blob: 54177564afbc7fc4399c7670a56f9e02849e99d3 [file] [log] [blame]
//===-- AMDGPUISelLowering.cpp - AMDGPU Common DAG lowering functions -----===//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
/// \file
/// This is the parent TargetLowering class for hardware code gen
/// targets.
#include "AMDGPUISelLowering.h"
#include "AMDGPU.h"
#include "AMDGPUInstrInfo.h"
#include "AMDGPUMachineFunction.h"
#include "GCNSubtarget.h"
#include "SIMachineFunctionInfo.h"
#include "llvm/CodeGen/Analysis.h"
#include "llvm/IR/DiagnosticInfo.h"
#include "llvm/IR/IntrinsicsAMDGPU.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Target/TargetMachine.h"
using namespace llvm;
#include ""
static cl::opt<bool> AMDGPUBypassSlowDiv(
cl::desc("Skip 64-bit divide for dynamic 32-bit values"),
// Find a larger type to do a load / store of a vector with.
EVT AMDGPUTargetLowering::getEquivalentMemType(LLVMContext &Ctx, EVT VT) {
unsigned StoreSize = VT.getStoreSizeInBits();
if (StoreSize <= 32)
return EVT::getIntegerVT(Ctx, StoreSize);
assert(StoreSize % 32 == 0 && "Store size not a multiple of 32");
return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
unsigned AMDGPUTargetLowering::numBitsUnsigned(SDValue Op, SelectionDAG &DAG) {
return DAG.computeKnownBits(Op).countMaxActiveBits();
unsigned AMDGPUTargetLowering::numBitsSigned(SDValue Op, SelectionDAG &DAG) {
// In order for this to be a signed 24-bit value, bit 23, must
// be a sign bit.
return DAG.ComputeMinSignedBits(Op);
AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
const AMDGPUSubtarget &STI)
: TargetLowering(TM), Subtarget(&STI) {
// Lower floating point store/load to integer store/load to reduce the number
// of patterns in tablegen.
setOperationAction(ISD::LOAD, MVT::f32, Promote);
AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
setOperationAction(ISD::LOAD, MVT::v3f32, Promote);
AddPromotedToType(ISD::LOAD, MVT::v3f32, MVT::v3i32);
setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
setOperationAction(ISD::LOAD, MVT::v5f32, Promote);
AddPromotedToType(ISD::LOAD, MVT::v5f32, MVT::v5i32);
setOperationAction(ISD::LOAD, MVT::v6f32, Promote);
AddPromotedToType(ISD::LOAD, MVT::v6f32, MVT::v6i32);
setOperationAction(ISD::LOAD, MVT::v7f32, Promote);
AddPromotedToType(ISD::LOAD, MVT::v7f32, MVT::v7i32);
setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
setOperationAction(ISD::LOAD, MVT::v32f32, Promote);
AddPromotedToType(ISD::LOAD, MVT::v32f32, MVT::v32i32);
setOperationAction(ISD::LOAD, MVT::i64, Promote);
AddPromotedToType(ISD::LOAD, MVT::i64, MVT::v2i32);
setOperationAction(ISD::LOAD, MVT::v2i64, Promote);
AddPromotedToType(ISD::LOAD, MVT::v2i64, MVT::v4i32);
setOperationAction(ISD::LOAD, MVT::f64, Promote);
AddPromotedToType(ISD::LOAD, MVT::f64, MVT::v2i32);
setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v4i32);
setOperationAction(ISD::LOAD, MVT::v3i64, Promote);
AddPromotedToType(ISD::LOAD, MVT::v3i64, MVT::v6i32);
setOperationAction(ISD::LOAD, MVT::v4i64, Promote);
AddPromotedToType(ISD::LOAD, MVT::v4i64, MVT::v8i32);
setOperationAction(ISD::LOAD, MVT::v3f64, Promote);
AddPromotedToType(ISD::LOAD, MVT::v3f64, MVT::v6i32);
setOperationAction(ISD::LOAD, MVT::v4f64, Promote);
AddPromotedToType(ISD::LOAD, MVT::v4f64, MVT::v8i32);
setOperationAction(ISD::LOAD, MVT::v8i64, Promote);
AddPromotedToType(ISD::LOAD, MVT::v8i64, MVT::v16i32);
setOperationAction(ISD::LOAD, MVT::v8f64, Promote);
AddPromotedToType(ISD::LOAD, MVT::v8f64, MVT::v16i32);
setOperationAction(ISD::LOAD, MVT::v16i64, Promote);
AddPromotedToType(ISD::LOAD, MVT::v16i64, MVT::v32i32);
setOperationAction(ISD::LOAD, MVT::v16f64, Promote);
AddPromotedToType(ISD::LOAD, MVT::v16f64, MVT::v32i32);
// There are no 64-bit extloads. These should be done as a 32-bit extload and
// an extension to 64-bit.
