summaryrefslogtreecommitdiff
path: root/lib/Target
diff options
context:
space:
mode:
Diffstat (limited to 'lib/Target')
-rw-r--r--lib/Target/AMDGPU/AMDGPU.td9
-rw-r--r--lib/Target/AMDGPU/AMDGPUFeatures.td9
-rw-r--r--lib/Target/AMDGPU/R600ISelLowering.cpp88
-rw-r--r--lib/Target/AMDGPU/R600ISelLowering.h8
-rw-r--r--lib/Target/AMDGPU/VOP3Instructions.td11
-rw-r--r--lib/Target/NVPTX/NVPTXISelDAGToDAG.cpp116
6 files changed, 160 insertions, 81 deletions
diff --git a/lib/Target/AMDGPU/AMDGPU.td b/lib/Target/AMDGPU/AMDGPU.td
index 16c2a366db285..445b69b35eb15 100644
--- a/lib/Target/AMDGPU/AMDGPU.td
+++ b/lib/Target/AMDGPU/AMDGPU.td
@@ -267,15 +267,6 @@ def FeatureD16PreservesUnusedBits : SubtargetFeature<
// Subtarget Features (options and debugging)
//===------------------------------------------------------------===//
-// Some instructions do not support denormals despite this flag. Using
-// fp32 denormals also causes instructions to run at the double
-// precision rate for the device.
-def FeatureFP32Denormals : SubtargetFeature<"fp32-denormals",
- "FP32Denormals",
- "true",
- "Enable single precision denormal handling"
->;
-
// Denormal handling for fp64 and fp16 is controlled by the same
// config register when fp16 supported.
// TODO: Do we need a separate f16 setting when not legal?
diff --git a/lib/Target/AMDGPU/AMDGPUFeatures.td b/lib/Target/AMDGPU/AMDGPUFeatures.td
index b375cae9018ea..3c7d8a8fc5509 100644
--- a/lib/Target/AMDGPU/AMDGPUFeatures.td
+++ b/lib/Target/AMDGPU/AMDGPUFeatures.td
@@ -19,6 +19,15 @@ def FeatureFMA : SubtargetFeature<"fmaf",
"Enable single precision FMA (not as fast as mul+add, but fused)"
>;
+// Some instructions do not support denormals despite this flag. Using
+// fp32 denormals also causes instructions to run at the double
+// precision rate for the device.
+def FeatureFP32Denormals : SubtargetFeature<"fp32-denormals",
+ "FP32Denormals",
+ "true",
+ "Enable single precision denormal handling"
+>;
+
class SubtargetFeatureLocalMemorySize <int Value> : SubtargetFeature<
"localmemorysize"#Value,
"LocalMemorySize",
diff --git a/lib/Target/AMDGPU/R600ISelLowering.cpp b/lib/Target/AMDGPU/R600ISelLowering.cpp
index 113d6249fa60a..e00dffc4be99a 100644
--- a/lib/Target/AMDGPU/R600ISelLowering.cpp
+++ b/lib/Target/AMDGPU/R600ISelLowering.cpp
@@ -903,7 +903,7 @@ SDValue R600TargetLowering::LowerImplicitParameter(SelectionDAG &DAG, EVT VT,
unsigned DwordOffset) const {
unsigned ByteOffset = DwordOffset * 4;
PointerType * PtrType = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
- AMDGPUASI.CONSTANT_BUFFER_0);
+ AMDGPUASI.PARAM_I_ADDRESS);
// We shouldn't be using an offset wider than 16-bits for implicit parameters.
assert(isInt<16>(ByteOffset));
@@ -1457,33 +1457,17 @@ SDValue R600TargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
return scalarizeVectorLoad(LoadNode, DAG);
}
+ // This is still used for explicit load from addrspace(8)
int ConstantBlock = ConstantAddressBlock(LoadNode->getAddressSpace());
if (ConstantBlock > -1 &&
((LoadNode->getExtensionType() == ISD::NON_EXTLOAD) ||
(LoadNode->getExtensionType() == ISD::ZEXTLOAD))) {
SDValue Result;
- if (isa<ConstantExpr>(LoadNode->getMemOperand()->getValue()) ||
- isa<Constant>(LoadNode->getMemOperand()->getValue()) ||
+ if (isa<Constant>(LoadNode->getMemOperand()->getValue()) ||
isa<ConstantSDNode>(Ptr)) {
- SDValue Slots[4];
- for (unsigned i = 0; i < 4; i++) {
- // We want Const position encoded with the following formula :
- // (((512 + (kc_bank << 12) + const_index) << 2) + chan)
- // const_index is Ptr computed by llvm using an alignment of 16.
