1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-22 18:54:02 +01:00
llvm-mirror/unittests/Analysis/VectorFunctionABITest.cpp

621 lines
23 KiB
C++
Raw Normal View History

//===------- VectorFunctionABITest.cpp - VFABI Unittests ---------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
#include "llvm/Analysis/VectorUtils.h"
#include "llvm/AsmParser/Parser.h"
#include "llvm/IR/InstIterator.h"
#include "gtest/gtest.h"
using namespace llvm;
namespace {
// Test fixture needed that holds the veariables needed by the parser.
class VFABIParserTest : public ::testing::Test {
private:
// Parser output.
VFInfo Info;
[llvm][VectorUtils] Tweak VFShape for scalable vector functions. Summary: This patch makes sure that the field VFShape.VF is greater than zero when demangling the vector function name of scalable vector functions encoded in the "vector-function-abi-variant" attribute. This change is required to be able to provide instances of VFShape that can be used to query the VFDatabase for the vectorization passes, as such passes always require a positive value for the Vectorization Factor (VF) needed by the vectorization process. It is not possible to extract the value of VFShape.VF from the mangled name of scalable vector functions, because it is encoded as `x`. Therefore, the VFABI demangling function has been modified to extract such information from the IR declaration of the vector function, under the assumption that _all_ vectors in the signature of the vector function have the same number of lanes. Such assumption is valid because it is also assumed by the Vector Function ABI specifications supported by the demangling function (x86, AArch64, and LLVM internal one). The unit tests that demangle scalable names have been modified by adding the IR module that carries the declaration of the vector function name being demangled. In particular, the demangling function fails in the following cases: 1. When the declaration of the scalable vector function is not present in the module. 2. When the value of VFSHape.VF is not greater than 0. Reviewers: jdoerfert, sdesmalen, andwar Reviewed By: jdoerfert Subscribers: mgorny, kristof.beyls, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73286
2020-01-22 23:34:27 +01:00
// Reset the data needed for the test.
void reset(const StringRef Name, const StringRef IRType) {
M = parseAssemblyString("declare void @dummy()", Err, Ctx);
EXPECT_NE(M.get(), nullptr) << "Loading an invalid module.\n "
<< Err.getMessage() << "\n";
Type *Ty = parseType(IRType, Err, *(M.get()));
FunctionType *FTy = dyn_cast<FunctionType>(Ty);
EXPECT_NE(FTy, nullptr) << "Invalid function type string: " << IRType
<< "\n"
<< Err.getMessage() << "\n";
FunctionCallee F = M->getOrInsertFunction(Name, FTy);
EXPECT_NE(F.getCallee(), nullptr)
<< "The function must be present in the module\n";
// Reset the VFInfo
Info = VFInfo();
}
[llvm][VectorUtils] Tweak VFShape for scalable vector functions. Summary: This patch makes sure that the field VFShape.VF is greater than zero when demangling the vector function name of scalable vector functions encoded in the "vector-function-abi-variant" attribute. This change is required to be able to provide instances of VFShape that can be used to query the VFDatabase for the vectorization passes, as such passes always require a positive value for the Vectorization Factor (VF) needed by the vectorization process. It is not possible to extract the value of VFShape.VF from the mangled name of scalable vector functions, because it is encoded as `x`. Therefore, the VFABI demangling function has been modified to extract such information from the IR declaration of the vector function, under the assumption that _all_ vectors in the signature of the vector function have the same number of lanes. Such assumption is valid because it is also assumed by the Vector Function ABI specifications supported by the demangling function (x86, AArch64, and LLVM internal one). The unit tests that demangle scalable names have been modified by adding the IR module that carries the declaration of the vector function name being demangled. In particular, the demangling function fails in the following cases: 1. When the declaration of the scalable vector function is not present in the module. 2. When the value of VFSHape.VF is not greater than 0. Reviewers: jdoerfert, sdesmalen, andwar Reviewed By: jdoerfert Subscribers: mgorny, kristof.beyls, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73286
2020-01-22 23:34:27 +01:00
// Data needed to load the optional IR passed to invokeParser
LLVMContext Ctx;
SMDiagnostic Err;
std::unique_ptr<Module> M;
// CallInst *CI;
protected:
// Referencies to the parser output field.
unsigned &VF = Info.Shape.VF;
VFISAKind &ISA = Info.ISA;
SmallVector<VFParameter, 8> &Parameters = Info.Shape.Parameters;
std::string &ScalarName = Info.ScalarName;
std::string &VectorName = Info.VectorName;
bool &IsScalable = Info.Shape.IsScalable;
[llvm][VectorUtils] Tweak VFShape for scalable vector functions. Summary: This patch makes sure that the field VFShape.VF is greater than zero when demangling the vector function name of scalable vector functions encoded in the "vector-function-abi-variant" attribute. This change is required to be able to provide instances of VFShape that can be used to query the VFDatabase for the vectorization passes, as such passes always require a positive value for the Vectorization Factor (VF) needed by the vectorization process. It is not possible to extract the value of VFShape.VF from the mangled name of scalable vector functions, because it is encoded as `x`. Therefore, the VFABI demangling function has been modified to extract such information from the IR declaration of the vector function, under the assumption that _all_ vectors in the signature of the vector function have the same number of lanes. Such assumption is valid because it is also assumed by the Vector Function ABI specifications supported by the demangling function (x86, AArch64, and LLVM internal one). The unit tests that demangle scalable names have been modified by adding the IR module that carries the declaration of the vector function name being demangled. In particular, the demangling function fails in the following cases: 1. When the declaration of the scalable vector function is not present in the module. 2. When the value of VFSHape.VF is not greater than 0. Reviewers: jdoerfert, sdesmalen, andwar Reviewed By: jdoerfert Subscribers: mgorny, kristof.beyls, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73286
2020-01-22 23:34:27 +01:00
// Invoke the parser. We need to make sure that a function exist in
// the module because the parser fails if such function don't
// exists. Every time this method is invoked the state of the test
// is reset.
//
// \p MangledName -> the string the parser has to demangle.
//
// \p VectorName -> optional vector name that the method needs to
// use to create the function in the module if it differs from the
// standard mangled name.
//
// \p IRType -> FunctionType string to be used for the signature of
// the vector function. The correct signature is needed by the
// parser only for scalable functions. For the sake of testing, the
// generic fixed-length case can use as signature `void()`.