for (MVT VT : MVT::integer_valuetypes()) {
setLoadExtAction(ISD::EXTLOAD, MVT::i64, VT, Expand);
setLoadExtAction(ISD::SEXTLOAD, MVT::i64, VT, Expand);
setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, VT, Expand);
for (MVT VT : MVT::integer_valuetypes()) {
if (VT == MVT::i64)
setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i8, Legal);
setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i16, Legal);
setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i32, Expand);
setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i8, Legal);
setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i16, Legal);
setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i32, Expand);
setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote);
setLoadExtAction(ISD::EXTLOAD, VT, MVT::i8, Legal);
setLoadExtAction(ISD::EXTLOAD, VT, MVT::i16, Legal);
setLoadExtAction(ISD::EXTLOAD, VT, MVT::i32, Expand);
for (MVT VT : MVT::integer_fixedlen_vector_valuetypes()) {
setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Expand);
setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Expand);
setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i8, Expand);
setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Expand);
setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Expand);
setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i8, Expand);
setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Expand);
setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Expand);
setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i16, Expand);
setLoadExtAction(ISD::EXTLOAD, VT, MVT::v3i16, Expand);
setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v3i16, Expand);
setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v3i16, Expand);
setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Expand);
setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Expand);
setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i16, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::v2f32, MVT::v2f16, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::v3f32, MVT::v3f16, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, MVT::v4f16, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::v8f32, MVT::v8f16, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::v16f32, MVT::v16f16, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::v32f32, MVT::v32f16, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f32, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::v3f64, MVT::v3f32, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f32, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::v8f64, MVT::v8f32, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::v16f64, MVT::v16f32, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f16, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::v3f64, MVT::v3f16, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f16, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::v8f64, MVT::v8f16, Expand);
setLoadExtAction(ISD::EXTLOAD, MVT::v16f64, MVT::v16f16, Expand);
setOperationAction(ISD::STORE, MVT::f32, Promote);
AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
setOperationAction(ISD::STORE, MVT::v2f32, Promote);
AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
setOperationAction(ISD::STORE, MVT::v3f32, Promote);
AddPromotedToType(ISD::STORE, MVT::v3f32, MVT::v3i32);
setOperationAction(ISD::STORE, MVT::v4f32, Promote);
AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
setOperationAction(ISD::STORE, MVT::v5f32, Promote);
AddPromotedToType(ISD::STORE, MVT::v5f32, MVT::v5i32);
setOperationAction(ISD::STORE, MVT::v6f32, Promote);
AddPromotedToType(ISD::STORE, MVT::v6f32, MVT::v6i32);
setOperationAction(ISD::STORE, MVT::v7f32, Promote);
AddPromotedToType(ISD::STORE, MVT::v7f32, MVT::v7i32);
setOperationAction(ISD::STORE, MVT::v8f32, Promote);
AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
setOperationAction(ISD::STORE, MVT::v16f32, Promote);
AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
setOperationAction(ISD::STORE, MVT::v32f32, Promote);
AddPromotedToType(ISD::STORE, MVT::v32f32, MVT::v32i32);
setOperationAction(ISD::STORE, MVT::i64, Promote);
AddPromotedToType(ISD::STORE, MVT::i64, MVT::v2i32);
setOperationAction(ISD::STORE, MVT::v2i64, Promote);
AddPromotedToType(ISD::STORE, MVT::v2i64, MVT::v4i32);
setOperationAction(ISD::STORE, MVT::f64, Promote);
AddPromotedToType(ISD::STORE, MVT::f64, MVT::v2i32);
setOperationAction(ISD::STORE, MVT::v2f64, Promote);
AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v4i32);
setOperationAction(ISD::STORE, MVT::v3i64, Promote);
AddPromotedToType(ISD::STORE, MVT::v3i64, MVT::v6i32);
setOperationAction(ISD::STORE, MVT::v3f64, Promote);
AddPromotedToType(ISD::STORE, MVT::v3f64, MVT::v6i32);
setOperationAction(ISD::STORE, MVT::v4i64, Promote);
AddPromotedToType(ISD::STORE, MVT::v4i64, MVT::v8i32);
setOperationAction(ISD::STORE, MVT::v4f64, Promote);
AddPromotedToType(ISD::STORE, MVT::v4f64, MVT::v8i32);
setOperationAction(ISD::STORE, MVT::v8i64, Promote);
AddPromotedToType(ISD::STORE, MVT::v8i64, MVT::v16i32);
setOperationAction(ISD::STORE, MVT::v8f64, Promote);
AddPromotedToType(ISD::STORE, MVT::v8f64, MVT::v16i32);
setOperationAction(ISD::STORE, MVT::v16i64, Promote);
AddPromotedToType(ISD::STORE, MVT::v16i64, MVT::v32i32);
setOperationAction(ISD::STORE, MVT::v16f64, Promote);
AddPromotedToType(ISD::STORE, MVT::v16f64, MVT::v32i32);
setTruncStoreAction(MVT::i64, MVT::i1, Expand);
setTruncStoreAction(MVT::i64, MVT::i8, Expand);
setTruncStoreAction(MVT::i64, MVT::i16, Expand);
setTruncStoreAction(MVT::i64, MVT::i32, Expand);
setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
setTruncStoreAction(MVT::v2i64, MVT::v2i8, Expand);
setTruncStoreAction(MVT::v2i64, MVT::v2i16, Expand);
setTruncStoreAction(MVT::v2i64, MVT::v2i32, Expand);
setTruncStoreAction(MVT::f32, MVT::f16, Expand);
setTruncStoreAction(MVT::v2f32, MVT::v2f16, Expand);
setTruncStoreAction(MVT::v3f32, MVT::v3f16, Expand);
setTruncStoreAction(MVT::v4f32, MVT::v4f16, Expand);
setTruncStoreAction(MVT::v8f32, MVT::v8f16, Expand);
setTruncStoreAction(MVT::v16f32, MVT::v16f16, Expand);
setTruncStoreAction(MVT::v32f32, MVT::v32f16, Expand);
setTruncStoreAction(MVT::f64, MVT::f16, Expand);
setTruncStoreAction(MVT::f64, MVT::f32, Expand);
setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
setTruncStoreAction(MVT::v2f64, MVT::v2f16, Expand);
setTruncStoreAction(MVT::v3i64, MVT::v3i32, Expand);
setTruncStoreAction(MVT::v3i64, MVT::v3i16, Expand);
setTruncStoreAction(MVT::v3f64, MVT::v3f32, Expand);
setTruncStoreAction(MVT::v3f64, MVT::v3f16, Expand);
setTruncStoreAction(MVT::v4i64, MVT::v4i32, Expand);
setTruncStoreAction(MVT::v4i64, MVT::v4i16, Expand);
setTruncStoreAction(MVT::v4f64, MVT::v4f32, Expand);
setTruncStoreAction(MVT::v4f64, MVT::v4f16, Expand);
setTruncStoreAction(MVT::v8f64, MVT::v8f32, Expand);
setTruncStoreAction(MVT::v8f64, MVT::v8f16, Expand);
setTruncStoreAction(MVT::v16f64, MVT::v16f32, Expand);
setTruncStoreAction(MVT::v16f64, MVT::v16f16, Expand);
setTruncStoreAction(MVT::v16i64, MVT::v16i16, Expand);
setTruncStoreAction(MVT::v16i64, MVT::v16i16, Expand);
setTruncStoreAction(MVT::v16i64, MVT::v16i8, Expand);
setTruncStoreAction(MVT::v16i64, MVT::v16i8, Expand);
setTruncStoreAction(MVT::v16i64, MVT::v16i1, Expand);
setOperationAction(ISD::Constant, MVT::i32, Legal);
setOperationAction(ISD::Constant, MVT::i64, Legal);
setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
setOperationAction(ISD::BR_JT, MVT::Other, Expand);
setOperationAction(ISD::BRIND, MVT::Other, Expand);
// This is totally unsupported, just custom lower to produce an error.
setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
// Library functions. These default to Expand, but we have instructions
// for them.
setOperationAction(ISD::FCEIL, MVT::f32, Legal);
setOperationAction(ISD::FEXP2, MVT::f32, Legal);
setOperationAction(ISD::FPOW, MVT::f32, Legal);
setOperationAction(ISD::FLOG2, MVT::f32, Legal);
setOperationAction(ISD::FABS, MVT::f32, Legal);
setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
setOperationAction(ISD::FRINT, MVT::f32, Legal);
setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
setOperationAction(ISD::FMINNUM, MVT::f32, Legal);
setOperationAction(ISD::FMAXNUM, MVT::f32, Legal);
setOperationAction(ISD::FROUND, MVT::f32, Custom);
setOperationAction(ISD::FROUND, MVT::f64, Custom);
setOperationAction(ISD::FLOG, MVT::f32, Custom);
setOperationAction(ISD::FLOG10, MVT::f32, Custom);
setOperationAction(ISD::FEXP, MVT::f32, Custom);
setOperationAction(ISD::FNEARBYINT, MVT::f32, Custom);
setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom);
setOperationAction(ISD::FREM, MVT::f16, Custom);
setOperationAction(ISD::FREM, MVT::f32, Custom);
setOperationAction(ISD::FREM, MVT::f64, Custom);
// Expand to fneg + fadd.
setOperationAction(ISD::FSUB, MVT::f64, Expand);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v3i32, Custom);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v3f32, Custom);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v5i32, Custom);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v5f32, Custom);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v6i32, Custom);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v6f32, Custom);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v7i32, Custom);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v7f32, Custom);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f16, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i16, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v3f32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v3i32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v5f32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v5i32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v6f32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v6i32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v7f32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v7i32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v16f32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v16i32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v32f32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v32i32, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f64, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i64, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v3f64, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v3i64, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f64, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i64, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f64, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i64, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v16f64, Custom);
setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v16i64, Custom);
setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
setOperationAction(ISD::FP_TO_FP16, MVT::f64, Custom);
setOperationAction(ISD::FP_TO_FP16, MVT::f32, Custom);
const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
for (MVT VT : ScalarIntVTs) {
// These should use [SU]DIVREM, so set them to expand
setOperationAction(ISD::SDIV, VT, Expand);
setOperationAction(ISD::UDIV, VT, Expand);
setOperationAction(ISD::SREM, VT, Expand);
setOperationAction(ISD::UREM, VT, Expand);
// GPU does not have divrem function for signed or unsigned.
setOperationAction(ISD::SDIVREM, VT, Custom);
setOperationAction(ISD::UDIVREM, VT, Custom);
// GPU does not have [S|U]MUL_LOHI functions as a single instruction.
setOperationAction(ISD::SMUL_LOHI, VT, Expand);
setOperationAction(ISD::UMUL_LOHI, VT, Expand);
setOperationAction(ISD::BSWAP, VT, Expand);
setOperationAction(ISD::CTTZ, VT, Expand);
setOperationAction(ISD::CTLZ, VT, Expand);
setOperationAction(ISD::ADDC, VT, Legal);
setOperationAction(ISD::SUBC, VT, Legal);
setOperationAction(ISD::ADDE, VT, Legal);
setOperationAction(ISD::SUBE, VT, Legal);
// The hardware supports 32-bit FSHR, but not FSHL.
setOperationAction(ISD::FSHR, MVT::i32, Legal);
// The hardware supports 32-bit ROTR, but not ROTL.
