This changes the interface of how targets describe how to legalize, see
the below description.
1. Interface for targets to describe how to legalize.
In GlobalISel, the API in the LegalizerInfo class is the main interface
for targets to specify which types are legal for which operations, and
what to do to turn illegal type/operation combinations into legal ones.
For each operation the type sizes that can be legalized without having
to change the size of the type are specified with a call to setAction.
This isn't different to how GlobalISel worked before. For example, for a
target that supports 32 and 64 bit adds natively:
for (auto Ty : {s32, s64})
setAction({G_ADD, 0, s32}, Legal);
or for a target that needs a library call for a 32 bit division:
setAction({G_SDIV, s32}, Libcall);
The main conceptual change to the LegalizerInfo API, is in specifying
how to legalize the type sizes for which a change of size is needed. For
example, in the above example, how to specify how all types from i1 to
i8388607 (apart from s32 and s64 which are legal) need to be legalized
and expressed in terms of operations on the available legal sizes
(again, i32 and i64 in this case). Before, the implementation only
allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0,
s128}, NarrowScalar). A worse limitation was that if you'd wanted to
specify how to legalize all the sized types as allowed by the LLVM-IR
LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times
and probably would need a lot of memory to store all of these
specifications.
Instead, the legalization actions that need to change the size of the
type are specified now using a "SizeChangeStrategy". For example:
setLegalizeScalarToDifferentSizeStrategy(
G_ADD, 0, widenToLargerAndNarrowToLargest);
This example indicates that for type sizes for which there is a larger
size that can be legalized towards, do it by Widening the size.
For example, G_ADD on s17 will be legalized by first doing WidenScalar
to make it s32, after which it's legal.
The "NarrowToLargest" indicates what to do if there is no larger size
that can be legalized towards. E.g. G_ADD on s92 will be legalized by
doing NarrowScalar to s64.
Another example, taken from the ARM backend is:
for (unsigned Op : {G_SDIV, G_UDIV}) {
setLegalizeScalarToDifferentSizeStrategy(Op, 0,
widenToLargerTypesUnsupportedOtherwise);
if (ST.hasDivideInARMMode())
setAction({Op, s32}, Legal);
else
setAction({Op, s32}, Libcall);
}
For this example, G_SDIV on s8, on a target without a divide
instruction, would be legalized by first doing action (WidenScalar,
s32), followed by (Libcall, s32).
The same principle is also followed for when the number of vector lanes
on vector data types need to be changed, e.g.:
setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal);
setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal);
setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal);
setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal);
setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal);
setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal);
setLegalizeVectorElementToDifferentSizeStrategy(
G_ADD, 0, widenToLargerTypesUnsupportedOtherwise);
As currently implemented here, vector types are legalized by first
making the vector element size legal, followed by then making the number
of lanes legal. The strategy to follow in the first step is set by a
call to setLegalizeVectorElementToDifferentSizeStrategy, see example
above. The strategy followed in the second step
"moreToWiderTypesAndLessToWidest" (see code for its definition),
indicating that vectors are widened to more elements so they map to
natively supported vector widths, or when there isn't a legal wider
vector, split the vector to map it to the widest vector supported.
Therefore, for the above specification, some example legalizations are:
* getAction({G_ADD, LLT::vector(3, 3)})
returns {WidenScalar, LLT::vector(3, 8)}
* getAction({G_ADD, LLT::vector(3, 8)})
then returns {MoreElements, LLT::vector(8, 8)}
* getAction({G_ADD, LLT::vector(20, 8)})
returns {FewerElements, LLT::vector(16, 8)}
2. Key implementation aspects.
How to legalize a specific (operation, type index, size) tuple is
represented by mapping intervals of integers representing a range of
size types to an action to take, e.g.:
setScalarAction({G_ADD, LLT:scalar(1)},
{{1, WidenScalar}, // bit sizes [ 1, 31[
{32, Legal}, // bit sizes [32, 33[
{33, WidenScalar}, // bit sizes [33, 64[
{64, Legal}, // bit sizes [64, 65[
{65, NarrowScalar} // bit sizes [65, +inf[
});
Please note that most of the code to do the actual lowering of
non-power-of-2 sized types is currently missing, this is just trying to
make it possible for targets to specify what is legal, and how non-legal
types should be legalized. Probably quite a bit of further work is
needed in the actual legalizing and the other passes in GlobalISel to
support non-power-of-2 sized types.
