Patch by Bill Seurer; some comment formatting changes by me.
There are a few PowerPC test cases for FastISel support that currently
fail with VSX support enabled. The temporary workaround under
discussion in http://reviews.llvm.org/D5362 helps, but the tests still
fail because they specify -fast-isel-abort, and the VSX workaround
punts back to SelectionDAG. We have plans to fix FastISel permanently
for VSX, but until that's in place these tests are preventing us from
enabling VSX by default. Therefore we are adding -mattr=-vsx to these
tests until the full support is ready.
llvm-svn: 220172
The VSX testing variant in test/CodeGen/PowerPC/fma.ll had to be
disabled because of unexpected behavior on many of the builders. I
tracked this down to a situation that occurs when the VSX attribute is
enabled for a target that disables the MI early scheduling pass. This
patch adds -mcpu=pwr7 to make this predictable. The other issue will
be addressed separately.
llvm-svn: 220171
With VSX enabled, LLVM crashes when compiling
test/CodeGen/PowerPC/fma.ll. I traced this to the liveness test
that's revised in this patch. The interval test is designed to only
work for virtual registers, but in this case the AddendSrcReg is
physical. Since there is already a walk of the MIs between the
AddendMI and the FMA, I added a check for def/kill of the AddendSrcReg
in that loop. At Hal Finkel's request, I converted the liveness test
to an assert restricted to virtual registers.
I've changed the fma.ll test to have VSX and non-VSX variants so we
can test both kinds of multiply-adds.
llvm-svn: 220090
Currently the VSX support enables use of lxvd2x and stxvd2x for 2x64
types, but does not yet use lxvw4x and stxvw4x for 4x32 types. This
patch adds that support.
As with lxvd2x/stxvd2x, this involves straightforward overriding of
the patterns normally recognized for lvx/stvx, with preference given
to the VSX patterns when VSX is enabled.
In addition, the logic for permitting misaligned memory accesses is
modified so that v4r32 and v4i32 are treated the same as v2f64 and
v2i64 when VSX is enabled. Finally, the DAG generation for unaligned
loads is changed to just use a normal LOAD (which will become lxvw4x)
on P8 and later hardware, where unaligned loads are preferred over
lvsl/lvx/lvx/vperm.
A number of tests now generate the VSX loads/stores instead of
lvx/stvx, so this patch adds VSX variants to those tests. I've also
added <4 x float> tests to the vsx.ll test case, and created a
vsx-p8.ll test case to be used for testing code generation for the
P8Vector feature. For now, that simply tests the unaligned load/store
behavior.
This has been tested along with a temporary patch to enable the VSX
and P8Vector features, with no new regressions encountered with or
without the temporary patch applied.
llvm-svn: 220047
Patch by Bill Seurer; committed on his behalf.
These test cases generate slightly different code sequences when VSX
is activated and thus fail. The update turns off VSX explicitly for
the existing checks and then adds a second set of checks for most of
them that test the VSX instruction output.
llvm-svn: 220019
This patch changes the fast-math implementation for calculating sqrt(x) from:
y = 1 / (1 / sqrt(x))
to:
y = x * (1 / sqrt(x))
This has 2 benefits: less code / faster code and one less estimate instruction
that may lose precision.
The only target that will be affected (until http://reviews.llvm.org/D5658 is approved)
is PPC. The difference in codegen for PPC is 2 less flops for a single-precision sqrtf
or vector sqrtf and 4 less flops for a double-precision sqrt.
We also eliminate a constant load and extra register usage.
Differential Revision: http://reviews.llvm.org/D5682
llvm-svn: 219445
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.
llvm-svn: 219441
This reverts commit r218918, effectively reapplying r218914 after fixing
an Ocaml bindings test and an Asan crash. The root cause of the latter
was a tightened-up check in `DILexicalBlock::Verify()`, so I'll file a
PR to investigate who requires the loose check (and why).
Original commit message follows.
--
This patch addresses the first stage of PR17891 by folding constant
arguments together into a single MDString. Integers are stringified and
a `\0` character is used as a separator.
Part of PR17891.
Note: I've attached my testcases upgrade scripts to the PR. If I've
just broken your out-of-tree testcases, they might help.
llvm-svn: 219010
The test Atomics-32.ll was both redundant (all operations are also checked by
atomics.ll at least) and not actually checking correctness (it was not using
FileCheck, just verifying that the compiler does not crash).
llvm-svn: 218997
Summary:
hwsync is only required for seq_cst fences, acquire and release one can use
the cheaper lwsync.
Test Plan: Added some cases to atomics.ll + make check-all
Reviewers: jfb, wschmidt
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D5317
llvm-svn: 218995
Older Book-E cores, such as the PPC 440, support only msync (which has the same
encoding as sync 0), but not any of the other sync forms. Newer Book-E cores,
however, do support sync, and for performance reasons we should allow the use
of the more-general form.
This refactors msync use into its own feature group so that it applies by
default only to older Book-E cores (of the relevant cores, we only have
definitions for the PPC440/450 currently).
llvm-svn: 218923
Summary:
Atomic loads and store of up to the native size (32 bits, or 64 for PPC64)
can be lowered to a simple load or store instruction (as the synchronization
is already handled by AtomicExpand, and the atomicity is guaranteed thanks to
the alignment requirements of atomic accesses). This is exactly what this patch
does. Previously, these were implemented by complex
load-linked/store-conditional loops.. an obvious performance problem.
For example, this patch turns
```
define void @store_i8_unordered(i8* %mem) {
store atomic i8 42, i8* %mem unordered, align 1
ret void
}
```
from
```
_store_i8_unordered: ; @store_i8_unordered
; BB#0:
rlwinm r2, r3, 3, 27, 28
li r4, 42
xori r5, r2, 24
rlwinm r2, r3, 0, 0, 29
li r3, 255
slw r4, r4, r5
slw r3, r3, r5
and r4, r4, r3
LBB4_1: ; =>This Inner Loop Header: Depth=1
lwarx r5, 0, r2
andc r5, r5, r3
or r5, r4, r5
stwcx. r5, 0, r2
bne cr0, LBB4_1
; BB#2:
blr
```
into
```
_store_i8_unordered: ; @store_i8_unordered
; BB#0:
li r2, 42
stb r2, 0(r3)
blr
```
which looks like a pretty clear win to me.
Test Plan:
fixed the tests + new test for indexed accesses + make check-all
Reviewers: jfb, wschmidt, hfinkel
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D5587
llvm-svn: 218922
This patch addresses the first stage of PR17891 by folding constant
arguments together into a single MDString. Integers are stringified and
a `\0` character is used as a separator.
Part of PR17891.
Note: I've attached my testcases upgrade scripts to the PR. If I've
just broken your out-of-tree testcases, they might help.
llvm-svn: 218914
argument of the llvm.dbg.declare/llvm.dbg.value intrinsics.
Previously, DIVariable was a variable-length field that has an optional
reference to a Metadata array consisting of a variable number of
complex address expressions. In the case of OpPiece expressions this is
wasting a lot of storage in IR, because when an aggregate type is, e.g.,
SROA'd into all of its n individual members, the IR will contain n copies
of the DIVariable, all alike, only differing in the complex address
reference at the end.
By making the complex address into an extra argument of the
dbg.value/dbg.declare intrinsics, all of the pieces can reference the
same variable and the complex address expressions can be uniqued across
the CU, too.
Down the road, this will allow us to move other flags, such as
"indirection" out of the DIVariable, too.
The new intrinsics look like this:
declare void @llvm.dbg.declare(metadata %storage, metadata %var, metadata %expr)
declare void @llvm.dbg.value(metadata %storage, i64 %offset, metadata %var, metadata %expr)
This patch adds a new LLVM-local tag to DIExpressions, so we can detect
and pretty-print DIExpression metadata nodes.
What this patch doesn't do:
This patch does not touch the "Indirect" field in DIVariable; but moving
that into the expression would be a natural next step.
http://reviews.llvm.org/D4919
rdar://problem/17994491
Thanks to dblaikie and dexonsmith for reviewing this patch!
Note: I accidentally committed a bogus older version of this patch previously.
llvm-svn: 218787
argument of the llvm.dbg.declare/llvm.dbg.value intrinsics.
Previously, DIVariable was a variable-length field that has an optional
reference to a Metadata array consisting of a variable number of
complex address expressions. In the case of OpPiece expressions this is
wasting a lot of storage in IR, because when an aggregate type is, e.g.,
SROA'd into all of its n individual members, the IR will contain n copies
of the DIVariable, all alike, only differing in the complex address
reference at the end.
By making the complex address into an extra argument of the
dbg.value/dbg.declare intrinsics, all of the pieces can reference the
same variable and the complex address expressions can be uniqued across
the CU, too.
Down the road, this will allow us to move other flags, such as
"indirection" out of the DIVariable, too.
The new intrinsics look like this:
declare void @llvm.dbg.declare(metadata %storage, metadata %var, metadata %expr)
declare void @llvm.dbg.value(metadata %storage, i64 %offset, metadata %var, metadata %expr)
This patch adds a new LLVM-local tag to DIExpressions, so we can detect
and pretty-print DIExpression metadata nodes.
What this patch doesn't do:
This patch does not touch the "Indirect" field in DIVariable; but moving
that into the expression would be a natural next step.
http://reviews.llvm.org/D4919
rdar://problem/17994491
Thanks to dblaikie and dexonsmith for reviewing this patch!
llvm-svn: 218778
This is purely refactoring. No functional changes intended. PowerPC is the only target
that is currently using this interface.
The ultimate goal is to allow targets other than PowerPC (certainly X86 and Aarch64) to turn this:
z = y / sqrt(x)
into:
z = y * rsqrte(x)
And:
z = y / x
into:
z = y * rcpe(x)
using whatever HW magic they can use. See http://llvm.org/bugs/show_bug.cgi?id=20900 .
There is one hook in TargetLowering to get the target-specific opcode for an estimate instruction
along with the number of refinement steps needed to make the estimate usable.
Differential Revision: http://reviews.llvm.org/D5484
llvm-svn: 218553
Summary:
This patch makes use of AtomicExpandPass in Power for inserting fences around
atomic as part of an effort to remove fence insertion from SelectionDAGBuilder.
As a big bonus, it lets us use sync 1 (lightweight sync, often used by the mnemonic
lwsync) instead of sync 0 (heavyweight sync) in many cases.
I also added a test, as there was no test for the barriers emitted by the Power
backend for atomic loads and stores.
Test Plan: new test + make check-all
Reviewers: jfb
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D5180
llvm-svn: 218331
We manage to generate all of the matching instructions (and a lot more) via
the reciprocal optimization function - even if we completely remove the square
root optimization. With CHECK_NEXT, we assure that we're executing the
expected square root optimization paths and not generating extra insts.
llvm-svn: 218284
The heuristic used by DAGCombine to form FMAs checks that the FMUL has only one
use, but this is overly-conservative on some systems. Specifically, if the FMA
and the FADD have the same latency (and the FMA does not compete for resources
with the FMUL any more than the FADD does), there is no need for the
restriction, and furthermore, forming the FMA leaving the FMUL can still allow
for higher overall throughput and decreased critical-path length.
Here we add a new TLI callback, enableAggressiveFMAFusion, false by default, to
elide the hasOneUse check. This is enabled for PowerPC by default, as most
PowerPC systems will benefit.
Patch by Olivier Sallenave, thanks!
llvm-svn: 218120
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.
llvm-svn: 217993
On MachO, and MachO only, we cannot have a truly empty function since that
breaks the linker logic for atomizing the section.
When we are emitting a frame pointer, the presence of an unreachable will
create a cfi instruction pointing past the last instruction. This is perfectly
fine. The FDE information encodes the pc range it applies to. If some tool
cannot handle this, we should explicitly say which bug we are working around
and only work around it when it is actually relevant (not for ELF for example).
Given the unreachable we could omit the .cfi_def_cfa_register, but then
again, we could also omit the entire function prologue if we wanted to.
llvm-svn: 217801
Inline asm may specify 'U' and 'X' constraints to print a 'u' for an
update-form memory reference, or an 'x' for an indexed-form memory
reference. However, these are really only useful in GCC internal code
generation. In inline asm the operand of the memory constraint is
typically just a register containing the address, so 'U' and 'X' make
no sense.
This patch quietly accepts 'U' and 'X' in inline asm patterns, but
otherwise does nothing. If we ever unexpectedly see a non-register,
we'll assert and sort it out afterwards.
I've added a new test for these constraints; the test case should be
used for other asm-constraints changes down the road.
llvm-svn: 217622
When I recommitted r208640 (in r216898) I added an exclusion for TargetConstant
offsets, as there is no guarantee that a backend can handle them on generic
ADDs (even if it generates them during address-mode matching) -- and,
specifically, applying this transformation directly with TargetConstants caused
a self-hosting failure on PPC64. Ignoring all TargetConstants, however, is less
than ideal. Instead, for non-opaque constants, we can convert them into regular
constants for use with the generated ADD (or SUB).
llvm-svn: 216908
Adds code generation support for dcbtst (data cache prefetch for write) and
icbt (instruction cache prefetch for read - Book E cores only).
We still end up with a 'cannot select' error for the non-supported prefetch
intrinsic forms. This will be fixed in a later commit.
Fixes PR20692.
llvm-svn: 216339
Note: This was originally reverted to track down a buildbot error. This commit
exposed a latent bug that was fixed in r215753. Therefore it is reapplied
without any modifications.
I run it through SPEC2k and SPEC2k6 for AArch64 and it didn't introduce any new
regeressions.
Original commit message:
This changes the order in which FastISel tries to materialize a constant.
Originally it would try to use a simple target-independent approach, which
can lead to the generation of inefficient code.
On X86 this would result in the use of movabsq to materialize any 64bit
integer constant - even for simple and small values such as 0 and 1. Also
some very funny floating-point materialization could be observed too.
On AArch64 it would materialize the constant 0 in a register even the
architecture has an actual "zero" register.
On ARM it would generate unnecessary mov instructions or not use mvn.
This change simply changes the order and always asks the target first if it
likes to materialize the constant. This doesn't fix all the issues
mentioned above, but it enables the targets to implement such
optimizations.
Related to <rdar://problem/17420988>.
llvm-svn: 216006
A byval object, even if allocated at a fixed offset (prescribed by the ABI) is
pointed to by IR values. Most fixed-offset stack objects are not pointed-to by
IR values, so the default is to assume this is not possible. However, we need
to override the default in this case (instruction scheduling can cause
miscompiles otherwise).
Fixes PR20280.
llvm-svn: 215795
This reverts:
r215595 "[FastISel][X86] Add large code model support for materializing floating-point constants."
r215594 "[FastISel][X86] Use XOR to materialize the "0" value."
r215593 "[FastISel][X86] Emit more efficient instructions for integer constant materialization."
r215591 "[FastISel][AArch64] Make use of the zero register when possible."
r215588 "[FastISel] Let the target decide first if it wants to materialize a constant."
r215582 "[FastISel][AArch64] Cleanup constant materialization code. NFCI."
llvm-svn: 215673
This changes the order in which FastISel tries to materialize a constant.
Originally it would try to use a simple target-independent approach, which
can lead to the generation of inefficient code.
On X86 this would result in the use of movabsq to materialize any 64bit
integer constant - even for simple and small values such as 0 and 1. Also
some very funny floating-point materialization could be observed too.
On AArch64 it would materialize the constant 0 in a register even the
architecture has an actual "zero" register.
On ARM it would generate unnecessary mov instructions or not use mvn.
This change simply changes the order and always asks the target first if it
likes to materialize the constant. This doesn't fix all the issues
mentioned above, but it enables the targets to implement such
optimizations.
Related to <rdar://problem/17420988>.
llvm-svn: 215588
This implements PPCTargetLowering::getTgtMemIntrinsic for Altivec load/store
intrinsics. As with the construction of the MachineMemOperands for the
intrinsic calls used for unaligned load/store lowering, the only slight
complication is that we need to represent a larger memory range than the
loaded/stored value-type size (because the address is rounded down to an
aligned address, and we need to conservatively represent the entire possible
range of the actual access). This required adding an extra size field to
TargetLowering::IntrinsicInfo, and this was done in a way that required no
modifications to other targets (the size defaults to the store size of the
provided memory data type).
This fixes test/CodeGen/PowerPC/unal-altivec-wint.ll (so it can be un-XFAILed).
llvm-svn: 215512
Unfortunately, our use of the SDNode class hierarchy for INTRINSIC_W_CHAIN and
INTRINSIC_VOID nodes is somewhat broken right now. These nodes sometimes are
used for memory intrinsics (those with MachineMemOperands), and sometimes not.
