------------------------------------------------------------------------
r223708 | hfinkel | 2014-12-08 22:54:22 +0000 (Mon, 08 Dec 2014) | 13 lines
[PowerPC] Don't use a non-allocatable register to implement the 'cc' alias
GCC accepts 'cc' as an alias for 'cr0', and we need to do the same when
processing inline asm constraints. This had previously been implemented using a
non-allocatable register, named 'cc', that was listed as an alias of 'cr0', but
the infrastructure does not seem to support this properly (neither the register
allocator nor the scheduler properly accounts for the alias). Instead, we can
just process this as a naming alias inside of the inline asm
constraint-processing code, so we'll do that instead.
There are two regression tests, one where the post-RA scheduler did the wrong
thing with the non-allocatable alias, and one where the register allocator did
the wrong thing. Fixes PR21742.
------------------------------------------------------------------------
git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_35@223748 91177308-0d34-0410-b5e6-96231b3b80d8
------------------------------------------------------------------------
r223328 | hfinkel | 2014-12-04 00:46:20 +0000 (Thu, 04 Dec 2014) | 8 lines
[PowerPC] 'cc' should be an alias only to 'cr0'
We had mistakenly believed that GCC's 'cc' referred to the entire
condition-code register (cr0 through cr7) -- and implemented this in r205630 to
fix PR19326, but 'cc' is actually an alias only to 'cr0'. This is causing LLVM
to clobber too much with legacy code with inline asm using the 'cc' clobber.
Fixes PR21451.
------------------------------------------------------------------------
git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_35@223747 91177308-0d34-0410-b5e6-96231b3b80d8
------------------------------------------------------------------------
r223318 | hfinkel | 2014-12-03 23:40:13 +0000 (Wed, 03 Dec 2014) | 12 lines
[PowerPC] Fix inline asm memory operands not to use r0
On PowerPC, inline asm memory operands might be expanded as 0($r), where $r is
a register containing the address. As a result, this register cannot be r0, and
we need to enforce this register subclass constraint to prevent miscompiling
the code (we'd get this constraint for free with the usual instruction
definitions, but that scheme has no knowledge of how we end up printing inline
asm memory operands, and so here we need to do it 'by hand'). We can accomplish
this within the current address-mode selection framework by introducing an
explicit COPY_TO_REGCLASS node.
Fixes PR21443.
------------------------------------------------------------------------
git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_35@223746 91177308-0d34-0410-b5e6-96231b3b80d8
------------------------------------------------------------------------
r223220 | hfinkel | 2014-12-03 09:37:50 +0000 (Wed, 03 Dec 2014) | 23 lines
[PowerPC] Print all inline-asm consts as signed numbers
Almost all immediates in PowerPC assembly (both 32-bit and 64-bit) are signed
numbers, and it is important that we print them as such. To make sure that
happens, we change PPCTargetLowering::LowerAsmOperandForConstraint so that it
does all intermediate checks on a signed-extended int64_t value, and then
creates the resulting target constant using MVT::i64. This will ensure that all
negative values are printed as negative values (mirroring what is done in other
backends to achieve the same sign-extension effect).
This came up in the context of inline assembly like this:
"add%I2 %0,%0,%2", ..., "Ir"(-1ll)
where we used to print:
addi 3,3,4294967295
and gcc would print:
addi 3,3,-1
and gas accepts both forms, but our builtin assembler (correctly) does not. Now
we print -1 like gcc does.
While here, I replaced a bunch of custom integer checks with isInt<16> and
friends from MathExtras.h.
Thanks to Paul Hargrove for the bug report.
------------------------------------------------------------------------
git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_35@223745 91177308-0d34-0410-b5e6-96231b3b80d8
------------------------------------------------------------------------
r222996 | foad | 2014-12-01 09:42:32 +0000 (Mon, 01 Dec 2014) | 19 lines
[PowerPC] Fix unwind info with dynamic stack realignment
Summary:
PowerPC DWARF unwind info defined CFA as SP + offset even in a function
where the stack had been dynamically realigned. This clearly doesn't
work because the offset from SP to CFA is not a constant. Fix it by
defining CFA as BP instead.
This was causing the AddressSanitizer null_deref test to fail 50% of
the time, depending on whether SP happened to be 32-byte aligned on
entry to a particular function or not.
Reviewers: willschm, uweigand, hfinkel
Reviewed By: hfinkel
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D6410
------------------------------------------------------------------------
git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_35@223744 91177308-0d34-0410-b5e6-96231b3b80d8
------------------------------------------------------------------------
r222672 | uweigand | 2014-11-24 18:09:47 +0000 (Mon, 24 Nov 2014) | 10 lines
[PowerPC] Fix PR 21652 - copy st_other bits on symbol assignment
When processing an assignment in the integrated assembler that sets
a symbol to the value of another symbol, we need to copy the st_other
bits that encode the local entry point offset.
Modeled after MipsTargetELFStreamer::emitAssignment handling of the
ELF::STO_MIPS_MICROMIPS flag.
------------------------------------------------------------------------
git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_35@223743 91177308-0d34-0410-b5e6-96231b3b80d8
------------------------------------------------------------------------
r221703 | wschmidt | 2014-11-11 20:44:09 +0000 (Tue, 11 Nov 2014) | 48 lines
[PowerPC] Replace foul hackery with real calls to __tls_get_addr
My original support for the general dynamic and local dynamic TLS
models contained some fairly obtuse hacks to generate calls to
__tls_get_addr when lowering a TargetGlobalAddress. Rather than
generating real calls, special GET_TLS_ADDR nodes were used to wrap
the calls and only reveal them at assembly time. I attempted to
provide correct parameter and return values by chaining CopyToReg and
CopyFromReg nodes onto the GET_TLS_ADDR nodes, but this was also not
fully correct. Problems were seen with two back-to-back stores to TLS
variables, where the call sequences ended up overlapping with unhappy
results. Additionally, since these weren't real calls, the proper
register side effects of a call were not recorded, so clobbered values
were kept live across the calls.
The proper thing to do is to lower these into calls in the first
place. This is relatively straightforward; see the changes to
PPCTargetLowering::LowerGlobalTLSAddress() in PPCISelLowering.cpp.
The changes here are standard call lowering, except that we need to
track the fact that these calls will require a relocation. This is
done by adding a machine operand flag of MO_TLSLD or MO_TLSGD to the
TargetGlobalAddress operand that appears earlier in the sequence.
The calls to LowerCallTo() eventually find their way to
LowerCall_64SVR4() or LowerCall_32SVR4(), which call FinishCall(),
which calls PrepareCall(). In PrepareCall(), we detect the calls to
__tls_get_addr and immediately snag the TargetGlobalTLSAddress with
the annotated relocation information. This becomes an extra operand
on the call following the callee, which is expected for nodes of type
tlscall. We change the call opcode to CALL_TLS for this case. Back
in FinishCall(), we change it again to CALL_NOP_TLS for 64-bit only,
since we require a TOC-restore nop following the call for the 64-bit
ABIs.
