aboutsummaryrefslogtreecommitdiff
path: root/llvm/lib/Target
diff options
context:
space:
mode:
Diffstat (limited to 'llvm/lib/Target')
-rw-r--r--llvm/lib/Target/AMDGPU/AMDGPUAttributor.cpp1
-rw-r--r--llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp93
-rw-r--r--llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp8
-rw-r--r--llvm/lib/Target/AMDGPU/FLATInstructions.td73
-rw-r--r--llvm/lib/Target/AMDGPU/GCNSubtarget.cpp57
-rw-r--r--llvm/lib/Target/AMDGPU/GCNSubtarget.h4
-rw-r--r--llvm/lib/Target/AMDGPU/SIISelLowering.cpp76
-rw-r--r--llvm/lib/Target/AMDGPU/SIISelLowering.h1
-rw-r--r--llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp29
-rw-r--r--llvm/lib/Target/AMDGPU/SIInstrInfo.cpp26
-rw-r--r--llvm/lib/Target/AMDGPU/SIInstrInfo.h6
-rw-r--r--llvm/lib/Target/AMDGPU/SILowerSGPRSpills.cpp5
-rw-r--r--llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp35
-rw-r--r--llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp61
-rw-r--r--llvm/lib/Target/AMDGPU/SIRegisterInfo.h5
-rw-r--r--llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp9
-rw-r--r--llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp3
-rw-r--r--llvm/lib/Target/ARM/ARMISelLowering.cpp11
-rw-r--r--llvm/lib/Target/AVR/MCTargetDesc/AVRMCAsmInfo.cpp2
-rw-r--r--llvm/lib/Target/AVR/MCTargetDesc/AVRMCAsmInfo.h4
-rw-r--r--llvm/lib/Target/BPF/BTFDebug.cpp74
-rw-r--r--llvm/lib/Target/BPF/MCTargetDesc/BPFMCAsmInfo.h5
-rw-r--r--llvm/lib/Target/Hexagon/HexagonCopyToCombine.cpp2
-rw-r--r--llvm/lib/Target/Hexagon/HexagonMask.cpp2
-rw-r--r--llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp37
-rw-r--r--llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchAsmBackend.cpp1
-rw-r--r--llvm/lib/Target/Mips/MipsAsmPrinter.cpp8
-rw-r--r--llvm/lib/Target/Mips/MipsISelLowering.cpp29
-rw-r--r--llvm/lib/Target/Mips/MipsISelLowering.h1
-rw-r--r--llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXMCAsmInfo.cpp2
-rw-r--r--llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXMCAsmInfo.h2
-rw-r--r--llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXTargetStreamer.cpp5
-rw-r--r--llvm/lib/Target/NVPTX/NVPTXISelDAGToDAG.cpp273
-rw-r--r--llvm/lib/Target/NVPTX/NVPTXISelDAGToDAG.h2
-rw-r--r--llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp629
-rw-r--r--llvm/lib/Target/NVPTX/NVPTXISelLowering.h10
-rw-r--r--llvm/lib/Target/NVPTX/NVPTXInstrInfo.td133
-rw-r--r--llvm/lib/Target/RISCV/Disassembler/RISCVDisassembler.cpp38
-rw-r--r--llvm/lib/Target/RISCV/MCTargetDesc/RISCVAsmBackend.cpp1
-rw-r--r--llvm/lib/Target/RISCV/RISCVFrameLowering.cpp3
-rw-r--r--llvm/lib/Target/RISCV/RISCVISelLowering.cpp23
-rw-r--r--llvm/lib/Target/RISCV/RISCVISelLowering.h1
-rw-r--r--llvm/lib/Target/RISCV/RISCVInterleavedAccess.cpp2
-rw-r--r--llvm/lib/Target/RISCV/RISCVTargetTransformInfo.h4
-rw-r--r--llvm/lib/Target/SystemZ/MCTargetDesc/SystemZHLASMAsmStreamer.cpp4
-rw-r--r--llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td2
46 files changed, 684 insertions, 1118 deletions
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUAttributor.cpp b/llvm/lib/Target/AMDGPU/AMDGPUAttributor.cpp
index 49d8b44..59cc1df 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUAttributor.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUAttributor.cpp
@@ -13,7 +13,6 @@
#include "AMDGPU.h"
#include "GCNSubtarget.h"
#include "Utils/AMDGPUBaseInfo.h"
-#include "llvm/CodeGen/TargetPassConfig.h"
#include "llvm/IR/IntrinsicsAMDGPU.h"
#include "llvm/IR/IntrinsicsR600.h"
#include "llvm/Target/TargetMachine.h"
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index e3ca09e..6118933 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -391,8 +391,9 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
// Library functions. These default to Expand, but we have instructions
// for them.
setOperationAction({ISD::FCEIL, ISD::FPOW, ISD::FABS, ISD::FFLOOR,
- ISD::FROUNDEVEN, ISD::FTRUNC, ISD::FMINNUM, ISD::FMAXNUM},
- MVT::f32, Legal);
+ ISD::FROUNDEVEN, ISD::FTRUNC},
+ {MVT::f16, MVT::f32}, Legal);
+ setOperationAction({ISD::FMINNUM, ISD::FMAXNUM}, MVT::f32, Legal);
setOperationAction(ISD::FLOG2, MVT::f32, Custom);
setOperationAction(ISD::FROUND, {MVT::f32, MVT::f64}, Custom);
@@ -412,9 +413,10 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
setOperationAction(ISD::FREM, {MVT::f16, MVT::f32, MVT::f64}, Custom);
- if (Subtarget->has16BitInsts())
+ if (Subtarget->has16BitInsts()) {
setOperationAction(ISD::IS_FPCLASS, {MVT::f16, MVT::f32, MVT::f64}, Legal);
- else {
+ setOperationAction({ISD::FLOG2, ISD::FEXP2}, MVT::f16, Legal);
+ } else {
setOperationAction(ISD::IS_FPCLASS, {MVT::f32, MVT::f64}, Legal);
setOperationAction({ISD::FLOG2, ISD::FEXP2}, MVT::f16, Custom);
}
@@ -4844,94 +4846,11 @@ AMDGPUTargetLowering::foldFreeOpFromSelect(TargetLowering::DAGCombinerInfo &DCI,
return SDValue();
}
-// Detect when CMP and SELECT use the same constant and fold them to avoid
-// loading the constant twice. Specifically handles patterns like:
-// %cmp = icmp eq i32 %val, 4242
-// %sel = select i1 %cmp, i32 4242, i32 %other
-// It can be optimized to reuse %val instead of 4242 in select.
-static SDValue
-foldCmpSelectWithSharedConstant(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
- const AMDGPUSubtarget *ST) {
- SDValue Cond = N->getOperand(0);
- SDValue TrueVal = N->getOperand(1);
- SDValue FalseVal = N->getOperand(2);
-
- // Check if condition is a comparison.
- if (Cond.getOpcode() != ISD::SETCC)
- return SDValue();
-
- SDValue LHS = Cond.getOperand(0);
- SDValue RHS = Cond.getOperand(1);
- ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
-
- bool isFloatingPoint = LHS.getValueType().isFloatingPoint();
- bool isInteger = LHS.getValueType().isInteger();
-
- // Handle simple floating-point and integer types only.
- if (!isFloatingPoint && !isInteger)
- return SDValue();
-
- bool isEquality = CC == (isFloatingPoint ? ISD::SETOEQ : ISD::SETEQ);
- bool isNonEquality = CC == (isFloatingPoint ? ISD::SETONE : ISD::SETNE);
- if (!isEquality && !isNonEquality)
- return SDValue();
-
- SDValue ArgVal, ConstVal;
- if ((isFloatingPoint && isa<ConstantFPSDNode>(RHS)) ||
- (isInteger && isa<ConstantSDNode>(RHS))) {
- ConstVal = RHS;
- ArgVal = LHS;
- } else if ((isFloatingPoint && isa<ConstantFPSDNode>(LHS)) ||
- (isInteger && isa<ConstantSDNode>(LHS))) {
- ConstVal = LHS;
- ArgVal = RHS;
- } else {
- return SDValue();
- }
-
- // Check if constant should not be optimized - early return if not.
- if (isFloatingPoint) {
- const APFloat &Val = cast<ConstantFPSDNode>(ConstVal)->getValueAPF();
- const GCNSubtarget *GCNST = static_cast<const GCNSubtarget *>(ST);
-
- // Only optimize normal floating-point values (finite, non-zero, and
- // non-subnormal as per IEEE 754), skip optimization for inlinable
- // floating-point constants.
- if (!Val.isNormal() || GCNST->getInstrInfo()->isInlineConstant(Val))
- return SDValue();
- } else {
- int64_t IntVal = cast<ConstantSDNode>(ConstVal)->getSExtValue();
-
- // Skip optimization for inlinable integer immediates.
- // Inlinable immediates include: -16 to 64 (inclusive).
- if (IntVal >= -16 && IntVal <= 64)
- return SDValue();
- }
-
- // For equality and non-equality comparisons, patterns:
- // select (setcc x, const), const, y -> select (setcc x, const), x, y
- // select (setccinv x, const), y, const -> select (setccinv x, const), y, x
- if (!(isEquality && TrueVal == ConstVal) &&
- !(isNonEquality && FalseVal == ConstVal))
- return SDValue();
-
- SDValue SelectLHS = (isEquality && TrueVal == ConstVal) ? ArgVal : TrueVal;
- SDValue SelectRHS =
- (isNonEquality && FalseVal == ConstVal) ? ArgVal : FalseVal;
- return DCI.DAG.getNode(ISD::SELECT, SDLoc(N), N->getValueType(0), Cond,
- SelectLHS, SelectRHS);
-}
-
SDValue AMDGPUTargetLowering::performSelectCombine(SDNode *N,
DAGCombinerInfo &DCI) const {
if (SDValue Folded = foldFreeOpFromSelect(DCI, SDValue(N, 0)))
return Folded;
- // Try to fold CMP + SELECT patterns with shared constants (both FP and
- // integer).
- if (SDValue Folded = foldCmpSelectWithSharedConstant(N, DCI, Subtarget))
- return Folded;
-
SDValue Cond = N->getOperand(0);
if (Cond.getOpcode() != ISD::SETCC)
return SDValue();
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
index c865082..38f9ee5 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
@@ -836,8 +836,10 @@ void AMDGPUTargetMachine::registerPassBuilderCallbacks(PassBuilder &PB) {
// When we are not using -fgpu-rdc, we can run accelerator code
// selection relatively early, but still after linking to prevent
// eager removal of potentially reachable symbols.
- if (EnableHipStdPar)
+ if (EnableHipStdPar) {
+ PM.addPass(HipStdParMathFixupPass());
PM.addPass(HipStdParAcceleratorCodeSelectionPass());
+ }
PM.addPass(AMDGPUPrintfRuntimeBindingPass());
}
@@ -916,8 +918,10 @@ void AMDGPUTargetMachine::registerPassBuilderCallbacks(PassBuilder &PB) {
// selection after linking to prevent, otherwise we end up removing
// potentially reachable symbols that were exported as external in other
// modules.
- if (EnableHipStdPar)
+ if (EnableHipStdPar) {
+ PM.addPass(HipStdParMathFixupPass());
PM.addPass(HipStdParAcceleratorCodeSelectionPass());
+ }
// We want to support the -lto-partitions=N option as "best effort".
// For that, we need to lower LDS earlier in the pipeline before the
// module is partitioned for codegen.
diff --git a/llvm/lib/Target/AMDGPU/FLATInstructions.td b/llvm/lib/Target/AMDGPU/FLATInstructions.td
index 7207c25..0f172e0d 100644
--- a/llvm/lib/Target/AMDGPU/FLATInstructions.td
+++ b/llvm/lib/Target/AMDGPU/FLATInstructions.td
@@ -369,31 +369,68 @@ multiclass FLAT_Global_Store_Pseudo_t16<string opName> {
}
}
-class FLAT_Global_Load_LDS_Pseudo <string opName, bit EnableSaddr = 0> : FLAT_Pseudo<
+// Async loads, introduced in gfx1250, will store directly
+// to a DS address in vdst (they will not use M0 for DS addess).
+class FLAT_Global_Load_LDS_Pseudo <string opName, bit EnableSaddr = 0, bit IsAsync = 0> : FLAT_Pseudo<
opName,
(outs ),
!con(
- !if(EnableSaddr, (ins SReg_64:$saddr, VGPR_32:$vaddr), (ins VReg_64:$vaddr)),
- (ins flat_offset:$offset, CPol_0:$cpol)),
- " $vaddr"#!if(EnableSaddr, ", $saddr", ", off")#"$offset$cpol"> {
- let LGKM_CNT = 1;
+ !if(IsAsync, (ins VGPR_32:$vdst), (ins)),
+ !if(EnableSaddr, (ins SReg_64:$saddr, VGPR_32:$vaddr), (ins VReg_64:$vaddr)),
+ (ins flat_offset:$offset, CPol_0:$cpol)),
+ !if(IsAsync, " $vdst,", "")#" $vaddr"#!if(EnableSaddr, ", $saddr", ", off")#"$offset$cpol"> {
+ let LGKM_CNT = !not(IsAsync);
+ let VM_CNT = !not(IsAsync);
+ let ASYNC_CNT = IsAsync;
let is_flat_global = 1;
let lds = 1;
let has_data = 0;
+ let has_vdst = IsAsync; // vdst for ds address with IsAsync
+ let mayLoad = 1;
+ let mayStore = 1;
+ let has_saddr = 1;
+ let enabled_saddr = EnableSaddr;
+ let VALU = 1;
+ let PseudoInstr = opName#!if(EnableSaddr, "_SADDR", "");
+ let Uses = !if(IsAsync, [EXEC, ASYNCcnt], [M0, EXEC]);
+ let Defs = !if(IsAsync, [ASYNCcnt], []);
+ let SchedRW = [WriteVMEM, WriteLDS];
+}
+
+multiclass FLAT_Global_Load_LDS_Pseudo<string opName, bit IsAsync = 0> {
+ def "" : FLAT_Global_Load_LDS_Pseudo<opName, 0, IsAsync>,
+ GlobalSaddrTable<0, opName>;
+ def _SADDR : FLAT_Global_Load_LDS_Pseudo<opName, 1, IsAsync>,
+ GlobalSaddrTable<1, opName>;
+}
+
+class FLAT_Global_STORE_LDS_Pseudo <string opName, bit EnableSaddr = 0> : FLAT_Pseudo<
+ opName,
+ (outs ),
+ !con(
+ !if(EnableSaddr, (ins SReg_64:$saddr, VGPR_32:$vaddr), (ins VReg_64:$vaddr)), (ins VGPR_32:$vdata),
+ (ins flat_offset:$offset, CPol_0:$cpol)),
+ " $vaddr, $vdata"#!if(EnableSaddr, ", $saddr", ", off")#"$offset$cpol"> {
+ let VM_CNT = 0;
+ let ASYNC_CNT = 1;
+ let is_flat_global = 1;
+ let lds = 1;
+ let has_data = 1; // vdata for ds address
let has_vdst = 0;
let mayLoad = 1;
let mayStore = 1;
let has_saddr = 1;
let enabled_saddr = EnableSaddr;
let VALU = 1;
- let Uses = [M0, EXEC];
+ let Uses = [EXEC, ASYNCcnt];
+ let Defs = [ASYNCcnt];
let SchedRW = [WriteVMEM, WriteLDS];
}
-multiclass FLAT_Global_Load_LDS_Pseudo<string opName> {
- def "" : FLAT_Global_Load_LDS_Pseudo<opName>,
+multiclass FLAT_Global_STORE_LDS_Pseudo<string opName> {
+ def "" : FLAT_Global_STORE_LDS_Pseudo<opName>,
GlobalSaddrTable<0, opName>;
- def _SADDR : FLAT_Global_Load_LDS_Pseudo<opName, 1>,
+ def _SADDR : FLAT_Global_STORE_LDS_Pseudo<opName, 1>,
GlobalSaddrTable<1, opName>;
}
@@ -1156,6 +1193,15 @@ let SubtargetPredicate = isGFX12Plus in {
let SubtargetPredicate = isGFX1250Plus in {
+defm GLOBAL_LOAD_ASYNC_TO_LDS_B8 : FLAT_Global_Load_LDS_Pseudo<"global_load_async_to_lds_b8", 1>;
+defm GLOBAL_LOAD_ASYNC_TO_LDS_B32 : FLAT_Global_Load_LDS_Pseudo<"global_load_async_to_lds_b32", 1>;
+defm GLOBAL_LOAD_ASYNC_TO_LDS_B64 : FLAT_Global_Load_LDS_Pseudo<"global_load_async_to_lds_b64", 1>;
+defm GLOBAL_LOAD_ASYNC_TO_LDS_B128 : FLAT_Global_Load_LDS_Pseudo<"global_load_async_to_lds_b128", 1>;
+defm GLOBAL_STORE_ASYNC_FROM_LDS_B8 : FLAT_Global_STORE_LDS_Pseudo<"global_store_async_from_lds_b8">;
+defm GLOBAL_STORE_ASYNC_FROM_LDS_B32 : FLAT_Global_STORE_LDS_Pseudo<"global_store_async_from_lds_b32">;
+defm GLOBAL_STORE_ASYNC_FROM_LDS_B64 : FLAT_Global_STORE_LDS_Pseudo<"global_store_async_from_lds_b64">;
+defm GLOBAL_STORE_ASYNC_FROM_LDS_B128 : FLAT_Global_STORE_LDS_Pseudo<"global_store_async_from_lds_b128">;
+
def TENSOR_SAVE : FLAT_Global_Tensor_Pseudo<"tensor_save", 1>;
def TENSOR_STOP : FLAT_Global_Tensor_Pseudo<"tensor_stop">;
} // End SubtargetPredicate = isGFX1250Plus
@@ -3374,6 +3420,15 @@ defm GLOBAL_LOAD_MONITOR_B32 : VFLAT_Real_AllAddr_gfx1250<0x070>;
defm GLOBAL_LOAD_MONITOR_B64 : VFLAT_Real_AllAddr_gfx1250<0x071>;
defm GLOBAL_LOAD_MONITOR_B128 : VFLAT_Real_AllAddr_gfx1250<0x072>;
+defm GLOBAL_LOAD_ASYNC_TO_LDS_B8 : VFLAT_Real_AllAddr_gfx1250<0x5f>;
+defm GLOBAL_LOAD_ASYNC_TO_LDS_B32 : VFLAT_Real_AllAddr_gfx1250<0x60>;
+defm GLOBAL_LOAD_ASYNC_TO_LDS_B64 : VFLAT_Real_AllAddr_gfx1250<0x61>;
+defm GLOBAL_LOAD_ASYNC_TO_LDS_B128 : VFLAT_Real_AllAddr_gfx1250<0x62>;
+defm GLOBAL_STORE_ASYNC_FROM_LDS_B8 : VFLAT_Real_AllAddr_gfx1250<0x63>;
+defm GLOBAL_STORE_ASYNC_FROM_LDS_B32 : VFLAT_Real_AllAddr_gfx1250<0x64>;
+defm GLOBAL_STORE_ASYNC_FROM_LDS_B64 : VFLAT_Real_AllAddr_gfx1250<0x65>;
+defm GLOBAL_STORE_ASYNC_FROM_LDS_B128 : VFLAT_Real_AllAddr_gfx1250<0x66>;
+
defm GLOBAL_LOAD_TR_B128_w32 : VFLAT_Real_AllAddr_gfx1250<0x057, "global_load_tr16_b128">;
defm GLOBAL_LOAD_TR_B64_w32 : VFLAT_Real_AllAddr_gfx1250<0x058, "global_load_tr8_b64">;
diff --git a/llvm/lib/Target/AMDGPU/GCNSubtarget.cpp b/llvm/lib/Target/AMDGPU/GCNSubtarget.cpp
index 9a2bab1..0a0a107 100644
--- a/llvm/lib/Target/AMDGPU/GCNSubtarget.cpp
+++ b/llvm/lib/Target/AMDGPU/GCNSubtarget.cpp
@@ -537,6 +537,63 @@ unsigned GCNSubtarget::getMaxNumVGPRs(const MachineFunction &MF) const {
return getMaxNumVGPRs(MF.getFunction());
}
+std::pair<unsigned, unsigned>
+GCNSubtarget::getMaxNumVectorRegs(const Function &F) const {
+ const unsigned MaxVectorRegs = getMaxNumVGPRs(F);
+
+ unsigned MaxNumVGPRs = MaxVectorRegs;
+ unsigned MaxNumAGPRs = 0;
+
+ // On GFX90A, the number of VGPRs and AGPRs need not be equal. Theoretically,
+ // a wave may have up to 512 total vector registers combining together both
+ // VGPRs and AGPRs. Hence, in an entry function without calls and without
+ // AGPRs used within it, it is possible to use the whole vector register
+ // budget for VGPRs.
