PowerPC supports pre-increment load/store instructions (except for Altivec/VSX
vector load/stores). Using these on embedded cores can be very important, but
most loops are not naturally set up to use them. We can often change that,
however, by placing loops into a non-canonical form. Generically, this means
transforming loops like this:
for (int i = 0; i < n; ++i)
array[i] = c;
to look like this:
T *p = array[-1];
for (int i = 0; i < n; ++i)
*++p = c;
the key point is that addresses accessed are pulled into dedicated PHIs and
"pre-decremented" in the loop preheader. This allows the use of pre-increment
load/store instructions without loop peeling.
A target-specific late IR-level pass (running post-LSR), PPCLoopPreIncPrep, is
introduced to perform this transformation. I've used this code out-of-tree for
generating code for the PPC A2 for over a year. Somewhat to my surprise,
running the test suite + externals on a P7 with this transformation enabled
showed no performance regressions, and one speedup:
External/SPEC/CINT2006/483.xalancbmk/483.xalancbmk
-2.32514% +/- 1.03736%
So I'm going to enable it on everything for now. I was surprised by this
because, on the POWER cores, these pre-increment load/store instructions are
cracked (and, thus, harder to schedule effectively). But seeing no regressions,
and feeling that it is generally easier to split instructions apart late than
it is to combine them late, this might be the better approach regardless.
In the future, we might want to integrate this functionality into LSR (but
currently LSR does not create new PHI nodes, so (for that and other reasons)
significant work would need to be done).
llvm-svn: 228328
PowerPC supports pre-increment floating-point load/store instructions, both r+r
and r+i, and we had patterns for them, but they were not marked as legal. Mark
them as legal (and add a test case).
llvm-svn: 228327
The combine that forms extloads used to be disabled on vector types,
because "None of the supported targets knows how to perform load and
sign extend on vectors in one instruction."
That's not entirely true, since at least SSE4.1 X86 knows how to do
those sextloads/zextloads (with PMOVS/ZX).
But there are several aspects to getting this right.
First, vector extloads are controlled by a profitability callback.
For instance, on ARM, several instructions have folded extload forms,
so it's not always beneficial to create an extload node (and trying to
match extloads is a whole 'nother can of worms).
The interesting optimization enables folding of s/zextloads to illegal
(splittable) vector types, expanding them into smaller legal extloads.
It's not ideal (it introduces some legalization-like behavior in the
combine) but it's better than the obvious alternative: form illegal
extloads, and later try to split them up. If you do that, you might
generate extloads that can't be split up, but have a valid ext+load
expansion. At vector-op legalization time, it's too late to generate
this kind of code, so you end up forced to scalarize. It's better to
just avoid creating egregiously illegal nodes.
This optimization is enabled unconditionally on X86.
Note that the splitting combine is happy with "custom" extloads. As
is, this bypasses the actual custom lowering, and just unrolls the
extload. But from what I've seen, this is still much better than the
current custom lowering, which does some kind of unrolling at the end
anyway (see for instance load_sext_4i8_to_4i64 on SSE2, and the added
FIXME).
Also note that the existing combine that forms extloads is now also
enabled on legal vectors. This doesn't have a big effect on X86
(because sext+load is usually combined to sext_inreg+aextload).
On ARM it fires on some rare occasions; that's for a separate commit.
Differential Revision: http://reviews.llvm.org/D6904
llvm-svn: 228325
The return value's address must be returned in %rax.
i.e. the callee needs to copy the sret argument (%rdi)
into the return value (%rax).
This probably won't manifest as a bug when the caller is LLVM-compiled
code. But it is an ABI guarantee and tools expect it.
llvm-svn: 228321
We should be setting UnrollingPreferences::MaxCount to MAX_UINT instead
of UnrollingPreferences::Count.
Count is a 'forced unrolling factor', while MaxCount sets an upper
limit to the unrolling factor.
Setting Count to MAX_UINT was causing the loop in the testcase to be
unrolled 15 times, when it only had a maximum of 4 iterations.
llvm-svn: 228303
The llvm.SI.end.cf intrinsic is used to mark the end of if-then blocks,
if-then-else blocks, and loops. It is responsible for updating the
exec mask to re-enable threads that had been masked during the preceding
control flow block. For example:
s_mov_b64 exec, 0x3 ; Initial exec mask
s_mov_b64 s[0:1], exec ; Saved exec mask
v_cmpx_gt_u32 exec, s[2:3], v0, 0 ; llvm.SI.if
do_stuff()
s_or_b64 exec, exec, s[0:1] ; llvm.SI.end.cf
The bug fixed by this patch was one where the llvm.SI.end.cf intrinsic
was being inserted into the header of loops. This would happen when
an if block terminated in a loop header and we would end up with
code like this:
s_mov_b64 exec, 0x3 ; Initial exec mask
s_mov_b64 s[0:1], exec ; Saved exec mask
v_cmpx_gt_u32 exec, s[2:3], v0, 0 ; llvm.SI.if
do_stuff()
LOOP: ; Start of loop header
s_or_b64 exec, exec, s[0:1] ; llvm.SI.end.cf <-BUG: The exec mask has the
same value at the beginning of each loop
iteration.
do_stuff();
s_cbranch_execnz LOOP
The fix is to create a new basic block before the loop and insert the
llvm.SI.end.cf there. This way the exec mask is restored before the
start of the loop instead of at the beginning of each iteration.
llvm-svn: 228302
Patch by Kit Barton.