During selection, patterns in PPCInstrInfo.td and PPCInstr64Bit.td
convert the CALL_TLS nodes into BL_TLS nodes, and convert the
CALL_NOP_TLS nodes into BL8_NOP_TLS nodes. This replaces the code
removed from PPCAsmPrinter.cpp, as the BL_TLS or BL8_NOP_TLS
nodes can now be emitted normally using their patterns and the
associated printTLSCall print method.
Finally, as a result of these changes, all references to get-tls-addr
in its various guises are no longer used, so they have been removed.
There are existing TLS tests to verify the changes haven't messed
anything up). I've added one new test that verifies that the problem
with the original code has been fixed.
------------------------------------------------------------------------
git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_35@223742 91177308-0d34-0410-b5e6-96231b3b80d8
------------------------------------------------------------------------
r220959 | uweigand | 2014-10-31 10:33:14 +0000 (Fri, 31 Oct 2014) | 13 lines
[PowerPC] Load BlockAddress values from the TOC in 64-bit SVR4 code
Since block address values can be larger than 2GB in 64-bit code, they
cannot be loaded simply using an @l / @ha pair, but instead must be
loaded from the TOC, just like GlobalAddress, ConstantPool, and
JumpTable values are.
The commit also fixes a bug in PPCLinuxAsmPrinter::doFinalization where
temporary labels could not be used as TOC values, since code would
attempt (and fail) to use GetOrCreateSymbol to create a symbol of the
same name as the temporary label.
------------------------------------------------------------------------
git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_35@223741 91177308-0d34-0410-b5e6-96231b3b80d8
------------------------------------------------------------------------
r219441 | sfantao | 2014-10-09 20:42:56 +0000 (Thu, 09 Oct 2014) | 3 lines
Fix bug in GPR to FPR moves in PPC64LE.
The current implementation of GPR->FPR register moves uses a stack slot. This mechanism writes a double word and reads a word. In big-endian the load address must be displaced by 4-bytes in order to get the right value. In little endian this is no longer required. This patch fixes the issue and adds LE regression tests to fast-isel-conversion which currently expose this problem.
------------------------------------------------------------------------
git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_35@223740 91177308-0d34-0410-b5e6-96231b3b80d8
------------------------------------------------------------------------
r217993 | sfantao | 2014-09-17 23:25:06 +0000 (Wed, 17 Sep 2014) | 5 lines
Fix FastISel bug in boolean returns for PowerPC.
For PPC targets, FastISel does not take the sign extension information into account when selecting return instructions whose operands are constants. A consequence of this is that the return of boolean values is not correct. This patch fixes the problem by evaluating the sign extension information also for constants, forwarding this information to PPCMaterializeInt which takes this information to drive the sign extension during the materialization.
------------------------------------------------------------------------
git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_35@223739 91177308-0d34-0410-b5e6-96231b3b80d8
------------------------------------------------------------------------
r214517 | uweigand | 2014-08-01 14:35:58 +0000 (Fri, 01 Aug 2014) | 8 lines
[PowerPC] PR20280 - Slots for byval parameters are not immutable
Found by inspection while looking at PR20280: code would mark slots
in the parameter save area where a byval parameter is passed as
"immutable". This is not correct since code is allowed to modify
byval parameters in place in the parameter save area.
------------------------------------------------------------------------
git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_35@223736 91177308-0d34-0410-b5e6-96231b3b80d8
------------------------------------------------------------------------
r215793 | hfinkel | 2014-08-16 00:16:29 +0000 (Sat, 16 Aug 2014) | 9 lines
[PowerPC] Darwin byval arguments are not immutable
On PPC/Darwin, byval arguments occur at fixed stack offsets in the callee's
frame, but are not immutable -- the pointer value is directly available to the
higher-level code as the address of the argument, and the value of the byval
argument can be modified at the IR level.
This is necessary, but not sufficient, to fix PR20280. When PR20280 is fixed in
a follow-up commit, its test case will cover this change.
------------------------------------------------------------------------
git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_35@223735 91177308-0d34-0410-b5e6-96231b3b80d8
As it turns out, the capture tracker named CaptureBefore used by AA, and now
available via the PointerMayBeCapturedBefore function, would have been
more-aptly named CapturedBeforeOrAt, because it considers captures at the
instruction provided. This is not always what one wants, and it is difficult to
get the strictly-before behavior given only the current interface. This adds an
additional parameter which controls whether or not you want to include
captures at the provided instruction. The default is not to include the
instruction provided, so that 'Before' matches its name.
No functionality change intended.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213582 91177308-0d34-0410-b5e6-96231b3b80d8
There were two generally-useful CaptureTracker classes defined in LLVM: the
simple tracker defined in CaptureTracking (and made available via the
PointerMayBeCaptured utility function), and the CapturesBefore tracker
available only inside of AA. This change moves the CapturesBefore tracker into
CaptureTracking, generalizes it slightly (by adding a ReturnCaptures
parameter), and makes it generally available via a PointerMayBeCapturedBefore
utility function.
This logic will be needed, for example, to perform noalias function parameter
attribute inference.
No functionality change intended.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213519 91177308-0d34-0410-b5e6-96231b3b80d8
The ability to identify function locals will exist outside of BasicAA (for
example, logic for inferring noalias function arguments will need this), so
make this concept generally accessible without code duplication.
No functionality change.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213514 91177308-0d34-0410-b5e6-96231b3b80d8
Prior to this change, the loop vectorizer did not make use of the alias
analysis infrastructure. Instead, it performed memory dependence analysis using
ScalarEvolution-based linear dependence checks within equivalence classes
derived from the results of ValueTracking's GetUnderlyingObjects.
Unfortunately, this meant that:
1. The loop vectorizer had logic that essentially duplicated that in BasicAA
for aliasing based on identified objects.
2. The loop vectorizer could not partition the space of dependency checks
based on information only easily available from within AA (TBAA metadata is
currently the prime example).
This means, for example, regardless of whether -fno-strict-aliasing was
provided, the vectorizer would only vectorize this loop with a runtime
memory-overlap check:
void foo(int *a, float *b) {
for (int i = 0; i < 1600; ++i)
a[i] = b[i];
}
This is suboptimal because the TBAA metadata already provides the information
necessary to show that this check unnecessary. Of course, the vectorizer has a
limit on the number of such checks it will insert, so in practice, ignoring
TBAA means not vectorizing more-complicated loops that we should.
This change causes the vectorizer to use an AliasSetTracker to keep track of
the pointers in the loop. The resulting alias sets are then used to partition
the space of dependency checks, and potential runtime checks; this results in
more-efficient vectorizations.
When pointer locations are added to the AliasSetTracker, two things are done:
1. The location size is set to UnknownSize (otherwise you'd not catch
inter-iteration dependencies)
2. For instructions in blocks that would need to be predicated, TBAA is
removed (because the metadata might have a control dependency on the condition
being speculated).