+ //
+ // TODO: it shall be possible to estimate maximum AGPR/VGPR pressure and split
+ // register file accordingly.
+ if (hasGFX90AInsts()) {
+ unsigned MinNumAGPRs = 0;
+ const unsigned TotalNumAGPRs = AMDGPU::AGPR_32RegClass.getNumRegs();
+ const unsigned TotalNumVGPRs = AMDGPU::VGPR_32RegClass.getNumRegs();
+
+ const std::pair<unsigned, unsigned> DefaultNumAGPR = {~0u, ~0u};
+
+ // TODO: The lower bound should probably force the number of required
+ // registers up, overriding amdgpu-waves-per-eu.
+ std::tie(MinNumAGPRs, MaxNumAGPRs) =
+ AMDGPU::getIntegerPairAttribute(F, "amdgpu-agpr-alloc", DefaultNumAGPR,
+ /*OnlyFirstRequired=*/true);
+
+ if (MinNumAGPRs == DefaultNumAGPR.first) {
+ // Default to splitting half the registers if AGPRs are required.
+ MinNumAGPRs = MaxNumAGPRs = MaxVectorRegs / 2;
+ } else {
+ // Align to accum_offset's allocation granularity.
+ MinNumAGPRs = alignTo(MinNumAGPRs, 4);
+
+ MinNumAGPRs = std::min(MinNumAGPRs, TotalNumAGPRs);
+ }
+
+ // Clamp values to be inbounds of our limits, and ensure min <= max.
+
+ MaxNumAGPRs = std::min(std::max(MinNumAGPRs, MaxNumAGPRs), MaxVectorRegs);
+ MinNumAGPRs = std::min(std::min(MinNumAGPRs, TotalNumAGPRs), MaxNumAGPRs);
+
+ MaxNumVGPRs = std::min(MaxVectorRegs - MinNumAGPRs, TotalNumVGPRs);
+ MaxNumAGPRs = std::min(MaxVectorRegs - MaxNumVGPRs, MaxNumAGPRs);
+
+ assert(MaxNumVGPRs + MaxNumAGPRs <= MaxVectorRegs &&
+ MaxNumAGPRs <= TotalNumAGPRs && MaxNumVGPRs <= TotalNumVGPRs &&
+ "invalid register counts");
+ } else if (hasMAIInsts()) {
+ // On gfx908 the number of AGPRs always equals the number of VGPRs.
+ MaxNumAGPRs = MaxNumVGPRs = MaxVectorRegs;
+ }
+
+ return std::pair(MaxNumVGPRs, MaxNumAGPRs);
+}
+
void GCNSubtarget::adjustSchedDependency(
SUnit *Def, int DefOpIdx, SUnit *Use, int UseOpIdx, SDep &Dep,
const TargetSchedModel *SchedModel) const {
diff --git a/llvm/lib/Target/AMDGPU/GCNSubtarget.h b/llvm/lib/Target/AMDGPU/GCNSubtarget.h
index 88a269f..785ede3 100644
--- a/llvm/lib/Target/AMDGPU/GCNSubtarget.h
+++ b/llvm/lib/Target/AMDGPU/GCNSubtarget.h
@@ -1667,6 +1667,10 @@ public:
return getMaxNumVGPRs(F);
}
+ /// Return a pair of maximum numbers of VGPRs and AGPRs that meet the number
+ /// of waves per execution unit required for the function \p MF.
+ std::pair<unsigned, unsigned> getMaxNumVectorRegs(const Function &F) const;
+
/// \returns Maximum number of VGPRs that meets number of waves per execution
/// unit requirement for function \p MF, or number of VGPRs explicitly
/// requested using "amdgpu-num-vgpr" attribute attached to function \p MF.
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 8d51ec6..9017f4f 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -15896,6 +15896,78 @@ SDValue SITargetLowering::performClampCombine(SDNode *N,
return SDValue(CSrc, 0);
}
+SDValue SITargetLowering::performSelectCombine(SDNode *N,
+ DAGCombinerInfo &DCI) const {
+
+ // Try to fold CMP + SELECT patterns with shared constants (both FP and
+ // integer).
+ // Detect when CMP and SELECT use the same constant and fold them to avoid
+ // loading the constant twice. Specifically handles patterns like:
+ // %cmp = icmp eq i32 %val, 4242
+ // %sel = select i1 %cmp, i32 4242, i32 %other
+ // It can be optimized to reuse %val instead of 4242 in select.
+ SDValue Cond = N->getOperand(0);
+ SDValue TrueVal = N->getOperand(1);
+ SDValue FalseVal = N->getOperand(2);
+
+ // Check if condition is a comparison.
+ if (Cond.getOpcode() != ISD::SETCC)
+ return SDValue();
+
+ SDValue LHS = Cond.getOperand(0);
+ SDValue RHS = Cond.getOperand(1);
+ ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
+
+ bool isFloatingPoint = LHS.getValueType().isFloatingPoint();
+ bool isInteger = LHS.getValueType().isInteger();
+
+ // Handle simple floating-point and integer types only.
+ if (!isFloatingPoint && !isInteger)
+ return SDValue();
+
+ bool isEquality = CC == (isFloatingPoint ? ISD::SETOEQ : ISD::SETEQ);
+ bool isNonEquality = CC == (isFloatingPoint ? ISD::SETONE : ISD::SETNE);
+ if (!isEquality && !isNonEquality)
+ return SDValue();
+
+ SDValue ArgVal, ConstVal;
+ if ((isFloatingPoint && isa<ConstantFPSDNode>(RHS)) ||
+ (isInteger && isa<ConstantSDNode>(RHS))) {
+ ConstVal = RHS;
+ ArgVal = LHS;
+ } else if ((isFloatingPoint && isa<ConstantFPSDNode>(LHS)) ||
+ (isInteger && isa<ConstantSDNode>(LHS))) {
+ ConstVal = LHS;
+ ArgVal = RHS;
+ } else {
+ return SDValue();
+ }
+
+ // Skip optimization for inlinable immediates.
+ if (isFloatingPoint) {
+ const APFloat &Val = cast<ConstantFPSDNode>(ConstVal)->getValueAPF();
+ if (!Val.isNormal() || Subtarget->getInstrInfo()->isInlineConstant(Val))
+ return SDValue();
+ } else {
+ if (AMDGPU::isInlinableIntLiteral(
+ cast<ConstantSDNode>(ConstVal)->getSExtValue()))
+ return SDValue();
+ }
+
+ // For equality and non-equality comparisons, patterns:
+ // select (setcc x, const), const, y -> select (setcc x, const), x, y
+ // select (setccinv x, const), y, const -> select (setccinv x, const), y, x
+ if (!(isEquality && TrueVal == ConstVal) &&
+ !(isNonEquality && FalseVal == ConstVal))
+ return SDValue();
+
+ SDValue SelectLHS = (isEquality && TrueVal == ConstVal) ? ArgVal : TrueVal;
+ SDValue SelectRHS =
+ (isNonEquality && FalseVal == ConstVal) ? ArgVal : FalseVal;
+ return DCI.DAG.getNode(ISD::SELECT, SDLoc(N), N->getValueType(0), Cond,
+ SelectLHS, SelectRHS);
+}
+
SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
DAGCombinerInfo &DCI) const {
switch (N->getOpcode()) {
@@ -15944,6 +16016,10 @@ SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
return performFMulCombine(N, DCI);
case ISD::SETCC:
return performSetCCCombine(N, DCI);
+ case ISD::SELECT:
+ if (auto Res = performSelectCombine(N, DCI))
+ return Res;
+ break;
case ISD::FMAXNUM:
case ISD::FMINNUM:
case ISD::FMAXNUM_IEEE:
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.h b/llvm/lib/Target/AMDGPU/SIISelLowering.h
index acf6158..dedd9ae 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.h
@@ -211,6 +211,7 @@ private:
SDValue performExtractVectorEltCombine(SDNode *N, DAGCombinerInfo &DCI) const;
SDValue performInsertVectorEltCombine(SDNode *N, DAGCombinerInfo &DCI) const;
SDValue performFPRoundCombine(SDNode *N, DAGCombinerInfo &DCI) const;
+ SDValue performSelectCombine(SDNode *N, DAGCombinerInfo &DCI) const;
SDValue reassociateScalarOps(SDNode *N, SelectionDAG &DAG) const;
unsigned getFusedOpcode(const SelectionDAG &DAG,
diff --git a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
index dd3f2fe..520c321 100644
--- a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
@@ -552,7 +552,7 @@ public:
(!Inst.mayLoad() || SIInstrInfo::isAtomicNoRet(Inst))) {
// FLAT and SCRATCH instructions may access scratch. Other VMEM
// instructions do not.
- if (SIInstrInfo::isFLAT(Inst) && mayAccessScratchThroughFlat(Inst))
+ if (TII->mayAccessScratchThroughFlat(Inst))
return SCRATCH_WRITE_ACCESS;
return VMEM_WRITE_ACCESS;
}
@@ -565,7 +565,6 @@ public:
bool mayAccessVMEMThroughFlat(const MachineInstr &MI) const;
bool mayAccessLDSThroughFlat(const MachineInstr &MI) const;
- bool mayAccessScratchThroughFlat(const MachineInstr &MI) const;
bool isVmemAccess(const MachineInstr &MI) const;
bool generateWaitcntInstBefore(MachineInstr &MI,
WaitcntBrackets &ScoreBrackets,
@@ -2160,32 +2159,6 @@ bool SIInsertWaitcnts::mayAccessLDSThroughFlat(const MachineInstr &MI) const {
return false;
}
-// This is a flat memory operation. Check to see if it has memory tokens for
-// either scratch or FLAT.
-bool SIInsertWaitcnts::mayAccessScratchThroughFlat(
- const MachineInstr &MI) const {
- assert(TII->isFLAT(MI));
-
- // SCRATCH instructions always access scratch.
- if (TII->isFLATScratch(MI))
- return true;
-
- // GLOBAL instructions never access scratch.
- if (TII->isFLATGlobal(MI))
- return false;
-
- // If there are no memory operands then conservatively assume the flat
- // operation may access scratch.
- if (MI.memoperands_empty())
- return true;
-
- // See if any memory operand specifies an address space that involves scratch.
- return any_of(MI.memoperands(), [](const MachineMemOperand *Memop) {
- unsigned AS = Memop->getAddrSpace();
- return AS == AMDGPUAS::PRIVATE_ADDRESS || AS == AMDGPUAS::FLAT_ADDRESS;
- });
-}
-
bool SIInsertWaitcnts::isVmemAccess(const MachineInstr &MI) const {
return (TII->isFLAT(MI) && mayAccessVMEMThroughFlat(MI)) ||
(TII->isVMEM(MI) && !AMDGPU::getMUBUFIsBufferInv(MI.getOpcode()));
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 8d6c1d0..2aa6b4e 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -4249,6 +4249,32 @@ bool SIInstrInfo::isAlwaysGDS(uint16_t Opcode) const {
Opcode == AMDGPU::DS_SUB_GS_REG_RTN || isGWS(Opcode);
}
+bool SIInstrInfo::mayAccessScratchThroughFlat(const MachineInstr &MI) const {
+ if (!isFLAT(MI) || isFLATGlobal(MI))
+ return false;
+
+ // If scratch is not initialized, we can never access it.
+ if (MI.getMF()->getFunction().hasFnAttribute("amdgpu-no-flat-scratch-init"))
+ return false;
+
+ // SCRATCH instructions always access scratch.
+ if (isFLATScratch(MI))
+ return true;
+
+ // If there are no memory operands then conservatively assume the flat
+ // operation may access scratch.
+ if (MI.memoperands_empty())
+ return true;
+
+ // TODO (?): Does this need to be taught how to read noalias.addrspace ?
+
+ // See if any memory operand specifies an address space that involves scratch.
+ return any_of(MI.memoperands(), [](const MachineMemOperand *Memop) {
+ unsigned AS = Memop->getAddrSpace();
+ return AS == AMDGPUAS::PRIVATE_ADDRESS || AS == AMDGPUAS::FLAT_ADDRESS;
+ });
+}
+
bool SIInstrInfo::modifiesModeRegister(const MachineInstr &MI) {
// Skip the full operand and register alias search modifiesRegister
// does. There's only a handful of instructions that touch this, it's only an
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.h b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
index 2ffb783..e042b59 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.h
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
@@ -678,6 +678,12 @@ public:
return get(Opcode).TSFlags & SIInstrFlags::FLAT;
}
+ /// \returns true for SCRATCH_ instructions, or FLAT_ instructions with
+ /// SCRATCH_ memory operands.
+ /// Conservatively correct; will return true if \p MI cannot be proven
+ /// to not hit scratch.
+ bool mayAccessScratchThroughFlat(const MachineInstr &MI) const;
+
static bool isBlockLoadStore(uint16_t Opcode) {
switch (Opcode) {
case AMDGPU::SI_BLOCK_SPILL_V1024_SAVE:
diff --git a/llvm/lib/Target/AMDGPU/SILowerSGPRSpills.cpp b/llvm/lib/Target/AMDGPU/SILowerSGPRSpills.cpp
index 9f61bf8..9509199 100644
--- a/llvm/lib/Target/AMDGPU/SILowerSGPRSpills.cpp
+++ b/llvm/lib/Target/AMDGPU/SILowerSGPRSpills.cpp
@@ -351,6 +351,7 @@ void SILowerSGPRSpills::determineRegsForWWMAllocation(MachineFunction &MF,
MachineRegisterInfo &MRI = MF.getRegInfo();
BitVector ReservedRegs = TRI->getReservedRegs(MF);
BitVector NonWwmAllocMask(TRI->getNumRegs());
+ const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
// FIXME: MaxNumVGPRsForWwmAllocation might need to be adjusted in the future
// to have a balanced allocation between WWM values and per-thread vector
@@ -359,7 +360,7 @@ void SILowerSGPRSpills::determineRegsForWWMAllocation(MachineFunction &MF,
NumRegs =
std::min(static_cast<unsigned>(MFI->getSGPRSpillVGPRs().size()), NumRegs);
- auto [MaxNumVGPRs, MaxNumAGPRs] = TRI->getMaxNumVectorRegs(MF);
+ auto [MaxNumVGPRs, MaxNumAGPRs] = ST.getMaxNumVectorRegs(MF.getFunction());
// Try to use the highest available registers for now. Later after
// vgpr-regalloc, they can be shifted to the lowest range.
unsigned I = 0;
@@ -376,7 +377,7 @@ void SILowerSGPRSpills::determineRegsForWWMAllocation(MachineFunction &MF,
// Reserve an arbitrary register and report the error.