When not, the nodes are not created as instances of MemIntrinsicSDNode, but
rather created as some other subclass of SDNode using DAG::getNode. When they
are memory intrinsics, they are created using DAG::getMemIntrinsicNode as
instances of MemIntrinsicSDNode. MemIntrinsicSDNode is a subclass of
MemSDNode, but prior to r214452, we had a non-self-consistent setup whereby
MemIntrinsicSDNode::classof on INTRINSIC_W_CHAIN and INTRINSIC_VOID would
return true but MemSDNode::classof on INTRINSIC_W_CHAIN and INTRINSIC_VOID
would return false. In r214452, MemSDNode::classof was changed to return true
for INTRINSIC_W_CHAIN and INTRINSIC_VOID, which is now self-consistent. The
problem is that neither the pre-r214452 logic and the post-r214452 logic are
really right. The truth is that not all INTRINSIC_W_CHAIN and INTRINSIC_VOID
nodes are instances of MemIntrinsicSDNode (or MemSDNode for that matter), and
the return value from classof needs to reflect that. This was broken before
r214452 (because MemIntrinsicSDNode::classof always returned true), and was
broken afterward (because MemSDNode::classof also always returned true), and
will now be correct.
The minimal solution is to grab one of the SubclassData bits (there is one left
for MemIntrinsicSDNode nodes) and use it to store whether or not a particular
INTRINSIC_W_CHAIN or INTRINSIC_VOID is really an instance of
MemIntrinsicSDNode or not. Doing this allows both MemIntrinsicSDNode::classof
and MemSDNode::classof to return the correct answer for the underlying object
for both the memory-intrinsic and non-memory-intrinsic cases.
This fixes the problem that r214452 created in the SelectionDAGDumper (thanks
to Matt Arsenault for pointing it out).
Because PowerPC does not implement getTgtMemIntrinsic, this change breaks
test/CodeGen/PowerPC/unal-altivec-wint.ll. I've XFAILed it for now, and will
fix it in a follow-up commit.
llvm-svn: 215511
Commits r213915 and r214718 fix recognition of shuffle masks for vmrg*
and vpku*um instructions for a little-endian target, by swapping the
input arguments. The vsldoi instruction requires similar treatment,
and also needs its shift count adjusted for little endian.
Reviewed by Ulrich Weigand.
This is a bug fix candidate for release 3.5 (and hopefully the last of
those for PowerPC).
llvm-svn: 214923
My original LE implementation of the vsldoi instruction, with its
altivec.h interfaces vec_sld and vec_vsldoi, produces incorrect
shufflevector operations in the LLVM IR. Correct code is generated
because the back end handles the incorrect shufflevector in a
consistent manner.
This patch and a companion patch for Clang correct this problem by
removing the fixup from altivec.h and the corresponding fixup from the
PowerPC back end. Several test cases are also modified to reflect the
now-correct LLVM IR.
llvm-svn: 214800
In commit r213915, Bill fixed little-endian usage of vmrgh* and vmrgl*
by swapping the input arguments. As it turns out, the exact same fix
is also required for the vpkuhum/vpkuwum patterns.
This fixes another regression in llvmpipe when vector support is
enabled.
Reviewed by Bill Schmidt.
llvm-svn: 214718
I ran into some test failures where common code changed vector division
by constant into a multiply-high operation (MULHU). But these are not
implemented by the back-end, so we failed to recognize the insn.
Fixed by marking MULHU/MULHS as Expand for vector types.
llvm-svn: 214716
This patch refactors code generation of vector comparisons.
This fixes a wrong code-gen bug for ISD::SETGE for floating-point types,
and improves generated code for vector comparisons in general.
Specifically, the patch moves all logic deciding how to implement vector
comparisons into getVCmpInst, which gets two extra boolean outputs
indicating to its caller whether its needs to swap the input operands
and/or negate the result of the comparison. Apart from implementing
these two modifications as directed by getVCmpInst, there is no need
to ever implement vector comparisons in any other manner; in particular,
there is never a need to perform two separate comparisons (e.g. one for
equal and one for greater-than, as code used to do before this patch).
Reviewed by Bill Schmidt.
llvm-svn: 214714
When generating unaligned vector loads, we need to search for other loads or
stores nearby offset by one vector width. If we find one, then we know that we
can safely generate another aligned load at that address. Otherwise, we must
generate the next load using an offset of the vector width minus one byte (so
we don't read off the end of the allocation if the base unaligned address
happened to be aligned at runtime). We had previously done this using only
other vector loads and stores, but did not consider the PowerPC-specific vector
load/store intrinsics. Now we'll also consider vector intrinsics. By itself,
this change is a feature enhancement, but is a necessary step toward fixing the
underlying problem behind PR19991.
llvm-svn: 214469
This is a follow-up to the activity in the bug at
http://llvm.org/bugs/show_bug.cgi?id=18663 . The underlying issue has
to do with how the KILL pseudo-instruction is handled. I defer to
Hal/Jakob/Uli for additional details and background.
This will disable the (bad?) assert, add an associated fixme comment,
and add a pair of tests.
The code change and the pr18663-2.ll test are copied from the referenced
bug. That test does not immediately fail in my environment, but I have
added the pr18663.ll test which does.
(Comment from Hal)
to provide everyone else with some context, this assert was not bad when
it was written. At that time, we only generated KILL pseudo instructions
around subregister copies. This logic, unfortunately, had its own problems.
In r199797, the relevant logic in MachineCopyPropagation was replaced to
generate KILLs for other kinds of copies too. This change in semantics broke
this now-problematic assumption in AggressiveAntiDepBreaker. The
AggressiveAntiDepBreaker really needs a proper cleanup to deal with the
change, but removing the assert (which just allows the function to return
false) is a safe conservative behavior, and should do for the time being.
llvm-svn: 214429
It seems that when I fixed this, almost exactly a year ago, I did not quite do
it correctly. When we have duplicate block predecessors, we can indeed not have
different incoming values for the same block, but we *must* have duplicate
entries. So, instead of skipping the duplicates, we explicitly add the
duplicate incoming values.
Fixes PR20442.
llvm-svn: 214423
Because the PowerPC vmrgh* and vmrgl* instructions have a built-in
big-endian bias, it is necessary to swap their inputs in little-endian
mode when using them to implement a vector shuffle. This was
previously missed in the vector LE implementation.
There was already logic to distinguish between unary and "normal"
vmrg* vector shuffles, so this patch extends that logic to use a third
option: "swapped" vmrg* vector shuffles that are used for little
endian in place of the "normal" ones.
I've updated the vec-shuffle-le.ll test to check for the expected
register ordering on the generated instructions.
This bug was discovered when testing the LE and ELFv2 patches for
safety if they were backported to 3.4. A different vectorization
decision was made in 3.4 than on mainline trunk, and that exposed the
problem. I've verified this fix takes care of that issue.
llvm-svn: 213915
insertions.
The old behavior could cause arbitrarily bad memory usage in the DAG
combiner if there was heavy traffic of adding nodes already on the
worklist to it. This commit switches the DAG combine worklist to work
the same way as the instcombine worklist where we null-out removed
entries and only add new entries to the worklist. My measurements of
codegen time shows slight improvement. The memory utilization is
unsurprisingly dominated by other factors (the IR and DAG itself
I suspect).
This change results in subtle, frustrating churn in the particular order
in which DAG combines are applied which causes a number of minor
regressions where we fail to match a pattern previously matched by
accident. AFAICT, all of these should be using AddToWorklist to directly
or should be written in a less brittle way. None of the changes seem
drastically bad, and a few of the changes seem distinctly better.
A major change required to make this work is to significantly harden the
way in which the DAG combiner handle nodes which become dead
(zero-uses). Previously, we relied on the ability to "priority-bump"
them on the combine worklist to achieve recursive deletion of these
nodes and ensure that the frontier of remaining live nodes all were
added to the worklist. Instead, I've introduced a routine to just
implement that precise logic with no indirection. It is a significantly
simpler operation than that of the combiner worklist proper. I suspect
this will also fix some other problems with the combiner.
I think the x86 changes are really minor and uninteresting, but the
avx512 change at least is hiding a "regression" (despite the test case
being just noise, not testing some performance invariant) that might be
looked into. Not sure if any of the others impact specific "important"
code paths, but they didn't look terribly interesting to me, or the
changes were really minor. The consensus in review is to fix any
regressions that show up after the fact here.
Thanks to the other reviewers for checking the output on other
architectures. There is a specific regression on ARM that Tim already
has a fix prepped to commit.
Differential Revision: http://reviews.llvm.org/D4616
llvm-svn: 213727
This patch adds infrastructure support for passing array types
directly. These can be used by the front-end to pass aggregate
types (coerced to an appropriate array type). The details of the
array type being used inform the back-end about ABI-relevant
properties. Specifically, the array element type encodes:
- whether the parameter should be passed in FPRs, VRs, or just
GPRs/stack slots (for float / vector / integer element types,
respectively)
- what the alignment requirements of the parameter are when passed in
GPRs/stack slots (8 for float / 16 for vector / the element type
size for integer element types) -- this corresponds to the
"byval align" field
Using the infrastructure provided by this patch, a companion patch
to clang will enable two features:
- In the ELFv2 ABI, pass (and return) "homogeneous" floating-point
or vector aggregates in FPRs and VRs (this is similar to the ARM
homogeneous aggregate ABI)
- As an optimization for both ELFv1 and ELFv2 ABIs, pass aggregates
that fit fully in registers without using the "byval" mechanism
The patch uses the functionArgumentNeedsConsecutiveRegisters callback
to encode that special treatment is required for all directly-passed
array types. The isInConsecutiveRegs / isInConsecutiveRegsLast bits set
as a results are then used to implement the required size and alignment
rules in CalculateStackSlotSize / CalculateStackSlotAlignment etc.
As a related change, the ABI routines have to be modified to support
passing floating-point types in GPRs. This is necessary because with
homogeneous aggregates of 4-byte float type we can now run out of FPRs
*before* we run out of the 64-byte argument save area that is shadowed
by GPRs. Any extra floating-point arguments that no longer fit in FPRs
must now be passed in GPRs until we run out of those too.
Note that there was already code to pass floating-point arguments in
GPRs used with vararg parameters, which was done by writing the argument
out to the argument save area first and then reloading into GPRs. The
patch re-implements this, however, in favor of code packing float arguments
directly via extension/truncation, BITCAST, and BUILD_PAIR operations.
This is required to support the ELFv2 ABI, since we cannot unconditionally
write to the argument save area (which the caller might not have allocated).
The change does, however, affect ELFv1 varags routines too; but even here
the overall effect should be advantageous: Instead of loading the argument
into the FPR, then storing the argument to the stack slot, and finally
reloading the argument from the stack slot into a GPR, the new code now
just loads the argument into the FPR, and subsequently loads the argument
into the GPR (via BITCAST). That BITCAST might imply a save/reload from
a stack temporary (in which case we're no worse than before); but it
might be implemented more efficiently in some cases.
The final part of the patch enables up to 8 FPRs and VRs for argument
return in PPCCallingConv.td; this is required to support returning
ELFv2 homogeneous aggregates. (Note that this doesn't affect other ABIs
since LLVM wil only look for which register to use if the parameter is
marked as "direct" return anyway.)
Reviewed by Hal Finkel.
llvm-svn: 213493
This is a minor improvement in the ELFv2 ABI. In ELFv1, DWARF CFI
would represent a saved CR word (holding CR fields CR2, CR3, and CR4)
using just a single CFI record refering to CR2. In ELFv2 instead,
each of the CR fields is represented by its own CFI record. The
advantage is that the compiler can now chose to save just a single
(or two) CR fields instead of all of them, if those are the only ones
that actually need saving. That can lead to more efficient code using
mf(o)crf instead of the (slow) mfcr instruction.
Note that this patch does not (yet) implement this more efficient
code generation, but it does implement the part that is required to
be ABI compliant: creating multiple CFI records if multiple CR fields
are saved.
Reviewed by Hal Finkel.
llvm-svn: 213492
The ELFv2 ABI reduces the amount of stack required to implement an
ABI-compliant function call in two ways:
* the "linkage area" is reduced from 48 bytes to 32 bytes by
eliminating two unused doublewords
* the 64-byte "parameter save area" is now optional and need not be
present in certain cases (it remains mandatory in functions with
variable arguments, and functions that have any parameter that is
passed on the stack)
The following patch implements this required changes:
- reducing the linkage area, and associated relocation of the TOC save
slot, in getLinkageSize / getTOCSaveOffset (this requires updating all
callers of these routines to pass in the isELFv2ABI flag).
- (partially) handling the case where the parameter save are is optional
This latter part requires some extra explanation: Currently, we still
always allocate the parameter save area when *calling* a function.
That is certainly always compliant with the ABI, but may cause code to
allocate stack unnecessarily. This can be addressed by a follow-on
optimization patch.
On the *callee* side, in LowerFormalArguments, we *must* track
correctly whether the ABI guarantees that the caller has allocated
the parameter save area for our use, and the patch does so. However,
there is one complication: the code that handles incoming "byval"
arguments will currently *always* write to the parameter save area,
because it has to force incoming register arguments to the stack since
it must return an *address* to implement the byval semantics.
To fix this, the patch changes the LowerFormalArguments code to write
arguments to a freshly allocated stack slot on the function's own stack
frame instead of the argument save area in those cases where that area
is not present.
Reviewed by Hal Finkel.
llvm-svn: 213490
This patch builds upon the two preceding MC changes to implement the
basic ELFv2 function call convention. In the ELFv1 ABI, a "function
descriptor" was associated with every function, pointing to both the
entry address and the related TOC base (and a static chain pointer
for nested functions). Function pointers would actually refer to that
descriptor, and the indirect call sequence needed to load up both entry
address and TOC base.
In the ELFv2 ABI, there are no more function descriptors, and function
pointers simply refer to the (global) entry point of the function code.
Indirect function calls simply branch to that address, after loading it
up into r12 (as required by the ABI rules for a global entry point).
Direct function calls continue to just do a "bl" to the target symbol;
this will be resolved by the linker to the local entry point of the
target function if it is local, and to a PLT stub if it is global.
That PLT stub would then load the (global) entry point address of the
final target into r12 and branch to it. Note that when performing a
local function call, r2 must be set up to point to the current TOC
base: if the target ends up local, the ABI requires that its local
entry point is called with r2 set up; if the target ends up global,
the PLT stub requires that r2 is set up.
This patch implements all LLVM changes to implement that scheme:
- No longer create a function descriptor when emitting a function
definition (in EmitFunctionEntryLabel)
- Emit two entry points *if* the function needs the TOC base (r2)
anywhere (this is done EmitFunctionBodyStart; note that this cannot
be done in EmitFunctionBodyStart because the global entry point
prologue code must be *part* of the function as covered by debug info).
- In order to make use tracking of r2 (as needed above) work correctly,
mark direct function calls as implicitly using r2.
- Implement the ELFv2 indirect function call sequence (no function
descriptors; load target address into r12).
- When creating an ELFv2 object file, emit the .abiversion 2 directive
to tell the linker to create the appropriate version of PLT stubs.
Reviewed by Hal Finkel.
llvm-svn: 213489
The PPCTargetLowering::SelectAddressRegImm routine needs to handle
FrameIndex nodes in a special manner, by tranlating them into a
TargetFrameIndex node. This was done in most cases, but seems to
have been neglected in one path: when the input tree has an OR of
the FrameIndex with an immediate. This can happen if the FrameIndex
can be proven to be sufficiently aligned that an OR of that immediate
is equivalent to an ADD.
The missing handling of FrameIndex in that case caused the SelectionDAG
instruction selection to miss opportunities to merge the OR back into
the FrameIndex node, leading to superfluous addi/ori instructions in
the final assembler output.
llvm-svn: 213482
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!
llvm-svn: 213427
This commit fixes a bug in PPCRegisterInfo::isFrameOffsetLegal that
could result in the LocalStackAlloc pass creating an MI instruction
out-of-range displacement:
%vreg17<def> = LD 33184, %vreg31; mem:LD8[%g](align=32)
%G8RC:%vreg17 G8RC_and_G8RC_NOX0:%vreg31
(In final assembler output the top bits are stripped off, resulting
in a negative offset loading from below the stack pointer.)
Common code expects the isFrameOffsetLegal routine to verify whether
adding a given offset to the offset already present in the instruction
results in a valid displacement. However, on PowerPC the routine
did not take the already present instruction offset into account.
This commit fixes isFrameOffsetLegal to add the instruction offset,
and updates a local caller (needsFrameBaseReg) to no longer add the
instruction offset itself before calling isFrameOffsetLegal.