For non-predicated blocks, you can leave the TBAA metadata. This is safe
because you can't have an iteration dependency on the TBAA metadata (if you
did, and you unrolled sufficiently, you'd end up with the same pointer value
used by two accesses that TBAA says should not alias, and that would yield
undefined behavior).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213486 91177308-0d34-0410-b5e6-96231b3b80d8
IRBuilder has CreateAligned(Load|Store) functions; use them and we don't need
to make a second call to setAlignment.
No functionality change intended.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213453 91177308-0d34-0410-b5e6-96231b3b80d8
There are some kinds of metadata that are safe to propagate from the scalar
instructions to the vector instructions (fpmath and tbaa currently).
Regarding TBAA, one might worry about propagating it on if-converted loads and
stores, because the metadata might have had a control dependency on the
condition, and thus actually aliased with some other non-speculated memory
access when the condition was false. However, this would be caught by the
runtime overlap checks.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213452 91177308-0d34-0410-b5e6-96231b3b80d8
All of the other similar functions in that part of the file look through
addrspacecast in addition to bitcast, and I see no reason why
stripAndAccumulateInBoundsConstantOffsets shouldn't do so also.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213449 91177308-0d34-0410-b5e6-96231b3b80d8
When we have a parameter (or call site return) with a dereferenceable
attribute, it can specify the size of an array pointed to by that parameter. If
we have a value for which we can accumulate a constant offset to such a
parameter, then we can use that offset in a direct comparison with the size
specified by the dereferenceable attribute.
This enables us to handle cases like this:
int foo(int a[static 3]) {
return a[2]; /* this is always dereferenceable */
}
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213447 91177308-0d34-0410-b5e6-96231b3b80d8
This adds initial support for PPC32 ELF PIC (Position Independent Code; the
-fPIC variety), thus rectifying a long-standing deficiency in the PowerPC
backend.
Patch by Justin Hibbits!
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213427 91177308-0d34-0410-b5e6-96231b3b80d8
This attribute indicates that the parameter or return pointer is
dereferenceable. Practically speaking, loads from such a pointer within the
associated byte range are safe to speculatively execute. Such pointer
parameters are common in source languages (C++ references, for example).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213385 91177308-0d34-0410-b5e6-96231b3b80d8
Currently the only kind of integer IR attributes that we have are alignment
attributes, and so the attribute kind that takes an integer parameter is called
AlignAttr, but that will change (we'll soon be adding a dereferenceable
attribute that also takes an integer value). Accordingly, rename AlignAttribute
to IntAttribute (class names, enums, etc.).
No functionality change intended.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213352 91177308-0d34-0410-b5e6-96231b3b80d8
This reverts, "r213024 - Revert r212572 "improve BasicAA CS-CS queries", it
causes PR20303." with a fix for the bug in pr20303. As it turned out, the
relevant code was both wrong and over-conservative (because, as with the code
it replaced, it would return the overall ModRef mask even if just Ref had been
implied by the argument aliasing results). Hopefully, this correctly fixes both
problems.
Thanks to Nick Lewycky for reducing the test case for pr20303 (which I've
cleaned up a little and added in DSE's test directory). The BasicAA test has
also been updated to check for this error.
Original commit message:
BasicAA contains knowledge of certain intrinsics, such as memcpy and memset,
and uses that information to form more-accurate answers to CallSite vs. Loc
ModRef queries. Unfortunately, it did not use this information when answering
CallSite vs. CallSite queries.
Generically, when an intrinsic takes one or more pointers and the intrinsic is
marked only to read/write from its arguments, the offset/size is unknown. As a
result, the generic code that answers CallSite vs. CallSite (and CallSite vs.
Loc) queries in AA uses UnknownSize when forming Locs from an intrinsic's
arguments. While BasicAA's CallSite vs. Loc override could use more-accurate
size information for some intrinsics, it did not do the same for CallSite vs.
CallSite queries.
This change refactors the intrinsic-specific logic in BasicAA into a generic AA
query function: getArgLocation, which is overridden by BasicAA to supply the
intrinsic-specific knowledge, and used by AA's generic implementation. This
allows the intrinsic-specific knowledge to be used by both CallSite vs. Loc and
CallSite vs. CallSite queries, and simplifies the BasicAA implementation.
Currently, only one function, Mac's memset_pattern16, is handled by BasicAA
(all the rest are intrinsics). As a side-effect of this refactoring, BasicAA's
getModRefBehavior override now also returns OnlyAccessesArgumentPointees for
this function (which is an improvement).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213219 91177308-0d34-0410-b5e6-96231b3b80d8
This is the one remaining place I see where passing
isSafeToSpeculativelyExecute a DataLayout pointer might matter (at least for
loads) -- I think I got the others in r212720. Most of the other remaining
callers of isSafeToSpeculativelyExecute only use it for call sites (or
otherwise exclude loads).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@212730 91177308-0d34-0410-b5e6-96231b3b80d8
isSafeToSpeculativelyExecute can optionally take a DataLayout pointer. In the
past, this was mainly used to make better decisions regarding divisions known
not to trap, and so was not all that important for users concerned with "cheap"
instructions. However, now it also helps look through bitcasts for
dereferencable loads, and will also be important if/when we add a
dereferencable pointer attribute.
This is some initial work to feed a DataLayout pointer through to callers of
isSafeToSpeculativelyExecute, generally where one was already available.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@212720 91177308-0d34-0410-b5e6-96231b3b80d8
isDereferenceablePointer should not give up upon encountering any bitcast. If
we're casting from a pointer to a larger type to a pointer to a small type, we
can continue by examining the bitcast's operand. This missing capability
was noted in a comment in the function.
In order for this to work, isDereferenceablePointer now takes an optional
DataLayout pointer (essentially all callers already had such a pointer
available). Most code uses isDereferenceablePointer though
isSafeToSpeculativelyExecute (which already took an optional DataLayout
pointer), and to enable the LICM test case, LICM needs to actually provide its DL
pointer to isSafeToSpeculativelyExecute (which it was not doing previously).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@212686 91177308-0d34-0410-b5e6-96231b3b80d8
BasicAA contains knowledge of certain intrinsics, such as memcpy and memset,
and uses that information to form more-accurate answers to CallSite vs. Loc
ModRef queries. Unfortunately, it did not use this information when answering
CallSite vs. CallSite queries.
Generically, when an intrinsic takes one or more pointers and the intrinsic is
marked only to read/write from its arguments, the offset/size is unknown. As a
result, the generic code that answers CallSite vs. CallSite (and CallSite vs.
Loc) queries in AA uses UnknownSize when forming Locs from an intrinsic's
arguments. While BasicAA's CallSite vs. Loc override could use more-accurate
size information for some intrinsics, it did not do the same for CallSite vs.
CallSite queries.
This change refactors the intrinsic-specific logic in BasicAA into a generic AA
query function: getArgLocation, which is overridden by BasicAA to supply the
intrinsic-specific knowledge, and used by AA's generic implementation. This
allows the intrinsic-specific knowledge to be used by both CallSite vs. Loc and
CallSite vs. CallSite queries, and simplifies the BasicAA implementation.