TRI->markSuperRegs(RegMask, AMDGPU::VGPR0);
MF.getFunction().getContext().emitError(
- "can't find enough VGPRs for wwm-regalloc");
+ "cannot find enough VGPRs for wwm-regalloc");
}
}
diff --git a/llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp b/llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp
index 0e8a420..f1262e11 100644
--- a/llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp
+++ b/llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp
@@ -321,7 +321,7 @@ public:
bool IsNonTemporal,
bool IsLastUse = false) const = 0;
- virtual bool expandSystemScopeStore(MachineBasicBlock::iterator &MI) const {
+ virtual bool finalizeStore(MachineInstr &MI, bool Atomic) const {
return false;
};
@@ -602,7 +602,7 @@ public:
bool IsVolatile, bool IsNonTemporal,
bool IsLastUse) const override;
- bool expandSystemScopeStore(MachineBasicBlock::iterator &MI) const override;
+ bool finalizeStore(MachineInstr &MI, bool Atomic) const override;
bool insertRelease(MachineBasicBlock::iterator &MI, SIAtomicScope Scope,
SIAtomicAddrSpace AddrSpace, bool IsCrossAddrSpaceOrdering,
@@ -2536,9 +2536,6 @@ bool SIGfx12CacheControl::enableVolatileAndOrNonTemporal(
if (IsVolatile) {
Changed |= setScope(MI, AMDGPU::CPol::SCOPE_SYS);
- if (Op == SIMemOp::STORE)
- Changed |= insertWaitsBeforeSystemScopeStore(MI);
-
// Ensure operation has completed at system scope to cause all volatile
// operations to be visible outside the program in a global order. Do not
// request cross address space as only the global address space can be
@@ -2551,11 +2548,24 @@ bool SIGfx12CacheControl::enableVolatileAndOrNonTemporal(
return Changed;
}
-bool SIGfx12CacheControl::expandSystemScopeStore(
- MachineBasicBlock::iterator &MI) const {
- MachineOperand *CPol = TII->getNamedOperand(*MI, OpName::cpol);
- if (CPol && ((CPol->getImm() & CPol::SCOPE) == CPol::SCOPE_SYS))
- return insertWaitsBeforeSystemScopeStore(MI);
+bool SIGfx12CacheControl::finalizeStore(MachineInstr &MI, bool Atomic) const {
+ MachineOperand *CPol = TII->getNamedOperand(MI, OpName::cpol);
+ if (!CPol)
+ return false;
+
+ const unsigned Scope = CPol->getImm() & CPol::SCOPE;
+
+ // GFX12.0 only: Extra waits needed before system scope stores.
+ if (!ST.hasGFX1250Insts()) {
+ if (!Atomic && Scope == CPol::SCOPE_SYS)
+ return insertWaitsBeforeSystemScopeStore(MI);
+ return false;
+ }
+
+ // GFX12.5 only: Require SCOPE_SE on stores that may hit the scratch address
+ // space.
+ if (TII->mayAccessScratchThroughFlat(MI) && Scope == CPol::SCOPE_CU)
+ return setScope(MI, CPol::SCOPE_SE);
return false;
}
@@ -2658,6 +2668,8 @@ bool SIMemoryLegalizer::expandStore(const SIMemOpInfo &MOI,
assert(!MI->mayLoad() && MI->mayStore());
bool Changed = false;
+ // FIXME: Necessary hack because iterator can lose track of the store.
+ MachineInstr &StoreMI = *MI;
if (MOI.isAtomic()) {
if (MOI.getOrdering() == AtomicOrdering::Monotonic ||
@@ -2674,6 +2686,7 @@ bool SIMemoryLegalizer::expandStore(const SIMemOpInfo &MOI,
MOI.getIsCrossAddressSpaceOrdering(),
Position::BEFORE);
+ Changed |= CC->finalizeStore(StoreMI, /*Atomic=*/true);
return Changed;
}
@@ -2686,7 +2699,7 @@ bool SIMemoryLegalizer::expandStore(const SIMemOpInfo &MOI,
// GFX12 specific, scope(desired coherence domain in cache hierarchy) is
// instruction field, do not confuse it with atomic scope.
- Changed |= CC->expandSystemScopeStore(MI);
+ Changed |= CC->finalizeStore(StoreMI, /*Atomic=*/false);
return Changed;
}
diff --git a/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp b/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
index 84cfa87..f3acc5c 100644
--- a/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
@@ -572,65 +572,6 @@ MCRegister SIRegisterInfo::reservedPrivateSegmentBufferReg(
return getAlignedHighSGPRForRC(MF, /*Align=*/4, &AMDGPU::SGPR_128RegClass);
}
-std::pair<unsigned, unsigned>
-SIRegisterInfo::getMaxNumVectorRegs(const MachineFunction &MF) const {
- const unsigned MaxVectorRegs = ST.getMaxNumVGPRs(MF);
-
- unsigned MaxNumVGPRs = MaxVectorRegs;
- unsigned MaxNumAGPRs = 0;
-
- // On GFX90A, the number of VGPRs and AGPRs need not be equal. Theoretically,
- // a wave may have up to 512 total vector registers combining together both
- // VGPRs and AGPRs. Hence, in an entry function without calls and without
- // AGPRs used within it, it is possible to use the whole vector register
- // budget for VGPRs.
- //
- // TODO: it shall be possible to estimate maximum AGPR/VGPR pressure and split
- // register file accordingly.
- if (ST.hasGFX90AInsts()) {
- unsigned MinNumAGPRs = 0;
- const unsigned TotalNumAGPRs = AMDGPU::AGPR_32RegClass.getNumRegs();
- const unsigned TotalNumVGPRs = AMDGPU::VGPR_32RegClass.getNumRegs();
-
- const std::pair<unsigned, unsigned> DefaultNumAGPR = {~0u, ~0u};
-
- // TODO: Move this logic into subtarget on IR function
- //
- // TODO: The lower bound should probably force the number of required
- // registers up, overriding amdgpu-waves-per-eu.
- std::tie(MinNumAGPRs, MaxNumAGPRs) = AMDGPU::getIntegerPairAttribute(
- MF.getFunction(), "amdgpu-agpr-alloc", DefaultNumAGPR,
- /*OnlyFirstRequired=*/true);
-
- if (MinNumAGPRs == DefaultNumAGPR.first) {
- // Default to splitting half the registers if AGPRs are required.
- MinNumAGPRs = MaxNumAGPRs = MaxVectorRegs / 2;
- } else {
- // Align to accum_offset's allocation granularity.
- MinNumAGPRs = alignTo(MinNumAGPRs, 4);
-
- MinNumAGPRs = std::min(MinNumAGPRs, TotalNumAGPRs);
- }
-
- // Clamp values to be inbounds of our limits, and ensure min <= max.
-
- MaxNumAGPRs = std::min(std::max(MinNumAGPRs, MaxNumAGPRs), MaxVectorRegs);
- MinNumAGPRs = std::min(std::min(MinNumAGPRs, TotalNumAGPRs), MaxNumAGPRs);
-
- MaxNumVGPRs = std::min(MaxVectorRegs - MinNumAGPRs, TotalNumVGPRs);
- MaxNumAGPRs = std::min(MaxVectorRegs - MaxNumVGPRs, MaxNumAGPRs);
-
- assert(MaxNumVGPRs + MaxNumAGPRs <= MaxVectorRegs &&
- MaxNumAGPRs <= TotalNumAGPRs && MaxNumVGPRs <= TotalNumVGPRs &&
- "invalid register counts");
- } else if (ST.hasMAIInsts()) {
- // On gfx908 the number of AGPRs always equals the number of VGPRs.
- MaxNumAGPRs = MaxNumVGPRs = MaxVectorRegs;
- }
-
- return std::pair(MaxNumVGPRs, MaxNumAGPRs);
-}
-
BitVector SIRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
BitVector Reserved(getNumRegs());
Reserved.set(AMDGPU::MODE);
@@ -742,7 +683,7 @@ BitVector SIRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
// Reserve VGPRs/AGPRs.
//
- auto [MaxNumVGPRs, MaxNumAGPRs] = getMaxNumVectorRegs(MF);
+ auto [MaxNumVGPRs, MaxNumAGPRs] = ST.getMaxNumVectorRegs(MF.getFunction());
for (const TargetRegisterClass *RC : regclasses()) {
if (RC->isBaseClass() && isVGPRClass(RC)) {
diff --git a/llvm/lib/Target/AMDGPU/SIRegisterInfo.h b/llvm/lib/Target/AMDGPU/SIRegisterInfo.h
index 0008e5f..5508f07 100644
--- a/llvm/lib/Target/AMDGPU/SIRegisterInfo.h
+++ b/llvm/lib/Target/AMDGPU/SIRegisterInfo.h
@@ -90,11 +90,6 @@ public:
/// spilling is needed.
MCRegister reservedPrivateSegmentBufferReg(const MachineFunction &MF) const;
- /// Return a pair of maximum numbers of VGPRs and AGPRs that meet the number
- /// of waves per execution unit required for the function \p MF.
- std::pair<unsigned, unsigned>
- getMaxNumVectorRegs(const MachineFunction &MF) const;
-
BitVector getReservedRegs(const MachineFunction &MF) const override;
bool isAsmClobberable(const MachineFunction &MF,
MCRegister PhysReg) const override;
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
index b5b3cc9..83e63ac 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
@@ -732,7 +732,14 @@ bool isGenericAtomic(unsigned Opc) {
}
bool isAsyncStore(unsigned Opc) {
- return false; // placeholder before async store implementation.
+ return Opc == GLOBAL_STORE_ASYNC_FROM_LDS_B8_gfx1250 ||
+ Opc == GLOBAL_STORE_ASYNC_FROM_LDS_B32_gfx1250 ||
+ Opc == GLOBAL_STORE_ASYNC_FROM_LDS_B64_gfx1250 ||
+ Opc == GLOBAL_STORE_ASYNC_FROM_LDS_B128_gfx1250 ||
+ Opc == GLOBAL_STORE_ASYNC_FROM_LDS_B8_SADDR_gfx1250 ||
+ Opc == GLOBAL_STORE_ASYNC_FROM_LDS_B32_SADDR_gfx1250 ||
+ Opc == GLOBAL_STORE_ASYNC_FROM_LDS_B64_SADDR_gfx1250 ||
+ Opc == GLOBAL_STORE_ASYNC_FROM_LDS_B128_SADDR_gfx1250;
}
bool isTensorStore(unsigned Opc) {
diff --git a/llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp b/llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp
index 50217c3..9e4dbec 100644
--- a/llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp
+++ b/llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp
@@ -4261,8 +4261,7 @@ std::optional<unsigned> ARMBaseInstrInfo::getOperandLatencyImpl(
// instructions).
if (Latency > 0 && Subtarget.isThumb2()) {
const MachineFunction *MF = DefMI.getParent()->getParent();
- // FIXME: Use Function::hasOptSize().
- if (MF->getFunction().hasFnAttribute(Attribute::OptimizeForSize))
+ if (MF->getFunction().hasOptSize())
--Latency;
}
return Latency;
diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index fca5dff..066b392 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -370,6 +370,11 @@ void ARMTargetLowering::addMVEVectorTypes(bool HasMVEFP) {
setOperationAction(ISD::FMINNUM, VT, Legal);
setOperationAction(ISD::FMAXNUM, VT, Legal);
setOperationAction(ISD::FROUND, VT, Legal);
+ setOperationAction(ISD::FROUNDEVEN, VT, Legal);
+ setOperationAction(ISD::FRINT, VT, Legal);
+ setOperationAction(ISD::FTRUNC, VT, Legal);
+ setOperationAction(ISD::FFLOOR, VT, Legal);
+ setOperationAction(ISD::FCEIL, VT, Legal);
setOperationAction(ISD::VECREDUCE_FADD, VT, Custom);
setOperationAction(ISD::VECREDUCE_FMUL, VT, Custom);
setOperationAction(ISD::VECREDUCE_FMIN, VT, Custom);
@@ -1507,6 +1512,12 @@ ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM_,
setOperationAction(ISD::FLOG2, MVT::f16, Promote);
setOperationAction(ISD::FROUND, MVT::f16, Legal);
+ setOperationAction(ISD::FROUNDEVEN, MVT::f16, Legal);
+ setOperationAction(ISD::FTRUNC, MVT::f16, Legal);
+ setOperationAction(ISD::FNEARBYINT, MVT::f16, Legal);
+ setOperationAction(ISD::FRINT, MVT::f16, Legal);
+ setOperationAction(ISD::FFLOOR, MVT::f16, Legal);
+ setOperationAction(ISD::FCEIL, MVT::f16, Legal);
}
if (Subtarget->hasNEON()) {
diff --git a/llvm/lib/Target/AVR/MCTargetDesc/AVRMCAsmInfo.cpp b/llvm/lib/Target/AVR/MCTargetDesc/AVRMCAsmInfo.cpp
index c2c1bb4..0615ec9 100644
--- a/llvm/lib/Target/AVR/MCTargetDesc/AVRMCAsmInfo.cpp
+++ b/llvm/lib/Target/AVR/MCTargetDesc/AVRMCAsmInfo.cpp
@@ -24,8 +24,6 @@ AVRMCAsmInfo::AVRMCAsmInfo(const Triple &TT, const MCTargetOptions &Options) {
CalleeSaveStackSlotSize = 2;
CommentString = ";";
SeparatorString = "$";
- PrivateGlobalPrefix = ".L";
- PrivateLabelPrefix = ".L";
UsesELFSectionDirectiveForBSS = true;
SupportsDebugInformation = true;
}
diff --git a/llvm/lib/Target/AVR/MCTargetDesc/AVRMCAsmInfo.h b/llvm/lib/Target/AVR/MCTargetDesc/AVRMCAsmInfo.h
index fab2713..1915fa8 100644
--- a/llvm/lib/Target/AVR/MCTargetDesc/AVRMCAsmInfo.h
+++ b/llvm/lib/Target/AVR/MCTargetDesc/AVRMCAsmInfo.h
@@ -14,7 +14,7 @@
#define LLVM_AVR_ASM_INFO_H
#include "MCTargetDesc/AVRMCExpr.h"
-#include "llvm/MC/MCAsmInfo.h"
+#include "llvm/MC/MCAsmInfoELF.h"
#include "llvm/MC/MCExpr.h"
namespace llvm {
@@ -22,7 +22,7 @@ namespace llvm {
class Triple;
/// Specifies the format of AVR assembly files.
-class AVRMCAsmInfo : public MCAsmInfo {
+class AVRMCAsmInfo : public MCAsmInfoELF {
public:
explicit AVRMCAsmInfo(const Triple &TT, const MCTargetOptions &Options);
void printSpecifierExpr(raw_ostream &OS,
diff --git a/llvm/lib/Target/BPF/BTFDebug.cpp b/llvm/lib/Target/BPF/BTFDebug.cpp
index a87b9a2..bed6bc9 100644
--- a/llvm/lib/Target/BPF/BTFDebug.cpp
+++ b/llvm/lib/Target/BPF/BTFDebug.cpp
@@ -957,47 +957,47 @@ void BTFDebug::visitMapDefType(const DIType *Ty, uint32_t &TypeId) {
return;
}
- // MapDef type may be a struct type or a non-pointer derived type
- const DIType *OrigTy = Ty;
- while (auto *DTy = dyn_cast<DIDerivedType>(Ty)) {
- auto Tag = DTy->getTag();
- if (Tag != dwarf::DW_TAG_typedef && Tag != dwarf::DW_TAG_const_type &&
- Tag != dwarf::DW_TAG_volatile_type &&
- Tag != dwarf::DW_TAG_restrict_type)
- break;
- Ty = DTy->getBaseType();
- }
-
- const auto *CTy = dyn_cast<DICompositeType>(Ty);
- if (!CTy)
- return;
-
- auto Tag = CTy->getTag();
- if (Tag != dwarf::DW_TAG_structure_type || CTy->isForwardDecl())
- return;
-
- // Visit all struct members to ensure their types are visited.
- const DINodeArray Elements = CTy->getElements();
- for (const auto *Element : Elements) {
- const auto *MemberType = cast<DIDerivedType>(Element);
- const DIType *MemberBaseType = MemberType->getBaseType();
-
- // If the member is a composite type, that may indicate the currently
- // visited composite type is a wrapper, and the member represents the
- // actual map definition.
- // In that case, visit the member with `visitMapDefType` instead of
- // `visitTypeEntry`, treating it specifically as a map definition rather
- // than as a regular composite type.
- const auto *MemberCTy = dyn_cast<DICompositeType>(MemberBaseType);
- if (MemberCTy) {
- visitMapDefType(MemberBaseType, TypeId);
- } else {
- visitTypeEntry(MemberBaseType);
+ uint32_t TmpId;
+ switch (Ty->getTag()) {
+ case dwarf::DW_TAG_typedef:
+ case dwarf::DW_TAG_const_type:
+ case dwarf::DW_TAG_volatile_type:
+ case dwarf::DW_TAG_restrict_type:
+ case dwarf::DW_TAG_pointer_type:
+ visitMapDefType(dyn_cast<DIDerivedType>(Ty)->getBaseType(), TmpId);
+ break;
+ case dwarf::DW_TAG_array_type:
+ // Visit nested map array and jump to the element type
+ visitMapDefType(dyn_cast<DICompositeType>(Ty)->getBaseType(), TmpId);
+ break;
+ case dwarf::DW_TAG_structure_type: {
+ // Visit all struct members to ensure their types are visited.
+ const auto *CTy = cast<DICompositeType>(Ty);
+ const DINodeArray Elements = CTy->getElements();
+ for (const auto *Element : Elements) {
+ const auto *MemberType = cast<DIDerivedType>(Element);
+ const DIType *MemberBaseType = MemberType->getBaseType();
+ // If the member is a composite type, that may indicate the currently
+ // visited composite type is a wrapper, and the member represents the
+ // actual map definition.
+ // In that case, visit the member with `visitMapDefType` instead of
+ // `visitTypeEntry`, treating it specifically as a map definition rather
+ // than as a regular composite type.