setOperationAction(ISD::ROTL, MVT::i32, Expand);
setOperationAction(ISD::ROTL, MVT::i64, Expand);
setOperationAction(ISD::ROTR, MVT::i64, Expand);
setOperationAction(ISD::MULHU, MVT::i16, Expand);
setOperationAction(ISD::MULHS, MVT::i16, Expand);
setOperationAction(ISD::MUL, MVT::i64, Expand);
setOperationAction(ISD::MULHU, MVT::i64, Expand);
setOperationAction(ISD::MULHS, MVT::i64, Expand);
setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
setOperationAction(ISD::SMIN, MVT::i32, Legal);
setOperationAction(ISD::UMIN, MVT::i32, Legal);
setOperationAction(ISD::SMAX, MVT::i32, Legal);
setOperationAction(ISD::UMAX, MVT::i32, Legal);
setOperationAction(ISD::CTTZ, MVT::i64, Custom);
setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Custom);
setOperationAction(ISD::CTLZ, MVT::i64, Custom);
setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
static const MVT::SimpleValueType VectorIntTypes[] = {
MVT::v2i32, MVT::v3i32, MVT::v4i32, MVT::v5i32, MVT::v6i32, MVT::v7i32};
for (MVT VT : VectorIntTypes) {
// Expand the following operations for the current type by default.
setOperationAction(ISD::ADD, VT, Expand);
setOperationAction(ISD::AND, VT, Expand);
setOperationAction(ISD::FP_TO_SINT, VT, Expand);
setOperationAction(ISD::FP_TO_UINT, VT, Expand);
setOperationAction(ISD::MUL, VT, Expand);
setOperationAction(ISD::MULHU, VT, Expand);
setOperationAction(ISD::MULHS, VT, Expand);
setOperationAction(ISD::OR, VT, Expand);
setOperationAction(ISD::SHL, VT, Expand);
setOperationAction(ISD::SRA, VT, Expand);
setOperationAction(ISD::SRL, VT, Expand);
setOperationAction(ISD::ROTL, VT, Expand);
setOperationAction(ISD::ROTR, VT, Expand);
setOperationAction(ISD::SUB, VT, Expand);
setOperationAction(ISD::SINT_TO_FP, VT, Expand);
setOperationAction(ISD::UINT_TO_FP, VT, Expand);
setOperationAction(ISD::SDIV, VT, Expand);
setOperationAction(ISD::UDIV, VT, Expand);
setOperationAction(ISD::SREM, VT, Expand);
setOperationAction(ISD::UREM, VT, Expand);
setOperationAction(ISD::SMUL_LOHI, VT, Expand);
setOperationAction(ISD::UMUL_LOHI, VT, Expand);
setOperationAction(ISD::SDIVREM, VT, Expand);
setOperationAction(ISD::UDIVREM, VT, Expand);
setOperationAction(ISD::SELECT, VT, Expand);
setOperationAction(ISD::VSELECT, VT, Expand);
setOperationAction(ISD::SELECT_CC, VT, Expand);
setOperationAction(ISD::XOR, VT, Expand);
setOperationAction(ISD::BSWAP, VT, Expand);
setOperationAction(ISD::CTPOP, VT, Expand);
setOperationAction(ISD::CTTZ, VT, Expand);
setOperationAction(ISD::CTLZ, VT, Expand);
setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
setOperationAction(ISD::SETCC, VT, Expand);
static const MVT::SimpleValueType FloatVectorTypes[] = {
MVT::v2f32, MVT::v3f32, MVT::v4f32, MVT::v5f32, MVT::v6f32, MVT::v7f32};
for (MVT VT : FloatVectorTypes) {
setOperationAction(ISD::FABS, VT, Expand);
setOperationAction(ISD::FMINNUM, VT, Expand);
setOperationAction(ISD::FMAXNUM, VT, Expand);
setOperationAction(ISD::FADD, VT, Expand);
setOperationAction(ISD::FCEIL, VT, Expand);
setOperationAction(ISD::FCOS, VT, Expand);
setOperationAction(ISD::FDIV, VT, Expand);
setOperationAction(ISD::FEXP2, VT, Expand);
setOperationAction(ISD::FEXP, VT, Expand);
setOperationAction(ISD::FLOG2, VT, Expand);
setOperationAction(ISD::FREM, VT, Expand);
setOperationAction(ISD::FLOG, VT, Expand);
setOperationAction(ISD::FLOG10, VT, Expand);
setOperationAction(ISD::FPOW, VT, Expand);
setOperationAction(ISD::FFLOOR, VT, Expand);
setOperationAction(ISD::FTRUNC, VT, Expand);
setOperationAction(ISD::FMUL, VT, Expand);
setOperationAction(ISD::FMA, VT, Expand);
setOperationAction(ISD::FRINT, VT, Expand);
setOperationAction(ISD::FNEARBYINT, VT, Expand);
setOperationAction(ISD::FSQRT, VT, Expand);
setOperationAction(ISD::FSIN, VT, Expand);
setOperationAction(ISD::FSUB, VT, Expand);
setOperationAction(ISD::FNEG, VT, Expand);
setOperationAction(ISD::VSELECT, VT, Expand);
setOperationAction(ISD::SELECT_CC, VT, Expand);
setOperationAction(ISD::FCOPYSIGN, VT, Expand);
setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
setOperationAction(ISD::SETCC, VT, Expand);
setOperationAction(ISD::FCANONICALIZE, VT, Expand);
// This causes using an unrolled select operation rather than expansion with
// bit operations. This is in general better, but the alternative using BFI
// instructions may be better if the select sources are SGPRs.
setOperationAction(ISD::SELECT, MVT::v2f32, Promote);
AddPromotedToType(ISD::SELECT, MVT::v2f32, MVT::v2i32);
setOperationAction(ISD::SELECT, MVT::v3f32, Promote);
AddPromotedToType(ISD::SELECT, MVT::v3f32, MVT::v3i32);
setOperationAction(ISD::SELECT, MVT::v4f32, Promote);
AddPromotedToType(ISD::SELECT, MVT::v4f32, MVT::v4i32);
setOperationAction(ISD::SELECT, MVT::v5f32, Promote);
AddPromotedToType(ISD::SELECT, MVT::v5f32, MVT::v5i32);
setOperationAction(ISD::SELECT, MVT::v6f32, Promote);
AddPromotedToType(ISD::SELECT, MVT::v6f32, MVT::v6i32);
setOperationAction(ISD::SELECT, MVT::v7f32, Promote);
AddPromotedToType(ISD::SELECT, MVT::v7f32, MVT::v7i32);
// There are no libcalls of any kind.