- // Thus we add (((512 + (kc_bank << 12)) + chan ) * 4 here and
- // then div by 4 at the ISel step
- SDValue NewPtr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
- DAG.getConstant(4 * i + ConstantBlock * 16, DL, MVT::i32));
- Slots[i] = DAG.getNode(AMDGPUISD::CONST_ADDRESS, DL, MVT::i32, NewPtr);
- }
- EVT NewVT = MVT::v4i32;
- unsigned NumElements = 4;
- if (VT.isVector()) {
- NewVT = VT;
- NumElements = VT.getVectorNumElements();
- }
- Result = DAG.getBuildVector(NewVT, DL, makeArrayRef(Slots, NumElements));
+ return constBufferLoad(LoadNode, LoadNode->getAddressSpace(), DAG);
} else {
+ //TODO: Does this even work?
// non-constant ptr can't be folded, keeps it as a v4f32 load
Result = DAG.getNode(AMDGPUISD::CONST_ADDRESS, DL, MVT::v4i32,
DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
@@ -1622,7 +1606,7 @@ SDValue R600TargetLowering::LowerFormalArguments(
}
PointerType *PtrTy = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
- AMDGPUASI.CONSTANT_BUFFER_0);
+ AMDGPUASI.PARAM_I_ADDRESS);
// i64 isn't a legal type, so the register type used ends up as i32, which
// isn't expected here. It attempts to create this sextload, but it ends up
@@ -1646,17 +1630,17 @@ SDValue R600TargetLowering::LowerFormalArguments(
unsigned ValBase = ArgLocs[In.getOrigArgIndex()].getLocMemOffset();
unsigned PartOffset = VA.getLocMemOffset();
+ unsigned Alignment = MinAlign(VT.getStoreSize(), PartOffset);
MachinePointerInfo PtrInfo(UndefValue::get(PtrTy), PartOffset - ValBase);
SDValue Arg = DAG.getLoad(
ISD::UNINDEXED, Ext, VT, DL, Chain,
DAG.getConstant(PartOffset, DL, MVT::i32), DAG.getUNDEF(MVT::i32),
PtrInfo,
- MemVT, /* Alignment = */ 4, MachineMemOperand::MONonTemporal |
+ MemVT, Alignment, MachineMemOperand::MONonTemporal |
MachineMemOperand::MODereferenceable |
MachineMemOperand::MOInvariant);
- // 4 is the preferred alignment for the CONSTANT memory space.
InVals.push_back(Arg);
}
return Chain;
@@ -1804,6 +1788,52 @@ SDValue R600TargetLowering::OptimizeSwizzle(SDValue BuildVector, SDValue Swz[4],
return BuildVector;
}
+SDValue R600TargetLowering::constBufferLoad(LoadSDNode *LoadNode, int Block,
+ SelectionDAG &DAG) const {
+ SDLoc DL(LoadNode);
+ EVT VT = LoadNode->getValueType(0);
+ SDValue Chain = LoadNode->getChain();
+ SDValue Ptr = LoadNode->getBasePtr();
+ assert (isa<ConstantSDNode>(Ptr));
+
+ //TODO: Support smaller loads
+ if (LoadNode->getMemoryVT().getScalarType() != MVT::i32 || !ISD::isNON_EXTLoad(LoadNode))
+ return SDValue();
+
+ if (LoadNode->getAlignment() < 4)
+ return SDValue();
+
+ int ConstantBlock = ConstantAddressBlock(Block);
+
+ SDValue Slots[4];
+ for (unsigned i = 0; i < 4; i++) {
+ // We want Const position encoded with the following formula :
+ // (((512 + (kc_bank << 12) + const_index) << 2) + chan)
+ // const_index is Ptr computed by llvm using an alignment of 16.