//
bool invokeParser(const StringRef MangledName,
const StringRef VectorName = "",
const StringRef IRType = "void()") {
StringRef Name = MangledName;
if (!VectorName.empty())
Name = VectorName;
// Reset the VFInfo and the Module to be able to invoke
// `invokeParser` multiple times in the same test.
reset(Name, IRType);
const auto OptInfo = VFABI::tryDemangleForVFABI(MangledName, *(M.get()));
if (OptInfo.hasValue()) {
Info = OptInfo.getValue();
return true;
}
return false;
}
// Checks that 1. the last Parameter in the Shape is of type
// VFParamKind::GlobalPredicate and 2. it is the only one of such
// type.
bool IsMasked() const {
const auto NGlobalPreds =
std::count_if(Info.Shape.Parameters.begin(),
Info.Shape.Parameters.end(), [](const VFParameter PK) {
return PK.ParamKind == VFParamKind::GlobalPredicate;
});
return NGlobalPreds == 1 && Info.Shape.Parameters.back().ParamKind ==
VFParamKind::GlobalPredicate;
}
};
} // unnamed namespace
// This test makes sure correct mangling occurs for given string.
TEST_F(VFABIParserTest, ManglingVectorTLINames) {
EXPECT_EQ(VFABI::mangleTLIVectorName("vec", "scalar", 3, 4),
"_ZGV_LLVM_N4vvv_scalar(vec)");
EXPECT_EQ(VFABI::mangleTLIVectorName("custom.call.v5", "custom.call", 1, 5),
"_ZGV_LLVM_N5v_custom.call(custom.call.v5)");
}
[llvm][VectorUtils] Tweak VFShape for scalable vector functions. Summary: This patch makes sure that the field VFShape.VF is greater than zero when demangling the vector function name of scalable vector functions encoded in the "vector-function-abi-variant" attribute. This change is required to be able to provide instances of VFShape that can be used to query the VFDatabase for the vectorization passes, as such passes always require a positive value for the Vectorization Factor (VF) needed by the vectorization process. It is not possible to extract the value of VFShape.VF from the mangled name of scalable vector functions, because it is encoded as `x`. Therefore, the VFABI demangling function has been modified to extract such information from the IR declaration of the vector function, under the assumption that _all_ vectors in the signature of the vector function have the same number of lanes. Such assumption is valid because it is also assumed by the Vector Function ABI specifications supported by the demangling function (x86, AArch64, and LLVM internal one). The unit tests that demangle scalable names have been modified by adding the IR module that carries the declaration of the vector function name being demangled. In particular, the demangling function fails in the following cases: 1. When the declaration of the scalable vector function is not present in the module. 2. When the value of VFSHape.VF is not greater than 0. Reviewers: jdoerfert, sdesmalen, andwar Reviewed By: jdoerfert Subscribers: mgorny, kristof.beyls, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73286
2020-01-22 23:34:27 +01:00
// This test makes sure that the demangling method succeeds only on
// valid values of the string.
TEST_F(VFABIParserTest, OnlyValidNames) {
// Incomplete string.
EXPECT_FALSE(invokeParser(""));
EXPECT_FALSE(invokeParser("_ZGV"));
EXPECT_FALSE(invokeParser("_ZGVn"));
EXPECT_FALSE(invokeParser("_ZGVnN"));
EXPECT_FALSE(invokeParser("_ZGVnN2"));
EXPECT_FALSE(invokeParser("_ZGVnN2v"));
EXPECT_FALSE(invokeParser("_ZGVnN2v_"));
// Missing parameters.
EXPECT_FALSE(invokeParser("_ZGVnN2_foo"));
// Missing _ZGV prefix.
EXPECT_FALSE(invokeParser("_ZVnN2v_foo"));
// Missing <isa>.
EXPECT_FALSE(invokeParser("_ZGVN2v_foo"));
// Missing <mask>.
EXPECT_FALSE(invokeParser("_ZGVn2v_foo"));
// Missing <vlen>.
EXPECT_FALSE(invokeParser("_ZGVnNv_foo"));
// Missing <scalarname>.
EXPECT_FALSE(invokeParser("_ZGVnN2v_"));
// Missing _ separator.
EXPECT_FALSE(invokeParser("_ZGVnN2vfoo"));
// Missing <vectorname>. Using `fakename` because the string being
// parsed is not a valid function name that `invokeParser` can add.
EXPECT_FALSE(invokeParser("_ZGVnN2v_foo()", "fakename"));
// Unterminated name. Using `fakename` because the string being
// parsed is not a valid function name that `invokeParser` can add.
EXPECT_FALSE(invokeParser("_ZGVnN2v_foo(bar", "fakename"));
}
TEST_F(VFABIParserTest, ParamListParsing) {
EXPECT_TRUE(invokeParser("_ZGVnN2vl16Ls32R3l_foo"));
EXPECT_EQ(Parameters.size(), (unsigned)5);
EXPECT_EQ(Parameters[0], VFParameter({0, VFParamKind::Vector, 0}));
EXPECT_EQ(Parameters[1], VFParameter({1, VFParamKind::OMP_Linear, 16}));
EXPECT_EQ(Parameters[2], VFParameter({2, VFParamKind::OMP_LinearValPos, 32}));
EXPECT_EQ(Parameters[3], VFParameter({3, VFParamKind::OMP_LinearRef, 3}));
EXPECT_EQ(Parameters[4], VFParameter({4, VFParamKind::OMP_Linear, 1}));
}
TEST_F(VFABIParserTest, ScalarNameAndVectorName_01) {
EXPECT_TRUE(invokeParser("_ZGVnM2v_sin"));
EXPECT_EQ(ScalarName, "sin");
EXPECT_EQ(VectorName, "_ZGVnM2v_sin");
}
TEST_F(VFABIParserTest, ScalarNameAndVectorName_02) {
EXPECT_TRUE(invokeParser("_ZGVnM2v_sin(UserFunc)", "UserFunc"));
EXPECT_EQ(ScalarName, "sin");
EXPECT_EQ(VectorName, "UserFunc");
}
TEST_F(VFABIParserTest, ScalarNameAndVectorName_03) {
EXPECT_TRUE(invokeParser("_ZGVnM2v___sin_sin_sin"));
EXPECT_EQ(ScalarName, "__sin_sin_sin");
EXPECT_EQ(VectorName, "_ZGVnM2v___sin_sin_sin");
}
TEST_F(VFABIParserTest, Parse) {
EXPECT_TRUE(invokeParser("_ZGVnN2vls2Ls27Us4Rs5l1L10U100R1000_sin"));
EXPECT_EQ(VF, (unsigned)2);
EXPECT_FALSE(IsMasked());
EXPECT_EQ(ISA, VFISAKind::AdvancedSIMD);
EXPECT_FALSE(IsScalable);
EXPECT_EQ(Parameters.size(), (unsigned)9);
EXPECT_EQ(Parameters[0], VFParameter({0, VFParamKind::Vector, 0}));
EXPECT_EQ(Parameters[1], VFParameter({1, VFParamKind::OMP_LinearPos, 2}));
EXPECT_EQ(Parameters[2], VFParameter({2, VFParamKind::OMP_LinearValPos, 27}));
EXPECT_EQ(Parameters[3], VFParameter({3, VFParamKind::OMP_LinearUValPos, 4}));
EXPECT_EQ(Parameters[4], VFParameter({4, VFParamKind::OMP_LinearRefPos, 5}));
EXPECT_EQ(Parameters[5], VFParameter({5, VFParamKind::OMP_Linear, 1}));
EXPECT_EQ(Parameters[6], VFParameter({6, VFParamKind::OMP_LinearVal, 10}));
EXPECT_EQ(Parameters[7], VFParameter({7, VFParamKind::OMP_LinearUVal, 100}));
EXPECT_EQ(Parameters[8], VFParameter({8, VFParamKind::OMP_LinearRef, 1000}));
EXPECT_EQ(ScalarName, "sin");
EXPECT_EQ(VectorName, "_ZGVnN2vls2Ls27Us4Rs5l1L10U100R1000_sin");
}
TEST_F(VFABIParserTest, ParseVectorName) {