+ const auto *MemberCTy = dyn_cast<DICompositeType>(MemberBaseType);
+ if (MemberCTy) {
+ visitMapDefType(MemberBaseType, TmpId);
+ } else {
+ visitTypeEntry(MemberBaseType);
+ }
}
+ break;
+ }
+ default:
+ break;
}
// Visit this type, struct or a const/typedef/volatile/restrict type
- visitTypeEntry(OrigTy, TypeId, false, false);
+ visitTypeEntry(Ty, TypeId, false, false);
}
/// Read file contents from the actual file or from the source
diff --git a/llvm/lib/Target/BPF/MCTargetDesc/BPFMCAsmInfo.h b/llvm/lib/Target/BPF/MCTargetDesc/BPFMCAsmInfo.h
index 7b21684..63d6e6f 100644
--- a/llvm/lib/Target/BPF/MCTargetDesc/BPFMCAsmInfo.h
+++ b/llvm/lib/Target/BPF/MCTargetDesc/BPFMCAsmInfo.h
@@ -13,18 +13,19 @@
#ifndef LLVM_LIB_TARGET_BPF_MCTARGETDESC_BPFMCASMINFO_H
#define LLVM_LIB_TARGET_BPF_MCTARGETDESC_BPFMCASMINFO_H
-#include "llvm/MC/MCAsmInfo.h"
+#include "llvm/MC/MCAsmInfoELF.h"
#include "llvm/TargetParser/Triple.h"
namespace llvm {
-class BPFMCAsmInfo : public MCAsmInfo {
+class BPFMCAsmInfo : public MCAsmInfoELF {
public:
explicit BPFMCAsmInfo(const Triple &TT, const MCTargetOptions &Options) {
if (TT.getArch() == Triple::bpfeb)
IsLittleEndian = false;
PrivateGlobalPrefix = ".L";
+ PrivateLabelPrefix = "L";
WeakRefDirective = "\t.weak\t";
UsesELFSectionDirectiveForBSS = true;
diff --git a/llvm/lib/Target/Hexagon/HexagonCopyToCombine.cpp b/llvm/lib/Target/Hexagon/HexagonCopyToCombine.cpp
index c86fa2b..54c3cea 100644
--- a/llvm/lib/Target/Hexagon/HexagonCopyToCombine.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonCopyToCombine.cpp
@@ -457,7 +457,7 @@ bool HexagonCopyToCombine::runOnMachineFunction(MachineFunction &MF) {
TII = ST->getInstrInfo();
const Function &F = MF.getFunction();
- bool OptForSize = F.hasFnAttribute(Attribute::OptimizeForSize);
+ bool OptForSize = F.hasOptSize();
// Combine aggressively (for code size)
ShouldCombineAggressively =
diff --git a/llvm/lib/Target/Hexagon/HexagonMask.cpp b/llvm/lib/Target/Hexagon/HexagonMask.cpp
index 6eccf80..9d7776d 100644
--- a/llvm/lib/Target/Hexagon/HexagonMask.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonMask.cpp
@@ -76,7 +76,7 @@ bool HexagonMask::runOnMachineFunction(MachineFunction &MF) {
HII = HST.getInstrInfo();
const Function &F = MF.getFunction();
- if (!F.hasFnAttribute(Attribute::OptimizeForSize))
+ if (!F.hasOptSize())
return false;
// Mask instruction is available only from v66
if (!HST.hasV66Ops())
diff --git a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
index 613cfb5..d96136c 100644
--- a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
@@ -2385,13 +2385,6 @@ SDValue LoongArchTargetLowering::lowerBF16_TO_FP(SDValue Op,
return Res;
}
-static bool isConstantBUILD_VECTOR(const BuildVectorSDNode *Op) {
- for (unsigned i = 0; i < Op->getNumOperands(); ++i)
- if (isIntOrFPConstant(Op->getOperand(i)))
- return true;
- return false;
-}
-
// Lower BUILD_VECTOR as broadcast load (if possible).
// For example:
// %a = load i8, ptr %ptr
@@ -2441,10 +2434,14 @@ SDValue LoongArchTargetLowering::lowerBUILD_VECTOR(SDValue Op,
SelectionDAG &DAG) const {
BuildVectorSDNode *Node = cast<BuildVectorSDNode>(Op);
EVT ResTy = Op->getValueType(0);
+ unsigned NumElts = ResTy.getVectorNumElements();
SDLoc DL(Op);
APInt SplatValue, SplatUndef;
unsigned SplatBitSize;
bool HasAnyUndefs;
+ bool IsConstant = false;
+ bool UseSameConstant = true;
+ SDValue ConstantValue;
bool Is128Vec = ResTy.is128BitVector();
bool Is256Vec = ResTy.is256BitVector();
@@ -2495,13 +2492,35 @@ SDValue LoongArchTargetLowering::lowerBUILD_VECTOR(SDValue Op,
if (DAG.isSplatValue(Op, /*AllowUndefs=*/false))
return Op;
- if (!isConstantBUILD_VECTOR(Node)) {
+ for (unsigned i = 0; i < NumElts; ++i) {
+ SDValue Opi = Node->getOperand(i);
+ if (isIntOrFPConstant(Opi)) {
+ IsConstant = true;
+ if (!ConstantValue.getNode())
+ ConstantValue = Opi;
+ else if (ConstantValue != Opi)
+ UseSameConstant = false;
+ }
+ }
+
+ // If the type of BUILD_VECTOR is v2f64, custom legalizing it has no benefits.
+ if (IsConstant && UseSameConstant && ResTy != MVT::v2f64) {
+ SDValue Result = DAG.getSplatBuildVector(ResTy, DL, ConstantValue);
+ for (unsigned i = 0; i < NumElts; ++i) {
+ SDValue Opi = Node->getOperand(i);
+ if (!isIntOrFPConstant(Opi))
+ Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ResTy, Result, Opi,
+ DAG.getConstant(i, DL, Subtarget.getGRLenVT()));
+ }
+ return Result;
+ }
+
+ if (!IsConstant) {
// Use INSERT_VECTOR_ELT operations rather than expand to stores.
// The resulting code is the same length as the expansion, but it doesn't
// use memory operations.
assert(ResTy.isVector());
- unsigned NumElts = ResTy.getVectorNumElements();
SDValue Op0 = Node->getOperand(0);
SDValue Vector = DAG.getUNDEF(ResTy);
diff --git a/llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchAsmBackend.cpp b/llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchAsmBackend.cpp
index 8fa72bc..d9ea88c 100644
--- a/llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchAsmBackend.cpp
+++ b/llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchAsmBackend.cpp
@@ -254,6 +254,7 @@ bool LoongArchAsmBackend::relaxAlign(MCFragment &F, unsigned &Size) {
MCFixup Fixup =
MCFixup::create(0, Expr, FirstLiteralRelocationKind + ELF::R_LARCH_ALIGN);
F.setVarFixups({Fixup});
+ F.setLinkerRelaxable();
F.getParent()->setLinkerRelaxable();
return true;
}
diff --git a/llvm/lib/Target/Mips/MipsAsmPrinter.cpp b/llvm/lib/Target/Mips/MipsAsmPrinter.cpp
index ca03310..a2e48ab 100644
--- a/llvm/lib/Target/Mips/MipsAsmPrinter.cpp
+++ b/llvm/lib/Target/Mips/MipsAsmPrinter.cpp
@@ -737,14 +737,18 @@ void MipsAsmPrinter::emitStartOfAsmFile(Module &M) {
if (FS.empty() && M.size() && F->hasFnAttribute("target-features"))
FS = F->getFnAttribute("target-features").getValueAsString();
+ std::string strFS = FS.str();
+ if (M.size() && F->getFnAttribute("use-soft-float").getValueAsBool())
+ strFS += strFS.empty() ? "+soft-float" : ",+soft-float";
+
// Compute MIPS architecture attributes based on the default subtarget
// that we'd have constructed.
// FIXME: For ifunc related functions we could iterate over and look
// for a feature string that doesn't match the default one.
StringRef CPU = MIPS_MC::selectMipsCPU(TT, TM.getTargetCPU());
const MipsTargetMachine &MTM = static_cast<const MipsTargetMachine &>(TM);
- const MipsSubtarget STI(TT, CPU, FS, MTM.isLittleEndian(), MTM,
- std::nullopt);
+ const MipsSubtarget STI(TT, CPU, StringRef(strFS), MTM.isLittleEndian(),
+ MTM, std::nullopt);
bool IsABICalls = STI.isABICalls();
const MipsABIInfo &ABI = MTM.getABI();
diff --git a/llvm/lib/Target/Mips/MipsISelLowering.cpp b/llvm/lib/Target/Mips/MipsISelLowering.cpp
index 0e581a7..ec6b382 100644
--- a/llvm/lib/Target/Mips/MipsISelLowering.cpp
+++ b/llvm/lib/Target/Mips/MipsISelLowering.cpp
@@ -522,9 +522,6 @@ MipsTargetLowering::MipsTargetLowering(const MipsTargetMachine &TM,
setOperationAction(ISD::TRAP, MVT::Other, Legal);
- setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
- setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
-
setTargetDAGCombine({ISD::SDIVREM, ISD::UDIVREM, ISD::SELECT, ISD::AND,
ISD::OR, ISD::ADD, ISD::SUB, ISD::AssertZext, ISD::SHL,
ISD::SIGN_EXTEND});
@@ -1360,8 +1357,6 @@ LowerOperation(SDValue Op, SelectionDAG &DAG) const
case ISD::FP_TO_SINT: return lowerFP_TO_SINT(Op, DAG);
case ISD::READCYCLECOUNTER:
return lowerREADCYCLECOUNTER(Op, DAG);
- case ISD::ConstantFP:
- return lowerConstantFP(Op, DAG);
}
return SDValue();
}
@@ -3019,30 +3014,6 @@ SDValue MipsTargetLowering::lowerFP_TO_SINT(SDValue Op,
return DAG.getNode(ISD::BITCAST, SDLoc(Op), Op.getValueType(), Trunc);
}
-SDValue MipsTargetLowering::lowerConstantFP(SDValue Op,
- SelectionDAG &DAG) const {
- SDLoc DL(Op);
- EVT VT = Op.getSimpleValueType();
- SDNode *N = Op.getNode();
- ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(N);
-
- if (!CFP->isNaN() || Subtarget.isNaN2008()) {
- return SDValue();
- }
-
- APFloat NaNValue = CFP->getValueAPF();
- auto &Sem = NaNValue.getSemantics();
-
- // The MSB of the mantissa should be zero for QNaNs in the MIPS legacy NaN
- // encodings, and one for sNaNs. Check every NaN constants and make sure
- // they are correctly encoded for legacy encodings.
- if (!NaNValue.isSignaling()) {
- APFloat RealQNaN = NaNValue.getSNaN(Sem);
- return DAG.getConstantFP(RealQNaN, DL, VT);
- }
- return SDValue();
-}
-
//===----------------------------------------------------------------------===//
// Calling Convention Implementation
//===----------------------------------------------------------------------===//
diff --git a/llvm/lib/Target/Mips/MipsISelLowering.h b/llvm/lib/Target/Mips/MipsISelLowering.h
index 31ac5d4..c65c76c 100644
--- a/llvm/lib/Target/Mips/MipsISelLowering.h
+++ b/llvm/lib/Target/Mips/MipsISelLowering.h
@@ -592,7 +592,6 @@ class TargetRegisterClass;
SDValue lowerEH_DWARF_CFA(SDValue Op, SelectionDAG &DAG) const;
SDValue lowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const;
SDValue lowerREADCYCLECOUNTER(SDValue Op, SelectionDAG &DAG) const;
- SDValue lowerConstantFP(SDValue Op, SelectionDAG &DAG) const;
/// isEligibleForTailCallOptimization - Check whether the call is eligible
/// for tail call optimization.
diff --git a/llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXMCAsmInfo.cpp b/llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXMCAsmInfo.cpp
index 614b321..ce9cd12 100644
--- a/llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXMCAsmInfo.cpp
+++ b/llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXMCAsmInfo.cpp
@@ -15,8 +15,6 @@
using namespace llvm;
-void NVPTXMCAsmInfo::anchor() {}
-
NVPTXMCAsmInfo::NVPTXMCAsmInfo(const Triple &TheTriple,
const MCTargetOptions &Options) {
if (TheTriple.getArch() == Triple::nvptx64) {
diff --git a/llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXMCAsmInfo.h b/llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXMCAsmInfo.h
index 77c4dae..f071406 100644
--- a/llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXMCAsmInfo.h
+++ b/llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXMCAsmInfo.h
@@ -19,8 +19,6 @@ namespace llvm {
class Triple;
class NVPTXMCAsmInfo : public MCAsmInfo {
- virtual void anchor();
-
public:
explicit NVPTXMCAsmInfo(const Triple &TheTriple,
const MCTargetOptions &Options);
diff --git a/llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXTargetStreamer.cpp b/llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXTargetStreamer.cpp
index 9f91143..329e3b5 100644
--- a/llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXTargetStreamer.cpp
+++ b/llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXTargetStreamer.cpp
@@ -97,10 +97,7 @@ void NVPTXTargetStreamer::changeSection(const MCSection *CurSection,
if (isDwarfSection(FI, Section)) {
// Emit DWARF .file directives in the outermost scope.
outputDwarfFileDirectives();
- OS << "\t.section";
- Section->printSwitchToSection(*getStreamer().getContext().getAsmInfo(),
- getStreamer().getContext().getTargetTriple(),
- OS, SubSection);
+ OS << "\t.section\t" << Section->getName() << '\n';
// DWARF sections are enclosed into braces - emit the open one.
OS << "\t{\n";
HasSections = true;
diff --git a/llvm/lib/Target/NVPTX/NVPTXISelDAGToDAG.cpp b/llvm/lib/Target/NVPTX/NVPTXISelDAGToDAG.cpp
index 65e7c56..96f52275 100644
--- a/llvm/lib/Target/NVPTX/NVPTXISelDAGToDAG.cpp
+++ b/llvm/lib/Target/NVPTX/NVPTXISelDAGToDAG.cpp
@@ -145,18 +145,6 @@ void NVPTXDAGToDAGISel::Select(SDNode *N) {
if (tryStoreVector(N))
return;
break;
- case NVPTXISD::LoadParam:
- case NVPTXISD::LoadParamV2:
- case NVPTXISD::LoadParamV4:
- if (tryLoadParam(N))
- return;
- break;
- case NVPTXISD::StoreParam:
- case NVPTXISD::StoreParamV2:
- case NVPTXISD::StoreParamV4:
- if (tryStoreParam(N))
- return;
- break;
case ISD::INTRINSIC_W_CHAIN:
if (tryIntrinsicChain(N))
return;
@@ -1462,267 +1450,6 @@ bool NVPTXDAGToDAGISel::tryStoreVector(SDNode *N) {
return true;
}
-bool NVPTXDAGToDAGISel::tryLoadParam(SDNode *Node) {
- SDValue Chain = Node->getOperand(0);
- SDValue Offset = Node->getOperand(2);
- SDValue Glue = Node->getOperand(3);
- SDLoc DL(Node);
- MemSDNode *Mem = cast<MemSDNode>(Node);
-
- unsigned VecSize;
- switch (Node->getOpcode()) {
- default:
- return false;
- case NVPTXISD::LoadParam:
- VecSize = 1;
- break;
- case NVPTXISD::LoadParamV2:
- VecSize = 2;
- break;
- case NVPTXISD::LoadParamV4:
- VecSize = 4;
- break;
- }
-
- EVT EltVT = Node->getValueType(0);
- EVT MemVT = Mem->getMemoryVT();
-
- std::optional<unsigned> Opcode;
-
- switch (VecSize) {
- default:
- return false;
- case 1:
- Opcode = pickOpcodeForVT(MemVT.getSimpleVT().SimpleTy,
- NVPTX::LoadParamMemI8, NVPTX::LoadParamMemI16,
- NVPTX::LoadParamMemI32, NVPTX::LoadParamMemI64);
- break;
- case 2:
- Opcode =
- pickOpcodeForVT(MemVT.getSimpleVT().SimpleTy, NVPTX::LoadParamMemV2I8,
- NVPTX::LoadParamMemV2I16, NVPTX::LoadParamMemV2I32,
- NVPTX::LoadParamMemV2I64);
- break;
- case 4:
- Opcode = pickOpcodeForVT(MemVT.getSimpleVT().SimpleTy,
- NVPTX::LoadParamMemV4I8, NVPTX::LoadParamMemV4I16,
- NVPTX::LoadParamMemV4I32, {/* no v4i64 */});
- break;
- }
- if (!Opcode)
- return false;
-
- SDVTList VTs;
- if (VecSize == 1) {
- VTs = CurDAG->getVTList(EltVT, MVT::Other, MVT::Glue);
- } else if (VecSize == 2) {
- VTs = CurDAG->getVTList(EltVT, EltVT, MVT::Other, MVT::Glue);
- } else {
- EVT EVTs[] = { EltVT, EltVT, EltVT, EltVT, MVT::Other, MVT::Glue };
- VTs = CurDAG->getVTList(EVTs);
- }
-
- unsigned OffsetVal = Offset->getAsZExtVal();
-
- SmallVector<SDValue, 2> Ops(
- {CurDAG->getTargetConstant(OffsetVal, DL, MVT::i32), Chain, Glue});
-
- ReplaceNode(Node, CurDAG->getMachineNode(*Opcode, DL, VTs, Ops));
- return true;
-}
-
-// Helpers for constructing opcode (ex: NVPTX::StoreParamV4F32_iiri)
-#define getOpcV2H(ty, opKind0, opKind1) \
- NVPTX::StoreParamV2##ty##_##opKind0##opKind1
-
-#define getOpcV2H1(ty, opKind0, isImm1) \
- (isImm1) ? getOpcV2H(ty, opKind0, i) : getOpcV2H(ty, opKind0, r)
-
-#define getOpcodeForVectorStParamV2(ty, isimm) \
- (isimm[0]) ? getOpcV2H1(ty, i, isimm[1]) : getOpcV2H1(ty, r, isimm[1])
-
-#define getOpcV4H(ty, opKind0, opKind1, opKind2, opKind3) \
- NVPTX::StoreParamV4##ty##_##opKind0##opKind1##opKind2##opKind3
-
-#define getOpcV4H3(ty, opKind0, opKind1, opKind2, isImm3) \
- (isImm3) ? getOpcV4H(ty, opKind0, opKind1, opKind2, i) \
- : getOpcV4H(ty, opKind0, opKind1, opKind2, r)
-
-#define getOpcV4H2(ty, opKind0, opKind1, isImm2, isImm3) \
- (isImm2) ? getOpcV4H3(ty, opKind0, opKind1, i, isImm3) \
- : getOpcV4H3(ty, opKind0, opKind1, r, isImm3)
-
-#define getOpcV4H1(ty, opKind0, isImm1, isImm2, isImm3) \
- (isImm1) ? getOpcV4H2(ty, opKind0, i, isImm2, isImm3) \
- : getOpcV4H2(ty, opKind0, r, isImm2, isImm3)
-
-#define getOpcodeForVectorStParamV4(ty, isimm) \
- (isimm[0]) ? getOpcV4H1(ty, i, isimm[1], isimm[2], isimm[3]) \
- : getOpcV4H1(ty, r, isimm[1], isimm[2], isimm[3])
-
-#define getOpcodeForVectorStParam(n, ty, isimm) \
- (n == 2) ? getOpcodeForVectorStParamV2(ty, isimm) \
- : getOpcodeForVectorStParamV4(ty, isimm)
-
-static unsigned pickOpcodeForVectorStParam(SmallVector<SDValue, 8> &Ops,
- unsigned NumElts,
- MVT::SimpleValueType MemTy,
- SelectionDAG *CurDAG, SDLoc DL) {
- // Determine which inputs are registers and immediates make new operators
- // with constant values
- SmallVector<bool, 4> IsImm(NumElts, false);
- for (unsigned i = 0; i < NumElts; i++) {
- IsImm[i] = (isa<ConstantSDNode>(Ops[i]) || isa<ConstantFPSDNode>(Ops[i]));
- if (IsImm[i]) {
- SDValue Imm = Ops[i];
- if (MemTy == MVT::f32 || MemTy == MVT::f64) {
- const ConstantFPSDNode *ConstImm = cast<ConstantFPSDNode>(Imm);
- const ConstantFP *CF = ConstImm->getConstantFPValue();
- Imm = CurDAG->getTargetConstantFP(*CF, DL, Imm->getValueType(0));
- } else {
- const ConstantSDNode *ConstImm = cast<ConstantSDNode>(Imm);
- const ConstantInt *CI = ConstImm->getConstantIntValue();
- Imm = CurDAG->getTargetConstant(*CI, DL, Imm->getValueType(0));
- }
- Ops[i] = Imm;
- }
- }
-
- // Get opcode for MemTy, size, and register/immediate operand ordering
- switch (MemTy) {
- case MVT::i8:
- return getOpcodeForVectorStParam(NumElts, I8, IsImm);
- case MVT::i16:
- return getOpcodeForVectorStParam(NumElts, I16, IsImm);
- case MVT::i32:
- return getOpcodeForVectorStParam(NumElts, I32, IsImm);
- case MVT::i64:
- assert(NumElts == 2 && "MVT too large for NumElts > 2");
- return getOpcodeForVectorStParamV2(I64, IsImm);
- case MVT::f32:
- return getOpcodeForVectorStParam(NumElts, F32, IsImm);
- case MVT::f64:
- assert(NumElts == 2 && "MVT too large for NumElts > 2");
- return getOpcodeForVectorStParamV2(F64, IsImm);
-
- // These cases don't support immediates, just use the all register version
- // and generate moves.