Reviewed by Hal Finkel.
llvm-svn: 212832
Arguments passed as "byval align" should get the specified alignment
in the parameter save area. There was some code in PPCISelLowering.cpp
that attempted to implement this, but this didn't work correctly:
while code did update the ArgOffset value, it neglected to update
the PtrOff value (which was already computed from the old ArgOffset),
and it also neglected to update GPR_idx -- fields skipped due to
alignment in the save area must likewise be skipped in GPRs.
This patch fixes and simplifies this logic by:
- handling argument offset alignment right at the beginning
of argument processing, using a new helper routine
CalculateStackSlotAlignment (this avoids having to update
PtrOff and other derived values later on)
- not tracking GPR_idx separately, but always computing the
correct GPR_idx for each argument *from* its ArgOffset
- removing some redundant computation in LowerFormalArguments:
MinReservedArea must equal ArgOffset after argument processing,
so there's no use in computing it twice.
[This doesn't change the behavior of the current clang front-end,
since that never creates "byval align" arguments at the moment.
This will change with a follow-on patch, however.]
llvm-svn: 212476
There were two issues here:
1. At the very least, scattered relocations cannot use the same code to
determine the corresponding symbol being referred to. For some reason we
pretend there is no symbol, even when one actually exists in the symtab, so to
match this behaviour getRelocationSymbol should simply return symbols_end for
scattered relocations.
2. Printing "-" when we can't get a symbol (including the scattered case, but
not exclusively), isn't that helpful. In both cases there *is* interesting
information in that field, so we should print it. As hex will do.
Small part of rdar://problem/17553104
llvm-svn: 212332
The PowerPC 128-bit long double data type (ppcf128 in LLVM) is in fact a
pair of two doubles, where one is considered the "high" or
more-significant part, and the other is considered the "low" or
less-significant part. When a ppcf128 value is stored in memory or a
register pair, the high part always comes first, i.e. at the lower
memory address or in the lower-numbered register, and the low part
always comes second. This is true both on big-endian and little-endian
PowerPC systems. (Similar to how with a complex number, the real part
always comes first and the imaginary part second, no matter the byte
order of the system.)
This was implemented incorrectly for little-endian systems in LLVM.
This commit fixes three related issues:
- When printing an immediate ppcf128 constant to assembler output
in emitGlobalConstantFP, emit the high part first on both big-
and little-endian systems.
- When lowering a ppcf128 type to a pair of f64 types in SelectionDAG
(which is used e.g. when generating code to load an argument into a
register pair), use correct low/high part ordering on little-endian
systems.
- In a related issue, because lowering ppcf128 into a pair of f64 must
operate differently from lowering an int128 into a pair of i64,
bitcasts between ppcf128 and int128 must not be optimized away by the
DAG combiner on little-endian systems, but must effect a word-swap.
Reviewed by Hal Finkel.
llvm-svn: 212274
I've run into a bug where current LLVM at -O0 (with fast-isel)
generated invalid code like:
ld 0, 20936(1) # 8-byte Folded Reload
stw 12, 10348(0)
stw 12, 10344(0)
The underlying vreg had been introduced as base register by the
Local Stack Slot Allocation pass. That register was constrained
to G8RC by PPCRegisterInfo::materializeFrameBaseRegister to match
the ADDI instruction used to set it, but it was *not* constrained
to G8RC_NOX0 to fit the *use* of the register in an address.
That should have happened in PPCRegisterInfo::resolveFrameIndex.
This patch adds an appropriate constrainRegClass call.
Reviewed by Hal Finkel.
llvm-svn: 211897
[LLVM part]
These patches rename the loop unrolling and loop vectorizer metadata
such that they have a common 'llvm.loop.' prefix. Metadata name
changes:
llvm.vectorizer.* => llvm.loop.vectorizer.*
llvm.loopunroll.* => llvm.loop.unroll.*
This was a suggestion from an earlier review
(http://reviews.llvm.org/D4090) which added the loop unrolling
metadata.
Patch by Mark Heffernan.
llvm-svn: 211710
PR20071 identifies a problem in PowerPC's fast-isel implementation for
floating-point conversion to integer. The fctiduz instruction was added in
Power ISA 2.06 (i.e., Power7 and later). However, this instruction is being
generated regardless of which 64-bit PowerPC target is selected.
The intent is for fast-isel to punt to DAG selection when this instruction is
not available. This patch implements that change. For testing purposes, the
existing fast-isel-conversion.ll test adds a RUN line for -mcpu=970 and tests
for the expected code generation. Additionally, the existing test
fast-isel-conversion-p5.ll was found to be incorrectly expecting the
unavailable instruction to be generated. I've removed these test variants
since we have adequate coverage in fast-isel-conversion.ll.
llvm-svn: 211627
The PPCFrameLowering::determineFrameLayout routine currently ensures
that every function that allocates a stack frame provides space for the
parameter save area (via PPCFrameLowering::getMinCallFrameSize).
This is actually not necessary. There may be functions that never call
another routine but still allocate a frame; those do not require the
parameter save area. In the future, with the ELFv2 ABI, even some
routines that do call other functions do not need to allocate the
parameter save area.
While it is not a bug to allocate the parameter area when it is not
needed, it is better to avoid it to save stack space.
Note that when any particular function call requires the parameter save
area, this space will already have been included by ABI code in the size
the CALLSEQ_START insn is annotated with, and therefore included in the
size returned by MFI->getMaxCallFrameSize().
This means that determineFrameLayout simply does not need to care about
the parameter save area. (It still needs to ensure that every frame
provides the linkage area.) This is implemented by this patch.
Note that this exposed a bug in the new fast-isel code where the parameter
area was *not* included in the CALLSEQ_START size; this is also fixed.
A couple of test cases needed to be adapted for the new (smaller) stack
frame size those tests now see.
llvm-svn: 211495
Current 64-bit SVR4 code seems to have some remnants of Darwin code
in AltiVec argument handing. This had the effect that AltiVec arguments
(or subsequent arguments) were not correctly placed in the parameter area
in some cases.
The correct behaviour with the 64-bit SVR4 ABI is:
- All AltiVec arguments take up space in the parameter area, just like
any other arguments, whether vararg or not.
- They are always 16-byte aligned, skipping a parameter area doubleword
(and the associated GPR, if any), if necessary.
This patch implements the correct behaviour and adds a test case.
(Verified against GCC behaviour via the ABI compat test suite.)
llvm-svn: 211492
When small arguments (structures < 8 bytes or "float") are passed in a
stack slot in the ppc64 SVR4 ABI, they must reside in the least
significant part of that slot. On BE, this means that an offset needs
to be added to the stack address of the parameter, but on LE, the least
significant part of the slot has the same address as the slot itself.
This changes the PowerPC back-end ABI code to only add the small
argument stack slot offset for BE. It also adds test cases to verify
the correct behavior on both BE and LE.
llvm-svn: 211368
During an indirect function call sequence on the 64-bit SVR4 ABI,
generate code must load and then restore the TOC register.
This does not use a regular LOAD instruction since the TOC
register r2 is marked as reserved. Instead, the are two
special instruction patterns:
let RST = 2, DS = 2 in
def LDinto_toc: DSForm_1a<58, 0, (outs), (ins g8rc:$reg),
"ld 2, 8($reg)", IIC_LdStLD,
[(PPCload_toc i64:$reg)]>, isPPC64;
let RST = 2, DS = 10, RA = 1 in
def LDtoc_restore : DSForm_1a<58, 0, (outs), (ins),
"ld 2, 40(1)", IIC_LdStLD,
[(PPCtoc_restore)]>, isPPC64;
Note that these not only restrict the destination of the
load to r2, but they also restrict the *source* of the
load to particular address combinations. The latter is
a problem when we want to support the ELFv2 ABI, since
there the TOC save slot is no longer at 40(1).
This patch replaces those two instructions with a single
instruction pattern that only hard-codes r2 as destination,
but supports generic addresses as source. This will allow
supporting the ELFv2 ABI, and also helps generate more
efficient code for calls to absolute addresses (allowing
simplification of the ppc64-calls.ll test case).
llvm-svn: 211193
As requested by Hal Finkel, this adds back a test for calls to
a known-constant function pointer value, and verifies that the
64-bit SVR4 indirect function call sequence is used.
llvm-svn: 211190
The PowerPC back-end uses BLA to implement calls to functions at
known-constant addresses, which is apparently used for certain
system routines on Darwin.
However, with the 64-bit SVR4 ABI, this is actually incorrect.
An immediate function pointer value on this platform is not
directly usable as a target address for BLA:
- in the ELFv1 ABI, the function pointer value refers to the
*function descriptor*, not the code address
- in the ELFv2 ABI, the function pointer value refers to the
global entry point, but BL(A) would only be correct when
calling the *local* entry point
This bug didn't show up since using immediate function pointer
values is not usually done in the 64-bit SVR4 ABI in the first
place. However, I ran into this issue with a certain use case
of LLVM as JIT, where immediate function pointer values were
uses to implement callbacks from JITted code to helpers in
statically compiled code.
Fixed by simply not using BLA with the 64-bit SVR4 ABI.
llvm-svn: 211174
Rafael opened http://llvm.org/bugs/show_bug.cgi?id=19893 to track non-optimal
code generation for forming a function address that is local to the compile
unit. The existing code was treating both local and non-local functions
identically.
This patch fixes the problem by properly identifying local functions and
generating the proper addis/addi code. I also noticed that Rafael's earlier
changes to correct the surrounding code in PPCISelLowering.cpp were also
needed for fast instruction selection in PPCFastISel.cpp, so this patch
fixes that code as well.
The existing test/CodeGen/PowerPC/func-addr.ll is modified to test the new
code generation. I've added a -O0 run line to test the fast-isel code as
well.
Tested on powerpc64[le]-unknown-linux-gnu with no regressions.
llvm-svn: 211056
This commit adds a weak variant of the cmpxchg operation, as described
in C++11. A cmpxchg instruction with this modifier is permitted to
fail to store, even if the comparison indicated it should.
As a result, cmpxchg instructions must return a flag indicating
success in addition to their original iN value loaded. Thus, for
uniformity *all* cmpxchg instructions now return "{ iN, i1 }". The
second flag is 1 when the store succeeded.
At the DAG level, a new ATOMIC_CMP_SWAP_WITH_SUCCESS node has been
added as the natural representation for the new cmpxchg instructions.
It is a strong cmpxchg.
By default this gets Expanded to the existing ATOMIC_CMP_SWAP during
Legalization, so existing backends should see no change in behaviour.
If they wish to deal with the enhanced node instead, they can call
setOperationAction on it. Beware: as a node with 2 results, it cannot
be selected from TableGen.
Currently, no use is made of the extra information provided in this
patch. Test updates are almost entirely adapting the input IR to the
new scheme.
Summary for out of tree users:
------------------------------
+ Legacy Bitcode files are upgraded during read.
+ Legacy assembly IR files will be invalid.
+ Front-ends must adapt to different type for "cmpxchg".
+ Backends should be unaffected by default.
llvm-svn: 210903
Various masks on shufflevector instructions are recognizable as
specific PowerPC instructions (vector pack, vector merge, etc.).
There is existing code in PPCISelLowering.cpp to recognize the correct
patterns for big endian code. The masks for these instructions are
different for little endian code due to the big-endian numbering
employed by these instructions. This patch adds the recognition code
for little endian.
I've added a new test case test/CodeGen/PowerPC/vec_shuffle_le.ll for
this. The existing recognizer test (vec_shuffle.ll) is unnecessarily
verbose and difficult to read, so I felt it was better to add a new
test rather than modify the old one.
llvm-svn: 210536
The code in PPCTargetLowering::PerformDAGCombine() that handles
unaligned Altivec vector loads generates a lvsl followed by a vperm.
As we've seen in numerous other places, the vperm instruction has a
big-endian bias, and this is fixed for little endian by complementing
the permute control vector and swapping the input operands. In this
case the lvsl is providing the permute control vector. Rather than
generating an lvsl and a complement operation, it is sufficient to
generate an lvsr instruction instead. Thus for LE code generation we
will generate an lvsr rather than an lvsl, and swap the other input
arguments on the vperm.
The existing test/CodeGen/PowerPC/vec_misalign.ll is updated to test
the code generation for PPC64 and PPC64LE, in addition to the existing
PPC32/G5 testing.
llvm-svn: 210493
The existing code in PPCTargetLowering::LowerMUL() for multiplying two
v16i8 values assumes that vector elements are numbered in big-endian
order. For little-endian targets, the vector element numbering is
reversed, but the vmuleub, vmuloub, and vperm instructions still
assume big-endian numbering. To account for this, we must adjust the
permute control vector and reverse the order of the input registers on
the vperm instruction.
The existing test/CodeGen/PowerPC/vec_mul.ll is updated to be executed
on powerpc64 and powerpc64le targets as well as the original powerpc
(32-bit) target.
llvm-svn: 210474
This patch fixes a couple of lowering issues for little endian
PowerPC. The code for lowering BUILD_VECTOR contains a number of
optimizations that are only valid for big endian. For now, we disable
those optimizations for correctness. In the future, we will add
analogous optimizations that are correct for little endian.
When lowering a SHUFFLE_VECTOR to a VPERM operation, we again need to
make the now-familiar transformation of swapping the input operands
and complementing the permute control vector. Correctness of this
transformation is tested by the accompanying test case.
llvm-svn: 210336
Chandler correctly pointed out that I need an LLVM IR test for
r210282, which modified the vperm -> shuffle transform for little
endian PowerPC. This patch provides that test.
llvm-svn: 210297
This seems to match what gcc does for ppc and what every other llvm
backend does.
This is a fixed version of r209638. The difference is to avoid any change
in behavior for functions. The logic for using constant pools for function
addresseses is spread over a few places and we have to keep them in sync.
llvm-svn: 209821
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)).
llvm-svn: 209745
In PPCISelLowering.cpp: PPCTargetLowering::LowerBUILD_VECTOR(), there
is an optimization for certain patterns to generate one or two vector
splats followed by a vector add or subtract. This operation is
represented by a VADD_SPLAT in the selection DAG. Prior to this
patch, it was possible for the VADD_SPLAT to be assigned the wrong
data type, causing incorrect code generation. This patch corrects the
problem.
Specifically, the code previously assigned the value type of the
BUILD_VECTOR node to the newly generated VADD_SPLAT node. This is
correct much of the time, but not always. The problem is that the
call to isConstantSplat() may return a SplatBitSize that is not the
same as the number of bits in the original element vector type. The
correct type to assign is a vector type with the same element bit size
as SplatBitSize.
The included test case shows an example of this, where the
BUILD_VECTOR node has a type of v16i8. The vector to be built is {0,
16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16}. isConstantSplat
detects that we can generate a splat of 16 for type v8i16, which is
the type we must assign to the VADD_SPLAT node. If we do not, we
generate a vspltisb of 8 and a vaddubm, which generates the incorrect
result {16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16,
16}. The correct code generation is a vspltish of 8 and a vadduhm.
This patch also corrected code generation for
CodeGen/PowerPC/2008-07-10-SplatMiscompile.ll, which had been marked
as an XFAIL, so we can remove the XFAIL from the test case.
llvm-svn: 209662
The SplitIndexingFromLoad changes exposed a latent isel bug in the PowerPC64
backend. We matched an immediate offset with STWX8 even though it only
supports register offset.
The culprit is the complex-pattern predicate, SelectAddrIdx, which decides
that if the offset is not ISD::Constant it must be a register.
Many thanks to Bill Schmidt for testing this.
llvm-svn: 209219
Many old tests using prior schemas still had some brokenness here (both
indirect arrays and arrays with single bogus elements). Fixed those up
so they don't hit the new assertions.
Also reduced nesting in some places, etc.
llvm-svn: 208817
Support for the intrinsics that read from and write to global named registers
is added for r1, r2 and r13 (depending on the subtarget).
llvm-svn: 208509
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.
llvm-svn: 208501
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.
llvm-svn: 206136
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.
llvm-svn: 206093
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).
llvm-svn: 206092
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.
llvm-svn: 205630
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.
llvm-svn: 205231
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.
llvm-svn: 205179
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.
llvm-svn: 205153
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.
llvm-svn: 205146
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.
llvm-svn: 205106
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.
llvm-svn: 205075
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.
llvm-svn: 205041
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.
llvm-svn: 204980
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.
llvm-svn: 204971
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.
llvm-svn: 204963
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.
llvm-svn: 204838
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.
llvm-svn: 204828
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.
llvm-svn: 204801
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
llvm-svn: 204768
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.
llvm-svn: 204559
When converting a signed 32-bit integer to double-precision floating point on
hardware without a lfiwax instruction, we have to instead use a lfd followed
by fcfid. We were erroneously offsetting the address by 4 bytes in
preparation for either a lfiwax or lfiwzx when generating the lfd. This fixes
that silly error.
