This patch fixes the following two bugs in `PPCInstrInfo::isSignOrZeroExtended` helper, which is used from sign-/zero-extension elimination in PPCMIPeephole pass.
- Registers defined by load with update (e.g. LBZU) were identified as already sign or zero-extended. But it is true only for the first def (loaded value) and not for the second def (i.e. updated pointer).
- Registers defined by ORIS/XORIS were identified as already sign-extended. But, it is not true for sign extension depending on the immediate (while it is ok for zero extension).
To handle the first case, the parameter for the helpers is changed from `MachineInstr` to a register number to distinguish first and second defs. Also, this patch moves the initialization of PPCMIPeepholePass to allow mir test case.
Reviewed By: nemanjai
Differential Revision: https://reviews.llvm.org/D40554
SPE doesn't have a fmadd instruction, so don't bother hoisting a
multiply and add sequence to this, as it'd become just a library call.
Hoisting happens too late for the CTR usability test to veto using the
CTR in a loop, and results in an assert "Invalid PPC CTR loop!".
Map hardware loop intrinsics loop_decrement and set_loop_iteration
to the new PowerPC pseudo instructions, so that the hardware loop
intrinsics will be expanded to normal cmp+branch form or ctrloop
form based on the CTR register usage on MIR level.
Reviewed By: lkail
Differential Revision: https://reviews.llvm.org/D123366
This patch ensures consistency in the construction of FP_ROUND nodes
such that they always use ISD::TargetConstant instead of ISD::Constant.
This additionally fixes a bug in the AArch64 SVE backend where patterns
were matching against TargetConstant nodes and sometimes failing when
passed a Constant node.
Reviewed By: paulwalker-arm
Differential Revision: https://reviews.llvm.org/D130370
I don't have any evidence these particular uses are actually causing any
issues, but we should avoid accidentally truncating immediate values
depending on the host.
Proposing to move the check for scalar MASS conversion from constructor
of PPCTargetLowering to the lowerLibCallBase function which decides
about the lowering.
The Target machine option Options.PPCGenScalarMASSEntries is set in
PPCTargetMachine.cpp. But an object of the class PPCTargetLowering
is created in one of the included header files. So, the constructor will run
before setting PPCGenScalarMASSEntries to correct value. So, we cannot
check this option in the constructor.
Differential: https://reviews.llvm.org/D128653
Reviewer: @bmahjour
opportunities
There are straight forward splat load opportunities blocked by
getNormalLoadInput(), since those cases involve consecutive bitcasts.
Improve by looking through bitcasts.
Reviewed By: nemanjai
Differential Revision: https://reviews.llvm.org/D128703
There are instances where using paired vector stores leads to significant
performance degradation due to issues with store forwarding.To avoid falling
into this trap with compiler - generated code, we will not emit these
instructions unless the user requests them explicitly(with a builtin or by
specifying the option).
Reviewed By : lei, amyk, saghir
Differential Revision: https://reviews.llvm.org/D127218
Currently in `combineVectorShuffle()`, we update the shuffle mask if either
input vector comes from a scalar_to_vector, and we keep the respective input
vectors in its permuted form by producing PPCISD::SCALAR_TO_VECTOR_PERMUTED.
However, it is possible that we end up in a situation where both input vectors
to the vector_shuffle are scalar_to_vector, and are different vector types.
In situations like this, the shuffle mask is updated incorrectly as the current
code assumes both scalar_to_vector inputs are the same vector type.
This patch skips the combines for vector_shuffle if both input vectors are
scalar_to_vector, and if they are of different vector types. A follow up patch
will focus on fixing this issue afterwards, in order to correctly update the
shuffle mask.