for (int I = 0; I < RTLIB::UNKNOWN_LIBCALL; ++I)
setLibcallName(static_cast<RTLIB::Libcall>(I), nullptr);
// FIXME: This is only partially true. If we have to do vector compares, any
// SGPR pair can be a condition register. If we have a uniform condition, we
// are better off doing SALU operations, where there is only one SCC. For now,
// we don't have a way of knowing during instruction selection if a condition
// will be uniform and we always use vector compares. Assume we are using
// vector compares until that is fixed.
PredictableSelectIsExpensive = false;
// We want to find all load dependencies for long chains of stores to enable
// merging into very wide vectors. The problem is with vectors with > 4
// elements. MergeConsecutiveStores will attempt to merge these because x8/x16
// vectors are a legal type, even though we have to split the loads
// usually. When we can more precisely specify load legality per address
// space, we should be able to make FindBetterChain/MergeConsecutiveStores
// smarter so that they can figure out what to do in 2 iterations without all
// N > 4 stores on the same chain.
GatherAllAliasesMaxDepth = 16;
// memcpy/memmove/memset are expanded in the IR, so we shouldn't need to worry
// about these during lowering.
MaxStoresPerMemcpy = 0xffffffff;
MaxStoresPerMemmove = 0xffffffff;
MaxStoresPerMemset = 0xffffffff;
// The expansion for 64-bit division is enormous.
if (AMDGPUBypassSlowDiv)
addBypassSlowDiv(64, 32);
bool AMDGPUTargetLowering::mayIgnoreSignedZero(SDValue Op) const {
if (getTargetMachine().Options.NoSignedZerosFPMath)
return true;
const auto Flags = Op.getNode()->getFlags();
if (Flags.hasNoSignedZeros())
return true;
return false;
// Target Information
static bool fnegFoldsIntoOp(unsigned Opc) {
switch (Opc) {
case ISD::FADD:
case ISD::FSUB:
case ISD::FMUL:
case ISD::FMA:
case ISD::FMAD:
case ISD::FSIN:
case ISD::FRINT:
// TODO: handle llvm.amdgcn.fma.legacy
return true;
return false;
/// \p returns true if the operation will definitely need to use a 64-bit
/// encoding, and thus will use a VOP3 encoding regardless of the source
/// modifiers.
static bool opMustUseVOP3Encoding(const SDNode *N, MVT VT) {
return N->getNumOperands() > 2 || VT == MVT::f64;
// Most FP instructions support source modifiers, but this could be refined
// slightly.
static bool hasSourceMods(const SDNode *N) {
if (isa<MemSDNode>(N))
return false;
switch (N->getOpcode()) {
case ISD::CopyToReg:
case ISD::FDIV:
case ISD::FREM:
// TODO: Should really be looking at the users of the bitcast. These are
// problematic because bitcasts are used to legalize all stores to integer
// types.
return false;
switch (cast<ConstantSDNode>(N->getOperand(0))->getZExtValue()) {
case Intrinsic::amdgcn_interp_p1:
case Intrinsic::amdgcn_interp_p2:
case Intrinsic::amdgcn_interp_mov:
case Intrinsic::amdgcn_interp_p1_f16:
case Intrinsic::amdgcn_interp_p2_f16:
return false;
return true;
return true;
bool AMDGPUTargetLowering::allUsesHaveSourceMods(const SDNode *N,
unsigned CostThreshold) {
// Some users (such as 3-operand FMA/MAD) must use a VOP3 encoding, and thus
// it is truly free to use a source modifier in all cases. If there are
// multiple users but for each one will necessitate using VOP3, there will be
// a code size increase. Try to avoid increasing code size unless we know it
// will save on the instruction count.
unsigned NumMayIncreaseSize = 0;
MVT VT = N->getValueType(0).getScalarType().getSimpleVT();
// XXX - Should this limit number of uses to check?
for (const SDNode *U : N->uses()) {
if (!hasSourceMods(U))
return false;
if (!opMustUseVOP3Encoding(U, VT)) {
if (++NumMayIncreaseSize > CostThreshold)
return false;
return true;
EVT AMDGPUTargetLowering::getTypeForExtReturn(LLVMContext &Context, EVT VT,
ISD::NodeType ExtendKind) const {
assert(!VT.isVector() && "only scalar expected");
// Round to the next multiple of 32-bits.
unsigned Size = VT.getSizeInBits();
if (Size <= 32)
return MVT::i32;
return EVT::getIntegerVT(Context, 32 * ((Size + 31) / 32));
MVT AMDGPUTargetLowering::getVectorIdxTy(const DataLayout &) const {
return MVT::i32;
bool AMDGPUTargetLowering::isSelectSupported(SelectSupportKind SelType) const {
return true;
// The backend supports 32 and 64 bit floating point immediates.
