Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
is only done if the add has one use. If the resulting constant
add can be folded into an addressing mode, force this to happen
for the pointer operand.
This ends up happening a lot because of how LDS objects are allocated.
Since the globals are allocated next to each other, acessing the first
element of the second object is directly indexed by a shifted pointer.
llvm-svn: 215739
The default assumes that a 16-bit signed offset is used.
LDS instruction use a 16-bit unsigned offset, so it wasn't
being used in some cases where it was assumed a negative offset
could be used.
More should be done here, but first isLegalAddressingMode needs
to gain an addressing mode argument. For now, copy most of the rest
of the default implementation with the immediate offset change.
llvm-svn: 215732
This for some reason fixes v1i64 kernel arguments on pre-SI. This
currently breaks some other cases in the kernel-args.ll test for R600,
but I'm not particularly confident in the new output. VTX_READ_* are not
used for some of the scalarized cases, and the code reading from the
constant buffer doesn't make much sense to me.
llvm-svn: 215564
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
v2: drop enum keyword
use correct extension mode
don't bother computing the sign in unsinged case
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 215462
v2: add tests
rename LowerSDIV24 to LowerSDIVREM24
handle the rem part in this function
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 215460
There are no variable values like registers encoded in the low 32 bits of MUBUF
instructions, so it is relatively easy to check these bits, and it will
help prevent us from introducing encoding bugs.
llvm-svn: 215397
This bit was left uninitialized, which was causing some random failures
of piglit tests.
NOTE: This is a candidate for the 3.5 branch.
llvm-svn: 215396
These tests were using SI-NOT: MOVREL to make sure concat vectors
weren't being lowered to stack loads and stores, but we are using
scratch buffers for the stack now instead of registers, so we need
to add an additional SI-NOT check for scratch buffers.
With this change I was able to uncover one broken test which will
be fixed in a future commit.
llvm-svn: 215269
This partially fixes weird looking load scheduling
in memcpy test. The load clustering doesn't seem
particularly smart, but this method seems to be partially
deprecated so it might not be worth trying to fix.
llvm-svn: 214943
This currently has a noticable effect on the kernel argument loads.
LDS and global loads are more problematic, I think because of how copies
are currently inserted to ensure that the address is a VGPR.
llvm-svn: 214942
This slipped in in r214467, so something like
V_MOV_B32_e32 v0, ... is now printed with 2 spaces
between the instruction name and first operand.
llvm-svn: 214660
Abs/neg folding has moved out of foldOperands and into the instruction
selection phase using complex patterns. As a consequence of this
change, we now prefer to select the 64-bit encoding for most
instructions and the modifier operands have been dropped from
integer VOP3 instructions.
llvm-svn: 214467
Rename to allowsMisalignedMemoryAccess.
On R600, 8 and 16 byte accesses are mostly OK with 4-byte alignment,
and don't need to be split into multiple accesses. Vector loads with
an alignment of the element type are not uncommon in OpenCL code.
llvm-svn: 214055
over each node in the worklist prior to combining.
This allows the combiner to produce new nodes which need to go back
through legalization. This is particularly useful when generating
operands to target specific nodes in a post-legalize DAG combine where
the operands are significantly easier to express as pre-legalized
operations. My immediate use case will be PSHUFB formation where we need
to build a constant shuffle mask with a build_vector node.
This also refactors the relevant functionality in the legalizer to
support this, and updates relevant tests. I've spoken to the R600 folks
and these changes look like improvements to them. The avx512 change
needs to be investigated, I suspect there is a disagreement between the
legalizer and the DAG combiner there, but it seems a minor issue so
leaving it to be re-evaluated after this patch.
Differential Revision: http://reviews.llvm.org/D4564
llvm-svn: 214020
which have successfully round-tripped through the combine phase, and use
this to ensure all operands to DAG nodes are visited by the combiner,
even if they are only added during the combine phase.
This is critical to have the combiner reach nodes that are *introduced*
during combining. Previously these would sometimes be visited and
sometimes not be visited based on whether they happened to end up on the
worklist or not. Now we always run them through the combiner.
This fixes quite a few bad codegen test cases lurking in the suite while
also being more principled. Among these, the TLS codegeneration is
particularly exciting for programs that have this in the critical path
like TSan-instrumented binaries (although I think they engineer to use
a different TLS that is faster anyways).
I've tried to check for compile-time regressions here by running llc
over a merged (but not LTO-ed) clang bitcode file and observed at most
a 3% slowdown in llc. Given that this is essentially a worst case (none
of opt or clang are running at this phase) I think this is tolerable.
The actual LTO case should be even less costly, and the cost in normal
compilation should be negligible.
With this combining logic, it is possible to re-legalize as we combine
which is necessary to implement PSHUFB formation on x86 as
a post-legalize DAG combine (my ultimate goal).
Differential Revision: http://reviews.llvm.org/D4638
llvm-svn: 213898
Use ComputeNumSignBits instead of checking for i8 / i16 which only
worked when AMDIL was lying about having legal i8 / i16.
