[GlobalISel][LLT] Introduce FPInfo for LLT (Enable bfloat, ppc128float and others in GlobalISel) #155107
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…ble extended types for llvm-ir lowering in TargetLoweringBase. Added flag for GISel emitter to enable/disable extended types in matcher tables
llvmbot
added
backend:AArch64
backend:AMDGPU
tablegen
llvm:codegen
llvm:globalisel
llvm:SelectionDAG
Member
|
@llvm/pr-subscribers-llvm-selectiondag @llvm/pr-subscribers-backend-amdgpu Author: Denis.G (DenisGZM) ChangesAdded extra information in LLT to support ambiguous fp types during GlobalISel. Original idea by @tgymnich Main differences from #122503 are:
Also MIRParser now may parse new type indetifiers. Patch is 97.84 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/155107.diff 50 Files Affected:
diff --git a/llvm/include/llvm/CodeGen/Analysis.h b/llvm/include/llvm/CodeGen/Analysis.h
index 98b52579d03b7..2aadc27930b7a 100644
--- a/llvm/include/llvm/CodeGen/Analysis.h
+++ b/llvm/include/llvm/CodeGen/Analysis.h
@@ -101,7 +101,8 @@ inline void ComputeValueVTs(const TargetLowering &TLI, const DataLayout &DL,
/// If Offsets is non-null, it points to a vector to be filled in
/// with the in-memory offsets of each of the individual values.
///
-void computeValueLLTs(const DataLayout &DL, Type &Ty,
+void computeValueLLTs(const TargetLowering &TLI,
+ const DataLayout &DL, Type &Ty,
SmallVectorImpl<LLT> &ValueTys,
SmallVectorImpl<uint64_t> *Offsets = nullptr,
uint64_t StartingOffset = 0);
diff --git a/llvm/include/llvm/CodeGen/GlobalISel/IRTranslator.h b/llvm/include/llvm/CodeGen/GlobalISel/IRTranslator.h
index 3d7ccd55ee042..5673fd5168477 100644
--- a/llvm/include/llvm/CodeGen/GlobalISel/IRTranslator.h
+++ b/llvm/include/llvm/CodeGen/GlobalISel/IRTranslator.h
@@ -642,6 +642,8 @@ class IRTranslator : public MachineFunctionPass {
StackProtectorDescriptor SPDescriptor;
+ bool mayTranslateUserTypes(const User &U) const;
+
/// Switch analysis and optimization.
class GISelSwitchLowering : public SwitchCG::SwitchLowering {
public:
diff --git a/llvm/include/llvm/CodeGen/GlobalISel/LegalizerInfo.h b/llvm/include/llvm/CodeGen/GlobalISel/LegalizerInfo.h
index fd72a3898562e..8412042a780e8 100644
--- a/llvm/include/llvm/CodeGen/GlobalISel/LegalizerInfo.h
+++ b/llvm/include/llvm/CodeGen/GlobalISel/LegalizerInfo.h
@@ -380,7 +380,7 @@ LLVM_ABI LegalizeMutation changeElementCountTo(unsigned TypeIdx,
/// Keep the same scalar or element type as \p TypeIdx, but take the number of
/// elements from \p Ty.
-LLVM_ABI LegalizeMutation changeElementCountTo(unsigned TypeIdx, LLT Ty);
+LLVM_ABI LegalizeMutation changeElementCountTo(unsigned TypeIdx, ElementCount EC);
/// Change the scalar size or element size to have the same scalar size as type
/// index \p FromIndex. Unlike changeElementTo, this discards pointer types and
diff --git a/llvm/include/llvm/CodeGen/LowLevelTypeUtils.h b/llvm/include/llvm/CodeGen/LowLevelTypeUtils.h
index 51a298eb8b247..1c8f34b1a8518 100644
--- a/llvm/include/llvm/CodeGen/LowLevelTypeUtils.h
+++ b/llvm/include/llvm/CodeGen/LowLevelTypeUtils.h
@@ -27,7 +27,7 @@ class Type;
struct fltSemantics;
/// Construct a low-level type based on an LLVM type.
-LLVM_ABI LLT getLLTForType(Type &Ty, const DataLayout &DL);
+LLVM_ABI LLT getLLTForType(Type &Ty, const DataLayout &DL, bool AllowExtendedLLT = false);
/// Get a rough equivalent of an MVT for a given LLT. MVT can't distinguish
/// pointers, so these will convert to a plain integer.
@@ -36,11 +36,11 @@ LLVM_ABI EVT getApproximateEVTForLLT(LLT Ty, LLVMContext &Ctx);
/// Get a rough equivalent of an LLT for a given MVT. LLT does not yet support
/// scalarable vector types, and will assert if used.
-LLVM_ABI LLT getLLTForMVT(MVT Ty);
+LLVM_ABI LLT getLLTForMVT(MVT Ty, bool AllowExtendedLLT = false);
/// Get the appropriate floating point arithmetic semantic based on the bit size
/// of the given scalar LLT.
LLVM_ABI const llvm::fltSemantics &getFltSemanticForLLT(LLT Ty);
-}
+} // namespace llvm
#endif // LLVM_CODEGEN_LOWLEVELTYPEUTILS_H
diff --git a/llvm/include/llvm/CodeGen/TargetLowering.h b/llvm/include/llvm/CodeGen/TargetLowering.h
index 4480ced637456..f4fdd8c69ea9e 100644
--- a/llvm/include/llvm/CodeGen/TargetLowering.h
+++ b/llvm/include/llvm/CodeGen/TargetLowering.h
@@ -416,6 +416,9 @@ class LLVM_ABI TargetLoweringBase {
/// amounts, returns MVT::i32.
EVT getShiftAmountTy(EVT LHSTy, const DataLayout &DL) const;
+ virtual LLT getLLTForType(Type &Ty, const DataLayout &DL) const;
+ virtual LLT getLLTForMVT(MVT Ty) const;
+
/// Return the preferred type to use for a shift opcode, given the shifted
/// amount type is \p ShiftValueTy.
LLVM_READONLY
diff --git a/llvm/include/llvm/CodeGenTypes/LowLevelType.h b/llvm/include/llvm/CodeGenTypes/LowLevelType.h
index d8e0848aff84d..0c217ba1fcf7e 100644
--- a/llvm/include/llvm/CodeGenTypes/LowLevelType.h
+++ b/llvm/include/llvm/CodeGenTypes/LowLevelType.h
@@ -39,68 +39,148 @@ class raw_ostream;
class LLT {
public:
+ enum class FPVariant {
+ IEEE_FLOAT = 0x0,
+ BRAIN_FLOAT = 0x1, // BRAIN_FLOAT
+ PPC128_FLOAT = 0x2, // PPC128_FLOAT
+ EXTENDED_FP80 = 0x3, // FP80
+ TENSOR_FLOAT32 = 0x4, // TENSOR_FLOAT32
+ VARIANT_FLOAT_5 = 0x5, // UNASSIGNED
+ VARIANT_FLOAT_6 = 0x6, // UNASSIGNED
+ VARIANT_FLOAT_7 = 0x7, // UNASSIGNED
+ };
+
+ enum class Kind : uint64_t {
+ INVALID = 0b0000,
+ ANY_SCALAR = 0b0001,
+ INTEGER = 0b0010,
+ FLOAT = 0b0011,
+ POINTER = 0b0100,
+ VECTOR_ANY = 0b0101,
+ VECTOR_INTEGER = 0b0110,
+ VECTOR_FLOAT = 0b0111,
+ VECTOR_POINTER = 0b1000,
+ };
+
+ constexpr static Kind toVector(Kind Ty) {
+ if (Ty == Kind::POINTER)
+ return Kind::VECTOR_POINTER;
+
+ if (Ty == Kind::INTEGER)
+ return Kind::VECTOR_INTEGER;
+
+ if (Ty == Kind::FLOAT)
+ return Kind::VECTOR_FLOAT;
+
+ return Kind::VECTOR_ANY;
+ }
+
+ constexpr static Kind toScalar(Kind Ty) {
+ if (Ty == Kind::VECTOR_POINTER)
+ return Kind::POINTER;
+
+ if (Ty == Kind::VECTOR_INTEGER)
+ return Kind::INTEGER;
+
+ if (Ty == Kind::VECTOR_FLOAT)
+ return Kind::FLOAT;
+
+ return Kind::ANY_SCALAR;
+ }
+
/// Get a low-level scalar or aggregate "bag of bits".
static constexpr LLT scalar(unsigned SizeInBits) {
- return LLT{/*isPointer=*/false, /*isVector=*/false, /*isScalar=*/true,
- ElementCount::getFixed(0), SizeInBits,
- /*AddressSpace=*/0};
+ return LLT{Kind::ANY_SCALAR, ElementCount::getFixed(0), SizeInBits,
+ /*AddressSpace=*/0, static_cast<FPVariant>(0)};
+ }
+
+ static constexpr LLT integer(unsigned SizeInBits) {
+ return LLT{Kind::INTEGER, ElementCount::getFixed(0), SizeInBits,
+ /*AddressSpace=*/0, static_cast<FPVariant>(0)};
+ }
+
+ static constexpr LLT floatingPoint(unsigned SizeInBits, FPVariant FP) {
+ return LLT{Kind::FLOAT, ElementCount::getFixed(0), SizeInBits,
+ /*AddressSpace=*/0, FP};
}
/// Get a low-level token; just a scalar with zero bits (or no size).