// FIXME: Why are we reporting vectors of FP immediates as legal?
bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT,
bool ForCodeSize) const {
EVT ScalarVT = VT.getScalarType();
return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64 ||
(ScalarVT == MVT::f16 && Subtarget->has16BitInsts()));
// We don't want to shrink f64 / f32 constants.
bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
EVT ScalarVT = VT.getScalarType();
return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
bool AMDGPUTargetLowering::shouldReduceLoadWidth(SDNode *N,
ISD::LoadExtType ExtTy,
EVT NewVT) const {
// TODO: This may be worth removing. Check regression tests for diffs.
if (!TargetLoweringBase::shouldReduceLoadWidth(N, ExtTy, NewVT))
return false;
unsigned NewSize = NewVT.getStoreSizeInBits();
// If we are reducing to a 32-bit load or a smaller multi-dword load,
// this is always better.
if (NewSize >= 32)
return true;
EVT OldVT = N->getValueType(0);
unsigned OldSize = OldVT.getStoreSizeInBits();
MemSDNode *MN = cast<MemSDNode>(N);
unsigned AS = MN->getAddressSpace();
// Do not shrink an aligned scalar load to sub-dword.
// Scalar engine cannot do sub-dword loads.
if (OldSize >= 32 && NewSize < 32 && MN->getAlignment() >= 4 &&
(isa<LoadSDNode>(N) &&
AS == AMDGPUAS::GLOBAL_ADDRESS && MN->isInvariant())) &&
return false;
// Don't produce extloads from sub 32-bit types. SI doesn't have scalar
// extloads, so doing one requires using a buffer_load. In cases where we
// still couldn't use a scalar load, using the wider load shouldn't really
// hurt anything.
// If the old size already had to be an extload, there's no harm in continuing
// to reduce the width.
return (OldSize < 32);
bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy, EVT CastTy,
const SelectionDAG &DAG,
const MachineMemOperand &MMO) const {
assert(LoadTy.getSizeInBits() == CastTy.getSizeInBits());
if (LoadTy.getScalarType() == MVT::i32)
return false;
unsigned LScalarSize = LoadTy.getScalarSizeInBits();
unsigned CastScalarSize = CastTy.getScalarSizeInBits();
if ((LScalarSize >= CastScalarSize) && (CastScalarSize < 32))
return false;
bool Fast = false;
return allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
CastTy, MMO, &Fast) &&
// SI+ has instructions for cttz / ctlz for 32-bit values. This is probably also
// profitable with the expansion for 64-bit since it's generally good to
// speculate things.
// FIXME: These should really have the size as a parameter.
bool AMDGPUTargetLowering::isCheapToSpeculateCttz() const {
return true;
bool AMDGPUTargetLowering::isCheapToSpeculateCtlz() const {
return true;
bool AMDGPUTargetLowering::isSDNodeAlwaysUniform(const SDNode *N) const {
switch (N->getOpcode()) {
case ISD::EntryToken:
case ISD::TokenFactor:
return true;
unsigned IntrID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
switch (IntrID) {
case Intrinsic::amdgcn_readfirstlane:
case Intrinsic::amdgcn_readlane:
return true;
return false;
case ISD::LOAD:
if (cast<LoadSDNode>(N)->getMemOperand()->getAddrSpace() ==
return true;
return false;
return false;
SDValue AMDGPUTargetLowering::getNegatedExpression(
SDValue Op, SelectionDAG &DAG, bool LegalOperations, bool ForCodeSize,
NegatibleCost &Cost, unsigned Depth) const {
switch (Op.getOpcode()) {
case ISD::FMA:
case ISD::FMAD: {
// Negating a fma is not free if it has users without source mods.
if (!allUsesHaveSourceMods(Op.getNode()))
return SDValue();
return TargetLowering::getNegatedExpression(Op, DAG, LegalOperations,
ForCodeSize, Cost, Depth);
// Target Properties
bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
// Packed operations do not have a fabs modifier.
return VT == MVT::f32 || VT == MVT::f64 ||
(Subtarget->has16BitInsts() && VT == MVT::f16);
bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
// Report this based on the end legalized type.
VT = VT.getScalarType();
return VT == MVT::f32 || VT == MVT::f64 || VT == MVT::f16;
bool AMDGPUTargetLowering:: storeOfVectorConstantIsCheap(EVT MemVT,
unsigned NumElem,
unsigned AS) const {
return true;
bool AMDGPUTargetLowering::aggressivelyPreferBuildVectorSources(EVT VecVT) const {
// There are few operations which truly have vector input operands. Any vector
// operation is going to involve operations on each component, and a
// build_vector will be a copy per element, so it always makes sense to use a
// build_vector input in place of the extracted element to avoid a copy into a
// super register.
// We should probably only do this if all users are extracts only, but this
// should be the common case.
return true;
bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
// Truncate is just accessing a subregister.
unsigned SrcSize = Source.getSizeInBits();
unsigned DestSize = Dest.getSizeInBits();
return DestSize < SrcSize && DestSize % 32 == 0 ;
bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
// Truncate is just accessing a subregister.
unsigned SrcSize = Source->getScalarSizeInBits();
unsigned DestSize = Dest->getScalarSizeInBits();
if (DestSize== 16 && Subtarget->has16BitInsts())
return SrcSize >= 32;
return DestSize < SrcSize && DestSize % 32 == 0;
bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
unsigned SrcSize = Src->getScalarSizeInBits();
unsigned DestSize = Dest->getScalarSizeInBits();
if (SrcSize == 16 && Subtarget->has16BitInsts())
return DestSize >= 32;
return SrcSize == 32 && DestSize == 64;
bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
// Any register load of a 64-bit value really requires 2 32-bit moves. For all
// practical purposes, the extra mov 0 to load a 64-bit is free. As used,
// this will enable reducing 64-bit operations the 32-bit, which is always
// good.