If an integer is known to fit in 24-bits, we can
do division faster with float ops.
llvm-svn: 213843
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
There are a few more cleanups to do, but I ran into some problems
with ext loads and trunc stores, when I tried to change some of the
vector loads and stores from custom to legal, so I wasn't able to
get rid of everything.
llvm-svn: 213552
This implements a solution for constant initializers suggested
by Vadim Girlin, where we store the data after the shader code
and then use the S_GETPC instruction to compute its address.
This saves use the trouble of creating a new buffer for constant data
and then having to pass the pointer to the kernel via user SGPRs or the
input buffer.
llvm-svn: 213530
These instructions can only take a limited input range, and return
the constant value 1 out of range. We should do range reduction to
be able to process arbitrary values. Use a FRACT instruction after
normalization to achieve this. Also add a test for constant folding
with the lowered code with unsafe-fp-math enabled.
v2: use DAG lowering instead of intrinsic, adapt test
v3: calculate constant, fold pattern into instruction definition
v4: misc style fixes, add sin-fold testcase, cosmetics
Patch by Grigori Goronzy
llvm-svn: 213458
Actual support for softening f16 operations is still limited, and can be added
when it's needed. But Soften is much closer to being a useful thing to try
than keeping it Legal when no registers can actually hold such values.
Longer term, we probably want something between Soften and Promote semantics
for most targets, it'll be more efficient to promote the 4 basic operations to
f32 than libcall them.
llvm-svn: 213372
Unfortunately, we don't seem to have a direct truncation, but the
extension can be legally split into two operations so we should
support that.
llvm-svn: 213357
This makes the two intrinsics @llvm.convert.from.f16 and
@llvm.convert.to.f16 accept types other than simple "float". This is
only strictly needed for the truncate operation, since otherwise
double rounding occurs and there's no way to represent the strict IEEE
conversion. However, for symmetry we allow larger types in the extend
too.
During legalization, we can expand an "fp16_to_double" operation into
two extends for convenience, but abort when the truncate isn't legal. A new
libcall is probably needed here.
Even after this commit, various target tweaks are needed to actually use the
extended intrinsics. I've put these into separate commits for clarity, so there
are no actual tests of f64 conversion here.
llvm-svn: 213248
v2: use ffbh/l if available
v3: Rebase on top of Matt's SI patches
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Tom Stellard <tom@stellard.net>
llvm-svn: 213072
This helps avoid redundant instructions to unpack, and repack
the vectors. Ideally we could recognize that pattern and eliminate
it. Currently v4i8 and other small element type vectors are scalarized,
so this has the added bonus of avoiding that.
llvm-svn: 213031
We need the intrinsics with offsets, so why not just add them all.
The R128 parameter will also be useful for reducing SGPR usage.
GL_ARB_image_load_store also adds some image GLSL modifiers like "coherent",
so Mesa will probably translate those to slc, glc, etc.
When LLVM 3.5 is released, I'll switch Mesa to these new intrinsics.
llvm-svn: 212830
Use alg. from LegalizeDAG.cpp
Move Expand setting to SIISellowering
v2: Extend existing tests instead of creating new ones
v3: use separate LowerFPTOSINT function
v4: use TargetLowering::expandFP_TO_SINT
add comment about using FP_TO_SINT for uints
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Tom Stellard <tom@stellard.net>
llvm-svn: 212773
The default rounding mode to initialize the mode register needs
to be reported to the runtime. Fill in other bits a kernel
may be interested in setting for future use.
llvm-svn: 211791
R600 was using a clamped version of rsq, but SI was not. Add a
new rsq_clamped intrinsic and use them consistently.
It's unclear to me from the documentation what behavior
the R600 instructions have, so I assume they have the legacy behavior
described by the SI documents. For R600, use RECIPSQRT_IEEE
for both llvm.AMDGPU.rsq.legacy and llvm.AMDGPU.rsq. R600 also
has RECIPSQRT_FF, which I'm not sure how it fits in here.
llvm-svn: 211637
These will be used for custom lowering and for library
implementations of various math functions, so it's useful
to expose these as builtins.
llvm-svn: 211247
The difference from rint isn't really relevant here,
so treat them as equivalent. OpenCL doesn't have nearbyint,
so this is sort of pointless other than for completeness.
llvm-svn: 211229
This contains all the previous patches + getlod support on top of it.
It doesn't use SDNodes anymore, so it's quite small.
It also adds v16i8 to SReg_128, which is used for the sampler descriptor.
Reviewed-by: Tom Stellard
llvm-svn: 211228
This would assert if a constant address space was extern
and therefore didn't have an initializer. If the initializer
was undef, it would hit the unreachable unhandled initializer case.
An extern global should never really occur since we don't have
machine linking, but bugpoint likes to remove initializers.
llvm-svn: 210967
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
Delete all unused ones, and add new AMDGPU named intrinsics for
the ones that are. Handle the old AMDIL names for comptability (although
remove their GCCBuiltin names) and add tests since there weren't any
for these before.
llvm-svn: 210827
This is the same problem fixed in r210664 for more types.
The test passes without this fix. For some reason
I'm only hitting this when creating selects lowered
to v2i32 selects.
llvm-svn: 210692
There seem to be only 2 places that produce these,
and it's kind of tricky to hit them.