static constexpr LLT token() {
- return LLT{/*isPointer=*/false, /*isVector=*/false,
- /*isScalar=*/true, ElementCount::getFixed(0),
+ return LLT{Kind::ANY_SCALAR, ElementCount::getFixed(0),
/*SizeInBits=*/0,
- /*AddressSpace=*/0};
+ /*AddressSpace=*/0, static_cast<FPVariant>(0)};
}
/// Get a low-level pointer in the given address space.
static constexpr LLT pointer(unsigned AddressSpace, unsigned SizeInBits) {
assert(SizeInBits > 0 && "invalid pointer size");
- return LLT{/*isPointer=*/true, /*isVector=*/false, /*isScalar=*/false,
- ElementCount::getFixed(0), SizeInBits, AddressSpace};
+ return LLT{Kind::POINTER, ElementCount::getFixed(0), SizeInBits,
+ AddressSpace, static_cast<FPVariant>(0)};
}
/// Get a low-level vector of some number of elements and element width.
static constexpr LLT vector(ElementCount EC, unsigned ScalarSizeInBits) {
assert(!EC.isScalar() && "invalid number of vector elements");
- return LLT{/*isPointer=*/false, /*isVector=*/true, /*isScalar=*/false,
- EC, ScalarSizeInBits, /*AddressSpace=*/0};
+ return LLT{Kind::VECTOR_ANY, EC, ScalarSizeInBits,
+ /*AddressSpace=*/0, static_cast<FPVariant>(0)};
}
/// Get a low-level vector of some number of elements and element type.
static constexpr LLT vector(ElementCount EC, LLT ScalarTy) {
assert(!EC.isScalar() && "invalid number of vector elements");
assert(!ScalarTy.isVector() && "invalid vector element type");
- return LLT{ScalarTy.isPointer(),
- /*isVector=*/true,
- /*isScalar=*/false,
- EC,
- ScalarTy.getSizeInBits().getFixedValue(),
- ScalarTy.isPointer() ? ScalarTy.getAddressSpace() : 0};
+
+ Kind Info = toVector(ScalarTy.Info);
+ return LLT{Info, EC, ScalarTy.getSizeInBits().getFixedValue(),
+ ScalarTy.isPointer() ? ScalarTy.getAddressSpace() : 0,
+ ScalarTy.isFloat() ? ScalarTy.getFPVariant()
+ : static_cast<FPVariant>(0)};
+ }
+ // Get a 8-bit brain float value.
+ static constexpr LLT bfloat8() {
+ return floatingPoint(8, FPVariant::BRAIN_FLOAT);
+ }
+
+ // Get a 16-bit brain float value.
+ static constexpr LLT bfloat16() {
+ return floatingPoint(16, FPVariant::BRAIN_FLOAT);
}
/// Get a 16-bit IEEE half value.
- /// TODO: Add IEEE semantics to type - This currently returns a simple `scalar(16)`.
static constexpr LLT float16() {
- return scalar(16);
+ return floatingPoint(16, FPVariant::IEEE_FLOAT);
}
/// Get a 32-bit IEEE float value.
static constexpr LLT float32() {
- return scalar(32);
+ return floatingPoint(32, FPVariant::IEEE_FLOAT);
}
/// Get a 64-bit IEEE double value.
static constexpr LLT float64() {
- return scalar(64);
+ return floatingPoint(64, FPVariant::IEEE_FLOAT);
+ }
+
+ /// Get a 80-bit X86 floating point value.
+ static constexpr LLT x86fp80() {
+ return floatingPoint(80, FPVariant::EXTENDED_FP80);
+ }
+
+ /// Get a 128-bit IEEE quad value.
+ static constexpr LLT float128() {
+ return floatingPoint(128, FPVariant::IEEE_FLOAT);
+ }
+
+ /// Get a 128-bit PowerPC double double value.
+ static constexpr LLT ppcf128() {
+ return floatingPoint(128, FPVariant::PPC128_FLOAT);
}
/// Get a low-level fixed-width vector of some number of elements and element
/// width.
- static constexpr LLT fixed_vector(unsigned NumElements,
- unsigned ScalarSizeInBits) {
- return vector(ElementCount::getFixed(NumElements), ScalarSizeInBits);
+ static constexpr LLT fixed_vector(unsigned NumElements, unsigned ScalarSizeInBits) {
+ return vector(ElementCount::getFixed(NumElements),
+ LLT::scalar(ScalarSizeInBits));
}
/// Get a low-level fixed-width vector of some number of elements and element
@@ -111,9 +191,9 @@ class LLT {
/// Get a low-level scalable vector of some number of elements and element
/// width.
- static constexpr LLT scalable_vector(unsigned MinNumElements,
- unsigned ScalarSizeInBits) {
- return vector(ElementCount::getScalable(MinNumElements), ScalarSizeInBits);
+ static constexpr LLT scalable_vector(unsigned MinNumElements, unsigned ScalarSizeInBits) {
+ return vector(ElementCount::getScalable(MinNumElements),
+ LLT::scalar(ScalarSizeInBits));
}
/// Get a low-level scalable vector of some number of elements and element
@@ -132,27 +212,79 @@ class LLT {
return scalarOrVector(EC, LLT::scalar(static_cast<unsigned>(ScalarSize)));
}
- explicit constexpr LLT(bool isPointer, bool isVector, bool isScalar,
- ElementCount EC, uint64_t SizeInBits,
- unsigned AddressSpace)
+ explicit constexpr LLT(Kind Info, ElementCount EC, uint64_t SizeInBits,
+ unsigned AddressSpace, FPVariant FP)
: LLT() {
- init(isPointer, isVector, isScalar, EC, SizeInBits, AddressSpace);
+ init(Info, EC, SizeInBits, AddressSpace, FP);
}
- explicit constexpr LLT()
- : IsScalar(false), IsPointer(false), IsVector(false), RawData(0) {}
- LLVM_ABI explicit LLT(MVT VT);
+ LLVM_ABI explicit LLT(MVT VT, bool AllowExtendedLLT = false);
+ explicit constexpr LLT() : Info(static_cast<Kind>(0)), RawData(0) {}
- constexpr bool isValid() const { return IsScalar || RawData != 0; }
- constexpr bool isScalar() const { return IsScalar; }
- constexpr bool isToken() const { return IsScalar && RawData == 0; };
- constexpr bool isVector() const { return isValid() && IsVector; }
+ constexpr bool isValid() const { return isToken() || RawData != 0; }
+ constexpr bool isScalar() const { return Info == Kind::ANY_SCALAR || Info == Kind::INTEGER || Info == Kind::FLOAT; }
+ constexpr bool isScalar(unsigned Size) const {
+ return isScalar() && getScalarSizeInBits() == Size;
+ }
+ constexpr bool isFloat() const { return isValid() && Info == Kind::FLOAT; }
+ constexpr bool isFloat(unsigned Size) const {
+ return isFloat() && getScalarSizeInBits() == Size;
+ }
+ constexpr bool isVariantFloat() const {
+ return isFloat() && getFPVariant() != FPVariant::IEEE_FLOAT;
+ }
+ constexpr bool isVariantFloat(FPVariant Variant) const {
+ return isFloat() && getFPVariant() == Variant;
+ }
+ constexpr bool isVariantFloat(unsigned Size, FPVariant Variant) const {
+ return isVariantFloat(Variant) && getScalarSizeInBits() == Size;
+ }
+ constexpr bool isFloatVector() const {
+ return isVector() && Info == Kind::VECTOR_FLOAT;
+ }
+ constexpr bool isIEEEFloat(unsigned Size) const {
+ return isVariantFloat(Size, FPVariant::IEEE_FLOAT);
+ }
+ constexpr bool isBFloat(unsigned Size) const {
+ return isVariantFloat(Size, FPVariant::BRAIN_FLOAT);
+ }
+ constexpr bool isX86FP80() const {
+ return isVariantFloat(80, FPVariant::EXTENDED_FP80);
+ }
+ constexpr bool isPPCF128() const {
+ return isVariantFloat(128, FPVariant::PPC128_FLOAT);
+ }
+ constexpr bool isToken() const {
+ return Info == Kind::ANY_SCALAR && RawData == 0;
+ }
+ constexpr bool isAnyScalar() const {
+ return isValid() && Info == Kind::ANY_SCALAR;
+ }
+ constexpr bool isVectorAny() const {
+ return isVector() && Info == Kind::VECTOR_ANY;
+ }
+ constexpr bool isInteger() const {
+ return isValid() && Info == Kind::INTEGER;
+ }
+ constexpr bool isInteger(unsigned Size) const {
+ return isInteger() && getScalarSizeInBits() == Size;
+ }
+ constexpr bool isIntegerVector() const {
+ return isVector() && Info == Kind::VECTOR_INTEGER;
+ }
+ constexpr bool isVector() const {
+ return isValid() &&
+ (Info == Kind::VECTOR_ANY || Info == Kind::VECTOR_INTEGER||
+ Info == Kind::VECTOR_FLOAT || Info == Kind::VECTOR_POINTER);
+ }
constexpr bool isPointer() const {
- return isValid() && IsPointer && !IsVector;
+ return isValid() && Info == Kind::POINTER;
+ }
+ constexpr bool isPointerVector() const {
+ return isVector() && Info == Kind::VECTOR_POINTER;
}
- constexpr bool isPointerVector() const { return IsPointer && isVector(); }
constexpr bool isPointerOrPointerVector() const {
- return IsPointer && isValid();
+ return isPointer() || isPointerVector();
}
/// Returns the number of elements in a vector LLT. Must only be called on
@@ -177,12 +309,18 @@ class LLT {
/// if the LLT is not a vector type.