Differential Revision: https://reviews.llvm.org/D127818
This patch changes the PowerPC backend to generate VSX load/store instructions
for all vector loads/stores on Power8 and earlier (LE) instead of VMX
load/store instructions. The reason for this change is because VMX instructions
require the vector to be 16-byte aligned. So, a vector load/store will fail with
VMX instructions if the vector is misaligned. Also, `gcc` generates VSX
instructions in this situation which allow for unaligned access but require a
swap instruction after loading/before storing. This is not an issue for BE
because we already emit VSX instructions since no swap is required. And this is
not an issue on Power9 and up since we have access to `lxv[x]`/`stxv[x]` which
allow for unaligned access and do not require swaps.
This patch also delays the VSX load/store for LE combines until after
LegalizeOps to prioritize other load/store combines.
Reviewed By: #powerpc, stefanp
Differential Revision: https://reviews.llvm.org/D127309
The combine step for shufflevector will sometimes replace undef in the mask
with a defined value. This can cause an infinite loop in some cases as another
combine will then put the undef back in the mask.
This patch fixes the issue so that undefs are not replaced when doing a combine.
Reviewed By: ZarkoCA, amyk, quinnp, saghir
Differential Revision: https://reviews.llvm.org/D127439
I can't remove the function just yet as it is used in the generated .inc files.
I would also like to provide a way to compare alignment with TypeSize since it came up a few times.
Differential Revision: https://reviews.llvm.org/D126910
If the mask is made up of elements that form a mask in the higher type
we can convert shuffle(bitcast into the bitcast type, simplifying the
instruction sequence. A v4i32 2,3,0,1 for example can be treated as a
1,0 v2i64 shuffle. This helps clean up some of the AArch64 concat load
combines, along with helping simplify a number of other tests.
The PowerPC combine for v16i8 splat vector loads needed some fixes to
keep it working for v16i8 vectors. This improves the handling of v2i64
shuffles to match too, hopefully improving them in general.
Differential Revision: https://reviews.llvm.org/D123801
AtomicExpandPass uses this variable to determine emitting libcalls or not. The default value is 1024 and if we don't specify it for PPC64 explicitly, AtomicExpandPass won't emit `__atomic_*` libcalls for those target unable to inline atomic ops and finally the backend emits `__sync_*` libcalls. Thanks @efriedma for pointing it out.
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D122868
Make 16-byte atomic type aligned to 16-byte on PPC64, thus consistent with GCC. Also enable inlining 16-byte atomics on non-AIX targets on PPC64.
Reviewed By: hubert.reinterpretcast
Differential Revision: https://reviews.llvm.org/D122377
Add support for builtin_[max|min] which has below prototype:
A builtin_max (A1, A2, A3, ...)
All arguments must have the same type; they must all be float, double, or long double.
Internally use SelectCC to get the result.
Reviewed By: qiucf
Differential Revision: https://reviews.llvm.org/D122478
To store a byval parameter the existing code would store as many 8 byte elements
as was required to store the full size of the byval parameter.
For example, a paramter of size 16 would store two element of 8 bytes.
A paramter of size 12 would also store two elements of 8 bytes.
This would sometimes store too many bytes as the size of the paramter is not
always a factor of 8.
This patch fixes that issue and now byval paramters are stored with the correct
number of bytes.
Reviewed By: nemanjai, #powerpc, quinnp, amyk
Differential Revision: https://reviews.llvm.org/D121430
After D108936, @llvm.smul.with.overflow.i64 was lowered to __multi3
instead of __mulodi4, which also doesn't exist on PowerPC 32-bit, not
even with compiler-rt. Block it as well so that we get inline code.
Because libgcc doesn't have __muloti4, we block that as well.
Fixes#54460.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D122090
We are going to remove the old 'perfect shuffle' optimization since it
brings performance penalty in hot loop around vectors. For example, in
following loop sharing the same mask:
%v.1 = shufflevector ... <0,1,2,3,8,9,10,11,16,17,18,19,24,25,26,27>
%v.2 = shufflevector ... <0,1,2,3,8,9,10,11,16,17,18,19,24,25,26,27>
The generated instructions will be `vmrglw-vmrghw-vmrglw-vmrghw` instead
of `vperm-vperm`. In some large loop cases, this causes 20%+ performance
penalty.