Currently, only one function, Mac's memset_pattern16, is handled by BasicAA
(all the rest are intrinsics). As a side-effect of this refactoring, BasicAA's
getModRefBehavior override now also returns OnlyAccessesArgumentPointees for
this function (which is an improvement).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@212572 91177308-0d34-0410-b5e6-96231b3b80d8
This reverts r208640 (I've just XFAILed the test) because it broke ppc64/Linux
self-hosting. Because nearly every regression test triggers a segfault, I hope
this will be easy to fix.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@209747 91177308-0d34-0410-b5e6-96231b3b80d8
This reverts commit r209638 because it broke self-hosting on ppc64/Linux. (the
Clang-compiled TableGen would segfault because it jumped to an invalid address
from within _ZNK4llvm17ManagedStaticBase21RegisterManagedStaticEPFPvvEPFvS1_E
(which is within the command-line parameter registration process)).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@209745 91177308-0d34-0410-b5e6-96231b3b80d8
Support for the intrinsics that read from and write to global named registers
is added for r1, r2 and r13 (depending on the subtarget).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208509 91177308-0d34-0410-b5e6-96231b3b80d8
We must validate the value type in TLI::getRegisterByName, because if we
don't and the wrong type was used with the IR intrinsic, then we'll assert
(because we won't be able to find a valid register class with which to
construct the requested copy operation). For PPC64, additionally, the type
information is necessary to decide between the 64-bit register and the 32-bit
subregister.
No functionality change.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208508 91177308-0d34-0410-b5e6-96231b3b80d8
The counter-loops formation pass needs to know what operations might be
function calls (because they can't appear in counter-based loops). On PPC32,
128-bit shifts might be runtime calls (even though you can't use __int128 on
PPC32, it seems that SROA might form them).
Fixes PR19709.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208501 91177308-0d34-0410-b5e6-96231b3b80d8
The old method used by X86TTI to determine partial-unrolling thresholds was
messy (because it worked by testing target features), and also would not
correctly identify the target CPU if certain target features were disabled.
After some discussions on IRC with Chandler et al., it was decided that the
processor scheduling models were the right containers for this information
(because it is often tied to special uop dispatch-buffer sizes).
This does represent a small functionality change:
- For generic x86-64 (which uses the SB model and, thus, will get some
unrolling).
- For AMD cores (because they still currently use the SB scheduling model)
- For Haswell (based on benchmarking by Louis Gerbarg, it was decided to bump
the default threshold to 50; we're working on a test case for this).
Otherwise, nothing has changed for any other targets. The logic, however, has
been moved into BasicTTI, so other targets may now also opt-in to this
functionality simply by setting LoopMicroOpBufferSize in their processor
model definitions.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208289 91177308-0d34-0410-b5e6-96231b3b80d8
The loop stream detector (LSD) on modern Intel cores, which optimizes the
execution of small loops, has limits on the number of taken branches in
addition to uop-count limits (modern AMD cores have similar limits).
Unfortunately, at the IR level, estimating the number of branches that will be
taken is difficult. For one thing, it strongly depends on later passes (block
placement, etc.). The original implementation took a conservative approach and
limited the maximal BB DFS depth of the loop. However, fairly-extensive
benchmarking by several of us has revealed that this is the wrong approach. In
fact, there are zero known cases where the branch limit prevents a detrimental
unrolling (but plenty of cases where it does prevent beneficial unrolling).
While we could improve the current branch counting logic by incorporating
branch probabilities, this further complication seems unjustified without a
motivating regression. Instead, unless and until a regression appears, the
branch counting will be removed.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208255 91177308-0d34-0410-b5e6-96231b3b80d8
The rationale for this artificial dependency seems to have been lost to the
ravages of time, it is covered by no regression tests, and has no impact on
test-suite performance numbers on either x86 or PPC.
For the test suite, on both x86 and PPC, I ran the test suite 10 times (both as
a baseline and with this change), and found no statistically-significant
changes. For PPC, I used a P7 box. For x86, I used an Intel Xeon E5430. Both
with -O3 -mcpu=native.
This was discussed on-list back in January, but I've not had a chance to run
the performance tests until today.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206795 91177308-0d34-0410-b5e6-96231b3b80d8
BasicTTI::getMemoryOpCost must explicitly check for non-simple types; setting
AllowUnknown=true with TLI->getSimpleValueType is not sufficient because, for
example, non-power-of-two vector types return non-simple EVTs (not MVT::Other).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206150 91177308-0d34-0410-b5e6-96231b3b80d8
Implements the various TTI functions to enable constant hoisting on PPC. The
only significant test-suite change is this:
MultiSource/Benchmarks/VersaBench/bmm/bmm - 20% speedup
(which essentially reverses the slowdown from r206120).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206141 91177308-0d34-0410-b5e6-96231b3b80d8
We had been using the known-zero values of the operand of the or to construct
the mask for an rlwimi; this is not quite correct, but fine when the mask is
constant. When the mask is constant, then the known zeros of the operand must
be a superset of the zeros in the mask. However, when the mask is not a
constant, then there might be bits in the operand that are not known to be zero
that, at runtime, might be zero in the mask. Therefore, we check that any bits
not known to be zero *are* known to be one in the mask. Otherwise, we can't
fold the mask with the or and shift.
This was revealed as a miscompile of
MultiSource/Benchmarks/BitBench/drop3/drop3 when I started experimenting with
constant hoisting.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206136 91177308-0d34-0410-b5e6-96231b3b80d8
We had disabled use of TBAA during CodeGen (even when otherwise using AA)
because the ptrtoint/inttoptr used by CGP for address sinking caused BasicAA to
miss basic type punning that it should catch (and, thus, we'd fail to override
TBAA when we should).
However, when AA is in use during CodeGen, CGP now uses normal GEPs and
bitcasts, instead of ptrtoint/inttoptr, when doing address sinking. As a
result, BasicAA should be able to make us do the right thing in the face of
type-punning, and it seems safe to enable use of TBAA again. self-hosting seems
fine on PPC64/Linux on the P7, with TBAA enabled and -misched=shuffle.
Note: We still don't update TBAA when merging stack slots, although because
BasicAA should now catch all such cases, this is no longer a blocking issue.
Nevertheless, I plan to commit code to deal with this properly in the near
future.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206093 91177308-0d34-0410-b5e6-96231b3b80d8
The current memory-instruction optimization logic in CGP, which sinks parts of
the address computation that can be adsorbed by the addressing mode, does this
by explicitly converting the relevant part of the address computation into
IR-level integer operations (making use of ptrtoint and inttoptr). For most
targets this is currently not a problem, but for targets wishing to make use of
IR-level aliasing analysis during CodeGen, the use of ptrtoint/inttoptr is a
problem for two reasons:
1. BasicAA becomes less powerful in the face of the ptrtoint/inttoptr
2. In cases where type-punning was used, and BasicAA was used
to override TBAA, BasicAA may no longer do so. (this had forced us to disable
all use of TBAA in CodeGen; something which we can now enable again)
This (use of GEPs instead of ptrtoint/inttoptr) is not currently enabled by
default (except for those targets that use AA during CodeGen), and so aside
from some PowerPC subtargets and SystemZ, there should be no change in
behavior. We may be able to switch completely away from the ptrtoint/inttoptr
sinking on all targets, but further testing is required.