constexpr bool isFixedVector() const { return isVector() && !isScalable(); }
+ constexpr bool isFixedVector(unsigned NumElements,
+ unsigned ScalarSize) const {
+ return isFixedVector() && getNumElements() == NumElements &&
+ getScalarSizeInBits() == ScalarSize;
+ }
+
/// Returns true if the LLT is a scalable vector. Returns false otherwise,
/// even if the LLT is not a vector type.
constexpr bool isScalableVector() const { return isVector() && isScalable(); }
constexpr ElementCount getElementCount() const {
- assert(IsVector && "cannot get number of elements on scalar/aggregate");
+ assert(isVector() && "cannot get number of elements on scalar/aggregate");
return ElementCount::get(getFieldValue(VectorElementsFieldInfo),
isScalable());
}
@@ -207,6 +345,15 @@ class LLT {
return isVector() ? getElementType() : *this;
}
+ constexpr FPVariant getFPVariant() const {
+ assert((isFloat() || isFloatVector()) &&
+ "cannot get FP info for non float type");
+
+ return FPVariant(getFieldValue(FPFieldInfo));
+ }
+
+ constexpr Kind getKind() const { return Info; }
+
/// If this type is a vector, return a vector with the same number of elements
/// but the new element type. Otherwise, return the new element type.
constexpr LLT changeElementType(LLT NewEltTy) const {
@@ -217,10 +364,14 @@ class LLT {
/// but the new element size. Otherwise, return the new element type. Invalid
/// for pointer types. For pointer types, use changeElementType.
constexpr LLT changeElementSize(unsigned NewEltSize) const {
- assert(!isPointerOrPointerVector() &&
+ assert(!isPointerOrPointerVector() && !(isFloat() || isFloatVector()) &&
"invalid to directly change element size for pointers");
- return isVector() ? LLT::vector(getElementCount(), NewEltSize)
- : LLT::scalar(NewEltSize);
+ return isVector()
+ ? LLT::vector(getElementCount(), getElementType().isInteger()
+ ? LLT::integer(NewEltSize)
+ : LLT::scalar(NewEltSize))
+ : isInteger() ? LLT::integer(NewEltSize)
+ : LLT::scalar(NewEltSize);
}
/// Return a vector or scalar with the same element type and the new element
@@ -229,6 +380,10 @@ class LLT {
return LLT::scalarOrVector(EC, getScalarType());
}
+ constexpr LLT changeElementCount(unsigned NumElements) const {
+ return changeElementCount(ElementCount::getFixed(NumElements));
+ }
+
/// Return a type that is \p Factor times smaller. Reduces the number of
/// elements if this is a vector, or the bitwidth for scalar/pointers. Does
/// not attempt to handle cases that aren't evenly divisible.
@@ -243,6 +398,9 @@ class LLT {
}
assert(getScalarSizeInBits() % Factor == 0);
+ if (isInteger())
+ return integer(getScalarSizeInBits() / Factor);
+
return scalar(getScalarSizeInBits() / Factor);
}
@@ -277,10 +435,26 @@ class LLT {
/// Returns the vector's element type. Only valid for vector types.
constexpr LLT getElementType() const {
assert(isVector() && "cannot get element type of scalar/aggregate");
- if (IsPointer)
+ if (isPointerVector())
return pointer(getAddressSpace(), getScalarSizeInBits());
- else
- return scalar(getScalarSizeInBits());
+
+ if (isFloatVector())
+ return floatingPoint(getScalarSizeInBits(), getFPVariant());
+
+ if (isIntegerVector())
+ return integer(getScalarSizeInBits());
+
+ return scalar(getScalarSizeInBits());
+ }
+
+ constexpr LLT changeToInteger() const {
+ if (isPointer() || isPointerVector())
+ return *this;
+
+ if (isVector())
+ return vector(getElementCount(), LLT::integer(getScalarSizeInBits()));
+
+ return integer(getSizeInBits());
}
LLVM_ABI void print(raw_ostream &OS) const;
@@ -290,8 +464,14 @@ class LLT {
#endif
constexpr bool operator==(const LLT &RHS) const {
- return IsPointer == RHS.IsPointer && IsVector == RHS.IsVector &&
- IsScalar == RHS.IsScalar && RHS.RawData == RawData;
+ if (isAnyScalar() || RHS.isAnyScalar()) {
+ return isScalar() == RHS.isScalar() && RawData == RHS.RawData;
+ }
+ if (isVector() && RHS.isVector()) {
+ return getElementType() == RHS.getElementType() &&
+ getElementCount() == RHS.getElementCount();
+ }
+ return Info == RHS.Info && RawData == RHS.RawData;
}
constexpr bool operator!=(const LLT &RHS) const { return !(*this == RHS); }
@@ -301,37 +481,33 @@ class LLT {
private:
/// LLT is packed into 64 bits as follows:
- /// isScalar : 1
- /// isPointer : 1
- /// isVector : 1
- /// with 61 bits remaining for Kind-specific data, packed in bitfields
- /// as described below. As there isn't a simple portable way to pack bits
- /// into bitfields, here the different fields in the packed structure is
+ /// Info : 4
+ /// RawData : 60
+ /// with 61 bits of RawData remaining for Kind-specific data, packed in
+ /// bitfields as described below. As there isn't a simple portable way to pack
+ /// bits into bitfields, here the different fields in the packed structure is
/// described in static const *Field variables. Each of these variables
/// is a 2-element array, with the first ele...
[truncated]
|
Member
|
@llvm/pr-subscribers-llvm-globalisel Author: Denis.G (DenisGZM) ChangesAdded extra information in LLT to support ambiguous fp types during GlobalISel. Original idea by @tgymnich Main differences from #122503 are:
Also MIRParser now may parse new type indetifiers. Patch is 97.84 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/155107.diff 50 Files Affected:
diff --git a/llvm/include/llvm/CodeGen/Analysis.h b/llvm/include/llvm/CodeGen/Analysis.h
index 98b52579d03b7..2aadc27930b7a 100644
--- a/llvm/include/llvm/CodeGen/Analysis.h
+++ b/llvm/include/llvm/CodeGen/Analysis.h
@@ -101,7 +101,8 @@ inline void ComputeValueVTs(const TargetLowering &TLI, const DataLayout &DL,
/// If Offsets is non-null, it points to a vector to be filled in
/// with the in-memory offsets of each of the individual values.
///
-void computeValueLLTs(const DataLayout &DL, Type &Ty,
+void computeValueLLTs(const TargetLowering &TLI,
+ const DataLayout &DL, Type &Ty,
SmallVectorImpl<LLT> &ValueTys,
SmallVectorImpl<uint64_t> *Offsets = nullptr,
uint64_t StartingOffset = 0);
diff --git a/llvm/include/llvm/CodeGen/GlobalISel/IRTranslator.h b/llvm/include/llvm/CodeGen/GlobalISel/IRTranslator.h
index 3d7ccd55ee042..5673fd5168477 100644
--- a/llvm/include/llvm/CodeGen/GlobalISel/IRTranslator.h
+++ b/llvm/include/llvm/CodeGen/GlobalISel/IRTranslator.h
@@ -642,6 +642,8 @@ class IRTranslator : public MachineFunctionPass {
StackProtectorDescriptor SPDescriptor;
+ bool mayTranslateUserTypes(const User &U) const;
+
/// Switch analysis and optimization.
class GISelSwitchLowering : public SwitchCG::SwitchLowering {
public:
diff --git a/llvm/include/llvm/CodeGen/GlobalISel/LegalizerInfo.h b/llvm/include/llvm/CodeGen/GlobalISel/LegalizerInfo.h
index fd72a3898562e..8412042a780e8 100644
--- a/llvm/include/llvm/CodeGen/GlobalISel/LegalizerInfo.h
+++ b/llvm/include/llvm/CodeGen/GlobalISel/LegalizerInfo.h
@@ -380,7 +380,7 @@ LLVM_ABI LegalizeMutation changeElementCountTo(unsigned TypeIdx,
/// Keep the same scalar or element type as \p TypeIdx, but take the number of
/// elements from \p Ty.
-LLVM_ABI LegalizeMutation changeElementCountTo(unsigned TypeIdx, LLT Ty);
+LLVM_ABI LegalizeMutation changeElementCountTo(unsigned TypeIdx, ElementCount EC);
/// Change the scalar size or element size to have the same scalar size as type
/// index \p FromIndex. Unlike changeElementTo, this discards pointer types and
diff --git a/llvm/include/llvm/CodeGen/LowLevelTypeUtils.h b/llvm/include/llvm/CodeGen/LowLevelTypeUtils.h
index 51a298eb8b247..1c8f34b1a8518 100644
--- a/llvm/include/llvm/CodeGen/LowLevelTypeUtils.h
+++ b/llvm/include/llvm/CodeGen/LowLevelTypeUtils.h
@@ -27,7 +27,7 @@ class Type;
struct fltSemantics;
/// Construct a low-level type based on an LLVM type.
-LLVM_ABI LLT getLLTForType(Type &Ty, const DataLayout &DL);
+LLVM_ABI LLT getLLTForType(Type &Ty, const DataLayout &DL, bool AllowExtendedLLT = false);
/// Get a rough equivalent of an MVT for a given LLT. MVT can't distinguish
/// pointers, so these will convert to a plain integer.
@@ -36,11 +36,11 @@ LLVM_ABI EVT getApproximateEVTForLLT(LLT Ty, LLVMContext &Ctx);
/// Get a rough equivalent of an LLT for a given MVT. LLT does not yet support
/// scalarable vector types, and will assert if used.
-LLVM_ABI LLT getLLTForMVT(MVT Ty);
+LLVM_ABI LLT getLLTForMVT(MVT Ty, bool AllowExtendedLLT = false);
/// Get the appropriate floating point arithmetic semantic based on the bit size
/// of the given scalar LLT.