+ // Thus we add (((512 + (kc_bank << 12)) + chan ) * 4 here and
+ // then div by 4 at the ISel step
+ SDValue NewPtr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
+ DAG.getConstant(4 * i + ConstantBlock * 16, DL, MVT::i32));
+ Slots[i] = DAG.getNode(AMDGPUISD::CONST_ADDRESS, DL, MVT::i32, NewPtr);
+ }
+ EVT NewVT = MVT::v4i32;
+ unsigned NumElements = 4;
+ if (VT.isVector()) {
+ NewVT = VT;
+ NumElements = VT.getVectorNumElements();
+ }
+ SDValue Result = DAG.getBuildVector(NewVT, DL, makeArrayRef(Slots, NumElements));
+ if (!VT.isVector()) {
+ Result = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, Result,
+ DAG.getConstant(0, DL, MVT::i32));
+ }
+ SDValue MergedValues[2] = {
+ Result,
+ Chain
+ };
+ return DAG.getMergeValues(MergedValues, DL);
+}
+
//===----------------------------------------------------------------------===//
// Custom DAG Optimizations
//===----------------------------------------------------------------------===//
@@ -2022,6 +2052,16 @@ SDValue R600TargetLowering::PerformDAGCombine(SDNode *N,
NewArgs[1] = OptimizeSwizzle(N->getOperand(1), &NewArgs[2], DAG, DL);
return DAG.getNode(AMDGPUISD::TEXTURE_FETCH, DL, N->getVTList(), NewArgs);
}
+
+ case ISD::LOAD: {
+ LoadSDNode *LoadNode = cast<LoadSDNode>(N);
+ SDValue Ptr = LoadNode->getBasePtr();
+ if (LoadNode->getAddressSpace() == AMDGPUAS::PARAM_I_ADDRESS &&
+ isa<ConstantSDNode>(Ptr))
+ return constBufferLoad(LoadNode, AMDGPUAS::CONSTANT_BUFFER_0, DAG);
+ break;
+ }
+
default: break;
}
diff --git a/lib/Target/AMDGPU/R600ISelLowering.h b/lib/Target/AMDGPU/R600ISelLowering.h
index 907d1f10e1519..767c3c7bd5bfe 100644
--- a/lib/Target/AMDGPU/R600ISelLowering.h
+++ b/lib/Target/AMDGPU/R600ISelLowering.h
@@ -98,9 +98,11 @@ private:
bool isHWTrueValue(SDValue Op) const;
bool isHWFalseValue(SDValue Op) const;
- bool FoldOperand(SDNode *ParentNode, unsigned SrcIdx, SDValue &Src,
- SDValue &Neg, SDValue &Abs, SDValue &Sel, SDValue &Imm,
- SelectionDAG &DAG) const;
+ bool FoldOperand(SDNode *ParentNode, unsigned SrcIdx, SDValue &Src,
+ SDValue &Neg, SDValue &Abs, SDValue &Sel, SDValue &Imm,
+ SelectionDAG &DAG) const;
+ SDValue constBufferLoad(LoadSDNode *LoadNode, int Block,
+ SelectionDAG &DAG) const;
SDNode *PostISelFolding(MachineSDNode *N, SelectionDAG &DAG) const override;
};
diff --git a/lib/Target/AMDGPU/VOP3Instructions.td b/lib/Target/AMDGPU/VOP3Instructions.td
index 17ae08dc62670..26bc5260e17f4 100644
--- a/lib/Target/AMDGPU/VOP3Instructions.td
+++ b/lib/Target/AMDGPU/VOP3Instructions.td
@@ -461,17 +461,6 @@ def : GCNPat <
(inst i16:$src0, i16:$src1, i16:$src2, (i1 0))
>;
-def : GCNPat<
- (i32 (op3 (op2 (op1 i16:$src0, i16:$src1), i16:$src2))),
- (inst i16:$src0, i16:$src1, i16:$src2, (i1 0))
->;
-
-def : GCNPat<
- (i64 (op3 (op2 (op1 i16:$src0, i16:$src1), i16:$src2))),