I've doubled-up on a number of existing tests that are sensitive to the
address sinking behavior (including some store-merging tests that are
sensitive to the order of the resulting ADD operations at the SDAG level).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206092 91177308-0d34-0410-b5e6-96231b3b80d8
PPC::isVSLDOIShuffleMask should return -1, not false, when the shuffle
predicate should be false.
Noticed by inspection; no test case (yet).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205787 91177308-0d34-0410-b5e6-96231b3b80d8
This provides more realistic costs for the insert/extractelement instructions
(which are load/store pairs), accounts for the cheap unaligned Altivec load
sequence, and for unaligned VSX load/stores.
Bad news:
MultiSource/Applications/sgefa/sgefa - 35% slowdown (this will require more investigation)
SingleSource/Benchmarks/McGill/queens - 20% slowdown (we no longer vectorize this, but it was a constant store that was scalarized)
MultiSource/Benchmarks/FreeBench/pcompress2/pcompress2 - 2% slowdown
Good news:
SingleSource/Benchmarks/Shootout/ary3 - 54% speedup
SingleSource/Benchmarks/Shootout-C++/ary - 40% speedup
MultiSource/Benchmarks/Ptrdist/ks/ks - 35% speedup
MultiSource/Benchmarks/FreeBench/neural/neural - 30% speedup
MultiSource/Benchmarks/TSVC/Symbolics-flt/Symbolics-flt - 20% speedup
Unfortunately, estimating the costs of the stack-based scalarization sequences
is hard, and adjusting these costs is like a game of whac-a-mole :( I'll
revisit this again after we have better codegen for vector extloads and
truncstores and unaligned load/stores.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205658 91177308-0d34-0410-b5e6-96231b3b80d8
gcc inline asm supports specifying "cc" as a clobber of all condition
registers. Add just enough modeling of the full register to make this work.
Fixed PR19326.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205630 91177308-0d34-0410-b5e6-96231b3b80d8
When a vector type legalizes to a larger vector type, and the target does not
support the associated extending load (or truncating store), then legalization
will scalarize the load (or store) resulting in an associated scalarization
cost. BasicTTI::getMemoryOpCost needs to account for this.
Between this, and r205487, PowerPC on the P7 with VSX enabled shows:
MultiSource/Benchmarks/PAQ8p/paq8p: 43% speedup
SingleSource/Benchmarks/BenchmarkGame/puzzle: 51% speedup
SingleSource/UnitTests/Vectorizer/gcc-loops 28% speedup
(some of these are new; some of these, such as PAQ8p, just reverse regressions
that VSX support would trigger)
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205495 91177308-0d34-0410-b5e6-96231b3b80d8
For an cast (extension, etc.), the currently logic predicts a low cost if the
associated operation (keyed on the destination type) is legal (or promoted).
This is not true when the number of values required to legalize the type is
changing. For example, <8 x i16> being sign extended by <8 x i32> is not
generically cheap on PPC with VSX, even though sign extension to v4i32 is
legal, because two output v4i32 values are required compared to the single
v8i16 input value, and without custom logic in the target, this conversion will
scalarize.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205487 91177308-0d34-0410-b5e6-96231b3b80d8
PPCTTI::getMemoryOpCost will now make use of BasicTTI::getMemoryOpCost to
calculate the base cost of the memory access, and then adjust on top of that.
There is no functionality change from this modification, but it will become
important so that PPCTTI can take advantage of scalarization information for which
BasicTTI::getMemoryOpCost will account in the near future.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205476 91177308-0d34-0410-b5e6-96231b3b80d8
For the purpose of calculating the cost of the loop at various vectorization
factors, we need to count dependencies of consecutive pointers as uniforms
(which means that the VF = 1 cost is used for all overall VF values).
For example, the TSVC benchmark function s173 has:
...
%3 = add nsw i64 %indvars.iv, 16000
%arrayidx8 = getelementptr inbounds %struct.GlobalData* @global_data, i64 0, i32 0, i64 %3
...
and we must realize that the add will be a scalar in order to correctly deduce
it to be profitable to vectorize this on PowerPC with VSX enabled. In fact, all
dependencies of a consecutive pointer must be a scalar (uniform), and so we
simply need to add all consecutive pointers to the worklist that currently
detects collects uniforms.
Fixes PR19296.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205387 91177308-0d34-0410-b5e6-96231b3b80d8
This provides an initial implementation of getUnrollingPreferences for x86.
getUnrollingPreferences is used by the generic (concatenation) unroller, which
is distinct from the unrolling done by the loop vectorizer. Many modern x86
cores have some kind of uop cache and loop-stream detector (LSD) used to
efficiently dispatch small loops, and taking full advantage of this requires
unrolling small loops (small here means 10s of uops).
These caches also have limits on the number of taken branches in the loop, and
so we also cap the loop unrolling factor based on the maximum "depth" of the
loop. This is currently calculated with a partial DFS traversal (partial
because it will stop early if the path length grows too much). This is still an
approximation, and one that is both conservative (because it does not account
for branches eliminated via block placement) and optimistic (because it is only
recording the maximum depth over minimum paths). Nevertheless, because the
loops that fit in these uop caches are so small, it is not clear how much the
details matter.
The original set of patches posted for review produced the following test-suite
performance results (from the TSVC benchmark) at that time:
ControlLoops-dbl - 13% speedup
ControlLoops-flt - 15% speedup
Reductions-dbl - 7.5% speedup
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205348 91177308-0d34-0410-b5e6-96231b3b80d8
In preparation for an upcoming commit implementing unrolling preferences for
x86, this adds additional fields to the UnrollingPreferences structure:
- PartialThreshold and PartialOptSizeThreshold - Like Threshold and
OptSizeThreshold, but used when not fully unrolling. These are necessary
because we need different thresholds for full unrolling from those used when
partially unrolling (the full unrolling thresholds are generally going to be
larger).
- MaxCount - A cap on the unrolling factor when partially unrolling. This can
be used by a target to prevent the unrolled loop from exceeding some
resource limit independent of the loop size (such as number of branches).
There should be no functionality change for any in-tree targets.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205347 91177308-0d34-0410-b5e6-96231b3b80d8
The implementation of getUserCost had duplicated (and hard-coded) the default
logic in getGEPCost. Instead, it is better to use getGEPCost directly, which
limits the default logic to the implementation of one function, and allows
targets to override the behavior.
No functionality change intended.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205346 91177308-0d34-0410-b5e6-96231b3b80d8
The generic (concatenation) loop unroller is currently placed early in the
standard optimization pipeline. This is a good place to perform full unrolling,
but not the right place to perform partial/runtime unrolling. However, most
targets don't enable partial/runtime unrolling, so this never mattered.