I hope the documentation in LegalizerInfo.h and the examples provided in the
various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well
enough how this is meant to be used.
This drops the need for LLT::{half,double}...Size().
Differential Revision: https://reviews.llvm.org/D30529
llvm-svn: 317560
Summary:
Calls using invoke in funclet based functions are assumed to clobber
all registers, which causes the stack adjustment using pops to consider
all registers not defined by the call to be undefined, which can
unfortunately include the base pointer, if one is needed.
To prevent this (and possibly other hazards), skip reserved registers
when looking for candidate registers.
This fixes issue #45034 in the Rust compiler.
Reviewers: mkuper
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D39636
llvm-svn: 317551
The EVEX to VEX pass is already assuming this is true under AVX512VL. We had special patterns to use zmm instructions if VLX and F16C weren't available.
Instead just make AVX512 imply F16C to make the EVEX to VEX behavior explicitly legal and remove the extra patterns.
All known CPUs with AVX512 have F16C so this should safe for now.
llvm-svn: 317521
Summary:
Print %subreg.<subregidxname> instead of just the subregister
index when printing immediate operands corresponding to subreg
indices in INSERT_SUBREG, EXTRACT_SUBREG, SUBREG_TO_REG and
REG_SEQUENCE.
Reviewers: qcolombet, MatzeB
Reviewed By: MatzeB
Subscribers: nhaehnle, javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D39696
llvm-svn: 317513
The backend assumes pointer in default addr space is 32 bit, which is not
true for the new addr space mapping and causes assertion for unresolved
functions.
This patch fixes that.
Differential Revision: https://reviews.llvm.org/D39643
llvm-svn: 317476
Added TESTM and TESTNM to the list of instructions that already zeroing unused upper bits
and does not need the redundant shift left and shift right instructions afterwards.
Added a pattern for TESTM and TESTNM in iselLowering, so now icmp(neq,and(X,Y), 0) goes folds into TESTM
and icmp(eq,and(X,Y), 0) goes folds into TESTNM
This commit is a preparation for lowering the test and testn X86 intrinsics to IR.
Differential Revision: https://reviews.llvm.org/D38732
llvm-svn: 317465
Summary:
Try to lower a BUILD_VECTOR composed of extract-extract chains that can be
reasoned to be a permutation of a vector by indices in a non-constant vector.
We saw this pattern created by ISPC, which resolts to creating it due to the
requirement that shufflevector's mask operand be a *constant* vector.
I didn't check this but we could possibly use this pattern for lowering the X86 permute
C-instrinsics instead of llvm.x86 instrinsics.
This change can be followed by more improvements:
1. Handle vectors with undef elements.
2. Utilize pshufb and zero-mask-blending to support more effiecient
construction of vectors with constant-0 elements.
3. Use smaller-element vectors of same width, and "interpolate" the indices,
when no native operation available.
Reviewers: RKSimon, craig.topper
Reviewed By: RKSimon
Subscribers: chandlerc, DavidKreitzer
Differential Revision: https://reviews.llvm.org/D39126
llvm-svn: 317463
This patch, together with a matching clang patch (https://reviews.llvm.org/D38683), implements the lowering of X86 broadcastm intrinsics to IR.
Differential Revision: https://reviews.llvm.org/D38684
Change-Id: I709ac0b34641095397e994c8ff7e15d1315b3540
llvm-svn: 317458
Use feature names instead of CPU names.
A future commit will add avx512vl command lines to demonstrate missed use of EVEX instructions.
llvm-svn: 317451
Summary:
AVX512 added RCP14 and RSQRT instructions which improve accuracy over the legacy RCP and RSQRT instruction, but not enough accuracy to remove the need for a Newton Raphson refinement.
Currently we use these new instructions for the legacy packed SSE instrinics, but not the scalar instrinsics. And we use it for fast math optimization of division and reciprocal sqrt.
I think switching the legacy instrinsics maybe surprising to the user since it changes the answer based on which processor you're using regardless of any fastmath settings. It's also weird that we did something different between scalar and packed.
As far at the reciprocal estimation, I think it creates unnecessary deltas in our output behavior (and prevents EVEX->VEX). A little playing around with gcc and icc and godbolt suggest they don't change which instructions they use here.
This patch adds new X86ISD nodes for the RCP14/RSQRT14 and uses those for the new intrinsics. Leaving the old intrinsics to use the old instructions.