[llvm][VectorUtils] Tweak VFShape for scalable vector functions. Summary: This patch makes sure that the field VFShape.VF is greater than zero when demangling the vector function name of scalable vector functions encoded in the "vector-function-abi-variant" attribute. This change is required to be able to provide instances of VFShape that can be used to query the VFDatabase for the vectorization passes, as such passes always require a positive value for the Vectorization Factor (VF) needed by the vectorization process. It is not possible to extract the value of VFShape.VF from the mangled name of scalable vector functions, because it is encoded as `x`. Therefore, the VFABI demangling function has been modified to extract such information from the IR declaration of the vector function, under the assumption that _all_ vectors in the signature of the vector function have the same number of lanes. Such assumption is valid because it is also assumed by the Vector Function ABI specifications supported by the demangling function (x86, AArch64, and LLVM internal one). The unit tests that demangle scalable names have been modified by adding the IR module that carries the declaration of the vector function name being demangled. In particular, the demangling function fails in the following cases: 1. When the declaration of the scalable vector function is not present in the module. 2. When the value of VFSHape.VF is not greater than 0. Reviewers: jdoerfert, sdesmalen, andwar Reviewed By: jdoerfert Subscribers: mgorny, kristof.beyls, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73286
2020-01-22 23:34:27 +01:00
EXPECT_TRUE(invokeParser("_ZGVnN2v_sin(my_v_sin)", "my_v_sin"));
EXPECT_EQ(VF, (unsigned)2);
EXPECT_FALSE(IsMasked());
EXPECT_FALSE(IsScalable);
EXPECT_EQ(ISA, VFISAKind::AdvancedSIMD);
EXPECT_EQ(Parameters.size(), (unsigned)1);
EXPECT_EQ(Parameters[0], VFParameter({0, VFParamKind::Vector, 0}));
EXPECT_EQ(ScalarName, "sin");
EXPECT_EQ(VectorName, "my_v_sin");
}
TEST_F(VFABIParserTest, LinearWithCompileTimeNegativeStep) {
EXPECT_TRUE(invokeParser("_ZGVnN2ln1Ln10Un100Rn1000_sin"));
EXPECT_EQ(VF, (unsigned)2);
EXPECT_FALSE(IsMasked());
EXPECT_EQ(ISA, VFISAKind::AdvancedSIMD);
EXPECT_FALSE(IsScalable);
EXPECT_EQ(Parameters.size(), (unsigned)4);
EXPECT_EQ(Parameters[0], VFParameter({0, VFParamKind::OMP_Linear, -1}));
EXPECT_EQ(Parameters[1], VFParameter({1, VFParamKind::OMP_LinearVal, -10}));
EXPECT_EQ(Parameters[2], VFParameter({2, VFParamKind::OMP_LinearUVal, -100}));
EXPECT_EQ(Parameters[3], VFParameter({3, VFParamKind::OMP_LinearRef, -1000}));
EXPECT_EQ(ScalarName, "sin");
EXPECT_EQ(VectorName, "_ZGVnN2ln1Ln10Un100Rn1000_sin");
}
TEST_F(VFABIParserTest, ParseScalableSVE) {
[llvm][VectorUtils] Tweak VFShape for scalable vector functions. Summary: This patch makes sure that the field VFShape.VF is greater than zero when demangling the vector function name of scalable vector functions encoded in the "vector-function-abi-variant" attribute. This change is required to be able to provide instances of VFShape that can be used to query the VFDatabase for the vectorization passes, as such passes always require a positive value for the Vectorization Factor (VF) needed by the vectorization process. It is not possible to extract the value of VFShape.VF from the mangled name of scalable vector functions, because it is encoded as `x`. Therefore, the VFABI demangling function has been modified to extract such information from the IR declaration of the vector function, under the assumption that _all_ vectors in the signature of the vector function have the same number of lanes. Such assumption is valid because it is also assumed by the Vector Function ABI specifications supported by the demangling function (x86, AArch64, and LLVM internal one). The unit tests that demangle scalable names have been modified by adding the IR module that carries the declaration of the vector function name being demangled. In particular, the demangling function fails in the following cases: 1. When the declaration of the scalable vector function is not present in the module. 2. When the value of VFSHape.VF is not greater than 0. Reviewers: jdoerfert, sdesmalen, andwar Reviewed By: jdoerfert Subscribers: mgorny, kristof.beyls, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73286
2020-01-22 23:34:27 +01:00
EXPECT_TRUE(invokeParser(
"_ZGVsMxv_sin(custom_vg)", "custom_vg",
"<vscale x 2 x i32>(<vscale x 2 x i32>, <vscale x 2 x i1>)"));
EXPECT_EQ(VF, (unsigned)2);
EXPECT_TRUE(IsMasked());
EXPECT_TRUE(IsScalable);
EXPECT_EQ(ISA, VFISAKind::SVE);
EXPECT_EQ(ScalarName, "sin");
[llvm][VectorUtils] Tweak VFShape for scalable vector functions. Summary: This patch makes sure that the field VFShape.VF is greater than zero when demangling the vector function name of scalable vector functions encoded in the "vector-function-abi-variant" attribute. This change is required to be able to provide instances of VFShape that can be used to query the VFDatabase for the vectorization passes, as such passes always require a positive value for the Vectorization Factor (VF) needed by the vectorization process. It is not possible to extract the value of VFShape.VF from the mangled name of scalable vector functions, because it is encoded as `x`. Therefore, the VFABI demangling function has been modified to extract such information from the IR declaration of the vector function, under the assumption that _all_ vectors in the signature of the vector function have the same number of lanes. Such assumption is valid because it is also assumed by the Vector Function ABI specifications supported by the demangling function (x86, AArch64, and LLVM internal one). The unit tests that demangle scalable names have been modified by adding the IR module that carries the declaration of the vector function name being demangled. In particular, the demangling function fails in the following cases: 1. When the declaration of the scalable vector function is not present in the module. 2. When the value of VFSHape.VF is not greater than 0. Reviewers: jdoerfert, sdesmalen, andwar Reviewed By: jdoerfert Subscribers: mgorny, kristof.beyls, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73286
2020-01-22 23:34:27 +01:00
EXPECT_EQ(VectorName, "custom_vg");
}
TEST_F(VFABIParserTest, ParseFixedWidthSVE) {
EXPECT_TRUE(invokeParser("_ZGVsM2v_sin"));
EXPECT_EQ(VF, (unsigned)2);
EXPECT_TRUE(IsMasked());
EXPECT_FALSE(IsScalable);
EXPECT_EQ(ISA, VFISAKind::SVE);
EXPECT_EQ(ScalarName, "sin");
EXPECT_EQ(VectorName, "_ZGVsM2v_sin");
}
TEST_F(VFABIParserTest, NotAVectorFunctionABIName) {
// Vector names should start with `_ZGV`.