However, even some x86 cores benefit from partial/runtime unrolling of very
small loops, and follow-up commits will enable this. First, we need to move
partial/runtime unrolling late in the optimization pipeline (importantly, this
is after SLP and loop vectorization, as vectorization can drastically change
the size of a loop), while keeping the full unrolling where it is now. This
change does just that.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205264 91177308-0d34-0410-b5e6-96231b3b80d8
This adds the ability to expand large (meaning with more than two unique
defined values) BUILD_VECTOR nodes in terms of SCALAR_TO_VECTOR and (legal)
vector shuffles. There is now no limit of the size we are capable of expanding
this way, although we don't currently do this for vectors with many unique
values because of the default implementation of TLI's
shouldExpandBuildVectorWithShuffles function.
There is currently no functional change to any existing targets because the new
capabilities are not used unless some target overrides the TLI
shouldExpandBuildVectorWithShuffles function. As a result, I've not included a
test case for the new functionality in this commit, but regression tests will
(at least) be added soon when I commit support for the PPC QPX vector
instruction set.
The benefit of committing this now is that it makes the
shouldExpandBuildVectorWithShuffles callback, which had to be added for other
reasons regardless, fully functional. I suspect that other targets will
also benefit from tuning the heuristic.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205243 91177308-0d34-0410-b5e6-96231b3b80d8
If we have two unique values for a v2i64 build vector, this will always result
in two vector loads if we expand using shuffles. Only one is necessary.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205231 91177308-0d34-0410-b5e6-96231b3b80d8
There are two general methods for expanding a BUILD_VECTOR node:
1. Use SCALAR_TO_VECTOR on the defined scalar values and then shuffle
them together.
2. Build the vector on the stack and then load it.
Currently, we use a fixed heuristic: If there are only one or two unique
defined values, then we attempt an expansion in terms of SCALAR_TO_VECTOR and
vector shuffles (provided that the required shuffle mask is legal). Otherwise,
always expand via the stack. Even when SCALAR_TO_VECTOR is not legal, this
can still be a good idea depending on what tricks the target can play when
lowering the resulting shuffle. If the target can't do anything special,
however, and if SCALAR_TO_VECTOR is expanded via the stack, this heuristic
leads to sub-optimal code (two stack loads instead of one).
Because only the target knows whether the SCALAR_TO_VECTORs and shuffles for a
build vector of a particular type are likely to be optimial, this adds a new
TLI function: shouldExpandBuildVectorWithShuffles which takes the vector type
and the count of unique defined values. If this function returns true, then
method (1) will be used, subject to the constraint that all of the necessary
shuffles are legal (as determined by isShuffleMaskLegal). If this function
returns false, then method (2) is always used.
This commit does not enhance the current code to support expanding a
build_vector with more than two unique values using shuffles, but I'll commit
an implementation of the more-general case shortly.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205230 91177308-0d34-0410-b5e6-96231b3b80d8
When the loop vectorizer vectorizes code that uses the loop induction variable,
we often end up with IR like this:
%b1 = insertelement <2 x i32> undef, i32 %v, i32 0
%b2 = shufflevector <2 x i32> %b1, <2 x i32> undef, <2 x i32> zeroinitializer
%i = add <2 x i32> %b2, <i32 2, i32 3>
If the add in this example is not legal (as is the case on PPC with VSX), it
will be scalarized, and we'll end up with a number of extract_vector_elt nodes
with the vector shuffle as the input operand, and that vector shuffle is fed by
one or more build_vector nodes. By the time that vector operations are
expanded, visitEXTRACT_VECTOR_ELT will not create new extract_vector_elt by
looking through the vector shuffle (to make sure that no illegal operations are
created), and so the extract_vector_elt -> vector shuffle -> build_vector is
never simplified to an operand of the build vector.
By looking at build_vectors through a shuffle we fix this particular situation,
preventing a vector from being built, only to be deconstructed again (for the
scalarized add) -- an expensive proposition when this all needs to be done via
the stack. We probably want a more comprehensive fix here where we look back
recursively through any shuffles to any build_vectors or scalar_to_vectors,
etc. but that can come later.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205179 91177308-0d34-0410-b5e6-96231b3b80d8
When expanding EXTRACT_VECTOR_ELT and EXTRACT_SUBVECTOR using
SelectionDAGLegalize::ExpandExtractFromVectorThroughStack, we store the entire
vector and then load the piece we want. This is fine in isolation, but
generating a new store (and corresponding stack slot) for each extraction ends
up producing code of poor quality. When we scalarize a vector operation (using
SelectionDAG::UnrollVectorOp for example) we generate one EXTRACT_VECTOR_ELT
for each element in the vector. This used to generate one stored copy of the
vector for each element in the vector. Now we search the uses of the vector for
a suitable store before generating a new one, which results in much more
efficient scalarization code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205153 91177308-0d34-0410-b5e6-96231b3b80d8
sitofp from v2i32 to v2f64 ends up generating a SIGN_EXTEND_INREG v2i64 node
(and similarly for v2i16 and v2i8). Even though there are no sign-extension (or
algebraic shifts) for v2i64 types, we can handle v2i32 sign extensions by
converting two and from v2i64. The small trick necessary here is to shift the
i32 elements into the right lanes before the i32 -> f64 step. This is because
of the big Endian nature of the system, we need the i32 portion in the high
word of the i64 elements.
For v2i16 and v2i8 we can do the same, but we first use the default Altivec
shift-based expansion from v2i16 or v2i8 to v2i32 (by casting to v4i32) and
then apply the above procedure.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205146 91177308-0d34-0410-b5e6-96231b3b80d8
v2i64 is a legal type under VSX, however we don't have native vector
comparisons. We can handle eq/ne by casting it to an Altivec type, but
everything else must be expanded.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205106 91177308-0d34-0410-b5e6-96231b3b80d8
The vector divide and sqrt instructions have high latencies, and the scalar
comparisons are like all of the others. On the P7, permutations take an extra
cycle over purely-simple vector ops.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205096 91177308-0d34-0410-b5e6-96231b3b80d8
We had stored both f64 values and v2f64, etc. values in the VSX registers. This
worked, but was suboptimal because we would always spill 16-byte values even
through we almost always had scalar 8-byte values. This resulted in an
increase in stack-size use, extra memory bandwidth, etc. To fix this, I've
added 64-bit subregisters of the Altivec registers, and combined those with the
existing scalar floating-point registers to form a class of VSX scalar
floating-point registers. The ABI code has also been enhanced to use this
register class and some other necessary improvements have been made.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205075 91177308-0d34-0410-b5e6-96231b3b80d8
Not only did I invert the indices when I wrote the code, but I also did the
same thing when I wrote the regression test. Oops.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205046 91177308-0d34-0410-b5e6-96231b3b80d8
v2[fi]64 values need to be explicitly passed in VSX registers. This is because
the code in TRI that finds the minimal register class given a register and a
value type will assert if given an Altivec register and a non-Altivec type.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205041 91177308-0d34-0410-b5e6-96231b3b80d8
As explained in r204976, because of how the allocation of VSX registers
interacts with the call-lowering code, we sometimes end up generating self VSX
copies. Specifically, things like this:
%VSL2<def> = COPY %F2, %VSL2<imp-use,kill>
(where %F2 is really a sub-register of %VSL2, and so this copy is a nop)
This adds a small cleanup pass to remove these prior to post-RA scheduling.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204980 91177308-0d34-0410-b5e6-96231b3b80d8
Because of how the allocation of VSX registers interacts with the call-lowering
code, we sometimes end up generating self VSX copies. Specifically, things like
this:
%VSL2<def> = COPY %F2, %VSL2<imp-use,kill>
(where %F2 is really a sub-register of %VSL2, and so this copy is a nop)
The problem is that ExpandPostRAPseudos always assumes that *some* instruction
has been inserted, and adds implicit defs to it. This is a problem if no copy
was inserted because it can cause subtle problems during post-RA scheduling.