if (Src == MVT::i16)
return Dest == MVT::i32 ||Dest == MVT::i64 ;
return Src == MVT::i32 && Dest == MVT::i64;
bool AMDGPUTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
return isZExtFree(Val.getValueType(), VT2);
bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
// There aren't really 64-bit registers, but pairs of 32-bit ones and only a
// limited number of native 64-bit operations. Shrinking an operation to fit
// in a single 32-bit register should always be helpful. As currently used,
// this is much less general than the name suggests, and is only used in
// places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
// not profitable, and may actually be harmful.
return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
// TargetLowering Callbacks
CCAssignFn *AMDGPUCallLowering::CCAssignFnForCall(CallingConv::ID CC,
bool IsVarArg) {
switch (CC) {
case CallingConv::AMDGPU_VS:
case CallingConv::AMDGPU_GS:
case CallingConv::AMDGPU_PS:
case CallingConv::AMDGPU_CS:
case CallingConv::AMDGPU_HS:
case CallingConv::AMDGPU_ES:
case CallingConv::AMDGPU_LS:
return CC_AMDGPU;
case CallingConv::C:
case CallingConv::Fast:
case CallingConv::Cold:
return CC_AMDGPU_Func;
case CallingConv::AMDGPU_Gfx:
return CC_SI_Gfx;
case CallingConv::AMDGPU_KERNEL:
case CallingConv::SPIR_KERNEL:
report_fatal_error("Unsupported calling convention for call");
CCAssignFn *AMDGPUCallLowering::CCAssignFnForReturn(CallingConv::ID CC,
bool IsVarArg) {
switch (CC) {
case CallingConv::AMDGPU_KERNEL:
case CallingConv::SPIR_KERNEL:
llvm_unreachable("kernels should not be handled here");
case CallingConv::AMDGPU_VS:
case CallingConv::AMDGPU_GS:
case CallingConv::AMDGPU_PS:
case CallingConv::AMDGPU_CS:
case CallingConv::AMDGPU_HS:
case CallingConv::AMDGPU_ES:
case CallingConv::AMDGPU_LS:
return RetCC_SI_Shader;
case CallingConv::AMDGPU_Gfx:
return RetCC_SI_Gfx;
case CallingConv::C:
case CallingConv::Fast:
case CallingConv::Cold:
return RetCC_AMDGPU_Func;
report_fatal_error("Unsupported calling convention.");
/// The SelectionDAGBuilder will automatically promote function arguments
/// with illegal types. However, this does not work for the AMDGPU targets
/// since the function arguments are stored in memory as these illegal types.
/// In order to handle this properly we need to get the original types sizes
/// from the LLVM IR Function and fixup the ISD:InputArg values before
/// passing them to AnalyzeFormalArguments()
/// When the SelectionDAGBuilder computes the Ins, it takes care of splitting
/// input values across multiple registers. Each item in the Ins array
/// represents a single value that will be stored in registers. Ins[x].VT is
/// the value type of the value that will be stored in the register, so
/// whatever SDNode we lower the argument to needs to be this type.
/// In order to correctly lower the arguments we need to know the size of each
/// argument. Since Ins[x].VT gives us the size of the register that will
/// hold the value, we need to look at Ins[x].ArgVT to see the 'real' type
/// for the original function argument so that we can deduce the correct memory
/// type to use for Ins[x]. In most cases the correct memory type will be
/// Ins[x].ArgVT. However, this will not always be the case. If, for example,
/// we have a kernel argument of type v8i8, this argument will be split into
/// 8 parts and each part will be represented by its own item in the Ins array.
/// For each part the Ins[x].ArgVT will be the v8i8, which is the full type of
/// the argument before it was split. From this, we deduce that the memory type
/// for each individual part is i8. We pass the memory type as LocVT to the
/// calling convention analysis function and the register type (Ins[x].VT) as
/// the ValVT.
void AMDGPUTargetLowering::analyzeFormalArgumentsCompute(
CCState &State,
const SmallVectorImpl<ISD::InputArg> &Ins) const {
const MachineFunction &MF = State.getMachineFunction();
const Function &Fn = MF.getFunction();
LLVMContext &Ctx = Fn.getParent()->getContext();
const AMDGPUSubtarget &ST = AMDGPUSubtarget::get(MF);
const unsigned ExplicitOffset = ST.getExplicitKernelArgOffset(Fn);
CallingConv::ID CC = Fn.getCallingConv();
Align MaxAlign = Align(1);
uint64_t ExplicitArgOffset = 0;
const DataLayout &DL = Fn.getParent()->getDataLayout();
unsigned InIndex = 0;
for (const Argument &Arg : Fn.args()) {
const bool IsByRef = Arg.hasByRefAttr();
Type *BaseArgTy = Arg.getType();
Type *MemArgTy = IsByRef ? Arg.getParamByRefType() : BaseArgTy;
MaybeAlign Alignment = IsByRef ? Arg.getParamAlign() : None;
if (!Alignment)
Alignment = DL.getABITypeAlign(MemArgTy);
MaxAlign = max(Alignment, MaxAlign);
uint64_t AllocSize = DL.getTypeAllocSize(MemArgTy);
uint64_t ArgOffset = alignTo(ExplicitArgOffset, Alignment) + ExplicitOffset;
ExplicitArgOffset = alignTo(ExplicitArgOffset, Alignment) + AllocSize;
// We're basically throwing away everything passed into us and starting over
// to get accurate in-memory offsets. The "PartOffset" is completely useless
// to us as computed in Ins.
// We also need to figure out what type legalization is trying to do to get
// the correct memory offsets.