Add the vector count leading zeros instruction for byte, halfword,
word, and doubleword sizes. This is a fairly straightforward addition
after the changes made for vpopcnt:
1. Add the correct definitions for the various instructions in
PPCInstrAltivec.td
2. Make the CTLZ operation legal on vector types when using P8Altivec
in PPCISelLowering.cpp
Test Plan
Created new test case in test/CodeGen/PowerPC/vec_clz.ll to check the
instructions are being generated when the CTLZ operation is used in
LLVM.
Check the encoding and decoding in test/MC/PowerPC/ppc_encoding_vmx.s
and test/Disassembler/PowerPC/ppc_encoding_vmx.txt respectively.
llvm-svn: 228301
Implement a BITCAST dag combine to transform i32->mmx conversion patterns
into a X86 specific node (MMX_MOVW2D) and guarantee that moves between
i32 and x86mmx are better handled, i.e., don't use store-load to do the
conversion..
llvm-svn: 228293
Summary: When evaluating floating point instructions in the inliner, ask the TTI whether it is an expensive operation. By default, it's not an expensive operation. This keeps the default behavior the same as before. The ARM TTI has been updated to return back TCC_Expensive for targets which don't have hardware floating point.
Reviewers: chandlerc, echristo
Reviewed By: echristo
Subscribers: t.p.northover, aemerson, llvm-commits
Differential Revision: http://reviews.llvm.org/D6936
llvm-svn: 228263
v2i32, i32, trunc i32 to i16, and truc i32 to i8 stores are legal for
all address spaces. We had marked them as custom in order to lower
them for the private address space, but this is no longer necessary.
This enables lowering of misaligned stores of these types in the
DAGLegalizer.
llvm-svn: 228189
This is a bug that was caused due to storing the feature bitset in a 32-bit
variable when it is a 64-bit mask, discarding the top half of the feature set.
llvm-svn: 228151
Currently, Cortex-A72 is modelled as an Cortex-A57 except the fp
load balancing pass isn't enabled for Cortex-A72 as it's not
profitable to have it enabled for this core.
Patch by Ranjeet Singh.
llvm-svn: 228140
This associates movss and movsd with the packed single and packed double
execution domains (resp.). While this is largely cosmetic, as we now
don't have weird ping-pong-ing between single and double precision, it
is also useful because it avoids the domain fixing algorithm from seeing
domain breaks that don't actually exist. It will also be much more
important if we have an execution domain default other than packed
single, as that would cause us to mix movss and movsd with integer
vector code on a regular basis, a very bad mixture.
llvm-svn: 228135
This reverts patches 223862, 224198, 224203, and 224754, which were all
related to the vector load/store combining and were reverted/reaplied
a few times due to the same alignment problems we're seeing now.
Further tests, mainly self-hosting Clang, will be needed to reapply this
patch in the future.
llvm-svn: 228129
This is the simplest form of bit-math based blending which only fires
when we are blending with zero and is relatively profitable. I've only
enabled this path on very specific lowering strategies. I'm planning to
widen its applicability in subsequent patches, but so far you'll notice
that even though we get fewer shufps instructions, we *still* do the bit
math in the FP execution port. I'm looking into why this is still
happening.
llvm-svn: 228124
Specifically, the existing patterns were scalar-only. These cover the
packed vector bitwise operations when specifically requested with pseudo
instructions. This is particularly important in SSE1 where we can't
actually emit a logical operation on a v2i64 as that isn't a legal type.
This will be tested in subsequent patches which form the floating point
and patterns in more places.
llvm-svn: 228123
The ARM assembler allows register alias redefinitions as long as it
targets the same register. r222319 broke that. In the AArch64 case
it would just produce a new warning, but in the ARM case it would
error out on previously accepted assembler.
llvm-svn: 228109
Patch to match cases where shuffle masks can be reduced to bit shifts. Similar to byte shift shuffle matching from D5699.
Differential Revision: http://reviews.llvm.org/D6649
llvm-svn: 228047
Patch by Kit Barton.
Add the vector population count instructions for byte, halfword, word,
and doubleword sizes. There are two major changes here:
PPCISelLowering.cpp: Make CTPOP legal for vector types.