EXPECT_FALSE(invokeParser("ZGVnN2v_sin"));
}
TEST_F(VFABIParserTest, LinearWithRuntimeStep) {
EXPECT_FALSE(invokeParser("_ZGVnN2ls_sin"))
<< "A number should be present after \"ls\".";
EXPECT_TRUE(invokeParser("_ZGVnN2ls2_sin"));
EXPECT_FALSE(invokeParser("_ZGVnN2Rs_sin"))
<< "A number should be present after \"Rs\".";
EXPECT_TRUE(invokeParser("_ZGVnN2Rs4_sin"));
EXPECT_FALSE(invokeParser("_ZGVnN2Ls_sin"))
<< "A number should be present after \"Ls\".";
EXPECT_TRUE(invokeParser("_ZGVnN2Ls6_sin"));
EXPECT_FALSE(invokeParser("_ZGVnN2Us_sin"))
<< "A number should be present after \"Us\".";
EXPECT_TRUE(invokeParser("_ZGVnN2Us8_sin"));
}
TEST_F(VFABIParserTest, LinearWithoutCompileTime) {
EXPECT_TRUE(invokeParser("_ZGVnN3lLRUlnLnRnUn_sin"));
EXPECT_EQ(Parameters.size(), (unsigned)8);
EXPECT_EQ(Parameters[0], VFParameter({0, VFParamKind::OMP_Linear, 1}));
EXPECT_EQ(Parameters[1], VFParameter({1, VFParamKind::OMP_LinearVal, 1}));
EXPECT_EQ(Parameters[2], VFParameter({2, VFParamKind::OMP_LinearRef, 1}));
EXPECT_EQ(Parameters[3], VFParameter({3, VFParamKind::OMP_LinearUVal, 1}));
EXPECT_EQ(Parameters[4], VFParameter({4, VFParamKind::OMP_Linear, -1}));
EXPECT_EQ(Parameters[5], VFParameter({5, VFParamKind::OMP_LinearVal, -1}));
EXPECT_EQ(Parameters[6], VFParameter({6, VFParamKind::OMP_LinearRef, -1}));
EXPECT_EQ(Parameters[7], VFParameter({7, VFParamKind::OMP_LinearUVal, -1}));
}
TEST_F(VFABIParserTest, ISA) {
EXPECT_TRUE(invokeParser("_ZGVqN2v_sin"));
EXPECT_EQ(ISA, VFISAKind::Unknown);
EXPECT_TRUE(invokeParser("_ZGVnN2v_sin"));
EXPECT_EQ(ISA, VFISAKind::AdvancedSIMD);
EXPECT_TRUE(invokeParser("_ZGVsN2v_sin"));
EXPECT_EQ(ISA, VFISAKind::SVE);
EXPECT_TRUE(invokeParser("_ZGVbN2v_sin"));
EXPECT_EQ(ISA, VFISAKind::SSE);
EXPECT_TRUE(invokeParser("_ZGVcN2v_sin"));
EXPECT_EQ(ISA, VFISAKind::AVX);
EXPECT_TRUE(invokeParser("_ZGVdN2v_sin"));
EXPECT_EQ(ISA, VFISAKind::AVX2);
EXPECT_TRUE(invokeParser("_ZGVeN2v_sin"));
EXPECT_EQ(ISA, VFISAKind::AVX512);
}
TEST_F(VFABIParserTest, LLVM_ISA) {
EXPECT_FALSE(invokeParser("_ZGV_LLVM_N2v_sin"));
[llvm][VectorUtils] Tweak VFShape for scalable vector functions. Summary: This patch makes sure that the field VFShape.VF is greater than zero when demangling the vector function name of scalable vector functions encoded in the "vector-function-abi-variant" attribute. This change is required to be able to provide instances of VFShape that can be used to query the VFDatabase for the vectorization passes, as such passes always require a positive value for the Vectorization Factor (VF) needed by the vectorization process. It is not possible to extract the value of VFShape.VF from the mangled name of scalable vector functions, because it is encoded as `x`. Therefore, the VFABI demangling function has been modified to extract such information from the IR declaration of the vector function, under the assumption that _all_ vectors in the signature of the vector function have the same number of lanes. Such assumption is valid because it is also assumed by the Vector Function ABI specifications supported by the demangling function (x86, AArch64, and LLVM internal one). The unit tests that demangle scalable names have been modified by adding the IR module that carries the declaration of the vector function name being demangled. In particular, the demangling function fails in the following cases: 1. When the declaration of the scalable vector function is not present in the module. 2. When the value of VFSHape.VF is not greater than 0. Reviewers: jdoerfert, sdesmalen, andwar Reviewed By: jdoerfert Subscribers: mgorny, kristof.beyls, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73286
2020-01-22 23:34:27 +01:00
EXPECT_TRUE(invokeParser("_ZGV_LLVM_N2v_sin_(vector_name)", "vector_name"));
EXPECT_EQ(ISA, VFISAKind::LLVM);
}
TEST_F(VFABIParserTest, InvalidMask) {
EXPECT_FALSE(invokeParser("_ZGVsK2v_sin"));
}
TEST_F(VFABIParserTest, InvalidParameter) {
EXPECT_FALSE(invokeParser("_ZGVsM2vX_sin"));
}
TEST_F(VFABIParserTest, Align) {
EXPECT_TRUE(invokeParser("_ZGVsN2l2a2_sin"));
EXPECT_EQ(Parameters.size(), (unsigned)1);
EXPECT_EQ(Parameters[0].Alignment, Align(2));
// Missing alignment value.
EXPECT_FALSE(invokeParser("_ZGVsM2l2a_sin"));
// Invalid alignment token "x".
EXPECT_FALSE(invokeParser("_ZGVsM2l2ax_sin"));
// Alignment MUST be associated to a paramater.