This was not caught in the test suite since the conversion tests were run with
-mcpu=pwr7, which implies availability of lfiwax. I've added another test
case for older hardware that checks the code we expect in the absence of
lfiwax and other flavors of fcfid. There are fewer tests in this test case
because we punt to DAG selection in more cases on older hardware. (We must
generate complex fiddly sequences in those cases, and there is marginal
benefit in duplicating that logic in fast-isel.)
llvm-svn: 204155
Commit r181723 introduced code to avoid placing initialized variables
needing relocations into the .rodata section, which avoid copy relocs
that do not work as expected on ppc64 function references.
The same treatment is also needed for *named* .rodata.XXX sections.
This patch changes PPC64LinuxTargetObjectFile::SelectSectionForGlobal
to modify "Kind" *before* calling the default SelectSectionForGlobal
routine, instead of first calling the default routine and then just
checking for the (main) .rodata section afterwards.
llvm-svn: 203921
These linkages were introduced some time ago, but it was never very
clear what exactly their semantics were or what they should be used
for. Some investigation found these uses:
* utf-16 strings in clang.
* non-unnamed_addr strings produced by the sanitizers.
It turns out they were just working around a more fundamental problem.
For some sections a MachO linker needs a symbol in order to split the
section into atoms, and llvm had no idea that was the case. I fixed
that in r201700 and it is now safe to use the private linkage. When
the object ends up in a section that requires symbols, llvm will use a
'l' prefix instead of a 'L' prefix and things just work.
With that, these linkages were already dead, but there was a potential
future user in the objc metadata information. I am still looking at
CGObjcMac.cpp, but at this point I am convinced that linker_private
and linker_private_weak are not what they need.
The objc uses are currently split in
* Regular symbols (no '\01' prefix). LLVM already directly provides
whatever semantics they need.
* Uses of a private name (start with "\01L" or "\01l") and private
linkage. We can drop the "\01L" and "\01l" prefixes as soon as llvm
agrees with clang on L being ok or not for a given section. I have two
patches in code review for this.
* Uses of private name and weak linkage.
The last case is the one that one could think would fit one of these
linkages. That is not the case. The semantics are
* the linker will merge these symbol by *name*.
* the linker will hide them in the final DSO.
Given that the merging is done by name, any of the private (or
internal) linkages would be a bad match. They allow llvm to rename the
symbols, and that is really not what we want. From the llvm point of
view, these objects should really be (linkonce|weak)(_odr)?.
For now, just keeping the "\01l" prefix is probably the best for these
symbols. If we one day want to have a more direct support in llvm,
IMHO what we should add is not a linkage, it is just a hidden_symbol
attribute. It would be applicable to multiple linkages. For example,
on weak it would produce the current behavior we have for objc
metadata. On internal, it would be equivalent to private (and we
should then remove private).
llvm-svn: 203866
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.
llvm-svn: 203768
The syntax for "cmpxchg" should now look something like:
cmpxchg i32* %addr, i32 42, i32 3 acquire monotonic
where the second ordering argument gives the required semantics in the case
that no exchange takes place. It should be no stronger than the first ordering
constraint and cannot be either "release" or "acq_rel" (since no store will
have taken place).
rdar://problem/15996804
llvm-svn: 203559
When copying an i1 value into a GPR for a vaarg call, we need to explicitly
zero-extend the i1 value (otherwise an invalid CRBIT -> GPR copy will be
generated).
llvm-svn: 203041
On cores without fpcvt support, we cannot promote int_to_fp i1 operations,
because there is nothing to promote them to. The most straightforward
implementation of this uses a select to choose between the two possible
resulting floating-point values (and that's what is done here).
llvm-svn: 203015
Now that the PowerPC backend can track individual CR bits as first-class
registers, we should also have a way of allocating them for inline asm
statements. Because these registers are only one bit, if an output variable is
implicitly cast to a larger integer size, we'll get an any_extend to that
larger type (this is part of the existing target-independent logic). As a
result, regardless of the size of the output type, only the first bit is
meaningful.
The constraint identifier "wc" has been chosen for this purpose. Although gcc
does not currently support allocating individual CR bits, this identifier
choice has been coordinated with the gcc PowerPC team, and will be marked as
reserved for this purpose in the gcc constraints.md file.
llvm-svn: 202657
This generalizes the code to eliminate extra truncs/exts around i1 bit
operations to also do the same on PPC64 for i32 bit operations. This eliminates
a fairly prevalent code wart:
int foo(int a) {
return a == 5 ? 7 : 8;
}
On PPC64, because of the extension implied by the ABI, this would generate:
cmplwi 0, 3, 5
li 12, 8
li 4, 7
isel 3, 4, 12, 2
rldicl 3, 3, 0, 32
blr
where the 'rldicl 3, 3, 0, 32', the extension, is completely unnecessary. At
least for the single-BB case (which is all that the DAG combine mechanism can
handle), this unnecessary extension is no longer generated.
llvm-svn: 202600
The PPC isel instruction can fold 0 into the first operand (thus eliminating
the need to materialize a zero-containing register when the 'true' result of
the isel is 0). When the isel is fed by a bit register operation that we can
invert, do so as part of the bit-register-operation peephole routine.
llvm-svn: 202469
This change enables tracking i1 values in the PowerPC backend using the
condition register bits. These bits can be treated on PowerPC as separate
registers; individual bit operations (and, or, xor, etc.) are supported.
Tracking booleans in CR bits has several advantages:
- Reduction in register pressure (because we no longer need GPRs to store
boolean values).
- Logical operations on booleans can be handled more efficiently; we used to
have to move all results from comparisons into GPRs, perform promoted
logical operations in GPRs, and then move the result back into condition
register bits to be used by conditional branches. This can be very
inefficient, because the throughput of these CR <-> GPR moves have high
latency and low throughput (especially when other associated instructions
are accounted for).
- On the POWER7 and similar cores, we can increase total throughput by using
the CR bits. CR bit operations have a dedicated functional unit.
Most of this is more-or-less mechanical: Adjustments were needed in the
calling-convention code, support was added for spilling/restoring individual
condition-register bits, and conditional branch instruction definitions taking
specific CR bits were added (plus patterns and code for generating bit-level
operations).
This is enabled by default when running at -O2 and higher. For -O0 and -O1,
where the ability to debug is more important, this feature is disabled by
default. Individual CR bits do not have assigned DWARF register numbers,
and storing values in CR bits makes them invisible to the debugger.
It is critical, however, that we don't move i1 values that have been promoted
to larger values (such as those passed as function arguments) into bit
registers only to quickly turn around and move the values back into GPRs (such
as happens when values are returned by functions). A pair of target-specific
DAG combines are added to remove the trunc/extends in:
trunc(binary-ops(binary-ops(zext(x), zext(y)), ...)
and:
zext(binary-ops(binary-ops(trunc(x), trunc(y)), ...)
In short, we only want to use CR bits where some of the i1 values come from
comparisons or are used by conditional branches or selects. To put it another
way, if we can do the entire i1 computation in GPRs, then we probably should
(on the POWER7, the GPR-operation throughput is higher, and for all cores, the
CR <-> GPR moves are expensive).
POWER7 test-suite performance results (from 10 runs in each configuration):
SingleSource/Benchmarks/Misc/mandel-2: 35% speedup
MultiSource/Benchmarks/Prolangs-C++/city/city: 21% speedup
MultiSource/Benchmarks/MiBench/automotive-susan: 23% speedup
SingleSource/Benchmarks/CoyoteBench/huffbench: 13% speedup
SingleSource/Benchmarks/Misc-C++/Large/sphereflake: 13% speedup
SingleSource/Benchmarks/Misc-C++/mandel-text: 10% speedup
SingleSource/Benchmarks/Misc-C++-EH/spirit: 10% slowdown
MultiSource/Applications/lemon/lemon: 8% slowdown
llvm-svn: 202451
We need to abort the formation of counter-register-based loops where there are
128-bit integer operations that might become function calls.
llvm-svn: 202192
r201608 made llvm corretly handle private globals with MachO. r201622 fixed
a bug in it and r201624 and r201625 were changes for using private linkage,
assuming that llvm would do the right thing.
They all got reverted because r201608 introduced a crash in LTO. This patch
includes a fix for that. The issue was that TargetLoweringObjectFile now has
to be initialized before we can mangle names of private globals. This is
trivially true during the normal codegen pipeline (the asm printer does it),
but LTO has to do it manually.
llvm-svn: 201700
The IR
@foo = private constant i32 42
is valid, but before this patch we would produce an invalid MachO from it. It
was invalid because it would use an L label in a section where the liker needs
the labels in order to atomize it.
One way of fixing it would be to just reject this IR in the backend, but that
would not be very front end friendly.
What this patch does is use an 'l' prefix in sections that we know the linker
requires symbols for atomizing them. This allows frontends to just use
private and not worry about which sections they go to or how the linker handles
them.
One small issue with this strategy is that now a symbol name depends on the
section, which is not available before codegen. This is not a problem in
practice. The reason is that it only happens with private linkage, which will
be ignored by the non codegen users (llvm-nm and llvm-ar).
llvm-svn: 201608
Summary:
AsmPrinter::EmitInlineAsm() will no longer use the EmitRawText() call for
targets with mature MC support. Such targets will always parse the inline
assembly (even when emitting assembly). Targets without mature MC support
continue to use EmitRawText() for assembly output.
The hasRawTextSupport() check in AsmPrinter::EmitInlineAsm() has been replaced
with MCAsmInfo::UseIntegratedAs which when true, causes the integrated assembler
to parse inline assembly (even when emitting assembly output). UseIntegratedAs
is set to true for targets that consider any failure to parse valid assembly
to be a bug. Target specific subclasses generally enable the integrated
assembler in their constructor. The default value can be overridden with
-no-integrated-as.
All tests that rely on inline assembly supporting invalid assembly (for example,
those that use mnemonics such as 'foo' or 'hello world') have been updated to
disable the integrated assembler.
Changes since review (and last commit attempt):
- Fixed test failures that were missed due to configuration of local build.
(fixes crash.ll and a couple others).
- Fixed tests that happened to pass because the local build was on X86
(should fix 2007-12-17-InvokeAsm.ll)
- mature-mc-support.ll's should no longer require all targets to be compiled.
(should fix ARM and PPC buildbots)
- Object output (-filetype=obj and similar) now forces the integrated assembler
to be enabled regardless of default setting or -no-integrated-as.
(should fix SystemZ buildbots)
Reviewers: rafael
Reviewed By: rafael
CC: llvm-commits
Differential Revision: http://llvm-reviews.chandlerc.com/D2686
llvm-svn: 201333
Summary:
AsmPrinter::EmitInlineAsm() will no longer use the EmitRawText() call for targets with mature MC support. Such targets will always parse the inline assembly (even when emitting assembly). Targets without mature MC support continue to use EmitRawText() for assembly output.
The hasRawTextSupport() check in AsmPrinter::EmitInlineAsm() has been replaced with MCAsmInfo::UseIntegratedAs which when true, causes the integrated assembler to parse inline assembly (even when emitting assembly output). UseIntegratedAs is set to true for targets that consider any failure to parse valid assembly to be a bug. Target specific subclasses generally enable the integrated assembler in their constructor. The default value can be overridden with -no-integrated-as.
All tests that rely on inline assembly supporting invalid assembly (for example, those that use mnemonics such as 'foo' or 'hello world') have been updated to disable the integrated assembler.
Reviewers: rafael
Reviewed By: rafael
CC: llvm-commits
Differential Revision: http://llvm-reviews.chandlerc.com/D2686
llvm-svn: 201237
A bunch of test cases needed to be cleaned up for this, many my fault -
when implementid imported modules I updated test cases by simply
duplicating the prior metadata field - which wasn't always the empty
metadata entry.
llvm-svn: 200731
GPRC_NOR0 is not a subclass of GPRC (because it also contains the ZERO pseudo
register). As a result, we also need to check for it in the spilling code.
llvm-svn: 200288
I disabled the use of TBAA in CodeGen in r200093. This adds a test case that
demonstrates the problems with inttoptr and TBAA in CodeGen (and, specifically,
the problem that causes LLVM to miscompile itself in Release mode). This test
will currently fail if -use-tbaa-in-sched-mi is enabled.
llvm-svn: 200097
For PPC64 SVR (and Darwin), the stores that take byval aggregate parameters
from registers into the stack frame had MachinePointerInfo objects with
incorrect offsets. These offsets are relative to the object itself, not to the
stack frame base.
This fixes self hosting on PPC64 when compiling with -enable-aa-sched-mi.
llvm-svn: 199763
This is a base implementation of the powerpc-apple-darwin asm parser dialect.
* Enables infrastructure (essentially isDarwin()) and fixes up the parsing of asm directives to separate out ELF and MachO/Darwin additions.
* Enables parsing of {r,f,v}XX as register identifiers.
* Enables parsing of lo16() hi16() and ha16() as modifiers.
The changes to the test case are from David Fang (fangism).
llvm-svn: 197324
Aside from a few minor latency corrections, the major change here is a new
hazard recognizer which focuses on better dispatch-group formation on the
POWER7. As with the PPC970's hazard recognizer, the most important thing it
does is avoid load-after-store hazards within the same dispatch group. It uses
the POWER7's special dispatch-group-terminating nop instruction (instead of
inserting multiple regular nop instructions). This new hazard recognizer makes
use of the scheduling dependency graph itself, built using AA information, to
robustly detect the possibility of load-after-store hazards.
significant test-suite performance changes (the error bars are 99.5% confidence
intervals based on 5 test-suite runs both with and without the change --
speedups are negative):
speedups:
MultiSource/Benchmarks/FreeBench/pcompress2/pcompress2
-0.55171% +/- 0.333168%
MultiSource/Benchmarks/TSVC/CrossingThresholds-dbl/CrossingThresholds-dbl
-17.5576% +/- 14.598%
MultiSource/Benchmarks/TSVC/Reductions-dbl/Reductions-dbl
-29.5708% +/- 7.09058%
MultiSource/Benchmarks/TSVC/Reductions-flt/Reductions-flt
-34.9471% +/- 11.4391%
SingleSource/Benchmarks/BenchmarkGame/puzzle
-25.1347% +/- 11.0104%
SingleSource/Benchmarks/Misc/flops-8
-17.7297% +/- 9.79061%
SingleSource/Benchmarks/Shootout-C++/ary3
-35.5018% +/- 23.9458%
SingleSource/Regression/C/uint64_to_float
-56.3165% +/- 25.4234%
SingleSource/UnitTests/Vectorizer/gcc-loops
-18.5309% +/- 6.8496%
regressions:
MultiSource/Benchmarks/ASCI_Purple/SMG2000/smg2000
18.351% +/- 12.156%
SingleSource/Benchmarks/Shootout-C++/methcall
27.3086% +/- 14.4733%
llvm-svn: 197099
For one predicate to subsume another, they must both check the same condition
register. Failure to check this prerequisite was causing miscompiles.
Fixes PR18003.
llvm-svn: 197089
Convert this test to FileCheck, and improve it to check for the instructions it
is trying to exclude instead of checking for register use (especially because
grepping for r1 can be thrown off, for example, by a use of r12).
llvm-svn: 195979
Some of these tests did not specify a cpu but were also sensitive to
instruction scheduling and/or register assignment choices. A few others
similarly-sensitive tests specified a cpu (often the POWER7), and while the P7
currently uses the default model for PPC64, this will soon change. For those
tests which should not really be cpu-dependent anyway, the cpu is set to the
generic 'ppc64'.
llvm-svn: 195977
We are going to drop debug info without a version number or with a different
version number, to make sure we don't crash when we see bitcode files with
different debug info metadata format.
llvm-svn: 195504
The instruction definitions incorrectly specified that popcntd and popcntw have
record forms; they do not. This mistake was causing invalid code generation.
llvm-svn: 195272
Masking operations (where only some number of the low bits are being kept) are
selected to rldicl(x, 0, mb). If x is a logical right shift (which would become
rldicl(y, 64-n, n)), we might be able to fold the two instructions together:
rldicl(rldicl(x, 64-n, n), 0, mb) -> rldicl(x, 64-n, mb) for n <= mb
The right shift is really a left rotate followed by a mask, and if the explicit
mask is a more-restrictive sub-mask of the mask implied by the shift, only one
rldicl is needed.
llvm-svn: 195185
Stop folding constant adds into GEP when the type size doesn't match.