The original attempt to resolve this is to pre-record masks of every
shufflevector operation in DAG, but that is somewhat complex and brings
unnecessary computation (to scan all nodes) in optimization. Here we
disable it by default. There're indeed some cases becoming worse after
this, which will be fixed in a more careful way in future patches.
Reviewed By: jsji
Differential Revision: https://reviews.llvm.org/D121082
Currently in Clang, we have two types of builtins for fnmsub operation:
one for float/double vector, they'll be transformed into IR operations;
one for float/double scalar, they'll generate corresponding intrinsics.
But for the vector version of builtin, the 3 op chain may be recognized
as expensive by some passes (like early cse). We need some way to keep
the fnmsub form until code generation.
This patch introduces ppc.fnmsub.* intrinsic to unify four fnmsub
intrinsics.
Reviewed By: shchenz
Differential Revision: https://reviews.llvm.org/D116015
Perfect shuffle was introduced into PowerPC backend years ago, and only
available in big-endian subtargets. This optimization has good effects
in simple cases, but brings serious negative impact in large programs
with many shuffle instructions sharing the same mask.
Here introduces a temporary backend hidden option to control it until we
implemented better way to fix the gap in vectorshuffle decomposition.
Reviewed By: jsji
Differential Revision: https://reviews.llvm.org/D120072
Power ISA 3.1 adds xsmaxcqp/xsmincqp for quad-precision type-c max/min selection,
and this opens the opportunity to improve instruction selection on: llvm.maxnum.f128,
llvm.minnum.f128, and select_cc ordered gt/lt and (don't care) gt/lt.
Reviewed By: nemanjai, shchenz, amyk
Differential Revision: https://reviews.llvm.org/D117006
This patch introduces the conversions from math function calls
to MASS library calls. To resolves calls generated with these conversions, one
need to link libxlopt.a library. This patch is tested on PowerPC Linux and AIX.
Differential: https://reviews.llvm.org/D101759
Reviewer: bmahjour
This patch updates the P10 patterns with a load feeding into an insertelt to
utilize the refactored load and store infrastructure, as well as updating any
tests that exhibit any codegen changes.
Furthermore, custom legalization is added for v4f32 on Power9 and above to not
only assist with adjusting the refactored load/stores for P10 vector insert,
but also it enables the utilization of direct moves.
Differential Revision: https://reviews.llvm.org/D115691
This patch updates how splat loads handled and is an extension of D106555.
Particularly, for v2i64/v4f32/v4i32 types, they are updated to handle only
non-extending loads. For v8i16/v16i8 types, they are updated to handle extending
loads only if the memory VT is the same vector element VT type.
A test case has been added to illustrate a scenario where a PPCISD::LD_SPLAT
node should not be produced. In this test, it depicts the following f64
extending load used in a v2f64 build vector, but the extending load is actually
used in more places other than the build vector (such as in t12 and t16).
```
Type-legalized selection DAG: %bb.0 'test:entry'
SelectionDAG has 20 nodes:
t0: ch = EntryToken
t4: i64,ch = CopyFromReg t0, Register:i64 %1
t6: i64,ch = CopyFromReg t0, Register:i64 %2
t11: f64,ch = load<(load (s64) from %ir.b, !tbaa !7)> t0, t4, undef:i64
t16: f64 = fadd t31, t37
t34: ch = store<(store (s64) into %ir.c, !tbaa !7)> t31:1, t16, t6, undef:i64
t36: ch = TokenFactor t34, t37:1
t27: v2f64 = BUILD_VECTOR t37, t37
t22: ch,glue = CopyToReg t36, Register:v2f64 $v2, t27
t12: f64 = fadd t11, t37
t28: ch = store<(store (s64) into %ir.b, !tbaa !7)> t11:1, t12, t4, undef:i64
t31: f64,ch = load<(load (s64) from %ir.c, !tbaa !7)> t28, t6, undef:i64
t2: i64,ch = CopyFromReg t0, Register:i64 %0
t37: f64,ch = load<(load (s32) from %ir.a, !tbaa !3), anyext from f32> t0, t2, undef:i64
t23: ch = PPCISD::RET_FLAG t22, Register:v2f64 $v2, t22:1
```
Differential Revision: https://reviews.llvm.org/D117803
This reverts commit ef82063207.