EXPECT_FALSE(invokeParser("_ZGVsM2a2_sin"));
// Alignment must be a power of 2.
EXPECT_FALSE(invokeParser("_ZGVsN2l2a0_sin"));
EXPECT_TRUE(invokeParser("_ZGVsN2l2a1_sin"));
EXPECT_FALSE(invokeParser("_ZGVsN2l2a3_sin"));
EXPECT_FALSE(invokeParser("_ZGVsN2l2a6_sin"));
}
TEST_F(VFABIParserTest, ParseUniform) {
EXPECT_TRUE(invokeParser("_ZGVnN2u_sin"));
EXPECT_EQ(VF, (unsigned)2);
EXPECT_FALSE(IsMasked());
EXPECT_EQ(ISA, VFISAKind::AdvancedSIMD);
EXPECT_FALSE(IsScalable);
EXPECT_EQ(Parameters.size(), (unsigned)1);
EXPECT_EQ(Parameters[0], VFParameter({0, VFParamKind::OMP_Uniform, 0}));
EXPECT_EQ(ScalarName, "sin");
EXPECT_EQ(VectorName, "_ZGVnN2u_sin");
// Uniform doesn't expect extra data.
EXPECT_FALSE(invokeParser("_ZGVnN2u0_sin"));
}
TEST_F(VFABIParserTest, ISAIndependentMangling) {
// This test makes sure that the mangling of the parameters in
// independent on the <isa> token.
const SmallVector<VFParameter, 8> ExpectedParams = {
VFParameter({0, VFParamKind::Vector, 0}),
VFParameter({1, VFParamKind::OMP_LinearPos, 2}),
VFParameter({2, VFParamKind::OMP_LinearValPos, 27}),
VFParameter({3, VFParamKind::OMP_LinearUValPos, 4}),
VFParameter({4, VFParamKind::OMP_LinearRefPos, 5}),
VFParameter({5, VFParamKind::OMP_Linear, 1}),
VFParameter({6, VFParamKind::OMP_LinearVal, 10}),
VFParameter({7, VFParamKind::OMP_LinearUVal, 100}),
VFParameter({8, VFParamKind::OMP_LinearRef, 1000}),
VFParameter({9, VFParamKind::OMP_Uniform, 0}),
};
#define __COMMON_CHECKS \
do { \
EXPECT_EQ(VF, (unsigned)2); \
EXPECT_FALSE(IsMasked()); \
EXPECT_FALSE(IsScalable); \
EXPECT_EQ(Parameters.size(), (unsigned)10); \
EXPECT_EQ(Parameters, ExpectedParams); \
EXPECT_EQ(ScalarName, "sin"); \
} while (0)
// Advanced SIMD: <isa> = "n"
EXPECT_TRUE(invokeParser("_ZGVnN2vls2Ls27Us4Rs5l1L10U100R1000u_sin"));
EXPECT_EQ(ISA, VFISAKind::AdvancedSIMD);
__COMMON_CHECKS;
EXPECT_EQ(VectorName, "_ZGVnN2vls2Ls27Us4Rs5l1L10U100R1000u_sin");
// SVE: <isa> = "s"
EXPECT_TRUE(invokeParser("_ZGVsN2vls2Ls27Us4Rs5l1L10U100R1000u_sin"));
EXPECT_EQ(ISA, VFISAKind::SVE);
__COMMON_CHECKS;
EXPECT_EQ(VectorName, "_ZGVsN2vls2Ls27Us4Rs5l1L10U100R1000u_sin");
// SSE: <isa> = "b"
EXPECT_TRUE(invokeParser("_ZGVbN2vls2Ls27Us4Rs5l1L10U100R1000u_sin"));
EXPECT_EQ(ISA, VFISAKind::SSE);
__COMMON_CHECKS;
EXPECT_EQ(VectorName, "_ZGVbN2vls2Ls27Us4Rs5l1L10U100R1000u_sin");
// AVX: <isa> = "c"
EXPECT_TRUE(invokeParser("_ZGVcN2vls2Ls27Us4Rs5l1L10U100R1000u_sin"));
EXPECT_EQ(ISA, VFISAKind::AVX);
__COMMON_CHECKS;
EXPECT_EQ(VectorName, "_ZGVcN2vls2Ls27Us4Rs5l1L10U100R1000u_sin");
// AVX2: <isa> = "d"
EXPECT_TRUE(invokeParser("_ZGVdN2vls2Ls27Us4Rs5l1L10U100R1000u_sin"));
EXPECT_EQ(ISA, VFISAKind::AVX2);
__COMMON_CHECKS;
EXPECT_EQ(VectorName, "_ZGVdN2vls2Ls27Us4Rs5l1L10U100R1000u_sin");
// AVX512: <isa> = "e"
EXPECT_TRUE(invokeParser("_ZGVeN2vls2Ls27Us4Rs5l1L10U100R1000u_sin"));
EXPECT_EQ(ISA, VFISAKind::AVX512);
__COMMON_CHECKS;
EXPECT_EQ(VectorName, "_ZGVeN2vls2Ls27Us4Rs5l1L10U100R1000u_sin");
// LLVM: <isa> = "_LLVM_" internal vector function.
[llvm][VectorUtils] Tweak VFShape for scalable vector functions. Summary: This patch makes sure that the field VFShape.VF is greater than zero when demangling the vector function name of scalable vector functions encoded in the "vector-function-abi-variant" attribute. This change is required to be able to provide instances of VFShape that can be used to query the VFDatabase for the vectorization passes, as such passes always require a positive value for the Vectorization Factor (VF) needed by the vectorization process. It is not possible to extract the value of VFShape.VF from the mangled name of scalable vector functions, because it is encoded as `x`. Therefore, the VFABI demangling function has been modified to extract such information from the IR declaration of the vector function, under the assumption that _all_ vectors in the signature of the vector function have the same number of lanes. Such assumption is valid because it is also assumed by the Vector Function ABI specifications supported by the demangling function (x86, AArch64, and LLVM internal one). The unit tests that demangle scalable names have been modified by adding the IR module that carries the declaration of the vector function name being demangled. In particular, the demangling function fails in the following cases: 1. When the declaration of the scalable vector function is not present in the module. 2. When the value of VFSHape.VF is not greater than 0. Reviewers: jdoerfert, sdesmalen, andwar Reviewed By: jdoerfert Subscribers: mgorny, kristof.beyls, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73286
2020-01-22 23:34:27 +01:00
EXPECT_TRUE(invokeParser(
"_ZGV_LLVM_N2vls2Ls27Us4Rs5l1L10U100R1000u_sin(vectorf)", "vectorf"));
EXPECT_EQ(ISA, VFISAKind::LLVM);
__COMMON_CHECKS;
EXPECT_EQ(VectorName, "vectorf");
// Unknown ISA (randomly using "q"). This test will need update if
// some targets decide to use "q" as their ISA token.