- case MVT::i1:
- return (NumElts == 2) ? NVPTX::StoreParamV2I8_rr
- : NVPTX::StoreParamV4I8_rrrr;
- case MVT::f16:
- case MVT::bf16:
- return (NumElts == 2) ? NVPTX::StoreParamV2I16_rr
- : NVPTX::StoreParamV4I16_rrrr;
- case MVT::v2f16:
- case MVT::v2bf16:
- case MVT::v2i16:
- case MVT::v4i8:
- return (NumElts == 2) ? NVPTX::StoreParamV2I32_rr
- : NVPTX::StoreParamV4I32_rrrr;
- default:
- llvm_unreachable("Cannot select st.param for unknown MemTy");
- }
-}
-
-bool NVPTXDAGToDAGISel::tryStoreParam(SDNode *N) {
- SDLoc DL(N);
- SDValue Chain = N->getOperand(0);
- SDValue Param = N->getOperand(1);
- unsigned ParamVal = Param->getAsZExtVal();
- SDValue Offset = N->getOperand(2);
- unsigned OffsetVal = Offset->getAsZExtVal();
- MemSDNode *Mem = cast<MemSDNode>(N);
- SDValue Glue = N->getOperand(N->getNumOperands() - 1);
-
- // How many elements do we have?
- unsigned NumElts;
- switch (N->getOpcode()) {
- default:
- llvm_unreachable("Unexpected opcode");
- case NVPTXISD::StoreParam:
- NumElts = 1;
- break;
- case NVPTXISD::StoreParamV2:
- NumElts = 2;
- break;
- case NVPTXISD::StoreParamV4:
- NumElts = 4;
- break;
- }
-
- // Build vector of operands
- SmallVector<SDValue, 8> Ops;
- for (unsigned i = 0; i < NumElts; ++i)
- Ops.push_back(N->getOperand(i + 3));
- Ops.append({CurDAG->getTargetConstant(ParamVal, DL, MVT::i32),
- CurDAG->getTargetConstant(OffsetVal, DL, MVT::i32), Chain, Glue});
-
- // Determine target opcode
- // If we have an i1, use an 8-bit store. The lowering code in
- // NVPTXISelLowering will have already emitted an upcast.
- std::optional<unsigned> Opcode;
- switch (NumElts) {
- default:
- llvm_unreachable("Unexpected NumElts");
- case 1: {
- MVT::SimpleValueType MemTy = Mem->getMemoryVT().getSimpleVT().SimpleTy;
- SDValue Imm = Ops[0];
- if (MemTy != MVT::f16 && MemTy != MVT::bf16 &&
- (isa<ConstantSDNode>(Imm) || isa<ConstantFPSDNode>(Imm))) {
- // Convert immediate to target constant
- if (MemTy == MVT::f32 || MemTy == MVT::f64) {
- const ConstantFPSDNode *ConstImm = cast<ConstantFPSDNode>(Imm);
- const ConstantFP *CF = ConstImm->getConstantFPValue();
- Imm = CurDAG->getTargetConstantFP(*CF, DL, Imm->getValueType(0));
- } else {
- const ConstantSDNode *ConstImm = cast<ConstantSDNode>(Imm);
- const ConstantInt *CI = ConstImm->getConstantIntValue();
- Imm = CurDAG->getTargetConstant(*CI, DL, Imm->getValueType(0));
- }
- Ops[0] = Imm;
- // Use immediate version of store param
- Opcode =
- pickOpcodeForVT(MemTy, NVPTX::StoreParamI8_i, NVPTX::StoreParamI16_i,
- NVPTX::StoreParamI32_i, NVPTX::StoreParamI64_i);
- } else
- Opcode = pickOpcodeForVT(Mem->getMemoryVT().getSimpleVT().SimpleTy,
- NVPTX::StoreParamI8_r, NVPTX::StoreParamI16_r,
- NVPTX::StoreParamI32_r, NVPTX::StoreParamI64_r);
- if (Opcode == NVPTX::StoreParamI8_r) {
- // Fine tune the opcode depending on the size of the operand.
- // This helps to avoid creating redundant COPY instructions in
- // InstrEmitter::AddRegisterOperand().
- switch (Ops[0].getSimpleValueType().SimpleTy) {
- default:
- break;
- case MVT::i32:
- Opcode = NVPTX::StoreParamI8TruncI32_r;
- break;
- case MVT::i64:
- Opcode = NVPTX::StoreParamI8TruncI64_r;
- break;
- }
- }
- break;
- }
- case 2:
- case 4: {
- MVT::SimpleValueType MemTy = Mem->getMemoryVT().getSimpleVT().SimpleTy;
- Opcode = pickOpcodeForVectorStParam(Ops, NumElts, MemTy, CurDAG, DL);
- break;
- }
- }
-
- SDVTList RetVTs = CurDAG->getVTList(MVT::Other, MVT::Glue);
- SDNode *Ret = CurDAG->getMachineNode(*Opcode, DL, RetVTs, Ops);
- MachineMemOperand *MemRef = cast<MemSDNode>(N)->getMemOperand();
- CurDAG->setNodeMemRefs(cast<MachineSDNode>(Ret), {MemRef});
-
- ReplaceNode(N, Ret);
- return true;
-}
-
/// SelectBFE - Look for instruction sequences that can be made more efficient
/// by using the 'bfe' (bit-field extract) PTX instruction
bool NVPTXDAGToDAGISel::tryBFE(SDNode *N) {
diff --git a/llvm/lib/Target/NVPTX/NVPTXISelDAGToDAG.h b/llvm/lib/Target/NVPTX/NVPTXISelDAGToDAG.h
index b99b4ef..e504a8f 100644
--- a/llvm/lib/Target/NVPTX/NVPTXISelDAGToDAG.h
+++ b/llvm/lib/Target/NVPTX/NVPTXISelDAGToDAG.h
@@ -78,8 +78,6 @@ private:
bool tryLDG(MemSDNode *N);
bool tryStore(SDNode *N);
bool tryStoreVector(SDNode *N);
- bool tryLoadParam(SDNode *N);
- bool tryStoreParam(SDNode *N);
bool tryFence(SDNode *N);
void SelectAddrSpaceCast(SDNode *N);
bool tryBFE(SDNode *N);
diff --git a/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp b/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
index ddcecc00..f79b862 100644
--- a/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
+++ b/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
@@ -1075,12 +1075,6 @@ const char *NVPTXTargetLowering::getTargetNodeName(unsigned Opcode) const {
MAKE_CASE(NVPTXISD::DeclareArrayParam)
MAKE_CASE(NVPTXISD::DeclareScalarParam)
MAKE_CASE(NVPTXISD::CALL)
- MAKE_CASE(NVPTXISD::LoadParam)
- MAKE_CASE(NVPTXISD::LoadParamV2)
- MAKE_CASE(NVPTXISD::LoadParamV4)
- MAKE_CASE(NVPTXISD::StoreParam)
- MAKE_CASE(NVPTXISD::StoreParamV2)
- MAKE_CASE(NVPTXISD::StoreParamV4)
MAKE_CASE(NVPTXISD::MoveParam)
MAKE_CASE(NVPTXISD::UNPACK_VECTOR)
MAKE_CASE(NVPTXISD::BUILD_VECTOR)
@@ -1318,105 +1312,6 @@ Align NVPTXTargetLowering::getArgumentAlignment(const CallBase *CB, Type *Ty,
return DL.getABITypeAlign(Ty);
}
-static bool adjustElementType(EVT &ElementType) {
- switch (ElementType.getSimpleVT().SimpleTy) {
- default:
- return false;
- case MVT::f16:
- case MVT::bf16:
- ElementType = MVT::i16;
- return true;
- case MVT::f32:
- case MVT::v2f16:
- case MVT::v2bf16:
- ElementType = MVT::i32;
- return true;
- case MVT::f64:
- ElementType = MVT::i64;
- return true;
- }
-}
-
-// Use byte-store when the param address of the argument value is unaligned.
-// This may happen when the return value is a field of a packed structure.
-//
-// This is called in LowerCall() when passing the param values.
-static SDValue LowerUnalignedStoreParam(SelectionDAG &DAG, SDValue Chain,
- uint64_t Offset, EVT ElementType,
- SDValue StVal, SDValue &InGlue,
- unsigned ArgID, const SDLoc &dl) {
- // Bit logic only works on integer types
- if (adjustElementType(ElementType))
- StVal = DAG.getNode(ISD::BITCAST, dl, ElementType, StVal);
-
- // Store each byte
- SDVTList StoreVTs = DAG.getVTList(MVT::Other, MVT::Glue);
- for (unsigned i = 0, n = ElementType.getSizeInBits() / 8; i < n; i++) {
- // Shift the byte to the last byte position
- SDValue ShiftVal = DAG.getNode(ISD::SRL, dl, ElementType, StVal,
- DAG.getConstant(i * 8, dl, MVT::i32));
- SDValue StoreOperands[] = {Chain, DAG.getConstant(ArgID, dl, MVT::i32),
- DAG.getConstant(Offset + i, dl, MVT::i32),
- ShiftVal, InGlue};
- // Trunc store only the last byte by using
- // st.param.b8
- // The register type can be larger than b8.
- Chain = DAG.getMemIntrinsicNode(
- NVPTXISD::StoreParam, dl, StoreVTs, StoreOperands, MVT::i8,
- MachinePointerInfo(), Align(1), MachineMemOperand::MOStore);
- InGlue = Chain.getValue(1);
- }
- return Chain;
-}
-
-// Use byte-load when the param adress of the returned value is unaligned.
-// This may happen when the returned value is a field of a packed structure.
-static SDValue
-LowerUnalignedLoadRetParam(SelectionDAG &DAG, SDValue &Chain, uint64_t Offset,
- EVT ElementType, SDValue &InGlue,
- SmallVectorImpl<SDValue> &TempProxyRegOps,
- const SDLoc &dl) {
- // Bit logic only works on integer types
- EVT MergedType = ElementType;
- adjustElementType(MergedType);
-
- // Load each byte and construct the whole value. Initial value to 0
- SDValue RetVal = DAG.getConstant(0, dl, MergedType);
- // LoadParamMemI8 loads into i16 register only
- SDVTList LoadVTs = DAG.getVTList(MVT::i16, MVT::Other, MVT::Glue);
- for (unsigned i = 0, n = ElementType.getSizeInBits() / 8; i < n; i++) {
- SDValue LoadOperands[] = {Chain, DAG.getConstant(1, dl, MVT::i32),
- DAG.getConstant(Offset + i, dl, MVT::i32),
- InGlue};
- // This will be selected to LoadParamMemI8
- SDValue LdVal =
- DAG.getMemIntrinsicNode(NVPTXISD::LoadParam, dl, LoadVTs, LoadOperands,
- MVT::i8, MachinePointerInfo(), Align(1));
- SDValue TmpLdVal = LdVal.getValue(0);
- Chain = LdVal.getValue(1);
- InGlue = LdVal.getValue(2);
-
- TmpLdVal = DAG.getNode(NVPTXISD::ProxyReg, dl,
- TmpLdVal.getSimpleValueType(), TmpLdVal);
- TempProxyRegOps.push_back(TmpLdVal);
-
- SDValue CMask = DAG.getConstant(255, dl, MergedType);
- SDValue CShift = DAG.getConstant(i * 8, dl, MVT::i32);
- // Need to extend the i16 register to the whole width.
- TmpLdVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MergedType, TmpLdVal);
- // Mask off the high bits. Leave only the lower 8bits.
- // Do this because we are using loadparam.b8.
- TmpLdVal = DAG.getNode(ISD::AND, dl, MergedType, TmpLdVal, CMask);
- // Shift and merge
- TmpLdVal = DAG.getNode(ISD::SHL, dl, MergedType, TmpLdVal, CShift);
- RetVal = DAG.getNode(ISD::OR, dl, MergedType, RetVal, TmpLdVal);
- }
- if (ElementType != MergedType)
- RetVal = DAG.getNode(ISD::BITCAST, dl, ElementType, RetVal);
-
- return RetVal;
-}
-
static bool shouldConvertToIndirectCall(const CallBase *CB,
const GlobalAddressSDNode *Func) {
if (!Func)
@@ -1483,10 +1378,8 @@ SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
SelectionDAG &DAG = CLI.DAG;
SDLoc dl = CLI.DL;
- SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
- SDValue Chain = CLI.Chain;
+ const SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
SDValue Callee = CLI.Callee;
- bool &isTailCall = CLI.IsTailCall;
ArgListTy &Args = CLI.getArgs();
Type *RetTy = CLI.RetTy;
const CallBase *CB = CLI.CB;
@@ -1496,6 +1389,36 @@ SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
return DAG.getConstant(I, dl, MVT::i32);
};
+ const unsigned UniqueCallSite = GlobalUniqueCallSite++;
+ const SDValue CallChain = CLI.Chain;
+ const SDValue StartChain =
+ DAG.getCALLSEQ_START(CallChain, UniqueCallSite, 0, dl);
+ SDValue DeclareGlue = StartChain.getValue(1);
+
+ SmallVector<SDValue, 16> CallPrereqs{StartChain};
+
+ const auto MakeDeclareScalarParam = [&](SDValue Symbol, unsigned Size) {
+ // PTX ABI requires integral types to be at least 32 bits in size. FP16 is
+ // loaded/stored using i16, so it's handled here as well.
+ const unsigned SizeBits = promoteScalarArgumentSize(Size * 8);
+ SDValue Declare =
+ DAG.getNode(NVPTXISD::DeclareScalarParam, dl, {MVT::Other, MVT::Glue},
+ {StartChain, Symbol, GetI32(SizeBits), DeclareGlue});
+ CallPrereqs.push_back(Declare);
+ DeclareGlue = Declare.getValue(1);
+ return Declare;
+ };
+
+ const auto MakeDeclareArrayParam = [&](SDValue Symbol, Align Align,
+ unsigned Size) {
+ SDValue Declare = DAG.getNode(
+ NVPTXISD::DeclareArrayParam, dl, {MVT::Other, MVT::Glue},
+ {StartChain, Symbol, GetI32(Align.value()), GetI32(Size), DeclareGlue});
+ CallPrereqs.push_back(Declare);
+ DeclareGlue = Declare.getValue(1);
+ return Declare;
+ };
+
// Variadic arguments.
