle the ARM format embedded
addresses in movw/movt pairs. By using stubs, the potentially
dllimported addresses can be touched up by the runtime pseudo relocation
framework.
Differential Revision: https://reviews.llvm.org/D51450
llvm-svn: 341176
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
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
The inline sequence is very long (about 70 bytes on Thumb1), so it's
not really a good idea to inline it, especially when optimizing for
size.
Differential Revision: https://reviews.llvm.org/D47917
llvm-svn: 340458
On Windows, movw+movt pairs with relocations are handled with a single
relocation that covers them both. Therefore we can't inject anything
between these instructions, otherwise the relocation (which in LLVM
only is treated as the movw instruction's relocation, while the movt
instruction's relocation is dropped) will end up bogus.
These instructions are bundled up until right before the constant
islands pass, making this effectively the only place that can split
them apart.
Differential Revision: https://reviews.llvm.org/D51032
llvm-svn: 340451
This avoids a potential infinite loop setting and unsetting bits in the
mask.
Reduced from a failure on the polly-aosp bot.
Differential Revision: https://reviews.llvm.org/D51066
llvm-svn: 340446
Add intrinsic isel patterns for sxtb16, sxtab16, uxtb16 and uxtab16
so that they can perform a ror.
Differential Revision: https://reviews.llvm.org/D51034
llvm-svn: 340405
During combining, ReduceLoadWdith is used to combine AND nodes that
mask loads into narrow loads. This patch allows the mask to be a
shifted constant. This results in a narrow load which is then left
shifted to compensate for the new offset.
Differential Revision: https://reviews.llvm.org/D50432
llvm-svn: 340261
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
While searching through the use-def tree, ignore GetElementPtrInst
instructions because they don't need promoting and neither do their
indices. Otherwise, the wide indices prevent the transformation from
happening.
Differential Revision: https://reviews.llvm.org/D50762
llvm-svn: 339871
Originally committed in r339755 which was reverted in r339806 due to
an asan issue. The issue was caused by my assumption that operands to
a CallInst mapped to the FunctionType Params. CallInsts are now
handled by iterating over their ArgOperands instead of Operands.
Original Message:
Treat signed icmps as 'sinks', allowing them to be in the use-def
tree, enabling more promotions to be performed. As a sink, any
promoted incoming values need to be truncated before being used by
the signed icmp.
Differential Revision: https://reviews.llvm.org/D50067
llvm-svn: 339858
We only try to promote types with are smaller than 16-bits, but we
also need to check that the type is not less than 8-bits.
Differential Revision: https://reviews.llvm.org/D50769
llvm-svn: 339770
Treat signed icmps as 'sinks', allowing them to be in the use-def
tree, enabling more promotions to be performed. As a sink, any
promoted incoming values need to be truncated before being used by
the signed icmp.
Differential Revision: https://reviews.llvm.org/D50067
llvm-svn: 339755
Add pointers to the list of allowed types, but don't try to promote
them. Also fixed a bug with the promotion of undef values, so a new
value is now created instead of mutating in place. We also now only
promote if there's an instruction in the use-def chains other than
the icmp, sinks and sources.
Differential Revision: https://reviews.llvm.org/D50054
llvm-svn: 339754
LLVM normally prefers to minimize the number of bits set in an AND
immediate, but that doesn't always match the available ARM instructions.
In Thumb1 mode, prefer uxtb or uxth where possible; otherwise, prefer
a two-instruction sequence movs+ands or movs+bics.
Some potential improvements outlined in
ARMTargetLowering::targetShrinkDemandedConstant, but seems to work
pretty well already.
The ARMISelDAGToDAG fix ensures we don't generate an invalid UBFX
instruction due to a larger-than-expected mask. (It's orthogonal, in
some sense, but as far as I can tell it's either impossible or nearly
impossible to reproduce the bug without this change.)
According to my testing, this seems to consistently improve codesize by
a small amount by forming bic more often for ISD::AND with an immediate.
Differential Revision: https://reviews.llvm.org/D50030
llvm-svn: 339472
Enabling ARMCodeGenPrepare by default caused a whole load of
failures. This is due to zexts and truncs not being handled properly.
ZExts are messy so it's just easier to disable for now and truncs
are allowed only as 'sinks'. I still need to figure out why allowing
them as 'sources' causes so many failures. The other main changes are
that we are explicit in the types that we converting to, it's now
always 'TypeSize'. Type support is also now performed while checking
for valid opcodes as it unnecessarily complicated having the checks
are different stages.
I've moved the tests around too, so we have the zext and truncs in
their own file as well as the overflowing opcode tests.
Differential Revision: https://reviews.llvm.org/D50518
llvm-svn: 339432
Peter Smith