EXPECT_TRUE(invokeParser("_ZGVqN2vls2Ls27Us4Rs5l1L10U100R1000u_sin"));
EXPECT_EQ(ISA, VFISAKind::Unknown);
__COMMON_CHECKS;
EXPECT_EQ(VectorName, "_ZGVqN2vls2Ls27Us4Rs5l1L10U100R1000u_sin");
#undef __COMMON_CHECKS
}
TEST_F(VFABIParserTest, MissingScalarName) {
EXPECT_FALSE(invokeParser("_ZGVnN2v_"));
}
TEST_F(VFABIParserTest, MissingVectorName) {
EXPECT_FALSE(invokeParser("_ZGVnN2v_foo()"));
}
TEST_F(VFABIParserTest, MissingVectorNameTermination) {
EXPECT_FALSE(invokeParser("_ZGVnN2v_foo(bar"));
}
TEST_F(VFABIParserTest, ParseMaskingNEON) {
EXPECT_TRUE(invokeParser("_ZGVnM2v_sin"));
EXPECT_EQ(VF, (unsigned)2);
EXPECT_TRUE(IsMasked());
EXPECT_FALSE(IsScalable);
EXPECT_EQ(ISA, VFISAKind::AdvancedSIMD);
EXPECT_EQ(Parameters.size(), (unsigned)2);
EXPECT_EQ(Parameters[0], VFParameter({0, VFParamKind::Vector}));
EXPECT_EQ(Parameters[1], VFParameter({1, VFParamKind::GlobalPredicate}));
EXPECT_EQ(ScalarName, "sin");
}
TEST_F(VFABIParserTest, ParseMaskingSVE) {
EXPECT_TRUE(invokeParser("_ZGVsM2v_sin"));
EXPECT_EQ(VF, (unsigned)2);
EXPECT_TRUE(IsMasked());
EXPECT_FALSE(IsScalable);
EXPECT_EQ(ISA, VFISAKind::SVE);
EXPECT_EQ(Parameters.size(), (unsigned)2);
EXPECT_EQ(Parameters[0], VFParameter({0, VFParamKind::Vector}));
EXPECT_EQ(Parameters[1], VFParameter({1, VFParamKind::GlobalPredicate}));
EXPECT_EQ(ScalarName, "sin");
}
TEST_F(VFABIParserTest, ParseMaskingSSE) {
EXPECT_TRUE(invokeParser("_ZGVbM2v_sin"));
EXPECT_EQ(VF, (unsigned)2);
EXPECT_TRUE(IsMasked());
EXPECT_FALSE(IsScalable);
EXPECT_EQ(ISA, VFISAKind::SSE);
EXPECT_EQ(Parameters.size(), (unsigned)2);
EXPECT_EQ(Parameters[0], VFParameter({0, VFParamKind::Vector}));
EXPECT_EQ(Parameters[1], VFParameter({1, VFParamKind::GlobalPredicate}));
EXPECT_EQ(ScalarName, "sin");
}
TEST_F(VFABIParserTest, ParseMaskingAVX) {
EXPECT_TRUE(invokeParser("_ZGVcM2v_sin"));
EXPECT_EQ(VF, (unsigned)2);
EXPECT_TRUE(IsMasked());
EXPECT_FALSE(IsScalable);
EXPECT_EQ(ISA, VFISAKind::AVX);
EXPECT_EQ(Parameters.size(), (unsigned)2);
EXPECT_EQ(Parameters[0], VFParameter({0, VFParamKind::Vector}));
EXPECT_EQ(Parameters[1], VFParameter({1, VFParamKind::GlobalPredicate}));
EXPECT_EQ(ScalarName, "sin");
}
TEST_F(VFABIParserTest, ParseMaskingAVX2) {
EXPECT_TRUE(invokeParser("_ZGVdM2v_sin"));
EXPECT_EQ(VF, (unsigned)2);
EXPECT_TRUE(IsMasked());
EXPECT_FALSE(IsScalable);
EXPECT_EQ(ISA, VFISAKind::AVX2);
EXPECT_EQ(Parameters.size(), (unsigned)2);
EXPECT_EQ(Parameters[0], VFParameter({0, VFParamKind::Vector}));
EXPECT_EQ(Parameters[1], VFParameter({1, VFParamKind::GlobalPredicate}));
EXPECT_EQ(ScalarName, "sin");
}
TEST_F(VFABIParserTest, ParseMaskingAVX512) {
EXPECT_TRUE(invokeParser("_ZGVeM2v_sin"));
EXPECT_EQ(VF, (unsigned)2);
EXPECT_TRUE(IsMasked());
EXPECT_FALSE(IsScalable);
EXPECT_EQ(ISA, VFISAKind::AVX512);
EXPECT_EQ(Parameters.size(), (unsigned)2);
EXPECT_EQ(Parameters[0], VFParameter({0, VFParamKind::Vector}));
EXPECT_EQ(Parameters[1], VFParameter({1, VFParamKind::GlobalPredicate}));
EXPECT_EQ(ScalarName, "sin");
}
TEST_F(VFABIParserTest, ParseMaskingLLVM) {
[llvm][VectorUtils] Tweak VFShape for scalable vector functions. Summary: This patch makes sure that the field VFShape.VF is greater than zero when demangling the vector function name of scalable vector functions encoded in the "vector-function-abi-variant" attribute. This change is required to be able to provide instances of VFShape that can be used to query the VFDatabase for the vectorization passes, as such passes always require a positive value for the Vectorization Factor (VF) needed by the vectorization process. It is not possible to extract the value of VFShape.VF from the mangled name of scalable vector functions, because it is encoded as `x`. Therefore, the VFABI demangling function has been modified to extract such information from the IR declaration of the vector function, under the assumption that _all_ vectors in the signature of the vector function have the same number of lanes. Such assumption is valid because it is also assumed by the Vector Function ABI specifications supported by the demangling function (x86, AArch64, and LLVM internal one). The unit tests that demangle scalable names have been modified by adding the IR module that carries the declaration of the vector function name being demangled. In particular, the demangling function fails in the following cases: 1. When the declaration of the scalable vector function is not present in the module. 2. When the value of VFSHape.VF is not greater than 0. Reviewers: jdoerfert, sdesmalen, andwar Reviewed By: jdoerfert Subscribers: mgorny, kristof.beyls, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73286
2020-01-22 23:34:27 +01:00
EXPECT_TRUE(invokeParser("_ZGV_LLVM_M2v_sin(custom_vector_sin)",
"custom_vector_sin"));
EXPECT_EQ(VF, (unsigned)2);
EXPECT_TRUE(IsMasked());
EXPECT_FALSE(IsScalable);
EXPECT_EQ(ISA, VFISAKind::LLVM);
EXPECT_EQ(Parameters.size(), (unsigned)2);
EXPECT_EQ(Parameters[0], VFParameter({0, VFParamKind::Vector}));
EXPECT_EQ(Parameters[1], VFParameter({1, VFParamKind::GlobalPredicate}));
EXPECT_EQ(ScalarName, "sin");
EXPECT_EQ(VectorName, "custom_vector_sin");
}
TEST_F(VFABIParserTest, ParseScalableMaskingLLVM) {