- (REG_SEQUENCE VReg_64,
- (inst i16:$src0, i16:$src1, i16:$src2, (i1 0)), sub0,
- (V_MOV_B32_e32 (i32 0)), sub1)
->;
}
defm: Ternary_i16_Pats<mul, add, V_MAD_U16, zext>;
diff --git a/lib/Target/NVPTX/NVPTXISelDAGToDAG.cpp b/lib/Target/NVPTX/NVPTXISelDAGToDAG.cpp
index 4dfa8477a362f..21939d836dc73 100644
--- a/lib/Target/NVPTX/NVPTXISelDAGToDAG.cpp
+++ b/lib/Target/NVPTX/NVPTXISelDAGToDAG.cpp
@@ -16,6 +16,7 @@
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/IR/GlobalValue.h"
#include "llvm/IR/Instructions.h"
+#include "llvm/Support/AtomicOrdering.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
@@ -81,10 +82,12 @@ void NVPTXDAGToDAGISel::Select(SDNode *N) {
switch (N->getOpcode()) {
case ISD::LOAD:
+ case ISD::ATOMIC_LOAD:
if (tryLoad(N))
return;
break;
case ISD::STORE:
+ case ISD::ATOMIC_STORE:
if (tryStore(N))
return;
break;
@@ -834,17 +837,27 @@ static Optional<unsigned> pickOpcodeForVT(
bool NVPTXDAGToDAGISel::tryLoad(SDNode *N) {
SDLoc dl(N);
- LoadSDNode *LD = cast<LoadSDNode>(N);
+ MemSDNode *LD = cast<MemSDNode>(N);
+ assert(LD->readMem() && "Expected load");
+ LoadSDNode *PlainLoad = dyn_cast<LoadSDNode>(N);
EVT LoadedVT = LD->getMemoryVT();
SDNode *NVPTXLD = nullptr;
// do not support pre/post inc/dec
- if (LD->isIndexed())
+ if (PlainLoad && PlainLoad->isIndexed())
return false;
if (!LoadedVT.isSimple())
return false;
+ AtomicOrdering Ordering = LD->getOrdering();
+ // In order to lower atomic loads with stronger guarantees we would need to
+ // use load.acquire or insert fences. However these features were only added
+ // with PTX ISA 6.0 / sm_70.
+ // TODO: Check if we can actually use the new instructions and implement them.
+ if (isStrongerThanMonotonic(Ordering))
+ return false;
+
// Address Space Setting
unsigned int CodeAddrSpace = getCodeAddrSpace(LD);
if (canLowerToLDG(LD, *Subtarget, CodeAddrSpace, MF)) {
@@ -855,8 +868,9 @@ bool NVPTXDAGToDAGISel::tryLoad(SDNode *N) {
CurDAG->getDataLayout().getPointerSizeInBits(LD->getAddressSpace());
// Volatile Setting
- // - .volatile is only availalble for .global and .shared
- bool isVolatile = LD->isVolatile();
+ // - .volatile is only available for .global and .shared
+ // - .volatile has the same memory synchronization semantics as .relaxed.sys
+ bool isVolatile = LD->isVolatile() || Ordering == AtomicOrdering::Monotonic;
if (CodeAddrSpace != NVPTX::PTXLdStInstCode::GLOBAL &&
CodeAddrSpace != NVPTX::PTXLdStInstCode::SHARED &&
CodeAddrSpace != NVPTX::PTXLdStInstCode::GENERIC)
@@ -882,7 +896,7 @@ bool NVPTXDAGToDAGISel::tryLoad(SDNode *N) {
fromTypeWidth = 32;
}
- if ((LD->getExtensionType() == ISD::SEXTLOAD))
+ if (PlainLoad && (PlainLoad->getExtensionType() == ISD::SEXTLOAD))
fromType = NVPTX::PTXLdStInstCode::Signed;
else if (ScalarVT.isFloatingPoint())
// f16 uses .b16 as its storage type.