Otherwise, the adds' operands are effectively being promoted, changing the
conditions of an overflow. Results are different when:
sext(a) + sext(b) != sext(a + b)
Problem originally found on x86-64, but also fixed issues with ARM and PPC,
which used similar code.
<rdar://problem/15292280>
Patch by Duncan Exon Smith!
llvm-svn: 194840
In ELF and COFF an alias is just another offset in a section. There is no way
to represent an alias to something in another file.
In MachO, the spec has the N_INDR type which should allow for exactly that, but
is not currently implemented. Given that it is specified but not implemented,
we error in codegen to avoid miscompiling but don't reject aliases to
declarations in the verifier to leave the option open of implementing it.
In the past we have used alias to declarations as a way of implementing
weakref, which is why it exists in some old tests which this patch updates.
llvm-svn: 194705
On non-Darwin PPC systems, we currently strip off the register name prefix
prior to instruction printing. So instead of something like this:
mr r3, r4
we print this:
mr 3, 4
The first form is the default on Darwin, and is understood by binutils, but not
yet understood by our integrated assembler. Once our integrated-as understands
full register names as well, this temporary option will be replaced by tying
this functionality to the verbose-asm option. The numeric-only form is
compatible with legacy assemblers and tools, and is also gcc's default on most
PPC systems. On the other hand, it is harder to read, and there are some
analysis tools that expect full register names.
llvm-svn: 194384
PR17168 describes a test case that fails when compiling for debug with
fast-isel. Investigation showed that the test was failing because a DBG_VALUE
machine instruction was placed prior to a PHI.
For this problem to occur requires the following:
* Compile for debug
* Compile with fast-isel
* In a block B, fast-isel must partially succeed before punting to DAG-isel
* B must start with a PHI
* The first unhandled node in the DAG must not generate a machine instruction
* A debug value with an order less than that of that first node exists
When all of these circumstances apply, the existing test that an instruction
was not inserted won't fire. Currently it tests whether the block is empty,
or whether the last instruction generated is a phi. When fast-isel has
partially succeeded, the last instruction generated will not be a phi.
Instead, we need to check whether the current insert position is immediately
following a phi. This patch adds that check, and adds the test case from the
PR as a regression test.
llvm-svn: 192976
When generating code for shared libraries, even local calls may be
intercepted, so we need a nop after the call for the linker to fix up the
TOC. Test case adapted from the one provided in PR17354.
llvm-svn: 191440
Large code model on PPC64 requires creating and referencing TOC entries when
using the addis/ld form of addressing. This was not being done in all cases.
The changes in this patch to PPCAsmPrinter::EmitInstruction() fix this. Two
test cases are also modified to reflect this requirement.
Fast-isel was not creating correct code for loading floating-point constants
using large code model. This also requires the addis/ld form of addressing.
Previously we were using the addis/lfd shortcut which is only applicable to
medium code model. One test case is modified to reflect this requirement.
llvm-svn: 190882
This is a re-commit of r190764, with an extra check to make sure that we're not
performing the transformation on illegal types (a small test case has been
added for this as well).
Original commit message:
The PPC backend uses a target-specific DAG combine to turn unaligned Altivec
loads into a permutation-based sequence when possible. Unfortunately, the
target-specific DAG combine is not always called on all loads of interest
(sometimes the routines in DAGCombine call CombineTo such that the new node and
users are not added to the worklist); allowing the combine to trigger early
(before type legalization) mitigates this problem. Because the autovectorizers
only create legal vector types, I don't expect a lot of cases where this
optimization is enabled by type legalization in practice.
llvm-svn: 190771
This is causing test-suite failures.
Original commit message:
The PPC backend uses a target-specific DAG combine to turn unaligned Altivec
loads into a permutation-based sequence when possible. Unfortunately, the
target-specific DAG combine is not always called on all loads of interest
(sometimes the routines in DAGCombine call CombineTo such that the new node and
users are not added to the worklist); allowing the combine to trigger early
(before type legalization) mitigates this problem. Because the autovectorizers
only create legal vector types, I don't expect a lot of cases where this
optimization is enabled by type legalization in practice.
llvm-svn: 190765
The PPC backend uses a target-specific DAG combine to turn unaligned Altivec
loads into a permutation-based sequence when possible. Unfortunately, the
target-specific DAG combine is not always called on all loads of interest
(sometimes the routines in DAGCombine call CombineTo such that the new node and
users are not added to the worklist); allowing the combine to trigger early
(before type legalization) mitigates this problem. Because the autovectorizers
only create legal vector types, I don't expect a lot of cases where this
optimization is enabled by type legalization in practice.
llvm-svn: 190764
DAGCombiner::isAlias can be called with SrcValue1 or SrcValue2 null, and we
can't use AA in this case (if we try, then the casting code in AA will assert).
llvm-svn: 190763
When a structure is passed by value, and that structure contains a vector
member, according to the PPC ABI, the structure will receive enhanced alignment
(so that the vector within the structure will always be aligned).
This should resolve PR16641.
llvm-svn: 190636
In fast-math mode sqrt(x) is calculated using the fast expansion of the
reciprocal of the reciprocal sqrt expansion. The reciprocal and reciprocal
sqrt expansions use the associated estimate instructions along with some Newton
iterations. Unfortunately, as a result, sqrt(0) was being calculated as NaN,
which is not correct. Now we explicitly return a result of zero if the input is
zero.
llvm-svn: 190624
Aggressive anti-dependency breaking is enabled by default for all PPC cores.
This provides a general speedup on the P7 and other platforms (among other
factors, the instruction group formation for the non-embedded PPC cores is done
during post-RA scheduling). In order to do this safely, the incompatibility
between uses of the MFOCRF instruction and anti-dependency breaking are
resolved by marking MFOCRF with hasExtraSrcRegAllocReq. As noted in the removed
FIXME, the problem was that MFOCRF's output is sensitive to the identify of the
source register, and always paired with a shift to undo this effect. Because
anti-dependency breaking is unaware of this hidden dependency of the shift
amount on the source register of the MFOCRF instruction, changing that register
must be inhibited.
Two test cases were adjusted: The SjLj test was made more insensitive to
register choices and scheduling; the saveCR test disabled anti-dependency
breaking because part of what it is testing is proper register reuse.
llvm-svn: 190587
Field 2 of DIType (Context), field 9 of DIDerivedType (TypeDerivedFrom),
field 12 of DICompositeType (ContainingType), fields 2, 7, 12 of DISubprogram
(Context, Type, ContainingType).
llvm-svn: 190205
This patch adds fast-isel support for calls (but not intrinsic calls
or varargs calls). It also removes a badly-formed assert. There are
some new tests just for calls, and also for folding loads into
arguments on calls to avoid extra extends.
llvm-svn: 189701
Yet another chunk of fast-isel code. This one handles various
conversions involving floating-point. (It also includes some
miscellaneous handling throughout the back end for LWA_32 and LWAX_32
that should have been part of the load-store patch.)
llvm-svn: 189677
This is the next big chunk of fast-isel code. The primary purpose is
to implement selection of loads and stores, but there is a lot of
drag-along to support this. The common code to analyze addresses for
both loads and stores is substantial. It's also necessary to add the
materialization code for global values.
Related to load-store processing is the code to fold loads into
integer extends, since otherwise we generate lots of redundant
instructions. We also need to add some overrides to some FastEmit
routines to ensure we don't assign GPR 0 to a virtual register when
this would change the meaning of an instruction.
I added handling selection of a few binary arithmetic instructions, to
enable committing some test cases I wrote a while back.
Finally, ap couple of miscellaneous changes:
* I cleaned up some poor style from a previous patch in
PPCISelLowering.cpp, pointed out by David Blaikie.
* I enlarged the Addr.Offset field to avoid sign problems with 32-bit
offsets.
llvm-svn: 189636
DICompositeType will have an identifier field at position 14. For now, the
field is set to null in DIBuilder.
For DICompositeTypes where the template argument field (the 13th field)
was optional, modify DIBuilder to make sure the template argument field is set.
Now DICompositeType has 15 fields.
Update DIBuilder to use NULL instead of "i32 0" for null value of a MDNode.
Update verifier to check that DICompositeType has 15 fields and the last
field is null or a MDString.
Update testing cases to include an extra field for DICompositeType.
The identifier field will be used by type uniquing so a front end can
genearte a DICompositeType with a unique identifer.
llvm-svn: 189282
Incremental improvement to fast-isel for PPC64. This allows us to
select on ret, sext, and zext. Filling in sext/zext improves some of
the existing logic in handling compare-immediates that needed extends.
A simplified return convention for fast-isel is also added to the
PPC64 calling conventions. All call/return processing for DAG
selection is handled with custom code, so there isn't an existing CC
to rely on here. The include of PPCGenCallingConv.inc causes compiler
warnings due to the 32-bit calling conventions that are not used, so
the dummy function "usePPC32CCs()" is added here to silence those.
Test cases for the return and extend logic are added.
llvm-svn: 189266
First chunk of actual fast-isel selection code. This handles direct
and indirect branches, as well as feeding compares for direct
branches. PPCFastISel::PPCEmitIntExt() is just roughed in and will be
expanded in a future patch. This also corrects a problem with
selection for constant pool entries in JIT mode or with small code
model.
llvm-svn: 189202
copysign/copysignf never become function calls (because the SDAG expansion code
does not lower to the corresponding function call, but rather directly
implements the associated logic), but copysignl almost always is lowered into a
call to the requested libm functon (and, thus, might clobber CTR).
llvm-svn: 188727
We had previously been asserting when faced with a FCOPYSIGN f64, ppcf128 node
because there was no way to expand the FCOPYSIGN node. Because ppcf128 is the
sum of two doubles, and the first double must have the larger magnitude, we
can take the sign from the first double. As a result, in addition to fixing the
crash, this is also an optimization.
llvm-svn: 188655
Modern PPC cores support a floating-point copysign instruction, and we can use
this to lower the FCOPYSIGN node (which is created from calls to the libm
copysign function). A couple of extra patterns are necessary because the
operand types of FCOPYSIGN need not agree.
llvm-svn: 188653
- Instead of setting the suffixes in a bunch of places, just set one master
list in the top-level config. We now only modify the suffix list in a few
suites that have one particular unique suffix (.ml, .mc, .yaml, .td, .py).
- Aside from removing the need for a bunch of lit.local.cfg files, this enables
4 tests that were inadvertently being skipped (one in
Transforms/BranchFolding, a .s file each in DebugInfo/AArch64 and
CodeGen/PowerPC, and one in CodeGen/SI which is now failing and has been
XFAILED).
- This commit also fixes a bunch of config files to use config.root instead of
older copy-pasted code.
llvm-svn: 188513
This is a follow-up to r187693, correcting that code to request the correct
register class. The previous version, with the wrong register class, was not
really correcting the constraints, but rather was removing them. Coincidentally,
this fixed the failing test case in r187693, but obviously created other
problems.
llvm-svn: 188407
Various tests had sprung up over the years which had --check-prefix=ABC on the
RUN line, but "CHECK-ABC:" later on. This happened to work before, but was
strictly incorrect. FileCheck is getting stricter soon though.
Patch by Ron Ofir.
llvm-svn: 188173
this records relocation entries in the mach-o object file
for PIC code generation.
tested on powerpc-darwin8, validated against darwin otool -rvV
llvm-svn: 188004
Making use of the recently-added ISD::FROUND, which allows for custom lowering
of round(), the PPC backend will now map frin to round(). Previously, we had
been using frin to lower nearbyint() (and rint() via some custom lowering to
handle the extra fenv flags requirements), but only in fast-math mode because
frin does not tie-to-even. Several users had complained about this behavior,
and this new mapping of frin to round is certainly more appropriate (and does
not require fast-math mode).
In effect, this reverts r178362 (and part of r178337, replacing the nearbyint
mapping with the round mapping).
llvm-svn: 187960
The PPC backend had been missing a pattern to generate mulli for 64-bit
multiples. We had been generating it only for 32-bit multiplies. Unfortunately,
generating li + mulld unnecessarily increases register pressure.
llvm-svn: 187807
Internally, the PowerPC backend names the 32-bit GPRs R[0-9]+, and names the
64-bit parent GPRs X[0-9]+. When matching inline assembly constraints with
explicit register names, on PPC64 when an i64 MVT has been requested, we need
to follow gcc's convention of using r[0-9]+ to refer to the 64-bit (parent)
registers.
At some point, we'll probably want to arrange things so that the generic code
in TargetLowering uses the AsmName fields declared in *RegisterInfo.td in order
to match these inline asm register constraints. If we do that, this change can
be reverted.
llvm-svn: 187693
Support for dynamic stack alignments in the PPC backend has been unfinished, in
part because it depends on dynamic stack realignment (which I only just
recently implemented fully). Now we can also support dynamic allocas with
higher than the default target stack alignment (16 bytes).
In order to round-up the requested size to the maximum requested alignment, we
need an additional register to hold the rounded-up size. We're already using one
scavenged register to hold the previous stack-pointer value (which needs to be
stored with the signal-safe stdux update), and so when we have dynamic allocas
and a large alignment, we allocate two emergency spill slots for the scavenger.
llvm-svn: 186562
First, this changes the base-pointer implementation to remove an unnecessary
complication (and one that is incompatible with how builtin SjLj is
implemented): instead of using r31 as the base pointer when it is not needed as
a frame pointer, now the base pointer will always be r30 when needed.
Second, we introduce another pseudo register, BP, which is used just like the FP
pseudo register to refer to the base register before we know for certain what
register it will be.
Third, we now save BP into the jmp_buf, and restore r30 from that slot in
longjmp. If the function that called setjmp did not use a base pointer, then
r30 will be overwritten by the setjmp-calling-function's restore code. FP
restoration (which is restored into r31) works the same way.
llvm-svn: 186545
Because the builtin longjmp implementation uses a CTR-based indirect jump, when
the control flow arrives at the builtin setjmp call, the CTR register has
necessarily been clobbered. Correspondingly, this adds CTR to the list of
implicit definitions of the builtin setjmp pseudo instruction.
We don't need to add CTR to the implicit definitions of builtin longjmp
because, even though it does clobber the CTR register, the control flow cannot
return to inside the loop unless there is also a builtin setjmp call.
llvm-svn: 186488
This builds on some frame-lowering code that has existed since 2005 (r24224)
but was disabled in 2008 (r48188) because it needed base pointer support to
function correctly. This implementation follows the strategy suggested by Dale
Johannesen in r48188 where the following comment was added:
This does not currently work, because the delta between old and new stack
pointers is added to offsets that reference incoming parameters after the
prolog is generated, and the code that does that doesn't handle a variable
delta. You don't want to do that anyway; a better approach is to reserve
another register that retains to the incoming stack pointer, and reference
parameters relative to that.
And now we do exactly that. If we don't need a frame pointer, then we use r31
as a base pointer. If we do need a frame pointer, then we use r30 as a base
pointer. The base pointer retains the value of the stack pointer before it was
decremented in the prologue. We then use the base pointer to resolve all
negative frame indicies. The basic scheme follows that for base pointers in the
X86 backend.
We use a base pointer when we need to dynamically realign the incoming stack
pointer. This currently applies only to static objects (dynamic allocas with
large alignments, and base-pointer support in SjLj lowering will come in future
commits).
llvm-svn: 186478
When truncating to a format with fewer mantissa bits, APFloat::convert
will perform a right shift of the mantissa by the difference of the
precision of the two formats. Usually, this will result in just the
mantissa bits needed for the target format.
One special situation is if the input number is denormal. In this case,
the right shift may discard significant bits. This is usually not a
problem, since truncating a denormal usually results in zero (underflow)
after normalization anyway, since the result format's exponent range is
usually smaller than the target format's.
However, there is one case where the latter property does not hold:
when truncating from ppc_fp128 to double. In particular, truncating
a ppc_fp128 whose first double of the pair is denormal should result
in just that first double, not zero. The current code however
performs an excessive right shift, resulting in lost result bits.
This is then caught in the APFloat::normalize call performed by
APFloat::convert and causes an assertion failure.
This patch checks for the scenario of truncating a denormal, and
attempts to (possibly partially) replace the initial mantissa
right shift by decrementing the exponent, if doing so will still
result in a valid *target format* exponent.