PPCRegisterInfo.td: Added v2i64 to the VRRC register
definition. This is needed for the doubleword variations of the
integer ops that were added in P8.
Test Plan
Test the instruction vpcnt* encoding/decoding in ppc64-encoding-vmx.s
Test the generation of the vpopcnt instructions for various vector
data types. When adding the v2i64 type to the Vector Register set, I
also needed to add the appropriate bit conversion patterns between
v2i64 and the existing vector types. Testing for these conversions
were also added in the test case by passing a different vector type as
a parameter into the test functions. There is also a run step that
will ensure the vpopcnt instructions are generated when the vsx
feature is disabled.
llvm-svn: 228046
What this does is that if you accidentally select these instructions on VI,
the code generation will fail, because the pseudo -> _vi mapping will be
undefined.
The idea is to be able to catch possible future bugs easily.
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
llvm-svn: 228038
This patch adds general shuffle pattern matching for the MOVQ zero-extend instruction (copy lower 64bits, zero upper) for all 128-bit integer vectors, it is added as a fallback test in lowerVectorShuffleAsZeroOrAnyExtend.
llvm-svn: 228022
Summary:
Straight-line strength reduction (SLSR) is implemented in GCC but not yet in
LLVM. It has proven to effectively simplify statements derived from an unrolled
loop, and can potentially benefit many other cases too. For example,
LLVM unrolls
#pragma unroll
foo (int i = 0; i < 3; ++i) {
sum += foo((b + i) * s);
}
into
sum += foo(b * s);
sum += foo((b + 1) * s);
sum += foo((b + 2) * s);
However, no optimizations yet reduce the internal redundancy of the three
expressions:
b * s
(b + 1) * s
(b + 2) * s
With SLSR, LLVM can optimize these three expressions into:
t1 = b * s
t2 = t1 + s
t3 = t2 + s
This commit is only an initial step towards implementing a series of such
optimizations. I will implement more (see TODO in the file commentary) in the
near future. This optimization is enabled for the NVPTX backend for now.
However, I am more than happy to push it to the standard optimization pipeline
after more thorough performance tests.
Test Plan: test/StraightLineStrengthReduce/slsr.ll
Reviewers: eliben, HaoLiu, meheff, hfinkel, jholewinski, atrick
Reviewed By: jholewinski, atrick
Subscribers: karthikthecool, jholewinski, llvm-commits
Differential Revision: http://reviews.llvm.org/D7310
llvm-svn: 228016
This patch detects consecutive vector loads using the existing
EltsFromConsecutiveLoads() logic. This fixes:
http://llvm.org/bugs/show_bug.cgi?id=22329
This patch effectively reverts the tablegen additions of D6492 /
http://reviews.llvm.org/rL224344 ...which in hindsight were a horrible hack.
The test cases that were added with that patch are simply modified to load
from varying offsets of a base pointer. These loads did not match the existing
tablegen patterns.
A happy side effect of doing this optimization earlier is that we can now fold
the load into a math op where possible; this is shown in some of the updated
checks in the test file.
Differential Revision: http://reviews.llvm.org/D7303
llvm-svn: 228006
This can happen when a REV instruction is commuted.
The trick is not to define the _vi versions of instructions, which has these
consequences:
- code generation will always fail if a pseudo cannot be lowered
(very useful to catch bugs where an unsupported instruction somehow makes
it to the printer)
- ability to query if a pseudo can be lowered, which is done in commuteOpcode
to prevent REV from commuting to non-REV on VI
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
llvm-svn: 227990
The getCommute* functions are only used with pseudos, so this commit doesn't
change anything.
The issue with missing non-rev versions of shift instructions on VI will fixed
separately.
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
llvm-svn: 227989
- V_MAC_LEGACY_F32 exists on VI, but it's VOP3-only.
- Define CVT_PK opcodes which are different between SI and VI. These are
unused. The idea is to define all chip differences.
v2: keep V_MUL_LO_U32
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
llvm-svn: 227988
These are VOP2 on SI and VOP3 on VI, and their pseudos are neither, which can
be a problem. In order to make isVOP2 and isVOP3 queries behave as expected,
the encoding must be determined first.
This doesn't fix any known issue, but better safe than sorry.
v2: add and use getMCOpcodeFromPseudo
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
llvm-svn: 227987
This fixes a hang when using an empty geometry shader.
v2: - don't add s_nop when followed by s_waitcnt
- comestic changes
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
llvm-svn: 227986
r224330 introduced a bug by misinterpreting the "FeatureVectorUAMem" bit.
The commit log says that change did not affect anything, but that's not correct.
That change allowed SSE instructions to have unaligned mem operands folded into
math ops, and that's not allowed in the default specification for any SSE variant.