These self copies will have to be removed some other way.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204976 91177308-0d34-0410-b5e6-96231b3b80d8
First, v2f64 vector extract had not been declared legal (and so the existing
patterns were not being used). Second, the patterns for that, and for
scalar_to_vector, should really be a regclass copy, not a subregister
operation, because the VSX registers directly hold both the vector and scalar data.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204971 91177308-0d34-0410-b5e6-96231b3b80d8
These operations need to be expanded during legalization so that isel does not
crash. In theory, we might be able to custom lower some of these. That,
however, would need to be follow-up work.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204963 91177308-0d34-0410-b5e6-96231b3b80d8
I've not yet updated PPCTTI because I'm not sure what the actual relative cost
is compared to the aligned uses.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204848 91177308-0d34-0410-b5e6-96231b3b80d8
These patterns are dead (because v4f32 stores are currently promoted to v4i32
and stored using Altivec instructions), and also are likely not correct
(because they'd store the vector elements in the opposite order from that
assumed by the rest of the Altivec code).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204839 91177308-0d34-0410-b5e6-96231b3b80d8
These instructions have access to the complete VSX register file. In addition,
they "swap" the order of the elements so that element 0 (the scalar part) comes
first in memory and element 1 follows at a higher address.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204838 91177308-0d34-0410-b5e6-96231b3b80d8
v2i64 needs to be a legal VSX type because it is the SetCC result type from
v2f64 comparisons. We need to expand all non-arithmetic v2i64 operations.
This fixes the lowering for v2f64 VSELECT.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204828 91177308-0d34-0410-b5e6-96231b3b80d8
With VSX there is a real vector select instruction, and so we should use it.
Note that VSELECT will still scalarize for v2f64 because the corresponding
SetCC result type (v2i64) is not currently a legal type.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204801 91177308-0d34-0410-b5e6-96231b3b80d8
These instructions are essentially the same as their Altivec counterparts, but
have access to the larger VSX register file.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204782 91177308-0d34-0410-b5e6-96231b3b80d8
The VSX instruction set has two types of FMA instructions: A-type (where the
addend is taken from the output register) and M-type (where one of the product
operands is taken from the output register). This adds a small pass that runs
just after MI scheduling (and, thus, just before register allocation) that
mutates A-type instructions (that are created during isel) into M-type
instructions when:
1. This will eliminate an otherwise-necessary copy of the addend
2. One of the product operands is killed by the instruction
The "right" moment to make this decision is in between scheduling and register
allocation, because only there do we know whether or not one of the product
operands is killed by any particular instruction. Unfortunately, this also
makes the implementation somewhat complicated, because the MIs are not in SSA
form and we need to preserve the LiveIntervals analysis.
As a simple example, if we have:
%vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9
%vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16,
%RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16
...
%vreg9<def,tied1> = XSMADDADP %vreg9<tied0>, %vreg17, %vreg19,
%RM<imp-use>; VSLRC:%vreg9,%vreg17,%vreg19
...
We can eliminate the copy by changing from the A-type to the
M-type instruction. This means:
%vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16,
%RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16
is replaced by:
%vreg16<def,tied1> = XSMADDMDP %vreg16<tied0>, %vreg18, %vreg9,
%RM<imp-use>; VSLRC:%vreg16,%vreg18,%vreg9
and we remove: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204768 91177308-0d34-0410-b5e6-96231b3b80d8
Although the first two operands are the ones that can be swapped, the tied
input operand is listed before them, so we need to adjust for that.
I have a test case for this, but it goes along with an upcoming commit (so it
will come soon).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204748 91177308-0d34-0410-b5e6-96231b3b80d8
TableGen will create a lookup table for the A-type FMA instructions providing
their corresponding M-form opcodes. This will be used by upcoming commits.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204746 91177308-0d34-0410-b5e6-96231b3b80d8
I'm under the impression that we used to infer the isCommutable flag from the
instruction-associated pattern. Regardless, we don't seem to do this (at least
by default) any more. I've gone through all of our instruction definitions, and
marked as commutative all of those that should be trivial to commute (by
exchanging the first two operands). There has been special code for the RL*
instructions, and that's not changed.
Before this change, we had the following commutative instructions:
RLDIMI
RLDIMIo
RLWIMI
RLWIMI8
RLWIMI8o
RLWIMIo
XSADDDP
XSMULDP
XVADDDP
XVADDSP
XVMULDP
XVMULSP
After:
ADD4
ADD4o
ADD8
ADD8o
ADDC
ADDC8
ADDC8o
ADDCo
ADDE
ADDE8
ADDE8o
ADDEo
AND
AND8
AND8o
ANDo
CRAND
CREQV
CRNAND
CRNOR
CROR
CRXOR
EQV
EQV8
EQV8o
EQVo
FADD
FADDS
FADDSo
FADDo
FMADD
FMADDS
FMADDSo
FMADDo
FMSUB
FMSUBS
FMSUBSo
FMSUBo
FMUL
FMULS
FMULSo
FMULo
FNMADD
FNMADDS
FNMADDSo
FNMADDo
FNMSUB
FNMSUBS
FNMSUBSo
FNMSUBo
MULHD
MULHDU
MULHDUo
MULHDo
MULHW
MULHWU
MULHWUo
MULHWo
MULLD
MULLDo
MULLW
MULLWo
NAND
NAND8
NAND8o
NANDo
NOR
NOR8
NOR8o
NORo
OR
OR8
OR8o
ORo
RLDIMI
RLDIMIo
RLWIMI
RLWIMI8
RLWIMI8o
RLWIMIo
VADDCUW
VADDFP
VADDSBS
VADDSHS
VADDSWS
VADDUBM
VADDUBS
VADDUHM
VADDUHS
VADDUWM
VADDUWS
VAND
VAVGSB
VAVGSH
VAVGSW
VAVGUB
VAVGUH
VAVGUW
VMADDFP
VMAXFP
VMAXSB
VMAXSH
VMAXSW
VMAXUB
VMAXUH
VMAXUW
VMHADDSHS
VMHRADDSHS
VMINFP
VMINSB
VMINSH
VMINSW
VMINUB
VMINUH
VMINUW
VMLADDUHM
VMULESB
VMULESH
VMULEUB
VMULEUH
VMULOSB
VMULOSH
VMULOUB
VMULOUH
VNMSUBFP
VOR
VXOR
XOR
XOR8
XOR8o
XORo
XSADDDP
XSMADDADP
XSMAXDP
XSMINDP
XSMSUBADP
XSMULDP
XSNMADDADP
XSNMSUBADP
XVADDDP
XVADDSP
XVMADDADP
XVMADDASP
XVMAXDP
XVMAXSP
XVMINDP
XVMINSP
XVMSUBADP
XVMSUBASP
XVMULDP
XVMULSP
XVNMADDADP
XVNMADDASP
XVNMSUBADP
XVNMSUBASP
XXLAND
XXLNOR
XXLOR
XXLXOR
This is a by-inspection change, and I'm not sure how to write a reliable test
case. I would like advice on this, however.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204609 91177308-0d34-0410-b5e6-96231b3b80d8
I've done some experimentation with this, and it looks like using the
lower-latency (but lower throughput) copy instruction is essentially always the
right thing to do.