//
// Normally, for each argument, we declare a param scalar or a param
@@ -1511,15 +1434,17 @@ SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
//
// After all vararg is processed, 'VAOffset' holds the size of the
// vararg byte array.
+ assert((CLI.IsVarArg || CLI.Args.size() == CLI.NumFixedArgs) &&
+ "Non-VarArg function with extra arguments");
- SDValue VADeclareParam; // vararg byte array
const unsigned FirstVAArg = CLI.NumFixedArgs; // position of first variadic
- unsigned VAOffset = 0; // current offset in the param array
+ unsigned VAOffset = 0; // current offset in the param array
- const unsigned UniqueCallSite = GlobalUniqueCallSite++;
- SDValue TempChain = Chain;
- Chain = DAG.getCALLSEQ_START(Chain, UniqueCallSite, 0, dl);
- SDValue InGlue = Chain.getValue(1);
+ const SDValue VADeclareParam =
+ CLI.Args.size() > FirstVAArg
+ ? MakeDeclareArrayParam(getCallParamSymbol(DAG, FirstVAArg, MVT::i32),
+ Align(STI.getMaxRequiredAlignment()), 0)
+ : SDValue();
// Args.size() and Outs.size() need not match.
// Outs.size() will be larger
@@ -1580,43 +1505,19 @@ SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
assert((!IsByVal || TypeSize == ArgOuts[0].Flags.getByValSize()) &&
"type size mismatch");
- const std::optional<SDValue> ArgDeclare = [&]() -> std::optional<SDValue> {
- if (IsVAArg) {
- if (ArgI == FirstVAArg) {
- VADeclareParam = DAG.getNode(
- NVPTXISD::DeclareArrayParam, dl, {MVT::Other, MVT::Glue},
- {Chain, ParamSymbol, GetI32(STI.getMaxRequiredAlignment()),
- GetI32(0), InGlue});
- return VADeclareParam;
- }
- return std::nullopt;
- }
- if (IsByVal || shouldPassAsArray(Arg.Ty)) {
- // declare .param .align <align> .b8 .param<n>[<size>];
- return DAG.getNode(NVPTXISD::DeclareArrayParam, dl,
- {MVT::Other, MVT::Glue},
- {Chain, ParamSymbol, GetI32(ArgAlign.value()),
- GetI32(TypeSize), InGlue});
- }
+ const SDValue ArgDeclare = [&]() {
+ if (IsVAArg)
+ return VADeclareParam;
+
+ if (IsByVal || shouldPassAsArray(Arg.Ty))
+ return MakeDeclareArrayParam(ParamSymbol, ArgAlign, TypeSize);
+
assert(ArgOuts.size() == 1 && "We must pass only one value as non-array");
- // declare .param .b<size> .param<n>;
-
- // PTX ABI requires integral types to be at least 32 bits in
- // size. FP16 is loaded/stored using i16, so it's handled
- // here as well.
- const unsigned PromotedSize =
- (ArgOuts[0].VT.isInteger() || ArgOuts[0].VT.isFloatingPoint())
- ? promoteScalarArgumentSize(TypeSize * 8)
- : TypeSize * 8;
-
- return DAG.getNode(NVPTXISD::DeclareScalarParam, dl,
- {MVT::Other, MVT::Glue},
- {Chain, ParamSymbol, GetI32(PromotedSize), InGlue});
+ assert((ArgOuts[0].VT.isInteger() || ArgOuts[0].VT.isFloatingPoint()) &&
+ "Only int and float types are supported as non-array arguments");
+
+ return MakeDeclareScalarParam(ParamSymbol, TypeSize);
}();
- if (ArgDeclare) {
- Chain = ArgDeclare->getValue(0);
- InGlue = ArgDeclare->getValue(1);
- }
// PTX Interoperability Guide 3.3(A): [Integer] Values shorter
// than 32-bits are sign extended or zero extended, depending on
@@ -1626,36 +1527,25 @@ SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
Arg.Ty->isIntegerTy() && DL.getTypeAllocSizeInBits(Arg.Ty) < 32;
const auto GetStoredValue = [&](const unsigned I, EVT EltVT,
- const Align PartAlign) {
- SDValue StVal;
+ const MaybeAlign PartAlign) {
if (IsByVal) {
SDValue Ptr = ArgOutVals[0];
auto MPI = refinePtrAS(Ptr, DAG, DL, *this);
SDValue SrcAddr =
DAG.getObjectPtrOffset(dl, Ptr, TypeSize::getFixed(Offsets[I]));
- StVal = DAG.getLoad(EltVT, dl, TempChain, SrcAddr, MPI, PartAlign);
- } else {
- StVal = ArgOutVals[I];
-
- auto PromotedVT = promoteScalarIntegerPTX(StVal.getValueType());
- if (PromotedVT != StVal.getValueType()) {
- StVal = DAG.getNode(getExtOpcode(ArgOuts[I].Flags), dl, PromotedVT,
- StVal);
- }
+ return DAG.getLoad(EltVT, dl, CallChain, SrcAddr, MPI, PartAlign);
}
+ SDValue StVal = ArgOutVals[I];
+ assert(promoteScalarIntegerPTX(StVal.getValueType()) ==
+ StVal.getValueType() &&
+ "OutVal type should always be legal");
- if (ExtendIntegerParam) {
- assert(VTs.size() == 1 && "Scalar can't have multiple parts.");
- // zext/sext to i32
- StVal =
- DAG.getNode(getExtOpcode(ArgOuts[I].Flags), dl, MVT::i32, StVal);
- } else if (EltVT.getSizeInBits() < 16) {
- // Use 16-bit registers for small stores as it's the
- // smallest general purpose register size supported by NVPTX.
- StVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, StVal);
- }
- return StVal;
+ const EVT VTI = promoteScalarIntegerPTX(VTs[I]);
+ const EVT StoreVT =
+ ExtendIntegerParam ? MVT::i32 : (VTI == MVT::i1 ? MVT::i8 : VTI);
+
+ return correctParamType(StVal, StoreVT, ArgOuts[I].Flags, DAG, dl);
};
const auto VectorInfo =
@@ -1664,23 +1554,7 @@ SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
unsigned J = 0;
for (const unsigned NumElts : VectorInfo) {
const int CurOffset = Offsets[J];
- EVT EltVT = promoteScalarIntegerPTX(VTs[J]);
- const Align PartAlign = commonAlignment(ArgAlign, CurOffset);
-
- // If we have a PVF_SCALAR entry, it may not be sufficiently aligned for a
- // scalar store. In such cases, fall back to byte stores.
- if (NumElts == 1 && !IsVAArg && PartAlign < DAG.getEVTAlign(EltVT)) {
-
- SDValue StVal = GetStoredValue(J, EltVT, PartAlign);
- Chain = LowerUnalignedStoreParam(DAG, Chain,
- CurOffset + (IsByVal ? VAOffset : 0),
- EltVT, StVal, InGlue, ArgI, dl);
-
- // LowerUnalignedStoreParam took care of inserting the necessary nodes
- // into the SDAG, so just move on to the next element.
- J++;
- continue;
- }
+ const EVT EltVT = promoteScalarIntegerPTX(VTs[J]);
if (IsVAArg && !IsByVal)
// Align each part of the variadic argument to their type.
@@ -1688,44 +1562,45 @@ SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
assert((IsVAArg || VAOffset == 0) &&
"VAOffset must be 0 for non-VA args");
- SmallVector<SDValue, 6> StoreOperands{
- Chain, GetI32(IsVAArg ? FirstVAArg : ArgI),
- GetI32(VAOffset + ((IsVAArg && !IsByVal) ? 0 : CurOffset))};
- // Record the values to store.
- for (const unsigned K : llvm::seq(NumElts))
- StoreOperands.push_back(GetStoredValue(J + K, EltVT, PartAlign));
- StoreOperands.push_back(InGlue);
+ const unsigned Offset =
+ (VAOffset + ((IsVAArg && !IsByVal) ? 0 : CurOffset));
+ SDValue Ptr =
+ DAG.getObjectPtrOffset(dl, ParamSymbol, TypeSize::getFixed(Offset));
- NVPTXISD::NodeType Op;
- switch (NumElts) {
- case 1:
- Op = NVPTXISD::StoreParam;
- break;
- case 2:
- Op = NVPTXISD::StoreParamV2;
- break;
- case 4:
- Op = NVPTXISD::StoreParamV4;
- break;
- default:
- llvm_unreachable("Invalid vector info.");
+ const MaybeAlign CurrentAlign = ExtendIntegerParam
+ ? MaybeAlign(std::nullopt)
+ : commonAlignment(ArgAlign, Offset);
+
+ SDValue Val;
+ if (NumElts == 1) {
+ Val = GetStoredValue(J, EltVT, CurrentAlign);
+ } else {
+ SmallVector<SDValue, 8> StoreVals;
+ for (const unsigned K : llvm::seq(NumElts)) {
+ SDValue ValJ = GetStoredValue(J + K, EltVT, CurrentAlign);
+ if (ValJ.getValueType().isVector())
+ DAG.ExtractVectorElements(ValJ, StoreVals);
+ else
+ StoreVals.push_back(ValJ);
+ }
+
+ EVT VT = EVT::getVectorVT(
+ *DAG.getContext(), StoreVals[0].getValueType(), StoreVals.size());
+ Val = DAG.getBuildVector(VT, dl, StoreVals);
}
- // Adjust type of the store op if we've extended the scalar
- // return value.
- EVT TheStoreType = ExtendIntegerParam ? MVT::i32 : EltVT;
- Chain = DAG.getMemIntrinsicNode(
- Op, dl, DAG.getVTList(MVT::Other, MVT::Glue), StoreOperands,
- TheStoreType, MachinePointerInfo(), PartAlign,
- MachineMemOperand::MOStore);
- InGlue = Chain.getValue(1);
+ SDValue StoreParam =
+ DAG.getStore(ArgDeclare, dl, Val, Ptr,
+ MachinePointerInfo(ADDRESS_SPACE_PARAM), CurrentAlign);
+ CallPrereqs.push_back(StoreParam);
// TODO: We may need to support vector types that can be passed
// as scalars in variadic arguments.
if (IsVAArg && !IsByVal) {
assert(NumElts == 1 &&
"Vectorization is expected to be disabled for variadics.");
+ const EVT TheStoreType = ExtendIntegerParam ? MVT::i32 : EltVT;
VAOffset +=
DL.getTypeAllocSize(TheStoreType.getTypeForEVT(*DAG.getContext()));
}
@@ -1736,33 +1611,21 @@ SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
VAOffset += TypeSize;
}
- GlobalAddressSDNode *Func = dyn_cast<GlobalAddressSDNode>(Callee.getNode());
-
// Handle Result
if (!Ins.empty()) {
- const SDValue RetDeclare = [&]() {
- const SDValue RetSymbol = DAG.getExternalSymbol("retval0", MVT::i32);
- const unsigned ResultSize = DL.getTypeAllocSizeInBits(RetTy);
- if (shouldPassAsArray(RetTy)) {
- const Align RetAlign = getArgumentAlignment(CB, RetTy, 0, DL);
- return DAG.getNode(NVPTXISD::DeclareArrayParam, dl,
- {MVT::Other, MVT::Glue},
- {Chain, RetSymbol, GetI32(RetAlign.value()),
- GetI32(ResultSize / 8), InGlue});
- }
- const auto PromotedResultSize = promoteScalarArgumentSize(ResultSize);
- return DAG.getNode(
- NVPTXISD::DeclareScalarParam, dl, {MVT::Other, MVT::Glue},
- {Chain, RetSymbol, GetI32(PromotedResultSize), InGlue});
- }();
- Chain = RetDeclare.getValue(0);
- InGlue = RetDeclare.getValue(1);
+ const SDValue RetSymbol = DAG.getExternalSymbol("retval0", MVT::i32);
+ const unsigned ResultSize = DL.getTypeAllocSize(RetTy);
+ if (shouldPassAsArray(RetTy)) {
+ const Align RetAlign = getArgumentAlignment(CB, RetTy, 0, DL);
+ MakeDeclareArrayParam(RetSymbol, RetAlign, ResultSize);
+ } else {
+ MakeDeclareScalarParam(RetSymbol, ResultSize);
+ }
}
- const bool HasVAArgs = CLI.IsVarArg && (CLI.Args.size() > CLI.NumFixedArgs);
// Set the size of the vararg param byte array if the callee is a variadic
// function and the variadic part is not empty.
- if (HasVAArgs) {
+ if (VADeclareParam) {
SDValue DeclareParamOps[] = {VADeclareParam.getOperand(0),
VADeclareParam.getOperand(1),
VADeclareParam.getOperand(2), GetI32(VAOffset),
@@ -1771,6 +1634,7 @@ SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
VADeclareParam->getVTList(), DeclareParamOps);
}
+ const auto *Func = dyn_cast<GlobalAddressSDNode>(Callee.getNode());
// If the type of the callsite does not match that of the function, convert
// the callsite to an indirect call.
const bool ConvertToIndirectCall = shouldConvertToIndirectCall(CB, Func);
@@ -1800,15 +1664,16 @@ SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
// instruction.
// The prototype is embedded in a string and put as the operand for a
// CallPrototype SDNode which will print out to the value of the string.
+ const bool HasVAArgs = CLI.IsVarArg && (CLI.Args.size() > CLI.NumFixedArgs);
std::string Proto =
getPrototype(DL, RetTy, Args, CLI.Outs,
HasVAArgs ? std::optional(FirstVAArg) : std::nullopt, *CB,
UniqueCallSite);
const char *ProtoStr = nvTM->getStrPool().save(Proto).data();
- Chain = DAG.getNode(
- NVPTXISD::CallPrototype, dl, {MVT::Other, MVT::Glue},
- {Chain, DAG.getTargetExternalSymbol(ProtoStr, MVT::i32), InGlue});
- InGlue = Chain.getValue(1);
+ const SDValue PrototypeDeclare = DAG.getNode(
+ NVPTXISD::CallPrototype, dl, MVT::Other,
+ {StartChain, DAG.getTargetExternalSymbol(ProtoStr, MVT::i32)});
+ CallPrereqs.push_back(PrototypeDeclare);
}
if (ConvertToIndirectCall) {
@@ -1826,24 +1691,15 @@ SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
const unsigned NumArgs =
std::min<unsigned>(CLI.NumFixedArgs + 1, Args.size());
/// CALL(Chain, IsConvergent, IsIndirectCall/IsUniform, NumReturns,
- /// NumParams, Callee, Proto, InGlue)
- Chain = DAG.getNode(NVPTXISD::CALL, dl, {MVT::Other, MVT::Glue},
- {Chain, GetI32(CLI.IsConvergent), GetI32(IsIndirectCall),
- GetI32(Ins.empty() ? 0 : 1), GetI32(NumArgs), Callee,
- GetI32(Proto), InGlue});
- InGlue = Chain.getValue(1);
-
+ /// NumParams, Callee, Proto)
+ const SDValue CallToken = DAG.getTokenFactor(dl, CallPrereqs);
+ const SDValue Call = DAG.getNode(
+ NVPTXISD::CALL, dl, MVT::Other,
+ {CallToken, GetI32(CLI.IsConvergent), GetI32(IsIndirectCall),
+ GetI32(Ins.empty() ? 0 : 1), GetI32(NumArgs), Callee, GetI32(Proto)});
+
+ SmallVector<SDValue, 16> LoadChains{Call};
SmallVector<SDValue, 16> ProxyRegOps;
- // An item of the vector is filled if the element does not need a ProxyReg
- // operation on it and should be added to InVals as is. ProxyRegOps and
- // ProxyRegTruncates contain empty/none items at the same index.
- SmallVector<SDValue, 16> RetElts;
- // A temporary ProxyReg operations inserted in `LowerUnalignedLoadRetParam()`
- // to use the values of `LoadParam`s and to be replaced later then
- // `CALLSEQ_END` is added.
- SmallVector<SDValue, 16> TempProxyRegOps;
-
- // Generate loads from param memory/moves from registers for result
if (!Ins.empty()) {
SmallVector<EVT, 16> VTs;
SmallVector<uint64_t, 16> Offsets;
@@ -1860,104 +1716,65 @@ SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
const auto VectorInfo = VectorizePTXValueVTs(VTs, Offsets, RetAlign);
unsigned I = 0;
- for (const unsigned VectorizedSize : VectorInfo) {
- EVT TheLoadType = promoteScalarIntegerPTX(VTs[I]);
- EVT EltType = Ins[I].VT;
- const Align EltAlign = commonAlignment(RetAlign, Offsets[I]);
-
- if (TheLoadType != VTs[I])
- EltType = TheLoadType;
-
- if (ExtendIntegerRetVal) {
- TheLoadType = MVT::i32;
- EltType = MVT::i32;
- } else if (TheLoadType.getSizeInBits() < 16) {
- EltType = MVT::i16;
- }
+ for (const unsigned NumElts : VectorInfo) {
+ const MaybeAlign CurrentAlign =
+ ExtendIntegerRetVal ? MaybeAlign(std::nullopt)
+ : commonAlignment(RetAlign, Offsets[I]);
- // If we have a PVF_SCALAR entry, it may not be sufficiently aligned for a
- // scalar load. In such cases, fall back to byte loads.