LLVM_ABI const llvm::fltSemantics &getFltSemanticForLLT(LLT Ty);
-}
+} // namespace llvm
#endif // LLVM_CODEGEN_LOWLEVELTYPEUTILS_H
diff --git a/llvm/include/llvm/CodeGen/TargetLowering.h b/llvm/include/llvm/CodeGen/TargetLowering.h
index 4480ced637456..f4fdd8c69ea9e 100644
--- a/llvm/include/llvm/CodeGen/TargetLowering.h
+++ b/llvm/include/llvm/CodeGen/TargetLowering.h
@@ -416,6 +416,9 @@ class LLVM_ABI TargetLoweringBase {
/// amounts, returns MVT::i32.
EVT getShiftAmountTy(EVT LHSTy, const DataLayout &DL) const;
+ virtual LLT getLLTForType(Type &Ty, const DataLayout &DL) const;
+ virtual LLT getLLTForMVT(MVT Ty) const;
+
/// Return the preferred type to use for a shift opcode, given the shifted
/// amount type is \p ShiftValueTy.
LLVM_READONLY
diff --git a/llvm/include/llvm/CodeGenTypes/LowLevelType.h b/llvm/include/llvm/CodeGenTypes/LowLevelType.h
index d8e0848aff84d..0c217ba1fcf7e 100644
--- a/llvm/include/llvm/CodeGenTypes/LowLevelType.h
+++ b/llvm/include/llvm/CodeGenTypes/LowLevelType.h
@@ -39,68 +39,148 @@ class raw_ostream;
class LLT {
public:
+ enum class FPVariant {
+ IEEE_FLOAT = 0x0,
+ BRAIN_FLOAT = 0x1, // BRAIN_FLOAT
+ PPC128_FLOAT = 0x2, // PPC128_FLOAT
+ EXTENDED_FP80 = 0x3, // FP80
+ TENSOR_FLOAT32 = 0x4, // TENSOR_FLOAT32
+ VARIANT_FLOAT_5 = 0x5, // UNASSIGNED
+ VARIANT_FLOAT_6 = 0x6, // UNASSIGNED
+ VARIANT_FLOAT_7 = 0x7, // UNASSIGNED
+ };
+
+ enum class Kind : uint64_t {
+ INVALID = 0b0000,
+ ANY_SCALAR = 0b0001,
+ INTEGER = 0b0010,
+ FLOAT = 0b0011,
+ POINTER = 0b0100,
+ VECTOR_ANY = 0b0101,
+ VECTOR_INTEGER = 0b0110,
+ VECTOR_FLOAT = 0b0111,
+ VECTOR_POINTER = 0b1000,
+ };
+
+ constexpr static Kind toVector(Kind Ty) {
+ if (Ty == Kind::POINTER)
+ return Kind::VECTOR_POINTER;
+
+ if (Ty == Kind::INTEGER)
+ return Kind::VECTOR_INTEGER;
+
+ if (Ty == Kind::FLOAT)
+ return Kind::VECTOR_FLOAT;
+
+ return Kind::VECTOR_ANY;
+ }
+
+ constexpr static Kind toScalar(Kind Ty) {
+ if (Ty == Kind::VECTOR_POINTER)
+ return Kind::POINTER;
+
+ if (Ty == Kind::VECTOR_INTEGER)
+ return Kind::INTEGER;
+
+ if (Ty == Kind::VECTOR_FLOAT)
+ return Kind::FLOAT;
+
+ return Kind::ANY_SCALAR;
+ }
+
/// Get a low-level scalar or aggregate "bag of bits".
static constexpr LLT scalar(unsigned SizeInBits) {
- return LLT{/*isPointer=*/false, /*isVector=*/false, /*isScalar=*/true,
- ElementCount::getFixed(0), SizeInBits,
- /*AddressSpace=*/0};
+ return LLT{Kind::ANY_SCALAR, ElementCount::getFixed(0), SizeInBits,
+ /*AddressSpace=*/0, static_cast<FPVariant>(0)};
+ }
+
+ static constexpr LLT integer(unsigned SizeInBits) {
+ return LLT{Kind::INTEGER, ElementCount::getFixed(0), SizeInBits,
+ /*AddressSpace=*/0, static_cast<FPVariant>(0)};
+ }
+
+ static constexpr LLT floatingPoint(unsigned SizeInBits, FPVariant FP) {
+ return LLT{Kind::FLOAT, ElementCount::getFixed(0), SizeInBits,
+ /*AddressSpace=*/0, FP};
}
/// Get a low-level token; just a scalar with zero bits (or no size).
static constexpr LLT token() {
- return LLT{/*isPointer=*/false, /*isVector=*/false,
- /*isScalar=*/true, ElementCount::getFixed(0),
+ return LLT{Kind::ANY_SCALAR, ElementCount::getFixed(0),
/*SizeInBits=*/0,
- /*AddressSpace=*/0};
+ /*AddressSpace=*/0, static_cast<FPVariant>(0)};
}
/// Get a low-level pointer in the given address space.
static constexpr LLT pointer(unsigned AddressSpace, unsigned SizeInBits) {
assert(SizeInBits > 0 && "invalid pointer size");
- return LLT{/*isPointer=*/true, /*isVector=*/false, /*isScalar=*/false,
- ElementCount::getFixed(0), SizeInBits, AddressSpace};
+ return LLT{Kind::POINTER, ElementCount::getFixed(0), SizeInBits,
+ AddressSpace, static_cast<FPVariant>(0)};
}
/// Get a low-level vector of some number of elements and element width.
static constexpr LLT vector(ElementCount EC, unsigned ScalarSizeInBits) {
assert(!EC.isScalar() && "invalid number of vector elements");
- return LLT{/*isPointer=*/false, /*isVector=*/true, /*isScalar=*/false,
- EC, ScalarSizeInBits, /*AddressSpace=*/0};
+ return LLT{Kind::VECTOR_ANY, EC, ScalarSizeInBits,
+ /*AddressSpace=*/0, static_cast<FPVariant>(0)};
}
/// Get a low-level vector of some number of elements and element type.
static constexpr LLT vector(ElementCount EC, LLT ScalarTy) {
assert(!EC.isScalar() && "invalid number of vector elements");
assert(!ScalarTy.isVector() && "invalid vector element type");
- return LLT{ScalarTy.isPointer(),
- /*isVector=*/true,
- /*isScalar=*/false,
- EC,
- ScalarTy.getSizeInBits().getFixedValue(),
- ScalarTy.isPointer() ? ScalarTy.getAddressSpace() : 0};
+
+ Kind Info = toVector(ScalarTy.Info);
+ return LLT{Info, EC, ScalarTy.getSizeInBits().getFixedValue(),
+ ScalarTy.isPointer() ? ScalarTy.getAddressSpace() : 0,
+ ScalarTy.isFloat() ? ScalarTy.getFPVariant()
+ : static_cast<FPVariant>(0)};
+ }
+ // Get a 8-bit brain float value.
+ static constexpr LLT bfloat8() {
+ return floatingPoint(8, FPVariant::BRAIN_FLOAT);
+ }
+
+ // Get a 16-bit brain float value.
+ static constexpr LLT bfloat16() {
+ return floatingPoint(16, FPVariant::BRAIN_FLOAT);
}
/// Get a 16-bit IEEE half value.
- /// TODO: Add IEEE semantics to type - This currently returns a simple `scalar(16)`.
static constexpr LLT float16() {
- return scalar(16);
+ return floatingPoint(16, FPVariant::IEEE_FLOAT);
}
/// Get a 32-bit IEEE float value.
static constexpr LLT float32() {
- return scalar(32);
+ return floatingPoint(32, FPVariant::IEEE_FLOAT);
}
/// Get a 64-bit IEEE double value.
static constexpr LLT float64() {
- return scalar(64);
+ return floatingPoint(64, FPVariant::IEEE_FLOAT);
+ }
+
+ /// Get a 80-bit X86 floating point value.
+ static constexpr LLT x86fp80() {
+ return floatingPoint(80, FPVariant::EXTENDED_FP80);
+ }
+
+ /// Get a 128-bit IEEE quad value.
+ static constexpr LLT float128() {
+ return floatingPoint(128, FPVariant::IEEE_FLOAT);
+ }
+
+ /// Get a 128-bit PowerPC double double value.
+ static constexpr LLT ppcf128() {
+ return floatingPoint(128, FPVariant::PPC128_FLOAT);
}
/// Get a low-level fixed-width vector of some number of elements and element
/// width.
- static constexpr LLT fixed_vector(unsigned NumElements,
- unsigned ScalarSizeInBits) {
- return vector(ElementCount::getFixed(NumElements), ScalarSizeInBits);
+ static constexpr LLT fixed_vector(unsigned NumElements, unsigned ScalarSizeInBits) {
+ return vector(ElementCount::getFixed(NumElements),
+ LLT::scalar(ScalarSizeInBits));
}
/// Get a low-level fixed-width vector of some number of elements and element
@@ -111,9 +191,9 @@ class LLT {
/// Get a low-level scalable vector of some number of elements and element
/// width.