Also fixes failure to bitcast between i64 and v2f32,
although this for some reason wasn't actually broken in the
simple bitcast testcase, but did in the scalar_to_vector one.
llvm-svn: 210664
Alias with unnamed_addr were in a strange state. It is stored in GlobalValue,
the language reference talks about "unnamed_addr aliases" but the verifier
was rejecting them.
It seems natural to allow unnamed_addr in aliases:
* It is a property of how it is accessed, not of the data itself.
* It is perfectly possible to write code that depends on the address
of an alias.
This patch then makes unname_addr legal for aliases. One side effect is that
the syntax changes for a corner case: In globals, unnamed_addr is now printed
before the address space.
llvm-svn: 210302
SI_IF and SI_ELSE are terminators which also produce a value. For
these instructions ISel always inserts a COPY to move their value
to another basic block. This COPY ends up between SI_(IF|ELSE)
and the S_BRANCH* instruction at the end of the block.
This breaks MachineBasicBlock::getFirstTerminator() and also the
machine verifier which assumes that terminators are grouped together at
the end of blocks.
To solve this we coalesce the copy away right after ISel to make sure
there are no instructions in between terminators at the end of blocks.
llvm-svn: 207591
SALU instructions ignore control flow, so it is not always safe to use
them within branches. This is a partial solution to this problem
until we can come up with something better.
llvm-svn: 207590
This is a squash of several optimization commits:
- calculate DIV_Lo and DIV_Hi separately
- use BFE_U32 if we are operating on 32bit values
- use precomputed constants instead of shifting in UDVIREM
- skip the first 32 iterations of udivrem
v2: Check whether BFE is supported before using it
Patch by: Jan Vesely
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
llvm-svn: 207589
Use scalar BFE with constant shift and offset when possible.
This is complicated by the fact that the scalar version packs
the two operands of the vector version into one.
llvm-svn: 206558
Having i128 as a legal type complicates the legalization phase. v4i32
is already a legal type, so we will use that instead.
This fixes several piglit tests.
llvm-svn: 206500
Print in decimal for inline immediates, and hex otherwise. Use hex
always for offsets in addressing offsets.
This approximately matches what the shader compiler does.
llvm-svn: 206335
The TargetLowering::expandMUL() helper contains lowering code extracted
from the DAGTypeLegalizer and allows the SelectionDAGLegalizer to expand more
ISD::MUL patterns without having to use a library call.
llvm-svn: 206037
Moving these patterns from TableGen files to PerformDAGCombine()
should allow us to generate better code by eliminating unnecessary
shifts and extensions earlier.
This also fixes a bug where the MAD pattern was calling
SimplifyDemandedBits with a 24-bit mask on the first operand
even when the full pattern wasn't being matched. This occasionally
resulted in some instructions being incorrectly deleted from the
program.
v2:
- Fix bug with 64-bit mul
llvm-svn: 205731
Acording to AMD documentation, the correct opcode for
BFE_INT is 0x5, not 0x4
Fixes Arithm/Absdiff.Mat/3 OpenCV test
Patch by: Bruno Jiménez
llvm-svn: 205562
Try to match scalar and first like the other instructions.
Expand 64-bit ands to a pair of 32-bit ands since that is not
available on the VALU.
llvm-svn: 204660
It isn't actually used now, and probably never will be, plus it makes
tests less annoying. I also think SC prints GDS instructions as a
separate instruction name.
llvm-svn: 204270
The type of the immediates should not matter as long as the encoding is
equivalent to the encoding of one of the legal inline constants.
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
llvm-svn: 204056
This instructions writes to an 32-bit SGPR. This change required adding
the 32-bit VCC_LO and VCC_HI registers, because the full VCC register
is 64 bits.
This fixes verifier errors on several of the indirect addressing piglit
tests.
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
llvm-svn: 204055
LDS instructions are pseudo instructions which model
the OQAP defs and uses within a single instruction.
This fixes a hang in the opencv MedianFilter tests.
llvm-svn: 203818
These are sometimes created by the shrink to boolean optimization in the
globalopt pass.
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
llvm-svn: 203280
If the SI_KILL operand is constant, we can either clear the exec mask if
the operand is negative, or do nothing otherwise.
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
llvm-svn: 202337
DS instructions that access local memory can only uses addresses that
are less than or equal to the value of M0. When M0 is uninitialized,
then we experience undefined behavior.
This patch also changes the behavior to emit S_WQM_B64 on pixel shaders
no matter what kind of DS instruction is used.
llvm-svn: 201097
There was a problem with the old pattern, so we were copying some
larger immediates into registers when we could have been encoding
them in the instruction.
llvm-svn: 200932
The OpenCL specs say: "The vector versions of the math functions operate
component-wise. The description is per-component."
Patch by: Jan Vesely
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 200773
V_ADD_F32 with source modifier does not produce -0.0 for this. Just
manipulate the sign bit directly instead.
Also add a pattern for (fneg (fabs ...)).
Fixes a bunch of bit encoding piglit tests with radeonsi.
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
llvm-svn: 200743