- if (VectorizedSize == 1 && RetTy->isAggregateType() &&
- EltAlign < DAG.getEVTAlign(TheLoadType)) {
- SDValue Ret = LowerUnalignedLoadRetParam(
- DAG, Chain, Offsets[I], TheLoadType, InGlue, TempProxyRegOps, dl);
- ProxyRegOps.push_back(SDValue());
- RetElts.resize(I);
- RetElts.push_back(Ret);
-
- I++;
- continue;
- }
+ const EVT VTI = promoteScalarIntegerPTX(VTs[I]);
+ const EVT LoadVT =
+ ExtendIntegerRetVal ? MVT::i32 : (VTI == MVT::i1 ? MVT::i8 : VTI);
- SmallVector<EVT, 6> LoadVTs(VectorizedSize, EltType);
- LoadVTs.append({MVT::Other, MVT::Glue});
+ const unsigned PackingAmt =
+ LoadVT.isVector() ? LoadVT.getVectorNumElements() : 1;
- NVPTXISD::NodeType Op;
- switch (VectorizedSize) {
- case 1:
- Op = NVPTXISD::LoadParam;
- break;
- case 2:
- Op = NVPTXISD::LoadParamV2;
- break;
- case 4:
- Op = NVPTXISD::LoadParamV4;
- break;
- default:
- llvm_unreachable("Invalid vector info.");
- }
+ const EVT VecVT = NumElts == 1 ? LoadVT
+ : EVT::getVectorVT(*DAG.getContext(),
+ LoadVT.getScalarType(),
+ NumElts * PackingAmt);
- SDValue LoadOperands[] = {Chain, GetI32(1), GetI32(Offsets[I]), InGlue};
- SDValue RetVal = DAG.getMemIntrinsicNode(
- Op, dl, DAG.getVTList(LoadVTs), LoadOperands, TheLoadType,
- MachinePointerInfo(), EltAlign, MachineMemOperand::MOLoad);
+ const SDValue RetSymbol = DAG.getExternalSymbol("retval0", MVT::i32);
+ SDValue Ptr =
+ DAG.getObjectPtrOffset(dl, RetSymbol, TypeSize::getFixed(Offsets[I]));
- for (const unsigned J : llvm::seq(VectorizedSize)) {
- ProxyRegOps.push_back(RetVal.getValue(J));
- }
+ SDValue R =
+ DAG.getLoad(VecVT, dl, Call, Ptr,
+ MachinePointerInfo(ADDRESS_SPACE_PARAM), CurrentAlign);
- Chain = RetVal.getValue(VectorizedSize);
- InGlue = RetVal.getValue(VectorizedSize + 1);
+ LoadChains.push_back(R.getValue(1));
- I += VectorizedSize;
+ if (NumElts == 1)
+ ProxyRegOps.push_back(R);
+ else
+ for (const unsigned J : llvm::seq(NumElts)) {
+ SDValue Elt = DAG.getNode(
+ LoadVT.isVector() ? ISD::EXTRACT_SUBVECTOR
+ : ISD::EXTRACT_VECTOR_ELT,
+ dl, LoadVT, R, DAG.getVectorIdxConstant(J * PackingAmt, dl));
+ ProxyRegOps.push_back(Elt);
+ }
+ I += NumElts;
}
}
- Chain =
- DAG.getCALLSEQ_END(Chain, UniqueCallSite, UniqueCallSite + 1, InGlue, dl);
- InGlue = Chain.getValue(1);
+ const SDValue EndToken = DAG.getTokenFactor(dl, LoadChains);
+ const SDValue CallEnd = DAG.getCALLSEQ_END(EndToken, UniqueCallSite,
+ UniqueCallSite + 1, SDValue(), dl);
// Append ProxyReg instructions to the chain to make sure that `callseq_end`
// will not get lost. Otherwise, during libcalls expansion, the nodes can become
// dangling.
- for (const unsigned I : llvm::seq(ProxyRegOps.size())) {
- if (I < RetElts.size() && RetElts[I]) {
- InVals.push_back(RetElts[I]);
- continue;
- }
-
- SDValue Ret =
- DAG.getNode(NVPTXISD::ProxyReg, dl, ProxyRegOps[I].getSimpleValueType(),
- {Chain, ProxyRegOps[I]});
-
- const EVT ExpectedVT = Ins[I].VT;
- if (!Ret.getValueType().bitsEq(ExpectedVT)) {
- Ret = DAG.getNode(ISD::TRUNCATE, dl, ExpectedVT, Ret);
- }
+ for (const auto [I, Reg] : llvm::enumerate(ProxyRegOps)) {
+ SDValue Proxy =
+ DAG.getNode(NVPTXISD::ProxyReg, dl, Reg.getValueType(), {CallEnd, Reg});
+ SDValue Ret = correctParamType(Proxy, Ins[I].VT, Ins[I].Flags, DAG, dl);
InVals.push_back(Ret);
}
- for (SDValue &T : TempProxyRegOps) {
- SDValue Repl = DAG.getNode(NVPTXISD::ProxyReg, dl, T.getSimpleValueType(),
- {Chain, T.getOperand(0)});
- DAG.ReplaceAllUsesWith(T, Repl);
- DAG.RemoveDeadNode(T.getNode());
- }
-
- // set isTailCall to false for now, until we figure out how to express
+ // set IsTailCall to false for now, until we figure out how to express
// tail call optimization in PTX
- isTailCall = false;
- return Chain;
+ CLI.IsTailCall = false;
+ return CallEnd;
}
SDValue NVPTXTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
@@ -5117,10 +4934,6 @@ combineUnpackingMovIntoLoad(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
Operands.push_back(DCI.DAG.getIntPtrConstant(
cast<LoadSDNode>(LD)->getExtensionType(), DL));
break;
- case NVPTXISD::LoadParamV2:
- OldNumOutputs = 2;
- Opcode = NVPTXISD::LoadParamV4;
- break;
case NVPTXISD::LoadV2:
OldNumOutputs = 2;
Opcode = NVPTXISD::LoadV4;
@@ -5201,12 +5014,6 @@ static SDValue combinePackingMovIntoStore(SDNode *N,
MemVT = ST->getMemoryVT();
Opcode = NVPTXISD::StoreV2;
break;
- case NVPTXISD::StoreParam:
- Opcode = NVPTXISD::StoreParamV2;
- break;
- case NVPTXISD::StoreParamV2:
- Opcode = NVPTXISD::StoreParamV4;
- break;
case NVPTXISD::StoreV2:
MemVT = ST->getMemoryVT();
Opcode = NVPTXISD::StoreV4;
@@ -5218,7 +5025,6 @@ static SDValue combinePackingMovIntoStore(SDNode *N,
return SDValue();
Opcode = NVPTXISD::StoreV8;
break;
- case NVPTXISD::StoreParamV4:
case NVPTXISD::StoreV8:
// PTX doesn't support the next doubling of operands
return SDValue();
@@ -5263,30 +5069,11 @@ static SDValue combinePackingMovIntoStore(SDNode *N,
MemVT, ST->getMemOperand());
}
-static SDValue PerformStoreCombineHelper(SDNode *N,
- TargetLowering::DAGCombinerInfo &DCI,
- unsigned Front, unsigned Back) {
- if (all_of(N->ops().drop_front(Front).drop_back(Back),
- [](const SDUse &U) { return U.get()->isUndef(); }))
- // Operand 0 is the previous value in the chain. Cannot return EntryToken
- // as the previous value will become unused and eliminated later.
- return N->getOperand(0);
-
- return combinePackingMovIntoStore(N, DCI, Front, Back);
-}
-
static SDValue PerformStoreCombine(SDNode *N,
TargetLowering::DAGCombinerInfo &DCI) {
return combinePackingMovIntoStore(N, DCI, 1, 2);
}
-static SDValue PerformStoreParamCombine(SDNode *N,
- TargetLowering::DAGCombinerInfo &DCI) {
- // Operands from the 3rd to the 2nd last one are the values to be stored.
- // {Chain, ArgID, Offset, Val, Glue}
- return PerformStoreCombineHelper(N, DCI, 3, 1);
-}
-
/// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
///
static SDValue PerformADDCombine(SDNode *N,
@@ -5942,6 +5729,86 @@ static SDValue combinePRMT(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
N->getConstantOperandAPInt(2),
N->getConstantOperandVal(3)),
SDLoc(N), N->getValueType(0));
+ return SDValue();
+}
+
+// During call lowering we wrap the return values in a ProxyReg node which
+// depend on the chain value produced by the completed call. This ensures that
+// the full call is emitted in cases where libcalls are used to legalize
+// operations. To improve the functioning of other DAG combines we pull all
+// operations we can through one of these nodes, ensuring that the ProxyReg
+// directly wraps a load. That is:
+//
+// (ProxyReg (zext (load retval0))) => (zext (ProxyReg (load retval0)))
+//
+static SDValue sinkProxyReg(SDValue R, SDValue Chain,
+ TargetLowering::DAGCombinerInfo &DCI) {
+ switch (R.getOpcode()) {
+ case ISD::TRUNCATE:
+ case ISD::ANY_EXTEND:
+ case ISD::SIGN_EXTEND:
+ case ISD::ZERO_EXTEND:
+ case ISD::BITCAST: {
+ if (SDValue V = sinkProxyReg(R.getOperand(0), Chain, DCI))
+ return DCI.DAG.getNode(R.getOpcode(), SDLoc(R), R.getValueType(), V);
+ return SDValue();
+ }
+ case ISD::SHL:
+ case ISD::SRL:
+ case ISD::SRA:
+ case ISD::OR: {
+ if (SDValue A = sinkProxyReg(R.getOperand(0), Chain, DCI))
+ if (SDValue B = sinkProxyReg(R.getOperand(1), Chain, DCI))
+ return DCI.DAG.getNode(R.getOpcode(), SDLoc(R), R.getValueType(), A, B);
+ return SDValue();
+ }
+ case ISD::Constant:
+ return R;
+ case ISD::LOAD:
+ case NVPTXISD::LoadV2:
+ case NVPTXISD::LoadV4: {
+ return DCI.DAG.getNode(NVPTXISD::ProxyReg, SDLoc(R), R.getValueType(),
+ {Chain, R});
+ }
+ case ISD::BUILD_VECTOR: {
+ if (DCI.isBeforeLegalize())
+ return SDValue();
+
+ SmallVector<SDValue, 16> Ops;
+ for (auto &Op : R->ops()) {
+ SDValue V = sinkProxyReg(Op, Chain, DCI);
+ if (!V)
+ return SDValue();
+ Ops.push_back(V);
+ }
+ return DCI.DAG.getNode(ISD::BUILD_VECTOR, SDLoc(R), R.getValueType(), Ops);
+ }
+ case ISD::EXTRACT_VECTOR_ELT: {
+ if (DCI.isBeforeLegalize())
+ return SDValue();
+
+ if (SDValue V = sinkProxyReg(R.getOperand(0), Chain, DCI))
+ return DCI.DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(R),
+ R.getValueType(), V, R.getOperand(1));
+ return SDValue();
+ }
+ default:
+ return SDValue();
+ }
+}
+
+static SDValue combineProxyReg(SDNode *N,
+ TargetLowering::DAGCombinerInfo &DCI) {
+
+ SDValue Chain = N->getOperand(0);
+ SDValue Reg = N->getOperand(1);
+
+ // If the ProxyReg is not wrapping a load, try to pull the operations through
+ // the ProxyReg.
+ if (Reg.getOpcode() != ISD::LOAD) {
+ if (SDValue V = sinkProxyReg(Reg, Chain, DCI))
+ return V;
+ }
return SDValue();
}
@@ -5965,7 +5832,6 @@ SDValue NVPTXTargetLowering::PerformDAGCombine(SDNode *N,
case ISD::FADD:
return PerformFADDCombine(N, DCI, OptLevel);
case ISD::LOAD:
- case NVPTXISD::LoadParamV2:
case NVPTXISD::LoadV2:
case NVPTXISD::LoadV4:
return combineUnpackingMovIntoLoad(N, DCI);
@@ -5973,6 +5839,8 @@ SDValue NVPTXTargetLowering::PerformDAGCombine(SDNode *N,
return PerformMULCombine(N, DCI, OptLevel);
case NVPTXISD::PRMT:
return combinePRMT(N, DCI, OptLevel);
+ case NVPTXISD::ProxyReg:
+ return combineProxyReg(N, DCI);
case ISD::SETCC:
return PerformSETCCCombine(N, DCI, STI.getSmVersion());
case ISD::SHL:
@@ -5980,10 +5848,6 @@ SDValue NVPTXTargetLowering::PerformDAGCombine(SDNode *N,
case ISD::SREM:
case ISD::UREM:
return PerformREMCombine(N, DCI, OptLevel);
- case NVPTXISD::StoreParam:
- case NVPTXISD::StoreParamV2:
- case NVPTXISD::StoreParamV4:
- return PerformStoreParamCombine(N, DCI);
case ISD::STORE:
case NVPTXISD::StoreV2:
case NVPTXISD::StoreV4:
@@ -6332,6 +6196,22 @@ static void ReplaceCopyFromReg_128(SDNode *N, SelectionDAG &DAG,
Results.push_back(NewValue.getValue(3));
}
+static void replaceProxyReg(SDNode *N, SelectionDAG &DAG,
+ const TargetLowering &TLI,
+ SmallVectorImpl<SDValue> &Results) {
+ SDValue Chain = N->getOperand(0);
+ SDValue Reg = N->getOperand(1);
+
+ MVT VT = TLI.getRegisterType(*DAG.getContext(), Reg.getValueType());
+
+ SDValue NewReg = DAG.getAnyExtOrTrunc(Reg, SDLoc(N), VT);
+ SDValue NewProxy =
+ DAG.getNode(NVPTXISD::ProxyReg, SDLoc(N), VT, {Chain, NewReg});
+ SDValue Res = DAG.getAnyExtOrTrunc(NewProxy, SDLoc(N), N->getValueType(0));
+
+ Results.push_back(Res);
+}
+
void NVPTXTargetLowering::ReplaceNodeResults(
SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
switch (N->getOpcode()) {
@@ -6349,6 +6229,9 @@ void NVPTXTargetLowering::ReplaceNodeResults(
case ISD::CopyFromReg:
ReplaceCopyFromReg_128(N, DAG, Results);
return;
+ case NVPTXISD::ProxyReg:
+ replaceProxyReg(N, DAG, *this, Results);
+ return;
}
}
diff --git a/llvm/lib/Target/NVPTX/NVPTXISelLowering.h b/llvm/lib/Target/NVPTX/NVPTXISelLowering.h
index 228e2aa..cf72a1e 100644
--- a/llvm/lib/Target/NVPTX/NVPTXISelLowering.h
+++ b/llvm/lib/Target/NVPTX/NVPTXISelLowering.h
@@ -38,7 +38,7 @@ enum NodeType : unsigned {
/// This node represents a PTX call instruction. It's operands are as follows:
///
/// CALL(Chain, IsConvergent, IsIndirectCall/IsUniform, NumReturns,
- /// NumParams, Callee, Proto, InGlue)
+ /// NumParams, Callee, Proto)
CALL,
MoveParam,
@@ -84,13 +84,7 @@ enum NodeType : unsigned {
StoreV2,
StoreV4,
StoreV8,
- LoadParam,
- LoadParamV2,
- LoadParamV4,
- StoreParam,
- StoreParamV2,
- StoreParamV4,
- LAST_MEMORY_OPCODE = StoreParamV4,
+ LAST_MEMORY_OPCODE = StoreV8,
};
}
diff --git a/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td b/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td
index 442b900..86d6f7c 100644
--- a/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td
+++ b/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td
@@ -1757,12 +1757,6 @@ def SDTDeclareArrayParam :
SDTypeProfile<0, 3, [SDTCisVT<0, i32>, SDTCisVT<1, i32>, SDTCisVT<2, i32>]>;
def SDTDeclareScalarParam :
SDTypeProfile<0, 2, [SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
-def SDTLoadParamProfile : SDTypeProfile<1, 2, [SDTCisInt<1>, SDTCisInt<2>]>;
-def SDTLoadParamV2Profile : SDTypeProfile<2, 2, [SDTCisSameAs<0, 1>, SDTCisInt<2>, SDTCisInt<3>]>;
-def SDTLoadParamV4Profile : SDTypeProfile<4, 2, [SDTCisInt<4>, SDTCisInt<5>]>;
-def SDTStoreParamProfile : SDTypeProfile<0, 3, [SDTCisInt<0>, SDTCisInt<1>]>;
-def SDTStoreParamV2Profile : SDTypeProfile<0, 4, [SDTCisInt<0>, SDTCisInt<1>]>;
-def SDTStoreParamV4Profile : SDTypeProfile<0, 6, [SDTCisInt<0>, SDTCisInt<1>]>;
def SDTMoveParamProfile : SDTypeProfile<1, 1, [SDTCisInt<0>, SDTCisSameAs<0, 1>]>;
def SDTProxyReg : SDTypeProfile<1, 1, [SDTCisSameAs<0, 1>]>;
@@ -1774,104 +1768,20 @@ def declare_array_param :
def declare_scalar_param :
SDNode<"NVPTXISD::DeclareScalarParam", SDTDeclareScalarParam,
[SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
-
-def LoadParam :
- SDNode<"NVPTXISD::LoadParam", SDTLoadParamProfile,
- [SDNPHasChain, SDNPMayLoad, SDNPOutGlue, SDNPInGlue]>;
-def LoadParamV2 :
- SDNode<"NVPTXISD::LoadParamV2", SDTLoadParamV2Profile,
- [SDNPHasChain, SDNPMayLoad, SDNPOutGlue, SDNPInGlue]>;
-def LoadParamV4 :
- SDNode<"NVPTXISD::LoadParamV4", SDTLoadParamV4Profile,
- [SDNPHasChain, SDNPMayLoad, SDNPOutGlue, SDNPInGlue]>;
-def StoreParam :
- SDNode<"NVPTXISD::StoreParam", SDTStoreParamProfile,
- [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
-def StoreParamV2 :
- SDNode<"NVPTXISD::StoreParamV2", SDTStoreParamV2Profile,
- [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