- static constexpr LLT scalable_vector(unsigned MinNumElements,
- unsigned ScalarSizeInBits) {
- return vector(ElementCount::getScalable(MinNumElements), ScalarSizeInBits);
+ static constexpr LLT scalable_vector(unsigned MinNumElements, unsigned ScalarSizeInBits) {
+ return vector(ElementCount::getScalable(MinNumElements),
+ LLT::scalar(ScalarSizeInBits));
}
/// Get a low-level scalable vector of some number of elements and element
@@ -132,27 +212,79 @@ class LLT {
return scalarOrVector(EC, LLT::scalar(static_cast<unsigned>(ScalarSize)));
}
- explicit constexpr LLT(bool isPointer, bool isVector, bool isScalar,
- ElementCount EC, uint64_t SizeInBits,
- unsigned AddressSpace)
+ explicit constexpr LLT(Kind Info, ElementCount EC, uint64_t SizeInBits,
+ unsigned AddressSpace, FPVariant FP)
: LLT() {
- init(isPointer, isVector, isScalar, EC, SizeInBits, AddressSpace);
+ init(Info, EC, SizeInBits, AddressSpace, FP);
}
- explicit constexpr LLT()
- : IsScalar(false), IsPointer(false), IsVector(false), RawData(0) {}
- LLVM_ABI explicit LLT(MVT VT);
+ LLVM_ABI explicit LLT(MVT VT, bool AllowExtendedLLT = false);
+ explicit constexpr LLT() : Info(static_cast<Kind>(0)), RawData(0) {}
- constexpr bool isValid() const { return IsScalar || RawData != 0; }
- constexpr bool isScalar() const { return IsScalar; }
- constexpr bool isToken() const { return IsScalar && RawData == 0; };
- constexpr bool isVector() const { return isValid() && IsVector; }
+ constexpr bool isValid() const { return isToken() || RawData != 0; }
+ constexpr bool isScalar() const { return Info == Kind::ANY_SCALAR || Info == Kind::INTEGER || Info == Kind::FLOAT; }
+ constexpr bool isScalar(unsigned Size) const {
+ return isScalar() && getScalarSizeInBits() == Size;
+ }
+ constexpr bool isFloat() const { return isValid() && Info == Kind::FLOAT; }
+ constexpr bool isFloat(unsigned Size) const {
+ return isFloat() && getScalarSizeInBits() == Size;
+ }
+ constexpr bool isVariantFloat() const {
+ return isFloat() && getFPVariant() != FPVariant::IEEE_FLOAT;
+ }
+ constexpr bool isVariantFloat(FPVariant Variant) const {
+ return isFloat() && getFPVariant() == Variant;
+ }
+ constexpr bool isVariantFloat(unsigned Size, FPVariant Variant) const {
+ return isVariantFloat(Variant) && getScalarSizeInBits() == Size;
+ }
+ constexpr bool isFloatVector() const {
+ return isVector() && Info == Kind::VECTOR_FLOAT;
+ }
+ constexpr bool isIEEEFloat(unsigned Size) const {
+ return isVariantFloat(Size, FPVariant::IEEE_FLOAT);
+ }
+ constexpr bool isBFloat(unsigned Size) const {
+ return isVariantFloat(Size, FPVariant::BRAIN_FLOAT);
+ }
+ constexpr bool isX86FP80() const {
+ return isVariantFloat(80, FPVariant::EXTENDED_FP80);
+ }
+ constexpr bool isPPCF128() const {
+ return isVariantFloat(128, FPVariant::PPC128_FLOAT);
+ }
+ constexpr bool isToken() const {
+ return Info == Kind::ANY_SCALAR && RawData == 0;
+ }
+ constexpr bool isAnyScalar() const {
+ return isValid() && Info == Kind::ANY_SCALAR;
+ }
+ constexpr bool isVectorAny() const {
+ return isVector() && Info == Kind::VECTOR_ANY;
+ }
+ constexpr bool isInteger() const {
+ return isValid() && Info == Kind::INTEGER;
+ }
+ constexpr bool isInteger(unsigned Size) const {
+ return isInteger() && getScalarSizeInBits() == Size;
+ }
+ constexpr bool isIntegerVector() const {
+ return isVector() && Info == Kind::VECTOR_INTEGER;
+ }
+ constexpr bool isVector() const {
+ return isValid() &&
+ (Info == Kind::VECTOR_ANY || Info == Kind::VECTOR_INTEGER||
+ Info == Kind::VECTOR_FLOAT || Info == Kind::VECTOR_POINTER);
+ }
constexpr bool isPointer() const {
- return isValid() && IsPointer && !IsVector;
+ return isValid() && Info == Kind::POINTER;
+ }
+ constexpr bool isPointerVector() const {
+ return isVector() && Info == Kind::VECTOR_POINTER;
}
- constexpr bool isPointerVector() const { return IsPointer && isVector(); }
constexpr bool isPointerOrPointerVector() const {
- return IsPointer && isValid();
+ return isPointer() || isPointerVector();
}
/// Returns the number of elements in a vector LLT. Must only be called on
@@ -177,12 +309,18 @@ class LLT {
/// if the LLT is not a vector type.
constexpr bool isFixedVector() const { return isVector() && !isScalable(); }
+ constexpr bool isFixedVector(unsigned NumElements,
+ unsigned ScalarSize) const {
+ return isFixedVector() && getNumElements() == NumElements &&
+ getScalarSizeInBits() == ScalarSize;
+ }
+
/// Returns true if the LLT is a scalable vector. Returns false otherwise,
/// even if the LLT is not a vector type.
constexpr bool isScalableVector() const { return isVector() && isScalable(); }
constexpr ElementCount getElementCount() const {
- assert(IsVector && "cannot get number of elements on scalar/aggregate");
+ assert(isVector() && "cannot get number of elements on scalar/aggregate");
return ElementCount::get(getFieldValue(VectorElementsFieldInfo),
isScalable());
}
@@ -207,6 +345,15 @@ class LLT {
return isVector() ? getElementType() : *this;
}
+ constexpr FPVariant getFPVariant() const {
+ assert((isFloat() || isFloatVector()) &&
+ "cannot get FP info for non float type");
+
+ return FPVariant(getFieldValue(FPFieldInfo));
+ }
+
+ constexpr Kind getKind() const { return Info; }
+
/// If this type is a vector, return a vector with the same number of elements
/// but the new element type. Otherwise, return the new element type.
constexpr LLT changeElementType(LLT NewEltTy) const {
@@ -217,10 +364,14 @@ class LLT {
/// but the new element size. Otherwise, return the new element type. Invalid
/// for pointer types. For pointer types, use changeElementType.
constexpr LLT changeElementSize(unsigned NewEltSize) const {
- assert(!isPointerOrPointerVector() &&
+ assert(!isPointerOrPointerVector() && !(isFloat() || isFloatVector()) &&
"invalid to directly change element size for pointers");
- return isVector() ? LLT::vector(getElementCount(), NewEltSize)
- : LLT::scalar(NewEltSize);
+ return isVector()
+ ? LLT::vector(getElementCount(), getElementType().isInteger()
+ ? LLT::integer(NewEltSize)
+ : LLT::scalar(NewEltSize))
+ : isInteger() ? LLT::integer(NewEltSize)
+ : LLT::scalar(NewEltSize);
}
/// Return a vector or scalar with the same element type and the new element
@@ -229,6 +380,10 @@ class LLT {
return LLT::scalarOrVector(EC, getScalarType());
}
+ constexpr LLT changeElementCount(unsigned NumElements) const {
+ return changeElementCount(ElementCount::getFixed(NumElements));
+ }
+
/// Return a type that is \p Factor times smaller. Reduces the number of
/// elements if this is a vector, or the bitwidth for scalar/pointers. Does
/// not attempt to handle cases that aren't evenly divisible.
@@ -243,6 +398,9 @@ class LLT {
}
assert(getScalarSizeInBits() % Factor == 0);
+ if (isInteger())
+ return integer(getScalarSizeInBits() / Factor);
+
return scalar(getScalarSizeInBits() / Factor);
}
@@ -277,10 +435,26 @@ class LLT {
/// Returns the vector's element type. Only valid for vector types.
constexpr LLT getElementType() const {
assert(isVector() && "cannot get element type of scalar/aggregate");
- if (IsPointer)
+ if (isPointerVector())
return pointer(getAddressSpace(), getScalarSizeInBits());
- else
- return scalar(getScalarSizeInBits());
+
+ if (isFloatVector())
+ return floatingPoint(getScalarSizeInBits(), getFPVariant());
+
+ if (isIntegerVector())
+ return integer(getScalarSizeInBits());
+
+ return scalar(getScalarSizeInBits());
+ }
+
+ constexpr LLT changeToInteger() const {
+ if (isPointer() || isPointerVector())
+ return *this;
+
+ if (isVector())
+ return vector(getElementCount(), LLT::integer(getScalarSizeInBits()));
+
+ return integer(getSizeInBits());
}
LLVM_ABI void print(raw_ostream &OS) const;
@@ -290,8 +464,14 @@ class LLT {
#endif
constexpr bool operator==(const LLT &RHS) const {
- return IsPointer == RHS.IsPointer && IsVector == RHS.IsVector &&
- IsScalar == RHS.IsScalar && RHS.RawData == RawData;
+ if (isAnyScalar() || RHS.isAnyScalar()) {
+ return isScalar() == RHS.isScalar() && RawData == RHS.RawData;
+ }
+ if (isVector() && RHS.isVector()) {
+ return getElementType() == RHS.getElementType() &&
+ getElementCount() == RHS.getElementCount();
+ }
+ return Info == RHS.Info && RawData == RHS.RawData;
}
constexpr bool operator!=(const LLT &RHS) const { return !(*this == RHS); }
@@ -301,37 +481,33 @@ class LLT {
private:
/// LLT is packed into 64 bits as follows:
- /// isScalar : 1
- /// isPointer : 1
- /// isVector : 1
- /// with 61 bits remaining for Kind-specific data, packed in bitfields
- /// as described below. As there isn't a simple portable way to pack bits
- /// into bitfields, here the different fields in the packed structure is
+ /// Info : 4
+ /// RawData : 60
+ /// with 61 bits of RawData remaining for Kind-specific data, packed in
+ /// bitfields as described below. As there isn't a simple portable way to pack
+ /// bits into bitfields, here the different fields in the packed structure is
/// described in static const *Field variables. Each of these variables
/// is a 2-element array, with the first ele...