My assumption is that, in order to be relatively sure that the higher-latency
copy will increase throughput, we'd want to have it unlikely to be in-flight
with its use. On the P7, the global completion table (GCT) can hold a maximum
of 120 instructions, shared among all active threads (up to 4), giving 30
instructions per thread. So specifically, I'd require at least that many
instructions between the copy and the use before the high-latency variant is
used.
Trying this, however, over the entire test suite resulted in zero cases where
the high-latency form would be preferable. This may be a consequence of the
fact that the scheduler views copies as free, and so they tend to end up close
to their uses. For this experiment I created a function:
unsigned chooseVSXCopy(MachineBasicBlock &MBB,
MachineBasicBlock::iterator I,
unsigned DestReg, unsigned SrcReg,
unsigned StartDist = 1,
unsigned Depth = 3) const;
with an implementation like:
if (!Depth)
return PPC::XXLOR;
const unsigned MaxDist = 30;
unsigned Dist = StartDist;
for (auto J = I, JE = MBB.end(); J != JE && Dist <= MaxDist; ++J) {
if (J->isTransient() && !J->isCopy())
continue;
if (J->isCall() || J->isReturn() || J->readsRegister(DestReg, TRI))
return PPC::XXLOR;
++Dist;
}
// We've exceeded the required distance for the high-latency form, use it.
if (Dist > MaxDist)
return PPC::XVCPSGNDP;
// If this is only an exit block, use the low-latency form.
if (MBB.succ_empty())
return PPC::XXLOR;
// We've reached the end of the block, check the successor blocks (up to some
// depth), and use the high-latency form if that is okay with all successors.
for (auto J = MBB.succ_begin(), JE = MBB.succ_end(); J != JE; ++J) {
if (chooseVSXCopy(**J, (*J)->begin(), DestReg, SrcReg,
Dist, --Depth) == PPC::XXLOR)
return PPC::XXLOR;
}
// All of our successor blocks seem okay with the high-latency variant, so
// we'll use it.
return PPC::XVCPSGNDP;
and then changed the copy opcode selection from:
Opc = PPC::XXLOR;
to:
Opc = chooseVSXCopy(MBB, std::next(I), DestReg, SrcReg);
In conclusion, I'm removing the FIXME from the comment, because I believe that
there is, at least absent other examples, nothing to fix.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204591 91177308-0d34-0410-b5e6-96231b3b80d8
When VSX is available, these instructions should be used in preference to the
older variants that only have access to the scalar floating-point registers.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204559 91177308-0d34-0410-b5e6-96231b3b80d8
v2f64 values, like other 128-bit values, are returned under VSX in register
vs34 (Altivec register v2).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204543 91177308-0d34-0410-b5e6-96231b3b80d8
When an instruction's operand list does not have a sufficient number of
operands to match with all of the variables that contribute to its
encoding, instead of asserting inside a call to getSubOperandNumber, produce an
informative error.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204542 91177308-0d34-0410-b5e6-96231b3b80d8
VSX is an ISA extension supported on the POWER7 and later cores that enhances
floating-point vector and scalar capabilities. Among other things, this adds
<2 x double> support and generally helps to reduce register pressure.
The interesting part of this ISA feature is the register configuration: there
are 64 new 128-bit vector registers, the 32 of which are super-registers of the
existing 32 scalar floating-point registers, and the second 32 of which overlap
with the 32 Altivec vector registers. This makes things like vector insertion
and extraction tricky: this can be free but only if we force a restriction to
the right register subclass when needed. A new "minipass" PPCVSXCopy takes care
of this (although it could do a more-optimal job of it; see the comment about
unnecessary copies below).
Please note that, currently, VSX is not enabled by default when targeting
anything because it is not yet ready for that. The assembler and disassembler
are fully implemented and tested. However:
- CodeGen support causes miscompiles; test-suite runtime failures:
MultiSource/Benchmarks/FreeBench/distray/distray
MultiSource/Benchmarks/McCat/08-main/main
MultiSource/Benchmarks/Olden/voronoi/voronoi
MultiSource/Benchmarks/mafft/pairlocalalign
MultiSource/Benchmarks/tramp3d-v4/tramp3d-v4
SingleSource/Benchmarks/CoyoteBench/almabench
SingleSource/Benchmarks/Misc/matmul_f64_4x4
- The lowering currently falls back to using Altivec instructions far more
than it should. Worse, there are some things that are scalarized through the
stack that shouldn't be.
- A lot of unnecessary copies make it past the optimizers, and this needs to
be fixed.
- Many more regression tests are needed.
Normally, I'd fix these things prior to committing, but there are some
students and other contributors who would like to work this, and so it makes
sense to move this development process upstream where it can be subject to the
regular code-review procedures.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203768 91177308-0d34-0410-b5e6-96231b3b80d8
There are currently two schemes for mapping instruction operands to
instruction-format variables for generating the instruction encoders and
decoders for the assembler and disassembler respectively: a) to map by name and
b) to map by position.
In the long run, we'd like to remove the position-based scheme and use only
name-based mapping. Unfortunately, the name-based scheme currently cannot deal
with complex operands (those with suboperands), and so we currently must use
the position-based scheme for those. On the other hand, the position-based
scheme cannot deal with (register) variables that are split into multiple
ranges. An upcoming commit to the PowerPC backend (adding VSX support) will
require this capability. While we could teach the position-based scheme to
handle that, since we'd like to move away from the position-based mapping
generally, it seems silly to teach it new tricks now. What makes more sense is
to allow for partial transitioning: use the name-based mapping when possible,
and only use the position-based scheme when necessary.
Now the problem is that mixing the two sensibly was not possible: the
position-based mapping would map based on position, but would not skip those
variables that were mapped by name. Instead, the two sets of assignments would
overlap. However, I cannot currently change the current behavior, because there
are some backends that rely on it [I think mistakenly, but I'll send a message
to llvmdev about that]. So I've added a new TableGen bit variable:
noNamedPositionallyEncodedOperands, that can be used to cause the
position-based mapping to skip variables mapped by name.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203767 91177308-0d34-0410-b5e6-96231b3b80d8
Hal Finkel