Index: test/CodeGen/PowerPC/pr16573.ll
===================================================================
--- test/CodeGen/PowerPC/pr16573.ll (revision 0)
+++ test/CodeGen/PowerPC/pr16573.ll (revision 0)
@@ -0,0 +1,11 @@
+; RUN: llc < %s | FileCheck %s
+
+target triple = "powerpc64-unknown-linux-gnu"
+
+define double @test() {
+ %1 = fptrunc ppc_fp128 0xM818F2887B9295809800000000032D000 to double
+ ret double %1
+}
+
+; CHECK: .quad -9111018957755033591
+
Index: lib/Support/APFloat.cpp
===================================================================
--- lib/Support/APFloat.cpp (revision 185817)
+++ lib/Support/APFloat.cpp (working copy)
@@ -1956,6 +1956,23 @@
X86SpecialNan = true;
}
+ // If this is a truncation of a denormal number, and the target semantics
+ // has larger exponent range than the source semantics (this can happen
+ // when truncating from PowerPC double-double to double format), the
+ // right shift could lose result mantissa bits. Adjust exponent instead
+ // of performing excessive shift.
+ if (shift < 0 && isFiniteNonZero()) {
+ int exponentChange = significandMSB() + 1 - fromSemantics.precision;
+ if (exponent + exponentChange < toSemantics.minExponent)
+ exponentChange = toSemantics.minExponent - exponent;
+ if (exponentChange < shift)
+ exponentChange = shift;
+ if (exponentChange < 0) {
+ shift -= exponentChange;
+ exponent += exponentChange;
+ }
+ }
+
// If this is a truncation, perform the shift before we narrow the storage.
if (shift < 0 && (isFiniteNonZero() || category==fcNaN))
lostFraction = shiftRight(significandParts(), oldPartCount, -shift);
llvm-svn: 186409
PPCInstrInfo::insertSelect and PPCInstrInfo::canInsertSelect were computing the
common subclass of the true and false inputs, and then selecting either the
32-bit or the 64-bit isel variant based on the result of calling
PPC::GPRCRegClass.hasSubClassEq(RC) and PPC::G8RCRegClass.hasSubClassEq(RC)
(where RC is the common subclass). Unfortunately, this is not quite right: if
we have something like this:
%vreg8<def> = SELECT_CC_I8 %vreg4<kill>, %vreg7<kill>, %vreg6<kill>, 76;
G8RC_and_G8RC_NOX0:%vreg8 CRRC:%vreg4 G8RC_NOX0:%vreg7,%vreg6
then the common subclass of G8RC_and_G8RC_NOX0 and G8RC_NOX0 is G8RC_NOX0, and
G8RC_NOX0 is not a subclass of G8RC (because it also contains the ZERO8
pseudo-register). As a result, we also need to check the common subclass
against GPRC_NOR0 and G8RC_NOX0 explicitly.
This had not been a problem for clients of insertSelect that called
canInsertSelect first (because it had a compensating mistake), but insertSelect
is also used by the PPC pseudo-instruction expander, and this error was causing
a problem in that context.
This problem was found by csmith.
llvm-svn: 186343
There is a comment at the top of DAGTypeLegalizer::PerformExpensiveChecks
which, in part, says:
// Note that these invariants may not hold momentarily when processing a node:
// the node being processed may be put in a map before being marked Processed.
Unfortunately, this assert would be valid only if the above-mentioned invariant
held unconditionally. This was causing llc to assert when, in fact,
everything was fine.
Thanks to Richard Sandiford for investigating this issue!
Fixes PR16562.
llvm-svn: 186338
This update was done with the following bash script:
find test/CodeGen -name "*.ll" | \
while read NAME; do
echo "$NAME"
if ! grep -q "^; *RUN: *llc.*debug" $NAME; then
TEMP=`mktemp -t temp`
cp $NAME $TEMP
sed -n "s/^define [^@]*@\([A-Za-z0-9_]*\)(.*$/\1/p" < $NAME | \
while read FUNC; do
sed -i '' "s/;\(.*\)\([A-Za-z0-9_-]*\):\( *\)$FUNC: *\$/;\1\2-LABEL:\3$FUNC:/g" $TEMP
done
sed -i '' "s/;\(.*\)-LABEL-LABEL:/;\1-LABEL:/" $TEMP
sed -i '' "s/;\(.*\)-NEXT-LABEL:/;\1-NEXT:/" $TEMP
sed -i '' "s/;\(.*\)-NOT-LABEL:/;\1-NOT:/" $TEMP
sed -i '' "s/;\(.*\)-DAG-LABEL:/;\1-DAG:/" $TEMP
mv $TEMP $NAME
fi
done
llvm-svn: 186280
This was done with the following sed invocation to catch label lines demarking function boundaries:
sed -i '' "s/^;\( *\)\([A-Z0-9_]*\):\( *\)test\([A-Za-z0-9_-]*\):\( *\)$/;\1\2-LABEL:\3test\4:\5/g" test/CodeGen/*/*.ll
which was written conservatively to avoid false positives rather than false negatives. I scanned through all the changes and everything looks correct.
llvm-svn: 186258
We had patterns to match v4i32 immAllZerosV -> V_SET0, but not patterns for
v8i16 (which occurs in the test case) or v16i8. The same was true for
V_SETALLONES (so I added the associated patterns for those as well).
Another bug found by llvm-stress.
llvm-svn: 186108
This fixes a bug (found by csmith) at -O0 where we attempt to create a RLWIMI
with an out-of-range operand. Most uses of the isRunOfOnes function are guarded
by a condition that the value is not zero. This was not true in two places, and
in both places a zero input would result in an out-of-rage MB value (= 32).
To fix this, isRunOfOnes returns false on a zero input (and I've remove one
now-redundant guard).
llvm-svn: 186101
When computing currently-live registers, the register scavenger excludes undef
uses. As a result, undef uses are ignored when computing the restore points of
registers spilled into the emergency slots. While the register scavenger
normally excludes from consideration, when scavenging, registers used by the
current instruction, we need to not exclude undef uses. Otherwise, we might end
up requiring more emergency spill slots than we have (in the case where the
undef use *is* the currently-spilled register).
Another bug found by llvm-stress.
llvm-svn: 186067
Because integer BUILD_VECTOR operands may have a larger type than the result's
vector element type, and all operands must have the same type, when widening a
BUILD_VECTOR node by adding UNDEFs, we cannot use the vector element type, but
rather must use the type of the existing operands.
Another bug found by llvm-stress.
llvm-svn: 185960
A more complete example of the bug in PR16556 was recently provided,
showing that the previous fix was not sufficient. The previous fix is
reverted herein.
The real problem is that ReplaceNodeResults() uses LowerFP_TO_INT as
custom lowering for FP_TO_SINT during type legalization, without
checking whether the input type is handled by that routine.
LowerFP_TO_INT requires the input to be f32 or f64, so we fail when
the input is ppcf128.
I'm leaving the test case from the initial fix (r185821) in place, and
adding the new test as another crash-only check.
llvm-svn: 185959
in-tree implementations of TargetLoweringBase::isFMAFasterThanMulAndAdd in
order to resolve the following issues with fmuladd (i.e. optional FMA)
intrinsics:
1. On X86(-64) targets, ISD::FMA nodes are formed when lowering fmuladd
intrinsics even if the subtarget does not support FMA instructions, leading
to laughably bad code generation in some situations.
2. On AArch64 targets, ISD::FMA nodes are formed for operations on fp128,
resulting in a call to a software fp128 FMA implementation.
3. On PowerPC targets, FMAs are not generated from fmuladd intrinsics on types
like v2f32, v8f32, v4f64, etc., even though they promote, split, scalarize,
etc. to types that support hardware FMAs.
The function has also been slightly renamed for consistency and to force a
merge/build conflict for any out-of-tree target implementing it. To resolve,
see comments and fixed in-tree examples.
llvm-svn: 185956
ScalarEvolution::getSignedRange uses ComputeNumSignBits from ValueTracking on
ashr instructions. ComputeNumSignBits can return zero, but this case was not
handled correctly by the code in getSignedRange which was calling:
APInt::getSignedMinValue(BitWidth).ashr(NS - 1)
with NS = 0, resulting in an assertion failure in APInt::ashr.
Now, we just return the conservative result (as with NS == 1).
Another bug found by llvm-stress.
llvm-svn: 185955
When folding sub x, x (and other similar constructs), where x is a vector, the
result is a vector of zeros. After type legalization, make sure that the input
zero elements have a legal type. This type may be larger than the result's
vector element type.
This was another bug found by llvm-stress.
llvm-svn: 185949
In the commit message to r185476 I wrote:
>The PowerPC-specific modifiers VK_PPC_TLSGD and VK_PPC_TLSLD
>correspond exactly to the generic modifiers VK_TLSGD and VK_TLSLD.
>This causes some confusion with the asm parser, since VK_PPC_TLSGD
>is output as @tlsgd, which is then read back in as VK_TLSGD.
>
>To avoid this confusion, this patch removes the PowerPC-specific
>modifiers and uses the generic modifiers throughout. (The only
>drawback is that the generic modifiers are printed in upper case
>while the usual convention on PowerPC is to use lower-case modifiers.
>But this is just a cosmetic issue.)
This was unfortunately incorrect, there is is fact another,
serious drawback to using the default VK_TLSLD/VK_TLSGD
variant kinds: using these causes ELFObjectWriter::RelocNeedsGOT
to return true, which in turn causes the ELFObjectWriter to emit
an undefined reference to _GLOBAL_OFFSET_TABLE_.
This is a problem on powerpc64, because it uses the TOC instead
of the GOT, and the linker does not provide _GLOBAL_OFFSET_TABLE_,
so the symbol remains undefined. This means shared libraries
using TLS built with the integrated assembler are currently
broken.
While the whole RelocNeedsGOT / _GLOBAL_OFFSET_TABLE_ situation
probably ought to be properly fixed at some point, for now I'm
simply reverting the r185476 commit. Now this in turn exposes
the breakage of handling @tlsgd/@tlsld in the asm parser that
this check-in was originally intended to fix.
To avoid this regression, I'm also adding a different fix for
this problem: while common code now parses @tlsgd as VK_TLSGD,
a special hack in the asm parser translates this code to the
platform-specific VK_PPC_TLSGD that the back-end now expects.
While this is not really pretty, it's self-contained and
shouldn't hurt anything else for now. One the underlying
problem is fixed, this hack can be reverted again.
llvm-svn: 185945
This fixes another bug found by llvm-stress!
If we happen to be doing an i64 load or store into a stack slot that has less
than a 4-byte alignment, then the frame-index elimination may need to use an
indexed load or store instruction (because the offset may not be a multiple of
4, a requirement of the STD/LD instructions). The extra register needed to hold
the offset comes from the register scavenger, and it is possible that the
scavenger will need to use an emergency spill slot. As a result, we need to
make sure that a spill slot is allocated when doing an i64 load/store into a
less-than-4-byte-aligned stack slot.
Because test cases for things like this tend to be fairly fragile, I've
concatenated a few small bugpoint-reduced test cases together to form the
regression test.
llvm-svn: 185907
A setting in MCAsmInfo defines the "assembler dialect" to use. This is used
by common code to choose between alternatives in a multi-alternative GNU
inline asm statement like the following:
__asm__ ("{sfe|subfe} %0,%1,%2" : "=r" (out) : "r" (in1), "r" (in2));
The meaning of these dialects is platform specific, and GCC defines those
for PowerPC to use dialect 0 for old-style (POWER) mnemonics and 1 for
new-style (PowerPC) mnemonics, like in the example above.
To be compatible with inline asm used with GCC, LLVM ought to do the same.
Specifically, this means we should always use assembler dialect 1 since
old-style mnemonics really aren't supported on any current platform.
However, the current LLVM back-end uses:
AssemblerDialect = 1; // New-Style mnemonics.
in PPCMCAsmInfoDarwin, and
AssemblerDialect = 0; // Old-Style mnemonics.
in PPCLinuxMCAsmInfo.
The Linux setting really isn't correct, we should be using new-style
mnemonics everywhere. This is changed by this commit.
Unfortunately, the setting of this variable is overloaded in the back-end
to decide whether or not we are on a Darwin target. This is done in
PPCInstPrinter (the "SyntaxVariant" is initialized from the MCAsmInfo
AssemblerDialect setting), and also in PPCMCExpr. Setting AssemblerDialect
to 1 for both Darwin and Linux no longer allows us to make this distinction.
Instead, this patch uses the MCSubtargetInfo passed to createPPCMCInstPrinter
to distinguish Darwin targets, and ignores the SyntaxVariant parameter.
As to PPCMCExpr, this patch adds an explicit isDarwin argument that needs
to be passed in by the caller when creating a target MCExpr. (To do so
this patch implicitly also reverts commit 184441.)
llvm-svn: 185858
Another bug found by llvm-stress! This fixes hitting
llvm_unreachable("Invalid integer vector compare condition");
at the end of getVCmpInst in PPCISelDAGToDAG.
llvm-svn: 185855
PPCTargetLowering::LowerFP_TO_INT() expects its source operand to be
either an f32 or f64, but this is not checked. A long double
(ppcf128) operand will normally be custom-lowered to a conversion to
f64 in this context. However, this isn't the case for an UNDEF node.
This patch recognizes a ppcf128 as a legal source operand for
FP_TO_INT only if it's an undef, in which case it creates an undef of
the target type.
At some point we might want to do a wholesale custom lowering of
ISD::UNDEF when the type is ppcf128, but it's not really clear that's
a great idea, and probably more work than it's worth for a situation
that only arises in the case of a programming error. At this point I
think simple is best.
The test case comes from PR16556, and is a crash-test only.
llvm-svn: 185821
This fixes a bug (found by llvm-stress) in
DAGTypeLegalizer::PromoteIntRes_BUILD_VECTOR where it assumed that the result
type would always be larger than the original operands. This is not always
true, however, with boolean vectors. For example, promoting a node of type v8i1
(where the operands will be of type i32, the type to which i1 is promoted) will
yield a node with a result vector element type of i16 (and operands of type
i32). As a result, we cannot blindly assume that we can ANY_EXTEND the operands
to the result type.
llvm-svn: 185794
Just as with mfocrf, it is also preferable to use mtocrf instead of
mtcrf when only a single CR register is to be written.
Current code however always emits mtcrf. This probably does not matter
when using an external assembler, since the GNU assembler will in fact
automatically replace mtcrf with mtocrf when possible. It does create
inefficient code with the integrated assembler, however.
To fix this, this patch adds MTOCRF/MTOCRF8 instruction patterns and
uses those instead of MTCRF/MTCRF8 everything. Just as done in the
MFOCRF patch committed as 185556, these patterns will be converted
back to MTCRF if MTOCRF is not available on the machine.
As a side effect, this allows to modify the MTCRF pattern to accept
the full range of mask operands for the benefit of the asm parser.
llvm-svn: 185561
The PowerPC-specific modifiers VK_PPC_TLSGD and VK_PPC_TLSLD
correspond exactly to the generic modifiers VK_TLSGD and VK_TLSLD.
This causes some confusion with the asm parser, since VK_PPC_TLSGD
is output as @tlsgd, which is then read back in as VK_TLSGD.
To avoid this confusion, this patch removes the PowerPC-specific
modifiers and uses the generic modifiers throughout. (The only
drawback is that the generic modifiers are printed in upper case
while the usual convention on PowerPC is to use lower-case modifiers.
But this is just a cosmetic issue.)
llvm-svn: 185476
There are a couple of (small) related changes here:
1. The printed name of the VRSAVE register has been changed from VRsave to
vrsave in order to match the name accepted by GNU binutils.
2. Support for parsing vrsave has been added to the asm parser (it seems that
there was no test case specifically covering this code, so I've added one).
3. The list of Altivec registers, which was common to all calling conventions,
has been separated out. This allows us to define the base CSR lists, and then
lists for each ABI with Altivec included. This allows SjLj, for example, to
work correctly on non-Altivec targets without using unnatural definitions of
the NoRegs CSR list.
4. VRSAVE is now always reserved on non-Darwin targets and all Altivec
registers are reserved when Altivec is disabled.
With these changes, it is now possible to compile a function containing
__builtin_unwind_init() on Linux/PPC64 with debugging information. This did not
work previously because GNU binutils assumes that all .cfi_offset offsets will
be 8-byte aligned on PPC64 (and errors out if you provide a non-8-byte-aligned
offset). This is not true for the vrsave register, however, because this
register is used only on Darwin, GCC does not bother printing a .cfi_offset
entry for it (even though there is a slot in the stack frame for it as
specified by the ABI). This change allows us to do the same: we will also not
print .cfi_offset directives for vrsave.
llvm-svn: 185409
When phis get lowered, destination copies are inserted using an iterator that is
determined once for all phis in the block, which BuildMI interprets as a request
to insert an instruction directly before the iterator. In the case of a cyclic
phi, source copies may also be inserted directly before this iterator, which can
cause source copies to be inserted before destination copies. The fix is to keep
an iterator to the last phi and then advance it while lowering each phi in order
to insert destination copies directly after the phis.
llvm-svn: 185363
Although you can't generate this from C on PPC64, if you have a loop using a
64-bit counter on PPC32 then you can't form a CTR-based loop for it. This had
been cauing the PPCCTRLoops pass to assert.