[truncated]
|
Member
|
@llvm/pr-subscribers-backend-aarch64 Author: Denis.G (DenisGZM) ChangesAdded extra information in LLT to support ambiguous fp types during GlobalISel. Original idea by @tgymnich Main differences from #122503 are:
Also MIRParser now may parse new type indetifiers. Patch is 97.84 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/155107.diff 50 Files Affected:
diff --git a/llvm/include/llvm/CodeGen/Analysis.h b/llvm/include/llvm/CodeGen/Analysis.h
index 98b52579d03b7..2aadc27930b7a 100644
--- a/llvm/include/llvm/CodeGen/Analysis.h
+++ b/llvm/include/llvm/CodeGen/Analysis.h
@@ -101,7 +101,8 @@ inline void ComputeValueVTs(const TargetLowering &TLI, const DataLayout &DL,
/// If Offsets is non-null, it points to a vector to be filled in
/// with the in-memory offsets of each of the individual values.
///
-void computeValueLLTs(const DataLayout &DL, Type &Ty,
+void computeValueLLTs(const TargetLowering &TLI,
+ const DataLayout &DL, Type &Ty,
SmallVectorImpl<LLT> &ValueTys,
SmallVectorImpl<uint64_t> *Offsets = nullptr,
uint64_t StartingOffset = 0);
diff --git a/llvm/include/llvm/CodeGen/GlobalISel/IRTranslator.h b/llvm/include/llvm/CodeGen/GlobalISel/IRTranslator.h
index 3d7ccd55ee042..5673fd5168477 100644
--- a/llvm/include/llvm/CodeGen/GlobalISel/IRTranslator.h
+++ b/llvm/include/llvm/CodeGen/GlobalISel/IRTranslator.h
@@ -642,6 +642,8 @@ class IRTranslator : public MachineFunctionPass {
StackProtectorDescriptor SPDescriptor;
+ bool mayTranslateUserTypes(const User &U) const;
+
/// Switch analysis and optimization.
class GISelSwitchLowering : public SwitchCG::SwitchLowering {
public:
diff --git a/llvm/include/llvm/CodeGen/GlobalISel/LegalizerInfo.h b/llvm/include/llvm/CodeGen/GlobalISel/LegalizerInfo.h
index fd72a3898562e..8412042a780e8 100644
--- a/llvm/include/llvm/CodeGen/GlobalISel/LegalizerInfo.h
+++ b/llvm/include/llvm/CodeGen/GlobalISel/LegalizerInfo.h
@@ -380,7 +380,7 @@ LLVM_ABI LegalizeMutation changeElementCountTo(unsigned TypeIdx,
/// Keep the same scalar or element type as \p TypeIdx, but take the number of
/// elements from \p Ty.
-LLVM_ABI LegalizeMutation changeElementCountTo(unsigned TypeIdx, LLT Ty);
+LLVM_ABI LegalizeMutation changeElementCountTo(unsigned TypeIdx, ElementCount EC);
/// Change the scalar size or element size to have the same scalar size as type
/// index \p FromIndex. Unlike changeElementTo, this discards pointer types and
diff --git a/llvm/include/llvm/CodeGen/LowLevelTypeUtils.h b/llvm/include/llvm/CodeGen/LowLevelTypeUtils.h
index 51a298eb8b247..1c8f34b1a8518 100644
--- a/llvm/include/llvm/CodeGen/LowLevelTypeUtils.h
+++ b/llvm/include/llvm/CodeGen/LowLevelTypeUtils.h
@@ -27,7 +27,7 @@ class Type;
struct fltSemantics;
/// Construct a low-level type based on an LLVM type.
-LLVM_ABI LLT getLLTForType(Type &Ty, const DataLayout &DL);
+LLVM_ABI LLT getLLTForType(Type &Ty, const DataLayout &DL, bool AllowExtendedLLT = false);
/// Get a rough equivalent of an MVT for a given LLT. MVT can't distinguish
/// pointers, so these will convert to a plain integer.
@@ -36,11 +36,11 @@ LLVM_ABI EVT getApproximateEVTForLLT(LLT Ty, LLVMContext &Ctx);
/// Get a rough equivalent of an LLT for a given MVT. LLT does not yet support
/// scalarable vector types, and will assert if used.
-LLVM_ABI LLT getLLTForMVT(MVT Ty);
+LLVM_ABI LLT getLLTForMVT(MVT Ty, bool AllowExtendedLLT = false);
/// Get the appropriate floating point arithmetic semantic based on the bit size
/// of the given scalar LLT.
LLVM_ABI const llvm::fltSemantics &getFltSemanticForLLT(LLT Ty);
-}
+} // namespace llvm
#endif // LLVM_CODEGEN_LOWLEVELTYPEUTILS_H
diff --git a/llvm/include/llvm/CodeGen/TargetLowering.h b/llvm/include/llvm/CodeGen/TargetLowering.h
index 4480ced637456..f4fdd8c69ea9e 100644
--- a/llvm/include/llvm/CodeGen/TargetLowering.h
+++ b/llvm/include/llvm/CodeGen/TargetLowering.h
@@ -416,6 +416,9 @@ class LLVM_ABI TargetLoweringBase {
/// amounts, returns MVT::i32.
EVT getShiftAmountTy(EVT LHSTy, const DataLayout &DL) const;
+ virtual LLT getLLTForType(Type &Ty, const DataLayout &DL) const;
+ virtual LLT getLLTForMVT(MVT Ty) const;
+
/// Return the preferred type to use for a shift opcode, given the shifted
/// amount type is \p ShiftValueTy.
LLVM_READONLY
diff --git a/llvm/include/llvm/CodeGenTypes/LowLevelType.h b/llvm/include/llvm/CodeGenTypes/LowLevelType.h
index d8e0848aff84d..0c217ba1fcf7e 100644
--- a/llvm/include/llvm/CodeGenTypes/LowLevelType.h
+++ b/llvm/include/llvm/CodeGenTypes/LowLevelType.h
@@ -39,68 +39,148 @@ class raw_ostream;
class LLT {
public:
+ enum class FPVariant {
+ IEEE_FLOAT = 0x0,
+ BRAIN_FLOAT = 0x1, // BRAIN_FLOAT
+ PPC128_FLOAT = 0x2, // PPC128_FLOAT
+ EXTENDED_FP80 = 0x3, // FP80
+ TENSOR_FLOAT32 = 0x4, // TENSOR_FLOAT32
+ VARIANT_FLOAT_5 = 0x5, // UNASSIGNED
+ VARIANT_FLOAT_6 = 0x6, // UNASSIGNED
+ VARIANT_FLOAT_7 = 0x7, // UNASSIGNED
+ };
+
+ enum class Kind : uint64_t {
+ INVALID = 0b0000,
+ ANY_SCALAR = 0b0001,
+ INTEGER = 0b0010,
+ FLOAT = 0b0011,
+ POINTER = 0b0100,
+ VECTOR_ANY = 0b0101,
+ VECTOR_INTEGER = 0b0110,
+ VECTOR_FLOAT = 0b0111,
+ VECTOR_POINTER = 0b1000,
+ };
+
+ constexpr static Kind toVector(Kind Ty) {
+ if (Ty == Kind::POINTER)
+ return Kind::VECTOR_POINTER;
+
+ if (Ty == Kind::INTEGER)
+ return Kind::VECTOR_INTEGER;
+
+ if (Ty == Kind::FLOAT)
+ return Kind::VECTOR_FLOAT;
+
+ return Kind::VECTOR_ANY;
+ }
+
+ constexpr static Kind toScalar(Kind Ty) {
+ if (Ty == Kind::VECTOR_POINTER)
+ return Kind::POINTER;
+
+ if (Ty == Kind::VECTOR_INTEGER)
+ return Kind::INTEGER;
+
+ if (Ty == Kind::VECTOR_FLOAT)
+ return Kind::FLOAT;
+
+ return Kind::ANY_SCALAR;
+ }
+
/// Get a low-level scalar or aggregate "bag of bits".
static constexpr LLT scalar(unsigned SizeInBits) {
- return LLT{/*isPointer=*/false, /*isVector=*/false, /*isScalar=*/true,
- ElementCount::getFixed(0), SizeInBits,
- /*AddressSpace=*/0};
+ return LLT{Kind::ANY_SCALAR, ElementCount::getFixed(0), SizeInBits,
+ /*AddressSpace=*/0, static_cast<FPVariant>(0)};
+ }
+
+ static constexpr LLT integer(unsigned SizeInBits) {
+ return LLT{Kind::INTEGER, ElementCount::getFixed(0), SizeInBits,
+ /*AddressSpace=*/0, static_cast<FPVariant>(0)};
+ }
+
+ static constexpr LLT floatingPoint(unsigned SizeInBits, FPVariant FP) {
+ return LLT{Kind::FLOAT, ElementCount::getFixed(0), SizeInBits,
+ /*AddressSpace=*/0, FP};
}
/// Get a low-level token; just a scalar with zero bits (or no size).
static constexpr LLT token() {
- return LLT{/*isPointer=*/false, /*isVector=*/false,
- /*isScalar=*/true, ElementCount::getFixed(0),
+ return LLT{Kind::ANY_SCALAR, ElementCount::getFixed(0),
/*SizeInBits=*/0,
- /*AddressSpace=*/0};
+ /*AddressSpace=*/0, static_cast<FPVariant>(0)};
}
/// Get a low-level pointer in the given address space.