Going forward I think our focus should be on
-Supporting 512-bit vectors, which will have to use the RCP14/RSQRT14.
-Using RSQRT28/RCP28 to remove the Newton Raphson step on processors with AVX512ER
-Supporting double precision.
Reviewers: zvi, DavidKreitzer, RKSimon
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D39583
llvm-svn: 317413
AMDGPULibFunc hardcodes address space values of the old address space mapping,
which causes invalid addrspacecast instructions and undefined functions in
APPSDK sample MonteCarloAsianDP.
This patch fixes that.
Differential Revision: https://reviews.llvm.org/D39616
llvm-svn: 317409
This allows masked operations to be used and allows the register allocator to use YMM16-31 if necessary.
As a follow up I'll look into teaching EVEX->VEX how to turn this back into PERM2X128 if any of the additional features don't work out.
llvm-svn: 317403
The number of iterations was incorrectly determined for DP FP vector types
and the tests were insufficient to flag this issue.
Differential revision: https://reviews.llvm.org/D39507
llvm-svn: 317349
Summary:
The current LICM allows sinking an instruction only when it is exposed to exit
blocks through a trivially replacable PHI of which all incoming values are the
same instruction. This change enhance LICM to sink a sinkable instruction
through non-trivially replacable PHIs by spliting predecessors of loop
exits.
Reviewers: hfinkel, majnemer, davidxl, bmakam, mcrosier, danielcdh, efriedma, jtony
Reviewed By: efriedma
Subscribers: nemanjai, dberlin, llvm-commits
Differential Revision: https://reviews.llvm.org/D37163
llvm-svn: 317335
Change the ISel matching of 'ins', 'dins[mu]' from tablegen code to
C++ code. This resolves an issue where ISel would select 'dins' instead
of 'dinsm' when the instructions size and position were individually in
range but their sum was out of range according to the ISA specification.
Reviewers: atanasyan
Differential Revision: https://reviews.llvm.org/D39117
llvm-svn: 317331
Similar to the existing code to lower to PACKSS, we can use PACKUS if the input vector's leading zero bits extend all the way to the packed/truncated value.
We have to account for pre-SSE41 targets not supporting PACKUSDW
llvm-svn: 317315
The GlobalISel TableGen backend didn't check for predicates on the
source children. This caused it to generate code for ARM patterns such
as SMLABB or similar, but without properly checking for the sext_16_node
part of the operands. This in turn meant that we would select SMLABB
instead of MLA for simple sequences such as s32 + s32 * s32, which is
wrong (we want a MLA on the full operands, not just their bottom 16
bits).
This patch forces TableGen to skip patterns with predicates on the src
children, so it doesn't generate code for SMLABB and other similar ARM
instructions at all anymore. AArch64 and X86 are not affected.
Differential Revision: https://reviews.llvm.org/D39554
llvm-svn: 317313
Ideally we should probably produce WinEH here as well, but until
then, we can use dwarf exceptions, without any further changes
required in clang, libunwind or libcxxabi.
Differential Revision: https://reviews.llvm.org/D39535
llvm-svn: 317304
This fixes http://llvm.org/PR32560. We were missing a description for
half floating point type and as a result were using the FPR 32 mapping.
Because of the size mismatch the generic code was complaining that the
default mapping is not appropriate. Fix the mapping description so that
the default mapping can be properly applied.
llvm-svn: 317287
undefined reference to `llvm::TargetPassConfig::ID' on
clang-ppc64le-linux-multistage
This reverts commit eea333c33fa73ad225ef28607795984829f65688.
llvm-svn: 317213
Summary:
This is mostly a noop (most of the test diffs are renamed blocks).
There are a few temporary register renames (eax<->ecx) and a few blocks are
shuffled around.
See the discussion in PR33325 for more details.
Reviewers: spatel
Subscribers: mgorny
Differential Revision: https://reviews.llvm.org/D39456
llvm-svn: 317211
When splitting a large load to smaller legally-typed loads, the last load should be padded to reach the size of the previous one so a CONCAT_VECTORS node could reunite them again.
The code currently pads the last load to reach the size of the first load (instead of the previous).
Differential Revision: https://reviews.llvm.org/D38495
Change-Id: Ib60b55ed26ce901fabf68108daf52683fbd5013f
llvm-svn: 317206