-def StoreParamV4 :
- SDNode<"NVPTXISD::StoreParamV4", SDTStoreParamV4Profile,
- [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
def MoveParam :
SDNode<"NVPTXISD::MoveParam", SDTMoveParamProfile, []>;
def proxy_reg :
SDNode<"NVPTXISD::ProxyReg", SDTProxyReg, [SDNPHasChain]>;
/// CALL(Chain, IsConvergent, IsIndirectCall/IsUniform, NumReturns,
- /// NumParams, Callee, Proto, InGlue)
+ /// NumParams, Callee, Proto)
def SDTCallProfile : SDTypeProfile<0, 6,
[SDTCisVT<0, i32>, SDTCisVT<1, i32>, SDTCisVT<2, i32>,
SDTCisVT<3, i32>, SDTCisVT<5, i32>]>;
-def call :
- SDNode<"NVPTXISD::CALL", SDTCallProfile,
- [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
-
-let mayLoad = true in {
- class LoadParamMemInst<NVPTXRegClass regclass, string opstr> :
- NVPTXInst<(outs regclass:$dst), (ins Offseti32imm:$b),
- !strconcat("ld.param", opstr, " \t$dst, [retval0$b];"),
- []>;
-
- class LoadParamV2MemInst<NVPTXRegClass regclass, string opstr> :
- NVPTXInst<(outs regclass:$dst, regclass:$dst2), (ins Offseti32imm:$b),
- !strconcat("ld.param.v2", opstr,
- " \t{{$dst, $dst2}}, [retval0$b];"), []>;
-
- class LoadParamV4MemInst<NVPTXRegClass regclass, string opstr> :
- NVPTXInst<(outs regclass:$dst, regclass:$dst2, regclass:$dst3,
- regclass:$dst4),
- (ins Offseti32imm:$b),
- !strconcat("ld.param.v4", opstr,
- " \t{{$dst, $dst2, $dst3, $dst4}}, [retval0$b];"),
- []>;
-}
-
-let mayStore = true in {
-
- multiclass StoreParamInst<NVPTXRegClass regclass, Operand IMMType, string opstr, bit support_imm = true> {
- foreach op = [IMMType, regclass] in
- if !or(support_imm, !isa<NVPTXRegClass>(op)) then
- def _ # !if(!isa<NVPTXRegClass>(op), "r", "i")
- : NVPTXInst<(outs),
- (ins op:$val, i32imm:$a, Offseti32imm:$b),
- "st.param" # opstr # " \t[param$a$b], $val;",
- []>;
- }
-
- multiclass StoreParamV2Inst<NVPTXRegClass regclass, Operand IMMType, string opstr> {
- foreach op1 = [IMMType, regclass] in
- foreach op2 = [IMMType, regclass] in
- def _ # !if(!isa<NVPTXRegClass>(op1), "r", "i")
- # !if(!isa<NVPTXRegClass>(op2), "r", "i")
- : NVPTXInst<(outs),
- (ins op1:$val1, op2:$val2,
- i32imm:$a, Offseti32imm:$b),
- "st.param.v2" # opstr # " \t[param$a$b], {{$val1, $val2}};",
- []>;
- }
-
- multiclass StoreParamV4Inst<NVPTXRegClass regclass, Operand IMMType, string opstr> {
- foreach op1 = [IMMType, regclass] in
- foreach op2 = [IMMType, regclass] in
- foreach op3 = [IMMType, regclass] in
- foreach op4 = [IMMType, regclass] in
- def _ # !if(!isa<NVPTXRegClass>(op1), "r", "i")
- # !if(!isa<NVPTXRegClass>(op2), "r", "i")
- # !if(!isa<NVPTXRegClass>(op3), "r", "i")
- # !if(!isa<NVPTXRegClass>(op4), "r", "i")
-
- : NVPTXInst<(outs),
- (ins op1:$val1, op2:$val2, op3:$val3, op4:$val4,
- i32imm:$a, Offseti32imm:$b),
- "st.param.v4" # opstr #
- " \t[param$a$b], {{$val1, $val2, $val3, $val4}};",
- []>;
- }
-}
+def call : SDNode<"NVPTXISD::CALL", SDTCallProfile, [SDNPHasChain, SDNPSideEffect]>;
/// CALL(Chain, IsConvergent, IsIndirectCall/IsUniform, NumReturns,
-/// NumParams, Callee, Proto, InGlue)
+/// NumParams, Callee, Proto)
def CallOperand : Operand<i32> { let PrintMethod = "printCallOperand"; }
@@ -1908,43 +1818,6 @@ foreach is_convergent = [0, 1] in {
(call_uni_inst $addr, imm:$rets, imm:$params)>;
}
-def LoadParamMemI64 : LoadParamMemInst<B64, ".b64">;
-def LoadParamMemI32 : LoadParamMemInst<B32, ".b32">;
-def LoadParamMemI16 : LoadParamMemInst<B16, ".b16">;
-def LoadParamMemI8 : LoadParamMemInst<B16, ".b8">;
-def LoadParamMemV2I64 : LoadParamV2MemInst<B64, ".b64">;
-def LoadParamMemV2I32 : LoadParamV2MemInst<B32, ".b32">;
-def LoadParamMemV2I16 : LoadParamV2MemInst<B16, ".b16">;
-def LoadParamMemV2I8 : LoadParamV2MemInst<B16, ".b8">;
-def LoadParamMemV4I32 : LoadParamV4MemInst<B32, ".b32">;
-def LoadParamMemV4I16 : LoadParamV4MemInst<B16, ".b16">;
-def LoadParamMemV4I8 : LoadParamV4MemInst<B16, ".b8">;
-
-defm StoreParamI64 : StoreParamInst<B64, i64imm, ".b64">;
-defm StoreParamI32 : StoreParamInst<B32, i32imm, ".b32">;
-defm StoreParamI16 : StoreParamInst<B16, i16imm, ".b16">;
-defm StoreParamI8 : StoreParamInst<B16, i8imm, ".b8">;
-
-defm StoreParamI8TruncI32 : StoreParamInst<B32, i8imm, ".b8", /* support_imm */ false>;
-defm StoreParamI8TruncI64 : StoreParamInst<B64, i8imm, ".b8", /* support_imm */ false>;
-
-defm StoreParamV2I64 : StoreParamV2Inst<B64, i64imm, ".b64">;
-defm StoreParamV2I32 : StoreParamV2Inst<B32, i32imm, ".b32">;
-defm StoreParamV2I16 : StoreParamV2Inst<B16, i16imm, ".b16">;
-defm StoreParamV2I8 : StoreParamV2Inst<B16, i8imm, ".b8">;
-
-defm StoreParamV4I32 : StoreParamV4Inst<B32, i32imm, ".b32">;
-defm StoreParamV4I16 : StoreParamV4Inst<B16, i16imm, ".b16">;
-defm StoreParamV4I8 : StoreParamV4Inst<B16, i8imm, ".b8">;
-
-defm StoreParamF32 : StoreParamInst<B32, f32imm, ".b32">;
-defm StoreParamF64 : StoreParamInst<B64, f64imm, ".b64">;
-
-defm StoreParamV2F32 : StoreParamV2Inst<B32, f32imm, ".b32">;
-defm StoreParamV2F64 : StoreParamV2Inst<B64, f64imm, ".b64">;
-
-defm StoreParamV4F32 : StoreParamV4Inst<B32, f32imm, ".b32">;
-
def DECLARE_PARAM_array :
NVPTXInst<(outs), (ins i32imm:$a, i32imm:$align, i32imm:$size),
".param .align $align .b8 \t$a[$size];", []>;
diff --git a/llvm/lib/Target/RISCV/Disassembler/RISCVDisassembler.cpp b/llvm/lib/Target/RISCV/Disassembler/RISCVDisassembler.cpp
index 5e54b82..67cc01e 100644
--- a/llvm/lib/Target/RISCV/Disassembler/RISCVDisassembler.cpp
+++ b/llvm/lib/Target/RISCV/Disassembler/RISCVDisassembler.cpp
@@ -534,16 +534,26 @@ static DecodeStatus decodeRTZArg(MCInst &Inst, uint32_t Imm, int64_t Address,
return MCDisassembler::Success;
}
-static DecodeStatus decodeXTHeadMemPair(MCInst &Inst, uint32_t Insn,
- uint64_t Address,
- const MCDisassembler *Decoder);
-
static DecodeStatus decodeZcmpRlist(MCInst &Inst, uint32_t Imm,
uint64_t Address,
- const MCDisassembler *Decoder);
+ const MCDisassembler *Decoder) {
+ bool IsRVE = Decoder->getSubtargetInfo().hasFeature(RISCV::FeatureStdExtE);
+ if (Imm < RISCVZC::RA || (IsRVE && Imm >= RISCVZC::RA_S0_S2))
+ return MCDisassembler::Fail;
+ Inst.addOperand(MCOperand::createImm(Imm));
+ return MCDisassembler::Success;
+}
static DecodeStatus decodeXqccmpRlistS0(MCInst &Inst, uint32_t Imm,
uint64_t Address,
+ const MCDisassembler *Decoder) {
+ if (Imm < RISCVZC::RA_S0)
+ return MCDisassembler::Fail;
+ return decodeZcmpRlist(Inst, Imm, Address, Decoder);
+}
+
+static DecodeStatus decodeXTHeadMemPair(MCInst &Inst, uint32_t Insn,
+ uint64_t Address,
const MCDisassembler *Decoder);
static DecodeStatus decodeCSSPushPopchk(MCInst &Inst, uint32_t Insn,
@@ -592,24 +602,6 @@ static DecodeStatus decodeXTHeadMemPair(MCInst &Inst, uint32_t Insn,
return S;
}
-static DecodeStatus decodeZcmpRlist(MCInst &Inst, uint32_t Imm,
- uint64_t Address,
- const MCDisassembler *Decoder) {
- bool IsRVE = Decoder->getSubtargetInfo().hasFeature(RISCV::FeatureStdExtE);
- if (Imm < RISCVZC::RA || (IsRVE && Imm >= RISCVZC::RA_S0_S2))
- return MCDisassembler::Fail;
- Inst.addOperand(MCOperand::createImm(Imm));
- return MCDisassembler::Success;
-}
-
-static DecodeStatus decodeXqccmpRlistS0(MCInst &Inst, uint32_t Imm,
- uint64_t Address,
- const MCDisassembler *Decoder) {
- if (Imm < RISCVZC::RA_S0)
- return MCDisassembler::Fail;
- return decodeZcmpRlist(Inst, Imm, Address, Decoder);
-}
-
// Add implied SP operand for C.*SP compressed instructions. The SP operand
// isn't explicitly encoded in the instruction.
void RISCVDisassembler::addSPOperands(MCInst &MI) const {
diff --git a/llvm/lib/Target/RISCV/MCTargetDesc/RISCVAsmBackend.cpp b/llvm/lib/Target/RISCV/MCTargetDesc/RISCVAsmBackend.cpp
index 2c37c3b..82e3b5c 100644
--- a/llvm/lib/Target/RISCV/MCTargetDesc/RISCVAsmBackend.cpp
+++ b/llvm/lib/Target/RISCV/MCTargetDesc/RISCVAsmBackend.cpp
@@ -320,6 +320,7 @@ bool RISCVAsmBackend::relaxAlign(MCFragment &F, unsigned &Size) {
MCFixup Fixup =
MCFixup::create(0, Expr, FirstLiteralRelocationKind + ELF::R_RISCV_ALIGN);
F.setVarFixups({Fixup});
+ F.setLinkerRelaxable();
F.getParent()->setLinkerRelaxable();
return true;
}
diff --git a/llvm/lib/Target/RISCV/RISCVFrameLowering.cpp b/llvm/lib/Target/RISCV/RISCVFrameLowering.cpp
index b1ab76a..9fc0d81 100644
--- a/llvm/lib/Target/RISCV/RISCVFrameLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVFrameLowering.cpp
@@ -1581,7 +1581,8 @@ void RISCVFrameLowering::determineCalleeSaves(MachineFunction &MF,
// Set the register and all its subregisters.
if (!MRI.def_empty(CSReg) || MRI.getUsedPhysRegsMask().test(CSReg)) {
SavedRegs.set(CSReg);
- llvm::for_each(SubRegs, [&](unsigned Reg) { return SavedRegs.set(Reg); });
+ for (unsigned Reg : SubRegs)
+ SavedRegs.set(Reg);
}
// Combine to super register if all of its subregisters are marked.
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index 54845e5..607edd3 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -2739,6 +2739,27 @@ bool RISCVTargetLowering::isLegalElementTypeForRVV(EVT ScalarTy) const {
}
}
+bool RISCVTargetLowering::isLegalLoadStoreElementTypeForRVV(
+ EVT ScalarTy) const {
+ if (!ScalarTy.isSimple())
+ return false;
+ switch (ScalarTy.getSimpleVT().SimpleTy) {
+ case MVT::iPTR:
+ return Subtarget.is64Bit() ? Subtarget.hasVInstructionsI64() : true;
+ case MVT::i8:
+ case MVT::i16:
+ case MVT::i32:
+ case MVT::f16:
+ case MVT::bf16:
+ case MVT::f32:
+ return true;
+ case MVT::i64:
+ case MVT::f64:
+ return Subtarget.hasVInstructionsI64();
+ default:
+ return false;
+ }
+}
unsigned RISCVTargetLowering::combineRepeatedFPDivisors() const {
return NumRepeatedDivisors;
@@ -24239,7 +24260,7 @@ bool RISCVTargetLowering::isLegalStridedLoadStore(EVT DataType,
return false;
EVT ScalarType = DataType.getScalarType();
- if (!isLegalElementTypeForRVV(ScalarType))
+ if (!isLegalLoadStoreElementTypeForRVV(ScalarType))
return false;
if (!Subtarget.enableUnalignedVectorMem() &&
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.h b/llvm/lib/Target/RISCV/RISCVISelLowering.h
index ca70c46..a788c0b7 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.h
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.h
@@ -384,6 +384,7 @@ public:
bool shouldRemoveExtendFromGSIndex(SDValue Extend, EVT DataVT) const override;
bool isLegalElementTypeForRVV(EVT ScalarTy) const;
+ bool isLegalLoadStoreElementTypeForRVV(EVT ScalarTy) const;
bool shouldConvertFpToSat(unsigned Op, EVT FPVT, EVT VT) const override;
diff --git a/llvm/lib/Target/RISCV/RISCVInterleavedAccess.cpp b/llvm/lib/Target/RISCV/RISCVInterleavedAccess.cpp
index 30d8f85..3cbe668 100644
--- a/llvm/lib/Target/RISCV/RISCVInterleavedAccess.cpp
+++ b/llvm/lib/Target/RISCV/RISCVInterleavedAccess.cpp
@@ -32,7 +32,7 @@ bool RISCVTargetLowering::isLegalInterleavedAccessType(
if (!isTypeLegal(VT))
return false;
- if (!isLegalElementTypeForRVV(VT.getScalarType()) ||
+ if (!isLegalLoadStoreElementTypeForRVV(VT.getScalarType()) ||
!allowsMemoryAccessForAlignment(VTy->getContext(), DL, VT, AddrSpace,
Alignment))
return false;
diff --git a/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.h b/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.h
index d62d99c..f0510ec 100644
--- a/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.h
+++ b/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.h
@@ -265,7 +265,7 @@ public:
if (!ST->enableUnalignedVectorMem() && Alignment < ElemType.getStoreSize())
return false;
- return TLI->isLegalElementTypeForRVV(ElemType);
+ return TLI->isLegalLoadStoreElementTypeForRVV(ElemType);
}
bool isLegalMaskedLoad(Type *DataType, Align Alignment,
@@ -297,7 +297,7 @@ public:
if (!ST->enableUnalignedVectorMem() && Alignment < ElemType.getStoreSize())
return false;
- return TLI->isLegalElementTypeForRVV(ElemType);
+ return TLI->isLegalLoadStoreElementTypeForRVV(ElemType);
}
bool isLegalMaskedGather(Type *DataType, Align Alignment) const override {
diff --git a/llvm/lib/Target/SystemZ/MCTargetDesc/SystemZHLASMAsmStreamer.cpp b/llvm/lib/Target/SystemZ/MCTargetDesc/SystemZHLASMAsmStreamer.cpp
index 3ef6030..72bb372 100644
--- a/llvm/lib/Target/SystemZ/MCTargetDesc/SystemZHLASMAsmStreamer.cpp
+++ b/llvm/lib/Target/SystemZ/MCTargetDesc/SystemZHLASMAsmStreamer.cpp
@@ -69,8 +69,8 @@ void SystemZHLASMAsmStreamer::EmitEOL() {
void SystemZHLASMAsmStreamer::changeSection(MCSection *Section,
uint32_t Subsection) {
- Section->printSwitchToSection(*MAI, getContext().getTargetTriple(), OS,
- Subsection);
+ MAI->printSwitchToSection(*Section, Subsection,
+ getContext().getTargetTriple(), OS);
MCStreamer::changeSection(Section, Subsection);
}
diff --git a/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td b/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
index d13862f..143298b 100644
--- a/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
+++ b/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
@@ -1540,6 +1540,8 @@ multiclass SIMDMADD<Vec vec, bits<32> simdopA, bits<32> simdopS, list<Predicate>
def : Pat<(fadd_contract (vec.vt V128:$a), (fmul_contract (vec.vt V128:$b), (vec.vt V128:$c))),
(!cast<Instruction>("MADD_"#vec) V128:$a, V128:$b, V128:$c)>, Requires<[HasRelaxedSIMD]>;
+ def : Pat<(fsub_contract (vec.vt V128:$a), (fmul_contract (vec.vt V128:$b), (vec.vt V128:$c))),
+ (!cast<Instruction>("NMADD_"#vec) V128:$a, V128:$b, V128:$c)>, Requires<[HasRelaxedSIMD]>;
}
defm "" : SIMDMADD<F32x4, 0x105, 0x106, [HasRelaxedSIMD]>;