static constexpr LLT pointer(unsigned AddressSpace, unsigned SizeInBits) {
assert(SizeInBits > 0 && "invalid pointer size");
- return LLT{/*isPointer=*/true, /*isVector=*/false, /*isScalar=*/false,
- ElementCount::getFixed(0), SizeInBits, AddressSpace};
+ return LLT{Kind::POINTER, ElementCount::getFixed(0), SizeInBits,
+ AddressSpace, static_cast<FPVariant>(0)};
}
/// Get a low-level vector of some number of elements and element width.
static constexpr LLT vector(ElementCount EC, unsigned ScalarSizeInBits) {
assert(!EC.isScalar() && "invalid number of vector elements");
- return LLT{/*isPointer=*/false, /*isVector=*/true, /*isScalar=*/false,
- EC, ScalarSizeInBits, /*AddressSpace=*/0};
+ return LLT{Kind::VECTOR_ANY, EC, ScalarSizeInBits,
+ /*AddressSpace=*/0, static_cast<FPVariant>(0)};
}
/// Get a low-level vector of some number of elements and element type.
static constexpr LLT vector(ElementCount EC, LLT ScalarTy) {
assert(!EC.isScalar() && "invalid number of vector elements");
assert(!ScalarTy.isVector() && "invalid vector element type");
- return LLT{ScalarTy.isPointer(),
- /*isVector=*/true,
- /*isScalar=*/false,
- EC,
- ScalarTy.getSizeInBits().getFixedValue(),
- ScalarTy.isPointer() ? ScalarTy.getAddressSpace() : 0};
+
+ Kind Info = toVector(ScalarTy.Info);
+ return LLT{Info, EC, ScalarTy.getSizeInBits().getFixedValue(),
+ ScalarTy.isPointer() ? ScalarTy.getAddressSpace() : 0,
+ ScalarTy.isFloat() ? ScalarTy.getFPVariant()
+ : static_cast<FPVariant>(0)};
+ }
+ // Get a 8-bit brain float value.
+ static constexpr LLT bfloat8() {
+ return floatingPoint(8, FPVariant::BRAIN_FLOAT);
+ }
+
+ // Get a 16-bit brain float value.
+ static constexpr LLT bfloat16() {
+ return floatingPoint(16, FPVariant::BRAIN_FLOAT);
}
/// Get a 16-bit IEEE half value.
- /// TODO: Add IEEE semantics to type - This currently returns a simple `scalar(16)`.
static constexpr LLT float16() {
- return scalar(16);
+ return floatingPoint(16, FPVariant::IEEE_FLOAT);
}
/// Get a 32-bit IEEE float value.
static constexpr LLT float32() {
- return scalar(32);
+ return floatingPoint(32, FPVariant::IEEE_FLOAT);
}
/// Get a 64-bit IEEE double value.
static constexpr LLT float64() {
- return scalar(64);
+ return floatingPoint(64, FPVariant::IEEE_FLOAT);
+ }
+
+ /// Get a 80-bit X86 floating point value.
+ static constexpr LLT x86fp80() {
+ return floatingPoint(80, FPVariant::EXTENDED_FP80);
+ }
+
+ /// Get a 128-bit IEEE quad value.
+ static constexpr LLT float128() {
+ return floatingPoint(128, FPVariant::IEEE_FLOAT);
+ }
+
+ /// Get a 128-bit PowerPC double double value.
+ static constexpr LLT ppcf128() {
+ return floatingPoint(128, FPVariant::PPC128_FLOAT);
}
/// Get a low-level fixed-width vector of some number of elements and element
/// width.
- static constexpr LLT fixed_vector(unsigned NumElements,
- unsigned ScalarSizeInBits) {
- return vector(ElementCount::getFixed(NumElements), ScalarSizeInBits);
+ static constexpr LLT fixed_vector(unsigned NumElements, unsigned ScalarSizeInBits) {
+ return vector(ElementCount::getFixed(NumElements),
+ LLT::scalar(ScalarSizeInBits));
}
/// Get a low-level fixed-width vector of some number of elements and element
@@ -111,9 +191,9 @@ class LLT {
/// Get a low-level scalable vector of some number of elements and element
/// width.
- static constexpr LLT scalable_vector(unsigned MinNumElements,
- unsigned ScalarSizeInBits) {
- return vector(ElementCount::getScalable(MinNumElements), ScalarSizeInBits);
+ static constexpr LLT scalable_vector(unsigned MinNumElements, unsigned ScalarSizeInBits) {
+ return vector(ElementCount::getScalable(MinNumElements),
+ LLT::scalar(ScalarSizeInBits));
}
/// Get a low-level scalable vector of some number of elements and element
@@ -132,27 +212,79 @@ class LLT {
return scalarOrVector(EC, LLT::scalar(static_cast<unsigned>(ScalarSize)));
}
- explicit constexpr LLT(bool isPointer, bool isVector, bool isScalar,
- ElementCount EC, uint64_t SizeInBits,
- unsigned AddressSpace)
+ explicit constexpr LLT(Kind Info, ElementCount EC, uint64_t SizeInBits,
+ unsigned AddressSpace, FPVariant FP)
: LLT() {
- init(isPointer, isVector, isScalar, EC, SizeInBits, AddressSpace);
+ init(Info, EC, SizeInBits, AddressSpace, FP);
}
- explicit constexpr LLT()
- : IsScalar(false), IsPointer(false), IsVector(false), RawData(0) {}
- LLVM_ABI explicit LLT(MVT VT);
+ LLVM_ABI explicit LLT(MVT VT, bool AllowExtendedLLT = false);
+ explicit constexpr LLT() : Info(static_cast<Kind>(0)), RawData(0) {}
- constexpr bool isValid() const { return IsScalar || RawData != 0; }
- constexpr bool isScalar() const { return IsScalar; }
- constexpr bool isToken() const { return IsScalar && RawData == 0; };
- constexpr bool isVector() const { return isValid() && IsVector; }
+ constexpr bool isValid() const { return isToken() || RawData != 0; }
+ constexpr bool isScalar() const { return Info == Kind::ANY_SCALAR || Info == Kind::INTEGER || Info == Kind::FLOAT; }
+ constexpr bool isScalar(unsigned Size) const {
+ return isScalar() && getScalarSizeInBits() == Size;
+ }
+ constexpr bool isFloat() const { return isValid() && Info == Kind::FLOAT; }
+ constexpr bool isFloat(unsigned Size) const {
+ return isFloat() && getScalarSizeInBits() == Size;
+ }
+ constexpr bool isVariantFloat() const {
+ return isFloat() && getFPVariant() != FPVariant::IEEE_FLOAT;
+ }
+ constexpr bool isVariantFloat(FPVariant Variant) const {
+ return isFloat() && getFPVariant() == Variant;
+ }
+ constexpr bool isVariantFloat(unsigned Size, FPVariant Variant) const {
+ return isVariantFloat(Variant) && getScalarSizeInBits() == Size;
+ }
+ constexpr bool isFloatVector() const {
+ return isVector() && Info == Kind::VECTOR_FLOAT;
+ }
+ constexpr bool isIEEEFloat(unsigned Size) const {
+ return isVariantFloat(Size, FPVariant::IEEE_FLOAT);
+ }
+ constexpr bool isBFloat(unsigned Size) const {
+ return isVariantFloat(Size, FPVariant::BRAIN_FLOAT);
+ }
+ constexpr bool isX86FP80() const {
+ return isVariantFloat(80, FPVariant::EXTENDED_FP80);
+ }
+ constexpr bool isPPCF128() const {
+ return isVariantFloat(128, FPVariant::PPC128_FLOAT);
+ }
+ constexpr bool isToken() const {
+ return Info == Kind::ANY_SCALAR && RawData == 0;
+ }
+ constexpr bool isAnyScalar() const {
+ return isValid() && Info == Kind::ANY_SCALAR;
+ }
+ constexpr bool isVectorAny() const {
+ return isVector() && Info == Kind::VECTOR_ANY;
+ }
+ constexpr bool isInteger() const {
+ return isValid() && Info == Kind::INTEGER;
+ }
+ constexpr bool isInteger(unsigned Size) const {
+ return isInteger() && getScalarSizeInBits() == Size;
+ }
+ constexpr bool isIntegerVector() const {
+ return isVector() && Info == Kind::VECTOR_INTEGER;
+ }
+ constexpr bool isVector() const {
+ return isValid() &&
+ (Info == Kind::VECTOR_ANY || Info == Kind::VECTOR_INTEGER||
+ Info == Kind::VECTOR_FLOAT || Info == Kind::VECTOR_POINTER);
+ }
constexpr bool isPointer() const {
- return isValid() && IsPointer && !IsVector;
+ return isValid() && Info == Kind::POINTER;
+ }
+ constexpr bool isPointerVector() const {
+ return isVector() && Info == Kind::VECTOR_POINTER;
}
- constexpr bool isPointerVector() const { return IsPointer && isVector(); }
constexpr bool isPointerOrPointerVector() const {
- return IsPointer && isValid();
+ return isPointer() || isPointerVector();
}
/// Returns the number of elements in a vector LLT. Must only be called on
@@ -177,12 +309,18 @@ class LLT {
/// if the LLT is not a vector type.
constexpr bool isFixedVector() const { return isVector() && !isScalable(); }
+ constexpr bool isFixedVector(unsigned NumElements,
+ unsigned ScalarSize) const {
+ return isFixedVector() && getNumElements() == NumElements &&
+ getScalarSizeInBits() == ScalarSize;
+ }
+
/// Returns true if the LLT is a scalable vector. Returns false otherwise,
/// even if the LLT is not a vector type.