@@ -1691,25 +1705,38 @@ bool NVPTXDAGToDAGISel::tryLDGLDU(SDNode *N) {
bool NVPTXDAGToDAGISel::tryStore(SDNode *N) {
SDLoc dl(N);
- StoreSDNode *ST = cast<StoreSDNode>(N);
+ MemSDNode *ST = cast<MemSDNode>(N);
+ assert(ST->writeMem() && "Expected store");
+ StoreSDNode *PlainStore = dyn_cast<StoreSDNode>(N);
+ AtomicSDNode *AtomicStore = dyn_cast<AtomicSDNode>(N);
+ assert((PlainStore || AtomicStore) && "Expected store");
EVT StoreVT = ST->getMemoryVT();
SDNode *NVPTXST = nullptr;
// do not support pre/post inc/dec
- if (ST->isIndexed())
+ if (PlainStore && PlainStore->isIndexed())
return false;
if (!StoreVT.isSimple())
return false;
+ AtomicOrdering Ordering = ST->getOrdering();
+ // In order to lower atomic loads with stronger guarantees we would need to
+ // use store.release or insert fences. However these features were only added
+ // with PTX ISA 6.0 / sm_70.
+ // TODO: Check if we can actually use the new instructions and implement them.
+ if (isStrongerThanMonotonic(Ordering))
+ return false;
+
// Address Space Setting
unsigned int CodeAddrSpace = getCodeAddrSpace(ST);
unsigned int PointerSize =
CurDAG->getDataLayout().getPointerSizeInBits(ST->getAddressSpace());
// Volatile Setting
- // - .volatile is only availalble for .global and .shared
- bool isVolatile = ST->isVolatile();
+ // - .volatile is only available for .global and .shared
+ // - .volatile has the same memory synchronization semantics as .relaxed.sys
+ bool isVolatile = ST->isVolatile() || Ordering == AtomicOrdering::Monotonic;
if (CodeAddrSpace != NVPTX::PTXLdStInstCode::GLOBAL &&
CodeAddrSpace != NVPTX::PTXLdStInstCode::SHARED &&
CodeAddrSpace != NVPTX::PTXLdStInstCode::GENERIC)
@@ -1739,41 +1766,53 @@ bool NVPTXDAGToDAGISel::tryStore(SDNode *N) {
toType = NVPTX::PTXLdStInstCode::Unsigned;
// Create the machine instruction DAG
- SDValue Chain = N->getOperand(0);
- SDValue N1 = N->getOperand(1);
- SDValue N2 = N->getOperand(2);
+ SDValue Chain = ST->getChain();
+ SDValue Value = PlainStore ? PlainStore->getValue() : AtomicStore->getVal();
+ SDValue BasePtr = ST->getBasePtr();
SDValue Addr;
SDValue Offset, Base;
Optional<unsigned> Opcode;
- MVT::SimpleValueType SourceVT = N1.getNode()->getSimpleValueType(0).SimpleTy;
+ MVT::SimpleValueType SourceVT =
+ Value.getNode()->getSimpleValueType(0).SimpleTy;
- if (SelectDirectAddr(N2, Addr)) {
+ if (SelectDirectAddr(BasePtr, Addr)) {
Opcode = pickOpcodeForVT(SourceVT, NVPTX::ST_i8_avar, NVPTX::ST_i16_avar,
NVPTX::ST_i32_avar, NVPTX::ST_i64_avar,
NVPTX::ST_f16_avar, NVPTX::ST_f16x2_avar,
NVPTX::ST_f32_avar, NVPTX::ST_f64_avar);
if (!Opcode)
return false;
- SDValue Ops[] = { N1, getI32Imm(isVolatile, dl),
- getI32Imm(CodeAddrSpace, dl), getI32Imm(vecType, dl),
- getI32Imm(toType, dl), getI32Imm(toTypeWidth, dl), Addr,
- Chain };
+ SDValue Ops[] = {Value,