Thanks to Joerg Sonnenberger for providing a test case!
llvm-svn: 185361
This fixes PR16418, which reports that a function calling
__builtin_unwind_init() asserts. The cause is that this generates a
spill/restore for VRSAVE, and we support that only on Darwin (because VRSAVE is
only really used on Darwin).
The test case checks only that we don't crash. We can add correctness checks
once someone verifies what behavior the function is supposed to have.
llvm-svn: 185235
On OpenBSD, the stack-smash protection transform uses "__guard_local"
and "__stack_smash_handler" instead of "__stack_chk_guard" and
"__stack_chk_fail". However, CodeGen/PowerPC/stack-protector.ll
doesn't specify a target OS, so on OpenBSD it fails.
Add -mtriple=ppc32-unknown-linux to make the test host-OS agnostic. While
there, convert to FileCheck.
Patch by Matthew Dempsky.
llvm-svn: 185206
Under certain (evidently rare) circumstances, this code used to convert OR(a,
AND(x, y)) into OR(a, x). This was incorrect.
While there, I've added a comment to the code immediately above.
llvm-svn: 185201
This is a preliminary patch for fast instruction selection on
PowerPC. Code generation can differ between DAG isel and fast isel.
Existing tests that specify -O0 were written to expect DAG isel. Make
this explicit by adding -fast-isel=false to the tests.
In some cases specifying -fast-isel=false produces different code even
when there isn't a fast instruction selector specified. This is
because TM.Options.EnableFastISel = 1 at -O0 whether or not a FastISel
object exists. Thus disabling fast isel can actually produce less
conservative code. Because of this, some of the expected code
generation in the -O0 tests needs to be adjusted.
In particular, handling of function arguments is less conservative
with -fast-isel=false (see isOnlyUsedInEntryBlock() in
SelectionDAGBuilder.cpp). This results in fewer stack accesses and,
in some cases, reduced stack size as uselessly loaded values are no
longer stored back to spill locations in the stack.
No functional change with this patch; test case adjustments only.
llvm-svn: 183939
On PPC32, [su]div,rem on i64 types are transformed into runtime library
function calls. As a result, they are not allowed in counter-based loops (the
counter-loops verification pass caught this error; this change fixes PR16169).
llvm-svn: 183581
For COFF and MachO, sections semantically have relocations that apply to them.
That is not the case on ELF.
In relocatable objects (.o), a section with relocations in ELF has offsets to
another section where the relocations should be applied.
In dynamic objects and executables, relocations don't have an offset, they have
a virtual address. The section sh_info may or may not point to another section,
but that is not actually used for resolving the relocations.
This patch exposes that in the ObjectFile API. It has the following advantages:
* Most (all?) clients can handle this more efficiently. They will normally walk
all relocations, so doing an effort to iterate in a particular order doesn't
save time.
* llvm-readobj now prints relocations in the same way the native readelf does.
* probably most important, relocations that don't point to any section are now
visible. This is the case of relocations in the rela.dyn section. See the
updated relocation-executable.test for example.
llvm-svn: 182908
When expanding unaligned Altivec loads, we use the decremented offset trick to
prevent page faults. Unfortunately, if we have a sequence of consecutive
unaligned loads, this leads to suboptimal code generation because the 'extra'
load from the first unaligned load can be combined with the base load from the
second (but only if the decremented offset trick is not used for the first).
Search up and down the chain, through loads and token factors, looking for
consecutive loads, and if one is found, don't use the offset reduction trick.
These duplicate loads are later combined to yield the desired sequence (in the
future, we might want a more-powerful chain search, but that will require some
changes to allow the combiner routines to access the AA object).
This should complete the initial implementation of the optimized unaligned
Altivec load expansion. There is some refactoring that should be done, but
that will happen when the unaligned store expansion is added.
llvm-svn: 182719
The lvsl permutation control instruction is a function only of the alignment of
the pointer operand (relative to the 16-byte natural alignment of Altivec
vectors). As a result, multiple lvsl intrinsics where the operands differ by a
multiple of 16 can be combined.
llvm-svn: 182708
Altivec only directly supports aligned loads, but the loads have a strange
property: If given an unaligned address, they truncate the address to the next
lower aligned address, and load from there. This property, along with an extra
load and some special-purpose permutation-control instructions that generate
the appropriate permutations from the original unaligned address, allow
efficient lowering of aligned loads. This code uses the trick explained in the
Apple Velocity Engine optimization overview document to prevent the needed
extra load from possibly causing a page fault if the original address happens
to be aligned.
As noted in the FIXMEs, there are several additional optimizations that can be
performed to reduce the cost of these loads even more. These will be
implemented in future commits.
llvm-svn: 182691
We don't need to reject all inline asm as using the counter register (most does
not). Only those that explicitly clobber the counter register need to prevent
the transformation.
llvm-svn: 182191
Some IR-level instructions (such as FP <-> i64 conversions) are not chained
w.r.t. the mtctr intrinsic and yet may become function calls that clobber the
counter register. At the selection-DAG level, these might be reordered with the
mtctr intrinsic causing miscompiles. To avoid this situation, if an existing
preheader has instructions that might use the counter register, create a new
preheader for the mtctr intrinsic. This extra block will be remerged with the
old preheader at the MI level, but will prevent unwanted reordering at the
selection-DAG level.
llvm-svn: 182045
This is the second part of the change to always return "true"
offset values from getPreIndexedAddressParts, tackling the
case of "memrix" type operands.
This is about instructions like LD/STD that only have a 14-bit
field to encode immediate offsets, which are implicitly extended
by two zero bits by the machine, so that in effect we can access
16-bit offsets as long as they are a multiple of 4.
The PowerPC back end currently handles such instructions by
carrying the 14-bit value (as it will get encoded into the
actual machine instructions) in the machine operand fields
for such instructions. This means that those values are
in fact not the true offset, but rather the offset divided
by 4 (and then truncated to an unsigned 14-bit value).
Like in the case fixed in r182012, this makes common code
operations on such offset values not work as expected.
Furthermore, there doesn't really appear to be any strong
reason why we should encode machine operands this way.
This patch therefore changes the encoding of "memrix" type
machine operands to simply contain the "true" offset value
as a signed immediate value, while enforcing the rules that
it must fit in a 16-bit signed value and must also be a
multiple of 4.
This change must be made simultaneously in all places that
access machine operands of this type. However, just about
all those changes make the code simpler; in many cases we
can now just share the same code for memri and memrix
operands.
llvm-svn: 182032
While testing some experimental code to add vector-scalar registers to
PowerPC, I noticed that a couple of independent instructions were
flipped by the scheduler. The new CHECK-DAG support is perfect for
avoiding this problem.
llvm-svn: 182020
DAGCombiner::CombineToPreIndexedLoadStore calls a target routine to
decompose a memory address into a base/offset pair. It expects the
offset (if constant) to be the true displacement value in order to
perform optional additional optimizations; in particular, to convert
other uses of the original pointer into uses of the new base pointer
after pre-increment.
The PowerPC implementation of getPreIndexedAddressParts, however,
simply calls SelectAddressRegImm, which returns a TargetConstant.
This value is appropriate for encoding into the instruction, but
it is not always usable as true displacement value:
- Its type is always MVT::i32, even on 64-bit, where addresses
ought to be i64 ... this causes the optimization to simply
always fail on 64-bit due to this line in DAGCombiner:
// FIXME: In some cases, we can be smarter about this.
if (Op1.getValueType() != Offset.getValueType()) {
- Its value is truncated to an unsigned 16-bit value if negative.
This causes the above opimization to generate wrong code.
This patch fixes both problems by simply returning the true
displacement value (in its original type). This doesn't
affect any other user of the displacement.
llvm-svn: 182012
Without this change nothing was covering this addFrameMove:
// For 64-bit SVR4 when we have spilled CRs, the spill location
// is SP+8, not a frame-relative slot.
if (Subtarget.isSVR4ABI()
&& Subtarget.isPPC64()
&& (PPC::CR2 <= Reg && Reg <= PPC::CR4)) {
MachineLocation CSDst(PPC::X1, 8);
MachineLocation CSSrc(PPC::CR2);
MMI.addFrameMove(Label, CSDst, CSSrc);
continue;
}
llvm-svn: 181976
The old PPCCTRLoops pass, like the Hexagon pass version from which it was
derived, could only handle some simple loops in canonical form. We cannot
directly adapt the new Hexagon hardware loops pass, however, because the
Hexagon pass contains a fundamental assumption that non-constant-trip-count
loops will contain a guard, and this is not always true (the result being that
incorrect negative counts can be generated). With this commit, we replace the
pass with a late IR-level pass which makes use of SE to calculate the
backedge-taken counts and safely generate the loop-count expressions (including
any necessary max() parts). This IR level pass inserts custom intrinsics that
are lowered into the desired decrement-and-branch instructions.
The most fragile part of this new implementation is that interfering uses of
the counter register must be detected on the IR level (and, on PPC, this also
includes any indirect branches in addition to function calls). Also, to make
all of this work, we need a variant of the mtctr instruction that is marked
as having side effects. Without this, machine-code level CSE, DCE, etc.
illegally transform the resulting code. Hopefully, this can be improved
in the future.
This new pass is smaller than the original (and much smaller than the new
Hexagon hardware loops pass), and can handle many additional cases correctly.
In addition, the preheader-creation code has been copied from LoopSimplify, and
after we decide on where it belongs, this code will be refactored so that it
can be explicitly shared (making this implementation even smaller).
The new test-case files ctrloop-{le,lt,ne}.ll have been adapted from tests for
the new Hexagon pass. There are a few classes of loops that this pass does not
transform (noted by FIXMEs in the files), but these deficiencies can be
addressed within the SE infrastructure (thus helping many other passes as well).
llvm-svn: 181927
The changes to CR spill handling missed a case for 32-bit PowerPC.
The code in PPCFrameLowering::processFunctionBeforeFrameFinalized()
checks whether CR spill has occurred using a flag in the function
info. This flag is only set by storeRegToStackSlot and
loadRegFromStackSlot. spillCalleeSavedRegisters does not call
storeRegToStackSlot, but instead produces MI directly. Thus we don't
see the CR is spilled when assigning frame offsets, and the CR spill
ends up colliding with some other location (generally the FP slot).
This patch sets the flag in spillCalleeSavedRegisters for PPC32 so
that the CR spill is properly detected and gets its own slot in the
stack frame.
llvm-svn: 181800
This fixes warning messages observed in the oggenc application test in
projects/test-suite. Special handling is needed for the 64-bit
PowerPC SVR4 ABI when a constant is initialized with a pointer to a
function in a shared library. Because a function address is
implemented as the address of a function descriptor, the use of copy
relocations can lead to problems with initialization. GNU ld
therefore replaces copy relocations with dynamic relocations to be
resolved by the dynamic linker. This means the constant cannot reside
in the read-only data section, but instead belongs in .data.rel.ro,
which is designed for constants containing dynamic relocations.
The implementation creates a class PPC64LinuxTargetObjectFile
inheriting from TargetLoweringObjectFileELF, which behaves like its
parent except to place constants of this sort into .data.rel.ro.
The test case is reduced from the oggenc application.
llvm-svn: 181723
The floating-point record forms on PPC don't set the condition register bits
based on a comparison with zero (like the integer record forms do), but rather
based on the exception status bits.
llvm-svn: 181423
First, taking advantage of the fact that the virtual base registers are allocated in order of the local frame offsets, remove the quadratic register-searching behavior. Because of the ordering, we only need to check the last virtual base register created.
Second, store the frame index in the FrameRef structure, and get the frame index and the local offset from this structure at the top of the loop iteration. This allows us to de-nest the loops in insertFrameReferenceRegisters (and I think makes the code cleaner). I also moved the needsFrameBaseReg check into the first loop over instructions so that we don't bother pushing FrameRefs for instructions that don't want a virtual base register anyway.
Lastly, and this is the only functionality change, avoid the creation of single-use virtual base registers. These are currently not useful because, in general, they end up replacing what would be one r+r instruction with an add and a r+i instruction. Committing this removes the XFAIL in CodeGen/PowerPC/2007-09-07-LoadStoreIdxForms.ll
Jim has okayed this off-list.
llvm-svn: 180799
When matching a compare with a subtract where the arguments of the compare are
swapped w.r.t. the arguments of the subtract, we need to negate the predicates
(or CR bit indices) of the users. This, however, is not the same as inverting
the predicate (negating LT -> GT, but inverting LT -> GE, for example). The ARM
backend seems to do this correctly, but when I adapted the code for the PPC
backend, I introduced an error in this logic.
Comparison optimization is now enabled again by default.
llvm-svn: 179899
Many PPC instructions have a so-called 'record form' which stores to a specific
condition register the result of comparing the result of the instruction with
zero (always as a signed comparison). For integer operations on PPC64, this is
always a 64-bit comparison.
This implementation is derived from the implementation in the ARM backend;
there are some differences because PPC condition registers are allocatable
virtual registers (although the record forms always use a specific one), and we
look for a matching subtraction instruction after the compare (but before the
first use) in addition to before it.
llvm-svn: 179802
This fixes an ABI bug for non-Darwin PPC64. For the callee-saved condition
registers, the spill location is specified relative to the stack pointer (SP +
8). However, this is not relative to the SP after the new stack frame is
established, but instead relative to the caller's stack pointer (it is stored
into the linkage area of the parent's stack frame).
So, like with the link register, we don't directly spill the CRs with other
callee-saved registers, but just mark them to be spilled during prologue
generation.
In practice, this reverts r179457 for PPC64 (but leaves it in place for PPC32).
llvm-svn: 179500
For functions that need to spill CRs, and have dynamic stack allocations, the
value of the SP during the restore is not what it was during the save, and so
we need to use the FP in these cases (as for all of the other spills and
restores, but the CR restore has a special code path because its reserved slot,
like the link register, is specified directly relative to the adjusted SP).
llvm-svn: 179457
In the simple and triangle if-conversion cases, when CopyAndPredicateBlock is
used because the to-be-predicated block has other predecessors, we need to
explicitly remove the old copied block from the successors list. Normally if
conversion relies on TII->AnalyzeBranch combined with BB->CorrectExtraCFGEdges
to cleanup the successors list, but if the predicated block contained an
un-analyzable branch (such as a now-predicated return), then this will fail.
These extra successors were causing a problem on PPC because it was causing
later passes (such as PPCEarlyReturm) to leave dead return-only basic blocks in
the code.
llvm-svn: 179227
This enables us to form predicated branches (which are the same conditional
branches we had before) and also a larger set of predicated returns (including
instructions like bdnzlr which is a conditional return and loop-counter
decrement all in one).
At the moment, if conversion does not capture all possible opportunities. A
simple example is provided in early-ret2.ll, where if conversion forms one
predicated return, and then the PPCEarlyReturn pass picks up the other one. So,
at least for now, we'll keep both mechanisms.
llvm-svn: 179134
On PowerPC, non-vector loads and stores have r+i forms; however, in functions
with large stack frames these were not being used to access slots far from the
stack pointer because such slots were out of range for the signed 16-bit
immediate offset field. This increases register pressure because we need a
separate register for each offset (when the r+r form is used). By enabling
virtual base registers, we can deal with large stack frames without unduly
increasing register pressure.
llvm-svn: 179105
PowerPC has a conditional branch to the link register (return) instruction: BCLR.
This should be used any time when we'd otherwise have a conditional branch to a
return. This adds a small pass, PPCEarlyReturn, which runs just prior to the
branch selection pass (and, importantly, after block placement) to generate
these conditional returns when possible. It will also eliminate unconditional
branches to returns (these happen rarely; most of the time these have already
been tail duplicated by the time PPCEarlyReturn is invoked). This is a nice
optimization for small functions that do not maintain a stack frame.
llvm-svn: 179026
First, we should not cheat: fsel-based lowering of select_cc is a
finite-math-only optimization (the ISA manual, section F.3 of v2.06, makes
this clear, as does a note in our own README).