- It conflicts with the existing llvm::size in STLExtras, which will now
never be called.
- Calling it without llvm:: breaks C++17 compat
In commit 1674d9b6b2, I fixed the bug where we didn't consider
both words of the result of the comparison. However, the logic
needs to be different for eq and ne.
Namely for eq, we need both words of the doubleword to equal so it
is an AND. OTOH for ne, we need either word to be unequal so it
is an OR.
Instead use either Type::getPointerElementType() or
Type::getNonOpaquePointerElementType().
This is part of D117885, in preparation for deprecating the API.
When doing the float to int conversion the strict conversion also needs to
retun a chain. This patch fixes that.
Reviewed By: nemanjai, #powerpc, qiucf
Differential Revision: https://reviews.llvm.org/D117464
The current code makes the assumption that equality
comparison can be performed with a word comparison
instruction. While this is true if the entire 64-bit
results are used, it does not generally work. It is
possible that the low order words and high order
words produce different results and a user of only
one will get the wrong result.
This patch adds an and of the result words so that
each word has the result of the comparison of the
entire doubleword that contains it.
Differential revision: https://reviews.llvm.org/D115678
Now we won't copy the byval parameter (bigger than 8 bytes) to
caller's parameter save area. Instead, we will only copy the
byval parameter when it can not be passed entirely in registers
which means we have to use parameter save area according to the
64 bit SVR4 ABI.
Reviewed By: jsji
Differential Revision: https://reviews.llvm.org/D111485
The load/store infrastructure previously made an incorrect assumption that
whenever it is used with a load/store intrinsic on Power10 - those intrinsics
would automatically be the lxvp/stxvp intrinsics introduced in Power10.
However, this is obviously not the case as there are multiple instances of
pre-P10 intrinsics that use the refactored load/store implementation.
This patch corrects this assumption, and produces the expected intrinsic on pre-P10.
Differential Revision: https://reviews.llvm.org/D114978
The XL implementation of vec_round for vector double uses
"round-to-nearest, ties to even" just as the vector float
`version does. However clang and gcc use "round-to-nearest-away"
for vector double and "round-to-nearest, ties to even"
for vector float.
The XL behaviour is implemented under the __XL_COMPAT_ALTIVEC__
macro similarly to other instances of incompatibility.
Differential revision: https://reviews.llvm.org/D113642
Similarly to what GCC does, we should allow scalars with
the "v" constraint rather than introducing unnecessary
new constraints for scalars in Altivec registers.
Differential revision: https://reviews.llvm.org/D113635
The platform independent ISD::INSERT_VECTOR_ELT take a element index,
but vins* instructions take a byte index. Update 32bit td patterns for
vector insert to handle the element index accordingly.
Since vector insert for non constant index are supported in
ISA3.1, there is no need to use platform specific ISD node,
PPCISD::VECINSERT. Update td pattern to directly use
ISD::INSERT_VECTOR_ELT instead.
Reviewed By: nemanjai, #powerpc
Differential Revision: https://reviews.llvm.org/D113802
Currently, the floating point instructions that depend on
rounding mode are correctly marked in the PPC back end with
an implicit use of the RM register. Similarly, instructions
that explicitly define the register are marked with an
implicit def of the same register. So for the most part,
RM-using code won't be moved across RM-setting instructions.
However, calls are not marked as RM-setting instructions so
code can be moved across calls. This is generally desired,
but so is the ability to turn off this behaviour with an
appropriate option - and -frounding-math really should be
that option.
This patch provides a set of call instructions (for direct
and indirect calls) that are marked with an implicit def of
the RM register. These will be used for calls that are marked
with the strictfp attribute.
Differential revision: https://reviews.llvm.org/D111433
Stefan Pintilie