The bug is exposed when compiling for an AVX-capable CPU that had this feature
flag but without enabling AVX codegen. Another mistake in r224330 was not adding
the feature flag to all AVX CPUs; the AMD chips were excluded.
This is part of the fix for PR22371 ( http://llvm.org/bugs/show_bug.cgi?id=22371 ).
This feature bit is SSE-specific, so I've renamed it to "FeatureSSEUnalignedMem".
Changed the existing test case for the feature bit to reflect the new name and
renamed the test file itself to better reflect the feature.
Added runs to fold-vex.ll to check for the failing codegen.
Note that the feature bit is not set by default on any CPU because it may require a
configuration register setting to enable the enhanced unaligned behavior.
llvm-svn: 227983
This patch is a third attempt to properly handle the local-dynamic and
global-dynamic TLS models.
In my original implementation, calls to __tls_get_addr were hidden
from view until the asm-printer phase, at which point the underlying
branch-and-link instruction was created with proper relocations. This
mostly worked well, but I used some repellent techniques to ensure
that the TLS_GET_ADDR nodes at the SD and MI levels correctly received
input from GPR3 and produced output into GPR3. This proved to work
badly in the presence of multiple TLS variable accesses, with the
copies to and from GPR3 being scheduled incorrectly and generally
creating havoc.
In r221703, I addressed that problem by representing the calls to
__tls_get_addr as true calls during instruction lowering. This had
the advantage of removing all of the bad hacks and relying on the
existing call machinery to properly glue the copies in place. It
looked like this was going to be the right way to go.
However, as a side effect of the recent discovery of problems with
linker optimizations for TLS, we discovered cases of suboptimal code
generation with this strategy. The problem comes when tls_get_addr is
called for the same address, and there is a resulting CSE
opportunity. It turns out that in such cases MachineCSE will common
the addis/addi instructions that set up the input value to
tls_get_addr, but will not common the calls themselves. MachineCSE
does not have any machinery to common idempotent calls. This is
perfectly sensible, since presumably this would be done at the IR
level, and introducing calls in the back end isn't commonplace. In
any case, we end up with two calls to __tls_get_addr when one would
suffice, and that isn't good.
I presumed that the original design would have allowed commoning of
the machine-specific nodes that hid the __tls_get_addr calls, so as
suggested by Ulrich Weigand, I went back to that design and cleaned it
up so that the copies were properly held together by glue
nodes. However, it turned out that this didn't work either...the
presence of copies to physical registers kept the machine-specific
nodes from being commoned also.
All of which leads to the design presented here. This is a return to
the original design, except that no attempt is made to introduce
copies to and from GPR3 during instruction lowering. Virtual registers
are used until prior to register allocation. At that point, a special
pass is run that identifies the machine-specific nodes that hide the
tls_get_addr calls and introduces the copies to and from GPR3 around
them. The register allocator then coalesces these copies away. With
this design, MachineCSE succeeds in commoning tls_get_addr calls where
possible, and we get nice optimal code generation (better than GCC at
the moment, which does not common these calls).
One additional problem must be dealt with: After introducing the
mentions of the physical register GPR3, the aggressive anti-dependence
breaker sees opportunities to improve scheduling by selecting a
different register instead. Flags must be used on the instruction
descriptions to tell the anti-dependence breaker to keep its hands in
its pockets.
One thing missing from the original design was recording a definition
of the link register on the GET_TLS_ADDR nodes. Doing this was found
to be insufficient to force a stack frame to be created, which led to
looping behavior because two different LR values were stored at the
same address. This appears to have been an oversight in
PPCFrameLowering::determineFrameLayout(), which is repaired here.
Because MustSaveLR() returns true for calls to builtin_return_address,
this changed the expected behavior of
test/CodeGen/PowerPC/retaddr2.ll, which now stacks a frame but
formerly did not. I've fixed the test case to reflect this.
There are existing TLS tests to catch regressions; the checks in
test/CodeGen/PowerPC/tls-store2.ll proved to be too restrictive in the
face of instruction scheduling with these changes, so I fixed that
up.
I've added a new test case based on the PrettyStackTrace module that
demonstrated the original problem. This checks that we get correct
code generation and that CSE of the calls to __get_tls_addr has taken
place.
llvm-svn: 227976
Improve EXTRACT_VECTOR_ELT DAG combine to catch conversion patterns
between x86mmx and i32 with more layers of indirection.
Before:
movq2dq %mm0, %xmm0
movd %xmm0, %eax
After:
movd %mm0, %eax
llvm-svn: 227969
Summary:
Hi Eric,
this patch cleans up the layering violation that you're fixing across backends.
Anything else I need to fix on bpf backend side?
Thanks
Reviewers: echristo
Reviewed By: echristo
Differential Revision: http://reviews.llvm.org/D7355
llvm-svn: 227865