This also adds fsel-based lowering of EQ and NE condition codes. As it turned
out, fsel generation was covered by a grand total of zero regression test
cases. I've added some test cases to cover the existing behavior (which is now
finite-math only), as well as the new EQ cases.
llvm-svn: 179000
There are certain PPC instructions into which we can fold a zero immediate
operand. We can detect such cases by looking at the register class required
by the using operand (so long as it is not otherwise constrained).
llvm-svn: 178961
On cores for which we know the misprediction penalty, and we have
the isel instruction, we can profitably perform early if conversion.
This enables us to replace some small branch sequences with selects
and avoid the potential stalls from mispredicting the branches.
Enabling this feature required implementing canInsertSelect and
insertSelect in PPCInstrInfo; isel code in PPCISelLowering was
refactored to use these functions as well.
llvm-svn: 178926
The DAGCombine logic that recognized a/sqrt(b) and transformed it into
a multiplication by the reciprocal sqrt did not handle cases where the
sqrt and the division were separated by an fpext or fptrunc.
llvm-svn: 178801
For this we need to use a libcall. Previously LLVM didn't implement
libcall support for frem, so I've added it in the usual
straightforward manner. A test case from the bug report is included.
llvm-svn: 178639
When unsafe FP math operations are enabled, we can use the fre[s] and
frsqrte[s] instructions, which generate reciprocal (sqrt) estimates, together
with some Newton iteration, in order to quickly generate floating-point
division and sqrt results. All of these instructions are separately optional,
and so each has its own feature flag (except for the Altivec instructions,
which are covered under the existing Altivec flag). Doing this is not only
faster than using the IEEE-compliant fdiv/fsqrt instructions, but allows these
computations to be pipelined with other computations in order to hide their
overall latency.
I've also added a couple of missing fnmsub patterns which turned out to be
missing (but are necessary for good code generation of the Newton iterations).
Altivec needs a similar fix, but that will probably be more complicated because
fneg is expanded for Altivec's v4f32.
llvm-svn: 178617
When doing a partword atomic operation, a lwarx was being paired with
a stdcx. instead of a stwcx. when compiling for a 64-bit target. The
target has nothing to do with it in this case; we always need a stwcx.
Thanks to Kai Nacke for reporting the problem.
llvm-svn: 178559
The P7 and A2 have additional floating-point conversion instructions which
allow a direct two-instruction sequence (plus load/store) to convert from all
combinations (signed/unsigned i32/i64) <--> (float/double) (on previous cores,
only some combinations were directly available).
llvm-svn: 178480
The popcntw instruction is available whenever the popcntd instruction is
available, and performs a separate popcnt on the lower and upper 32-bits.
Ignoring the high-order count, this can be used for the 32-bit input case
(saving on the explicit zero extension otherwise required to use popcntd).
llvm-svn: 178470
This instruction is available on modern PPC64 CPUs, and is now used
to improve the SINT_TO_FP lowering (by eliminating the need for the
separate sign extension instruction and decreasing the amount of
needed stack space).
llvm-svn: 178446
The existing SINT_TO_FP code for i32 -> float/double conversion was disabled
because it relied on broken EXTSW_32/STD_32 instruction definitions. The
original intent had been to enable these 64-bit instructions to be used on CPUs
that support them even in 32-bit mode. Unfortunately, this form of lying to
the infrastructure was buggy (as explained in the FIXME comment) and had
therefore been disabled.
This re-enables this functionality, using regular DAG nodes, but only when
compiling in 64-bit mode. The old STD_32/EXTSW_32 definitions (which were dead)
are removed.
llvm-svn: 178438
Like nearbyint, rint can be implemented on PPC using the frin instruction. The
complication comes from the fact that rint needs to set the FE_INEXACT flag
when the result does not equal the input value (and frin does not do that). As
a result, we use a custom inserter which, after the rounding, compares the
rounded value with the original, and if they differ, explicitly sets the XX bit
in the FPSCR register (which corresponds to FE_INEXACT).
Once LLVM has better modeling of the floating-point environment we should be
able to (often) eliminate this extra complexity.
llvm-svn: 178362
These instructions are available on the P5x (and later) and on the A2. They
implement the standard floating-point rounding operations (floor, trunc, etc.).
One caveat: frin (round to nearest) does not implement "ties to even", and so
is only enabled in fast-math mode.
llvm-svn: 178337
Compiling in 32-bit mode on a P7 would assert after 64-bit DAG combines were
added for bswap with load/store. This is because these combines are really only
valid in 64-bit mode, regardless of the CPU (and this was not being checked).
llvm-svn: 178286
These are 64-bit load/store with byte-swap, and available on the P7 and the A2.
Like the similar instructions for 16- and 32-bit words, these are matched in the
target DAG-combine phase against load/store-bswap pairs.
llvm-svn: 178276
PPC ISA 2.06 (P7, A2, etc.) has a popcntd instruction. Add this instruction and
tell TTI about it so that popcount-loop recognition will know about it.
llvm-svn: 178233
There were a few places where kill flags were not being set correctly, and
where 32-bit instruction variants were being used with 64-bit registers. After
r178180, this code was being triggered causing llc to assert.
llvm-svn: 178220
Some implementation detail in the forgotten past required the link
register to be placed in the GPRC and G8RC register classes. This is
just wrong on the face of it, and causes several extra intersection
register classes to be generated. I found this was having evil
effects on instruction scheduling, by causing the wrong register class
to be consulted for register pressure decisions.
No code generation changes are expected, other than some minor changes
in instruction order. Seven tests in the test bucket required minor
tweaks to adjust to the new normal.
llvm-svn: 178114
As Bill Schmidt pointed out to me, only on Darwin do we need to spill/restore
VRSAVE in the SjLj code. For non-Darwin, don't spill/restore VRSAVE (and I've
added some asserts to make sure that we're not).
As it turns out, we're not currently handling the Darwin case correctly (I've
added a FIXME in the test case). I've tried adding various implied register
definitions/uses to force the spill without success, so I'll need to address
this later.
llvm-svn: 178096
Now that the register scavenger can support multiple spill slots, and PEI can
use virtual-register-based scavenging for multiple simultaneous registers, we
can use a virtual register for the transfer register in the CR spilling code.
This should eliminate the last place (outside of the prologue/epilogue) where
we depend on the unconditional availability of the r0 register. We will soon be
able to allocate it (in a somewhat restricted sense) as a GPR.
llvm-svn: 178060
Thanks to Jakob for isolating the underlying problem from the
test case in r177423. The original commit had introduced
asymmetric copy operations, but these turned out to be a work-around
to the real problem (the use of == instead of hasSubClassEq in PPCCTRLoops).
llvm-svn: 177679
This implements SJLJ lowering on PPC, making the Clang functions
__builtin_{setjmp/longjmp} functional on PPC platforms. The implementation
strategy is similar to that on X86, with the exception that a branch-and-link
variant is used to get the right jump address. Credit goes to Bill Schmidt for
suggesting the use of the unconditional bcl form (instead of the regular bl
instruction) to limit return-address-cache pollution.
Benchmarking the speed at -O3 of:
static jmp_buf env_sigill;
void foo() {
__builtin_longjmp(env_sigill,1);
}
main() {
...
for (int i = 0; i < c; ++i) {
if (__builtin_setjmp(env_sigill)) {
goto done;
} else {
foo();
}
done:;
}
...
}
vs. the same code using the libc setjmp/longjmp functions on a P7 shows that
this builtin implementation is ~4x faster with Altivec enabled and ~7.25x
faster with Altivec disabled. This comparison is somewhat unfair because the
libc version must also save/restore the VSX registers which we don't yet
support.
llvm-svn: 177666
Although there is only one Altivec VRSAVE register, it is a member of
a register class, and we need the ability to spill it. Because this
register is normally callee-preserved and handled by special code this
has never before been necessary. However, this capability will be required by
a forthcoming commit adding SjLj support.
llvm-svn: 177654
The old code used to lower FRAMEADDR tried to replicate the logic in the real
frame-lowering code that determines whether or not the frame pointer (r31) will
be used. When it seemed as through the frame pointer would not be used, the
stack pointer (r1) was used instead. Unfortunately, because the stack size is
not yet known, this does not work. Instead, this change introduces new
always-reserved pseudo-registers (FP and FP8) that are replaced during prologue
insertion with the real frame-pointer register (either r1 or r31).
It is important that this intrinsic always return a valid frame address because
it is used by Clang to store the frame address as part of code generation for
__builtin_setjmp.
llvm-svn: 177653
Currently, pre-increment store patterns are written to use two separate
operands to represent address base and displacement:
stwu $rS, $ptroff($ptrreg)
This causes problems when implementing the assembler parser, so this
commit changes the patterns to use standard (complex) memory operands
like in all other memory access instruction patterns:
stwu $rS, $dst
To still match those instructions against the appropriate pre_store
SelectionDAG nodes, the patch uses the new feature that allows a Pat
to match multiple DAG operands against a single (complex) instruction
operand.
Approved by Hal Finkel.
llvm-svn: 177429
Currently the PPC r0 register is unconditionally reserved. There are two reasons
for this:
1. r0 is treated specially (as the constant 0) by certain instructions, and so
cannot be used with those instructions as a regular register.
2. r0 is used as a temporary register in the CR-register spilling process
(where, under some circumstances, we require two GPRs).
This change addresses the first reason by introducing a restricted register
class (without r0) for use by those instructions that treat r0 specially. These
register classes have a new pseudo-register, ZERO, which represents the r0-as-0
use. This has the side benefit of making the existing target code simpler (and
easier to understand), and will make it clear to the register allocator that
uses of r0 as 0 don't conflict will real uses of the r0 register.
Once the CR spilling code is improved, we'll be able to allocate r0.
Adding these extra register classes, for some reason unclear to me, causes
requests to the target to copy 32-bit registers to 64-bit registers. The
resulting code seems correct (and causes no test-suite failures), and the new
test case covers this new kind of asymmetric copy.
As r0 is still reserved, no functionality change intended.
llvm-svn: 177423
Remove an accidentally-added instruction definition and add a comment in the
test case. This is in response to a post-commit review by Bill Schmidt.
No functionality change intended.
llvm-svn: 177404
PPC64 supports unaligned loads and stores of 64-bit values, but
in order to use the r+i forms, the offset must be a multiple of 4.
Unfortunately, this cannot always be determined by examining the
immediate itself because it might be available only via a TOC entry.
In order to get around this issue, we additionally predicate the
selection of the r+i form on the alignment of the load or store
(forcing it to be at least 4 in order to select the r+i form).
llvm-svn: 177338
Hal Finkel recently added code to allow unaligned memory references
for PowerPC. Two tests were temporarily modified with
-disable-ppc-unaligned to keep them from failing. This patch adjusts
the expected code generation for the unaligned references.
llvm-svn: 177328
Apparently my final cleanup to use a relevant suffix for these tests before
committing r176831 caused them to stop running since lit wasn't configured to
run tests with that suffix in those directories (why don't we just have a
global suffix list?). So, add the suffix to the relevant directories & fix the
test that has bitrotted over the last week due to my debug info schema changes.
llvm-svn: 177315
This commit fixes an assert that would occur on loops with large constant counts
(like looping for ((uint32_t) -1) iterations on PPC64). The existing code did
not handle counts that it computed to be negative (asserting instead), but
these can be created with valid inputs.
This bug was discovered by bugpoint while I was attempting to isolate a
completely different problem.
Also, in writing test cases for the negative-count problem, I discovered that
the ori/lsi handling was broken (there was a typo which caused the logic that
was supposed to detect these pairs and extract the iteration count to always
fail). This has now also been corrected (and is covered by one of the new test
cases).
llvm-svn: 177295
Because the initial-value constants had not been added to the list
of instructions considered for DCE the resulting code had redundant
constant-materialization instructions.
llvm-svn: 177294
This change cleans up two issues with Altivec register spilling:
1. The spilling code was inefficient (using two instructions, and add and a
load, when just one would do)
2. The code assumed that r0 would always be available (true for now, but this
will change)
The new code handles VR spilling just like GPR spills but forced into r+r mode.
As a result, when any VR spills are present, we must now always allocate the
register-scavenger spill slot.
llvm-svn: 177231
Unaligned access is supported on PPC for non-vector types, and is generally
more efficient than manually expanding the loads and stores.
A few of the existing test cases were using expanded unaligned loads and stores
to test other features (like load/store with update), and for these test cases,
unaligned access remains disabled.
llvm-svn: 177160
In preparation for the addition of other SIMD ISA extensions (such as QPX) we
need to make sure that all Altivec patterns are properly predicated on having
Altivec support.
No functionality change intended (one test case needed to be updated b/c it
assumed that Altivec intrinsics would be supported without enabling Altivec
support).
llvm-svn: 177152
For spills into a large stack frame, the FI-elimination code uses the register
scavenger to obtain a free GPR for use with an r+r-addressed load or store.
When there are no available GPRs, the scavenger gets one by using its spill
slot. Previously, we were not always allocating that spill slot and the RS
would assert when the spill slot was needed.
I don't currently have a small test that triggered the assert, but I've
created a small regression test that verifies that the spill slot is now
added when the stack frame is sufficiently large.
llvm-svn: 177140
We used to add a spill slot for the register scavenger whenever the function
has a frame pointer. This is unnecessarily conservative: We may need the spill
slot for dynamic stack allocations, and functions with dynamic stack
allocations always have a FP, but we might also have a FP for other reasons
(such as the user explicitly disabling frame-pointer elimination), and we don't
necessarily need a spill slot for those functions.
The structsinregs test needed adjustment because it disables FP elimination.
llvm-svn: 177106
This is the next step towards making the metadata for DIScopes have a common
prefix rather than having to delegate based on their tag type.
llvm-svn: 176913
This could be 'null' or the empty string, DIDescriptor::getStringField
coalesces the two cases anyway so it's just a matter of legible/efficient
representation.
The change in behavior of the DICompileUnit::get* functions could be
subsumed by the full verification check - but ideally that should just be an
assertion if we could front-load the actual debug info metadata failure paths.
llvm-svn: 176907
Now that only the register-scavenger version of the CR spilling code remains,
we no longer need the Darwin R2 hack. Darwin can use R0 as a spare register in
any case where the System V ABI uses it (R0 is special architecturally, and so
is reserved under all common ABIs).
A few test cases needed to be updated to reflect the register-allocation changes.
llvm-svn: 176868
Summary:
Statistics are still available in Release+Asserts (any +Asserts builds),
and stats can also be turned on with LLVM_ENABLE_STATS.
Move some of the FastISel stats that were moved under DEBUG()
back out of DEBUG(), since stats are disabled across the board now.
Many tests depend on grepping "-stats" output. Move those into
a orig_dir/Stats/. so that they can be marked as unsupported
when building without statistics.
Differential Revision: http://llvm-reviews.chandlerc.com/D486
llvm-svn: 176733
There's no need to generate a stack frame for PPC32 SVR4 when there are
no local variables assigned to the stack, i.e., when no red zone is needed.
(PPC64 supports a red zone, but PPC32 does not.)
llvm-svn: 176124
The PowerPC TLS relocation types were not previously added to the
necessary list in MCELFStreamer::fixSymbolsInTLSFixups(). Now they are!
llvm-svn: 176094
Large code model is identical to medium code model except that the
addis/addi sequence for "local" accesses is never used. All accesses
use the addis/ld sequence.
The coding changes are straightforward; most of the patch is taken up
with creating variants of the medium model tests for large model.
llvm-svn: 175767
This patch implements the PPCDAGToDAGISel::PostprocessISelDAG virtual
method to perform post-selection peephole optimizations on the DAG
representation.
One optimization is implemented here: folds to clean up complex
addressing expressions for thread-local storage and medium code
model. It will also be useful for large code model sequences when
those are added later. I originally thought about doing this on the
MI representation prior to register assignment, but it's difficult to
do effective global dead code elimination at that point. DCE is
trivial on the DAG representation.
A typical example of a candidate code sequence in assembly:
addis 3, 2, globalvar@toc@ha
addi 3, 3, globalvar@toc@l
lwz 5, 0(3)
When the final instruction is a load or store with an immediate offset
of zero, the offset from the add-immediate can replace the zero,
provided the relocation information is carried along:
addis 3, 2, globalvar@toc@ha
lwz 5, globalvar@toc@l(3)
Since the addi can in general have multiple uses, we need to only
delete the instruction when the last use is removed.
llvm-svn: 175697
This handles the cases where the 6-bit splat element is odd, converting
to a three-instruction sequence to add or subtract two splats. With this
fix, the XFAIL in test/CodeGen/PowerPC/vec_constants.ll is removed.
llvm-svn: 175663
The PPC backend doesn't handle these correctly. This patch uses logic
similar to that in the X86 and ARM backends to track these arguments
properly.
llvm-svn: 175635