[llvm][VectorUtils] Tweak VFShape for scalable vector functions. Summary: This patch makes sure that the field VFShape.VF is greater than zero when demangling the vector function name of scalable vector functions encoded in the "vector-function-abi-variant" attribute. This change is required to be able to provide instances of VFShape that can be used to query the VFDatabase for the vectorization passes, as such passes always require a positive value for the Vectorization Factor (VF) needed by the vectorization process. It is not possible to extract the value of VFShape.VF from the mangled name of scalable vector functions, because it is encoded as `x`. Therefore, the VFABI demangling function has been modified to extract such information from the IR declaration of the vector function, under the assumption that _all_ vectors in the signature of the vector function have the same number of lanes. Such assumption is valid because it is also assumed by the Vector Function ABI specifications supported by the demangling function (x86, AArch64, and LLVM internal one). The unit tests that demangle scalable names have been modified by adding the IR module that carries the declaration of the vector function name being demangled. In particular, the demangling function fails in the following cases: 1. When the declaration of the scalable vector function is not present in the module. 2. When the value of VFSHape.VF is not greater than 0. Reviewers: jdoerfert, sdesmalen, andwar Reviewed By: jdoerfert Subscribers: mgorny, kristof.beyls, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73286
2020-01-22 23:34:27 +01:00
EXPECT_TRUE(invokeParser(
"_ZGV_LLVM_Mxv_sin(custom_vector_sin)", "custom_vector_sin",
"<vscale x 2 x i32> (<vscale x 2 x i32>, <vscale x 2 x i1>)"));
EXPECT_TRUE(IsMasked());
[llvm][VectorUtils] Tweak VFShape for scalable vector functions. Summary: This patch makes sure that the field VFShape.VF is greater than zero when demangling the vector function name of scalable vector functions encoded in the "vector-function-abi-variant" attribute. This change is required to be able to provide instances of VFShape that can be used to query the VFDatabase for the vectorization passes, as such passes always require a positive value for the Vectorization Factor (VF) needed by the vectorization process. It is not possible to extract the value of VFShape.VF from the mangled name of scalable vector functions, because it is encoded as `x`. Therefore, the VFABI demangling function has been modified to extract such information from the IR declaration of the vector function, under the assumption that _all_ vectors in the signature of the vector function have the same number of lanes. Such assumption is valid because it is also assumed by the Vector Function ABI specifications supported by the demangling function (x86, AArch64, and LLVM internal one). The unit tests that demangle scalable names have been modified by adding the IR module that carries the declaration of the vector function name being demangled. In particular, the demangling function fails in the following cases: 1. When the declaration of the scalable vector function is not present in the module. 2. When the value of VFSHape.VF is not greater than 0. Reviewers: jdoerfert, sdesmalen, andwar Reviewed By: jdoerfert Subscribers: mgorny, kristof.beyls, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73286
2020-01-22 23:34:27 +01:00
EXPECT_EQ(VF, (unsigned)2);
EXPECT_TRUE(IsScalable);
EXPECT_EQ(ISA, VFISAKind::LLVM);
EXPECT_EQ(Parameters.size(), (unsigned)2);
EXPECT_EQ(Parameters[0], VFParameter({0, VFParamKind::Vector}));
EXPECT_EQ(Parameters[1], VFParameter({1, VFParamKind::GlobalPredicate}));
EXPECT_EQ(ScalarName, "sin");
EXPECT_EQ(VectorName, "custom_vector_sin");
}
TEST_F(VFABIParserTest, ParseScalableMaskingLLVMSincos) {
[llvm][VectorUtils] Tweak VFShape for scalable vector functions. Summary: This patch makes sure that the field VFShape.VF is greater than zero when demangling the vector function name of scalable vector functions encoded in the "vector-function-abi-variant" attribute. This change is required to be able to provide instances of VFShape that can be used to query the VFDatabase for the vectorization passes, as such passes always require a positive value for the Vectorization Factor (VF) needed by the vectorization process. It is not possible to extract the value of VFShape.VF from the mangled name of scalable vector functions, because it is encoded as `x`. Therefore, the VFABI demangling function has been modified to extract such information from the IR declaration of the vector function, under the assumption that _all_ vectors in the signature of the vector function have the same number of lanes. Such assumption is valid because it is also assumed by the Vector Function ABI specifications supported by the demangling function (x86, AArch64, and LLVM internal one). The unit tests that demangle scalable names have been modified by adding the IR module that carries the declaration of the vector function name being demangled. In particular, the demangling function fails in the following cases: 1. When the declaration of the scalable vector function is not present in the module. 2. When the value of VFSHape.VF is not greater than 0. Reviewers: jdoerfert, sdesmalen, andwar Reviewed By: jdoerfert Subscribers: mgorny, kristof.beyls, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73286
2020-01-22 23:34:27 +01:00
EXPECT_TRUE(invokeParser("_ZGV_LLVM_Mxvl8l8_sincos(custom_vector_sincos)",
"custom_vector_sincos",
"void(<vscale x 2 x double>, double *, double *)"));
EXPECT_EQ(VF, (unsigned)2);
EXPECT_TRUE(IsMasked());
EXPECT_TRUE(IsScalable);
EXPECT_EQ(ISA, VFISAKind::LLVM);
EXPECT_EQ(Parameters.size(), (unsigned)4);
EXPECT_EQ(Parameters[0], VFParameter({0, VFParamKind::Vector}));
EXPECT_EQ(Parameters[1], VFParameter({1, VFParamKind::OMP_Linear, 8}));
EXPECT_EQ(Parameters[2], VFParameter({2, VFParamKind::OMP_Linear, 8}));
EXPECT_EQ(Parameters[3], VFParameter({3, VFParamKind::GlobalPredicate}));
EXPECT_EQ(ScalarName, "sincos");
EXPECT_EQ(VectorName, "custom_vector_sincos");
}
class VFABIAttrTest : public testing::Test {
protected:
void SetUp() override {
M = parseAssemblyString(IR, Err, Ctx);
// Get the only call instruction in the block, which is the first
// instruction.