constexpr bool isScalableVector() const { return isVector() && isScalable(); }
constexpr ElementCount getElementCount() const {
- assert(IsVector && "cannot get number of elements on scalar/aggregate");
+ assert(isVector() && "cannot get number of elements on scalar/aggregate");
return ElementCount::get(getFieldValue(VectorElementsFieldInfo),
isScalable());
}
@@ -207,6 +345,15 @@ class LLT {
return isVector() ? getElementType() : *this;
}
+ constexpr FPVariant getFPVariant() const {
+ assert((isFloat() || isFloatVector()) &&
+ "cannot get FP info for non float type");
+
+ return FPVariant(getFieldValue(FPFieldInfo));
+ }
+
+ constexpr Kind getKind() const { return Info; }
+
/// If this type is a vector, return a vector with the same number of elements
/// but the new element type. Otherwise, return the new element type.
constexpr LLT changeElementType(LLT NewEltTy) const {
@@ -217,10 +364,14 @@ class LLT {
/// but the new element size. Otherwise, return the new element type. Invalid
/// for pointer types. For pointer types, use changeElementType.
constexpr LLT changeElementSize(unsigned NewEltSize) const {
- assert(!isPointerOrPointerVector() &&
+ assert(!isPointerOrPointerVector() && !(isFloat() || isFloatVector()) &&
"invalid to directly change element size for pointers");
- return isVector() ? LLT::vector(getElementCount(), NewEltSize)
- : LLT::scalar(NewEltSize);
+ return isVector()
+ ? LLT::vector(getElementCount(), getElementType().isInteger()
+ ? LLT::integer(NewEltSize)
+ : LLT::scalar(NewEltSize))
+ : isInteger() ? LLT::integer(NewEltSize)
+ : LLT::scalar(NewEltSize);
}
/// Return a vector or scalar with the same element type and the new element
@@ -229,6 +380,10 @@ class LLT {
return LLT::scalarOrVector(EC, getScalarType());
}
+ constexpr LLT changeElementCount(unsigned NumElements) const {
+ return changeElementCount(ElementCount::getFixed(NumElements));
+ }
+
/// Return a type that is \p Factor times smaller. Reduces the number of
/// elements if this is a vector, or the bitwidth for scalar/pointers. Does
/// not attempt to handle cases that aren't evenly divisible.
@@ -243,6 +398,9 @@ class LLT {
}
assert(getScalarSizeInBits() % Factor == 0);
+ if (isInteger())
+ return integer(getScalarSizeInBits() / Factor);
+
return scalar(getScalarSizeInBits() / Factor);
}
@@ -277,10 +435,26 @@ class LLT {
/// Returns the vector's element type. Only valid for vector types.
constexpr LLT getElementType() const {
assert(isVector() && "cannot get element type of scalar/aggregate");
- if (IsPointer)
+ if (isPointerVector())
return pointer(getAddressSpace(), getScalarSizeInBits());
- else
- return scalar(getScalarSizeInBits());
+
+ if (isFloatVector())
+ return floatingPoint(getScalarSizeInBits(), getFPVariant());
+
+ if (isIntegerVector())
+ return integer(getScalarSizeInBits());
+
+ return scalar(getScalarSizeInBits());
+ }
+
+ constexpr LLT changeToInteger() const {
+ if (isPointer() || isPointerVector())
+ return *this;
+
+ if (isVector())
+ return vector(getElementCount(), LLT::integer(getScalarSizeInBits()));
+
+ return integer(getSizeInBits());
}
LLVM_ABI void print(raw_ostream &OS) const;
@@ -290,8 +464,14 @@ class LLT {
#endif
constexpr bool operator==(const LLT &RHS) const {
- return IsPointer == RHS.IsPointer && IsVector == RHS.IsVector &&
- IsScalar == RHS.IsScalar && RHS.RawData == RawData;
+ if (isAnyScalar() || RHS.isAnyScalar()) {
+ return isScalar() == RHS.isScalar() && RawData == RHS.RawData;
+ }
+ if (isVector() && RHS.isVector()) {
+ return getElementType() == RHS.getElementType() &&
+ getElementCount() == RHS.getElementCount();
+ }
+ return Info == RHS.Info && RawData == RHS.RawData;
}
constexpr bool operator!=(const LLT &RHS) const { return !(*this == RHS); }
@@ -301,37 +481,33 @@ class LLT {
private:
/// LLT is packed into 64 bits as follows:
- /// isScalar : 1
- /// isPointer : 1
- /// isVector : 1
- /// with 61 bits remaining for Kind-specific data, packed in bitfields
- /// as described below. As there isn't a simple portable way to pack bits
- /// into bitfields, here the different fields in the packed structure is
+ /// Info : 4
+ /// RawData : 60
+ /// with 61 bits of RawData remaining for Kind-specific data, packed in
+ /// bitfields as described below. As there isn't a simple portable way to pack
+ /// bits into bitfields, here the different fields in the packed structure is
/// described in static const *Field variables. Each of these variables
/// is a 2-element array, with the first ele...
[truncated]
|
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
Contributor
Author
DenisGZM
changed the title
…WLLTData`. Added simple test for GlobalSelEmitter
Member
|
I once contributed to LLVM for a few months, and I recall the check-in process is like when you get enough number of approvals, it is good to be checked in. Given that this is an important feature that will unblock many downstream backends, can it be checked in now? If you still believe @arsenm review is necessary, I will write an email to them. |
Contributor
Author
Yes, please. I think any fresh look would be helpful. I don't mind landing this, but the final decision is up to maintainers |
Member
|
Hi @aemerson , @tgymnich , @Pierre-vh , @DenisGZM , I sent an email to @arsenm two days ago but didn't get reply yet. Is there an alternate reviewer who can do the final approval? |
Contributor
|
I have started attempting to use these changes and I think there is an issue preventing this from compiling. When I try to compile I get the following: I believe this is because the CMAKE in |
Contributor
Author
Added linking |
Contributor
|
Hi, I was wondering if you had any thoughts on an issue I'm encountering downstream of this patch. A lot of tablegen patterns that used to just rely on size, now care about type as well. This is causing a lot of tests to fail in the I'm wondering if we need some kind of machinery to compare only by size when comparing a generic scalar type in the match table? Maybe this should be the case more generally too (for example when comparing an s32 with an fp32)? |
spavloff
reviewed
Nov 11, 2025
Contributor
Author
I haven't tried to support this so far. I'll try to explore this issue asap. I think we definitely shouldn't duplicate patterns, but we may have to ensure we legalized new LLT types before selection.
Right now, compares (==) with scalar (new kind ANY_SCALAR) are processed only by size. I will check if tables could properly use it |
Contributor
|
Thank you for your thoughts and help.
All types should be legal before selection right? So I imagine you can operate on the assumption that any backend wanting to enable this will have gone through the efforts to legalize these types.
If I understand correctly, the |
Contributor
Author
I think not only types that comes from IRTranslator must be legal, but all other types are used during selection must also transform to more specific variants (integers, floating points). But I understand that it would be great if targets could continue use any_scalars.
You're right, we would compare rawdata. RawData is separate from type, but for floating points it also contained FPVariant which would break comparison. In current patch I fixed this logic. I checked GlobalISelMatcher after compare fix. I guess I found the main problem with it. Let say:
I this case we would like to apply any rule in either i32 or f32 groups. But as soon as we match the root type with any of them, we will abort the matching if the selection cannot be applied in the corresponding group. As I see, if only use extended types with match tables generated for extended llt, this problem may be avoided. Or we need to update logic in MatchTableEmitter to allow continue matching. |
Contributor
Author
|
I reworked FPVariants. Right now we rely on APFloat::Semantics to choose between different floating point variants.
Please, have a look once more |
tgymnich
reviewed
Nov 16, 2025
Member
There was a problem hiding this comment.
I have a few minor concerns with using APFloat::Semantics:
Using APFloat::Semantics couples LLT to APFloat::Semantics with a small chance that APFloat::Semantics might outgrow the 7-bits in LLT one day.
Additionally APFloat::Semantics does not currently represent all possible hardware floating point types (e.g. nvfp4).
I also have some reservations about having ANY_SCALAR and ANY_VECTOR kinds.
I assume it's ok to have them if we only use them in rules and forbid them as register types.
Overall this looks good to me and I want to see how this plays out once merged.
spavloff
reviewed
Nov 26, 2025
| using namespace gi; | ||
|
|
||
| static cl::opt<bool> | ||
| AllowExtendedLLT("gisel-extended-llt", |
Collaborator
There was a problem hiding this comment.
It looks like this option is used only in one test?
Contributor
Author
There was a problem hiding this comment.
Yes, you're right. This is because the patch didn't mean to enable extended llt in all available targets. Target maintainers can decide if they want to support such changes and if they are ready for it.
Or the question is, why are there so few tests?
Collaborator
There was a problem hiding this comment.
I don't unerstand how the target differentiation (supports extended LLT - does not support) works. If I want to enable ExtLLT in some target, what should do with TableGen?
Maybe it worth providing support in some target (SPIRV?) to demonstrate how this should work?
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
10 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Added extra information in LLT to support ambiguous fp types during GlobalISel. Original idea by @tgymnich
Main differences from #122503 are:
TargetOption::EnableGlobalISelExtendedLLT(disabled by default)IRTranslatoruseTargetLoweringInfofor appropriateLLTgeneration.LLT::float32for float types as they are real now. AddedTODOfor such cases.Also MIRParser now may parse new type indentifiers.