+ getI32Imm(isVolatile, dl),
+ getI32Imm(CodeAddrSpace, dl),
+ getI32Imm(vecType, dl),
+ getI32Imm(toType, dl),
+ getI32Imm(toTypeWidth, dl),
+ Addr,
+ Chain};
NVPTXST = CurDAG->getMachineNode(Opcode.getValue(), dl, MVT::Other, Ops);
- } else if (PointerSize == 64 ? SelectADDRsi64(N2.getNode(), N2, Base, Offset)
- : SelectADDRsi(N2.getNode(), N2, Base, Offset)) {
+ } else if (PointerSize == 64
+ ? SelectADDRsi64(BasePtr.getNode(), BasePtr, Base, Offset)
+ : SelectADDRsi(BasePtr.getNode(), BasePtr, Base, Offset)) {
Opcode = pickOpcodeForVT(SourceVT, NVPTX::ST_i8_asi, NVPTX::ST_i16_asi,
NVPTX::ST_i32_asi, NVPTX::ST_i64_asi,
NVPTX::ST_f16_asi, NVPTX::ST_f16x2_asi,
NVPTX::ST_f32_asi, NVPTX::ST_f64_asi);
if (!Opcode)
return false;
- SDValue Ops[] = { N1, getI32Imm(isVolatile, dl),
- getI32Imm(CodeAddrSpace, dl), getI32Imm(vecType, dl),
- getI32Imm(toType, dl), getI32Imm(toTypeWidth, dl), Base,
- Offset, Chain };
+ SDValue Ops[] = {Value,
+ getI32Imm(isVolatile, dl),
+ getI32Imm(CodeAddrSpace, dl),
+ getI32Imm(vecType, dl),
+ getI32Imm(toType, dl),
+ getI32Imm(toTypeWidth, dl),
+ Base,
+ Offset,
+ Chain};
NVPTXST = CurDAG->getMachineNode(Opcode.getValue(), dl, MVT::Other, Ops);
- } else if (PointerSize == 64 ? SelectADDRri64(N2.getNode(), N2, Base, Offset)
- : SelectADDRri(N2.getNode(), N2, Base, Offset)) {
+ } else if (PointerSize == 64
+ ? SelectADDRri64(BasePtr.getNode(), BasePtr, Base, Offset)
+ : SelectADDRri(BasePtr.getNode(), BasePtr, Base, Offset)) {
if (PointerSize == 64)
Opcode = pickOpcodeForVT(
SourceVT, NVPTX::ST_i8_ari_64, NVPTX::ST_i16_ari_64,
@@ -1787,10 +1826,15 @@ bool NVPTXDAGToDAGISel::tryStore(SDNode *N) {
if (!Opcode)
return false;
- SDValue Ops[] = { N1, getI32Imm(isVolatile, dl),
- getI32Imm(CodeAddrSpace, dl), getI32Imm(vecType, dl),
- getI32Imm(toType, dl), getI32Imm(toTypeWidth, dl), Base,
- Offset, Chain };
+ SDValue Ops[] = {Value,
+ getI32Imm(isVolatile, dl),
+ getI32Imm(CodeAddrSpace, dl),
+ getI32Imm(vecType, dl),
+ getI32Imm(toType, dl),
+ getI32Imm(toTypeWidth, dl),
+ Base,
+ Offset,
+ Chain};
NVPTXST = CurDAG->getMachineNode(Opcode.getValue(), dl, MVT::Other, Ops);
} else {
if (PointerSize == 64)
@@ -1806,10 +1850,14 @@ bool NVPTXDAGToDAGISel::tryStore(SDNode *N) {
NVPTX::ST_f32_areg, NVPTX::ST_f64_areg);
if (!Opcode)
return false;
- SDValue Ops[] = { N1, getI32Imm(isVolatile, dl),
- getI32Imm(CodeAddrSpace, dl), getI32Imm(vecType, dl),
- getI32Imm(toType, dl), getI32Imm(toTypeWidth, dl), N2,
- Chain };
+ SDValue Ops[] = {Value,
+ getI32Imm(isVolatile, dl),
+ getI32Imm(CodeAddrSpace, dl),
+ getI32Imm(vecType, dl),
+ getI32Imm(toType, dl),
+ getI32Imm(toTypeWidth, dl),
+ BasePtr,
+ Chain};
NVPTXST = CurDAG->getMachineNode(Opcode.getValue(), dl, MVT::Other, Ops);
}
generated by cgit v1.2.3 (git 2.25.1) at 2025-08-12 15:09:01 +0000