CI = dyn_cast<CallInst>(&*(instructions(M->getFunction("f")).begin()));
}
const char *IR = "define i32 @f(i32 %a) {\n"
" %1 = call i32 @g(i32 %a) #0\n"
" ret i32 %1\n"
"}\n"
"declare i32 @g(i32)\n"
"declare <2 x i32> @custom_vg(<2 x i32>)"
"declare <4 x i32> @_ZGVnN4v_g(<4 x i32>)"
"declare <8 x i32> @_ZGVnN8v_g(<8 x i32>)"
"attributes #0 = { "
"\"vector-function-abi-variant\"=\""
"_ZGVnN2v_g(custom_vg),_ZGVnN4v_g\" }";
LLVMContext Ctx;
SMDiagnostic Err;
std::unique_ptr<Module> M;
CallInst *CI;
SmallVector<std::string, 8> Mappings;
};
TEST_F(VFABIAttrTest, Read) {
VFABI::getVectorVariantNames(*CI, Mappings);
SmallVector<std::string, 8> Exp;
Exp.push_back("_ZGVnN2v_g(custom_vg)");
Exp.push_back("_ZGVnN4v_g");
EXPECT_EQ(Mappings, Exp);
}
TEST_F(VFABIParserTest, LLVM_InternalISA) {
EXPECT_FALSE(invokeParser("_ZGV_LLVM_N2v_sin"));
[llvm][VectorUtils] Tweak VFShape for scalable vector functions. Summary: This patch makes sure that the field VFShape.VF is greater than zero when demangling the vector function name of scalable vector functions encoded in the "vector-function-abi-variant" attribute. This change is required to be able to provide instances of VFShape that can be used to query the VFDatabase for the vectorization passes, as such passes always require a positive value for the Vectorization Factor (VF) needed by the vectorization process. It is not possible to extract the value of VFShape.VF from the mangled name of scalable vector functions, because it is encoded as `x`. Therefore, the VFABI demangling function has been modified to extract such information from the IR declaration of the vector function, under the assumption that _all_ vectors in the signature of the vector function have the same number of lanes. Such assumption is valid because it is also assumed by the Vector Function ABI specifications supported by the demangling function (x86, AArch64, and LLVM internal one). The unit tests that demangle scalable names have been modified by adding the IR module that carries the declaration of the vector function name being demangled. In particular, the demangling function fails in the following cases: 1. When the declaration of the scalable vector function is not present in the module. 2. When the value of VFSHape.VF is not greater than 0. Reviewers: jdoerfert, sdesmalen, andwar Reviewed By: jdoerfert Subscribers: mgorny, kristof.beyls, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73286
2020-01-22 23:34:27 +01:00
EXPECT_TRUE(invokeParser("_ZGV_LLVM_N2v_sin_(vector_name)", "vector_name"));
EXPECT_EQ(ISA, VFISAKind::LLVM);
}
TEST_F(VFABIParserTest, IntrinsicsInLLVMIsa) {
[llvm][VectorUtils] Tweak VFShape for scalable vector functions. Summary: This patch makes sure that the field VFShape.VF is greater than zero when demangling the vector function name of scalable vector functions encoded in the "vector-function-abi-variant" attribute. This change is required to be able to provide instances of VFShape that can be used to query the VFDatabase for the vectorization passes, as such passes always require a positive value for the Vectorization Factor (VF) needed by the vectorization process. It is not possible to extract the value of VFShape.VF from the mangled name of scalable vector functions, because it is encoded as `x`. Therefore, the VFABI demangling function has been modified to extract such information from the IR declaration of the vector function, under the assumption that _all_ vectors in the signature of the vector function have the same number of lanes. Such assumption is valid because it is also assumed by the Vector Function ABI specifications supported by the demangling function (x86, AArch64, and LLVM internal one). The unit tests that demangle scalable names have been modified by adding the IR module that carries the declaration of the vector function name being demangled. In particular, the demangling function fails in the following cases: 1. When the declaration of the scalable vector function is not present in the module. 2. When the value of VFSHape.VF is not greater than 0. Reviewers: jdoerfert, sdesmalen, andwar Reviewed By: jdoerfert Subscribers: mgorny, kristof.beyls, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73286
2020-01-22 23:34:27 +01:00
EXPECT_TRUE(invokeParser("_ZGV_LLVM_N4vv_llvm.pow.f32(__svml_powf4)",
"__svml_powf4"));
EXPECT_EQ(VF, (unsigned)4);
EXPECT_FALSE(IsMasked());
EXPECT_FALSE(IsScalable);
EXPECT_EQ(ISA, VFISAKind::LLVM);
EXPECT_EQ(Parameters.size(), (unsigned)2);
EXPECT_EQ(Parameters[0], VFParameter({0, VFParamKind::Vector}));
EXPECT_EQ(Parameters[1], VFParameter({1, VFParamKind::Vector}));
EXPECT_EQ(ScalarName, "llvm.pow.f32");
}
[llvm][VectorUtils] Tweak VFShape for scalable vector functions. Summary: This patch makes sure that the field VFShape.VF is greater than zero when demangling the vector function name of scalable vector functions encoded in the "vector-function-abi-variant" attribute. This change is required to be able to provide instances of VFShape that can be used to query the VFDatabase for the vectorization passes, as such passes always require a positive value for the Vectorization Factor (VF) needed by the vectorization process. It is not possible to extract the value of VFShape.VF from the mangled name of scalable vector functions, because it is encoded as `x`. Therefore, the VFABI demangling function has been modified to extract such information from the IR declaration of the vector function, under the assumption that _all_ vectors in the signature of the vector function have the same number of lanes. Such assumption is valid because it is also assumed by the Vector Function ABI specifications supported by the demangling function (x86, AArch64, and LLVM internal one). The unit tests that demangle scalable names have been modified by adding the IR module that carries the declaration of the vector function name being demangled. In particular, the demangling function fails in the following cases: 1. When the declaration of the scalable vector function is not present in the module. 2. When the value of VFSHape.VF is not greater than 0. Reviewers: jdoerfert, sdesmalen, andwar Reviewed By: jdoerfert Subscribers: mgorny, kristof.beyls, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73286
2020-01-22 23:34:27 +01:00
TEST_F(VFABIParserTest, ParseScalableRequiresDeclaration) {
const char *MangledName = "_ZGVsMxv_sin(custom_vg)";
// The parser succeds only when the correct function definition of
// `custom_vg` is added to the module.
EXPECT_FALSE(invokeParser(MangledName));
EXPECT_TRUE(invokeParser(
MangledName, "custom_vg",
"<vscale x 4 x double>(<vscale x 4 x double>, <vscale x 4 x i1>)"));
}
TEST_F(VFABIParserTest, ZeroIsInvalidVLEN) {
EXPECT_FALSE(invokeParser("_ZGVeM0v_sin"));
EXPECT_FALSE(invokeParser("_ZGVeN0v_sin"));
EXPECT_FALSE(invokeParser("_ZGVsM0v_sin"));
EXPECT_FALSE(invokeParser("_ZGVsN0v_sin"));
}