SmallVector<EVT, 16> ValueVTs;
SmallVector<uint64_t, 16> Offsets;
ComputeValueVTs(*this, DL, BaseArgTy, ValueVTs, &Offsets, ArgOffset);
for (unsigned Value = 0, NumValues = ValueVTs.size();
Value != NumValues; ++Value) {
uint64_t BasePartOffset = Offsets[Value];
EVT ArgVT = ValueVTs[Value];
EVT MemVT = ArgVT;
MVT RegisterVT = getRegisterTypeForCallingConv(Ctx, CC, ArgVT);
unsigned NumRegs = getNumRegistersForCallingConv(Ctx, CC, ArgVT);
if (NumRegs == 1) {
// This argument is not split, so the IR type is the memory type.
if (ArgVT.isExtended()) {
// We have an extended type, like i24, so we should just use the
// register type.
MemVT = RegisterVT;
} else {
MemVT = ArgVT;
} else if (ArgVT.isVector() && RegisterVT.isVector() &&
ArgVT.getScalarType() == RegisterVT.getScalarType()) {
assert(ArgVT.getVectorNumElements() > RegisterVT.getVectorNumElements());
// We have a vector value which has been split into a vector with
// the same scalar type, but fewer elements. This should handle
// all the floating-point vector types.
MemVT = RegisterVT;
} else if (ArgVT.isVector() &&
ArgVT.getVectorNumElements() == NumRegs) {
// This arg has been split so that each element is stored in a separate
// register.
MemVT = ArgVT.getScalarType();
} else if (ArgVT.isExtended()) {
// We have an extended type, like i65.
MemVT = RegisterVT;
} else {
unsigned MemoryBits = ArgVT.getStoreSizeInBits() / NumRegs;
assert(ArgVT.getStoreSizeInBits() % NumRegs == 0);
if (RegisterVT.isInteger()) {
MemVT = EVT::getIntegerVT(State.getContext(), MemoryBits);
} else if (RegisterVT.isVector()) {
unsigned NumElements = RegisterVT.getVectorNumElements();
assert(MemoryBits % NumElements == 0);
// This vector type has been split into another vector type with
// a different elements size.
EVT ScalarVT = EVT::getIntegerVT(State.getContext(),
MemoryBits / NumElements);
MemVT = EVT::getVectorVT(State.getContext(), ScalarVT, NumElements);
} else {
llvm_unreachable("cannot deduce memory type.");
// Convert one element vectors to scalar.
if (MemVT.isVector() && MemVT.getVectorNumElements() == 1)
MemVT = MemVT.getScalarType();
// Round up vec3/vec5 argument.
if (MemVT.isVector() && !MemVT.isPow2VectorType()) {
assert(MemVT.getVectorNumElements() == 3 ||
MemVT.getVectorNumElements() == 5);
MemVT = MemVT.getPow2VectorType(State.getContext());
} else if (!MemVT.isSimple() && !MemVT.isVector()) {
MemVT = MemVT.getRoundIntegerType(State.getContext());
unsigned PartOffset = 0;
for (unsigned i = 0; i != NumRegs; ++i) {
State.addLoc(CCValAssign::getCustomMem(InIndex++, RegisterVT,
BasePartOffset + PartOffset,
PartOffset += MemVT.getStoreSize();
SDValue AMDGPUTargetLowering::LowerReturn(
SDValue Chain, CallingConv::ID CallConv,
bool isVarArg,
const SmallVectorImpl<ISD::OutputArg> &Outs,
const SmallVectorImpl<SDValue> &OutVals,
const SDLoc &DL, SelectionDAG &DAG) const {
// FIXME: Fails for r600 tests
//assert(!isVarArg && Outs.empty() && OutVals.empty() &&
// "wave terminate should not have return values");
return DAG.getNode(AMDGPUISD::ENDPGM, DL, MVT::Other, Chain);
// Target specific lowering
/// Selects the correct CCAssignFn for a given CallingConvention value.
CCAssignFn *AMDGPUTargetLowering::CCAssignFnForCall(CallingConv::ID CC,
bool IsVarArg) {
return AMDGPUCallLowering::CCAssignFnForCall(CC, IsVarArg);
CCAssignFn *AMDGPUTargetLowering::CCAssignFnForReturn(CallingConv::ID CC,
bool IsVarArg) {
return AMDGPUCallLowering::CCAssignFnForReturn(CC, IsVarArg);
SDValue AMDGPUTargetLowering::addTokenForArgument(SDValue Chain,
SelectionDAG &DAG,
MachineFrameInfo &MFI,
int ClobberedFI) const {
SmallVector<SDValue, 8> ArgChains;
int64_t FirstByte = MFI.getObjectOffset(ClobberedFI);
int64_t LastByte = FirstByte + MFI.getObjectSize(ClobberedFI) - 1;
// Include the original chain at the beginning of the list. When this is
// used by target LowerCall hooks, this helps legalize find the
// Add a chain value for each stack argument corresponding
for (SDNode *U : DAG.getEntryNode().getNode()->uses()) {
if (LoadSDNode *L = dyn_cast<LoadSDNode>(U)) {
if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr())) {
if (FI->getIndex() < 0) {
int64_t InFirstByte = MFI.getObjectOffset(FI->getIndex());
int64_t InLastByte = InFirstByte;
InLastByte += MFI.getObjectSize(FI->getIndex()) - 1;
if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
(FirstByte <= InFirstByte && InFirstByte <= LastByte))
ArgChains.push_back(SDValue(L, 1));
// Build a tokenfactor for all the chains.
return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
SDValue AMDGPUTargetLowering::lowerUnhandledCall(CallLoweringInfo &CLI,
SmallVectorImpl<SDValue> &InVals,
StringRef Reason) const {
SDValue Callee = CLI.Callee;
SelectionDAG &DAG = CLI.DAG;
const Function &Fn = DAG.getMachineFunction().getFunction();
StringRef FuncName("<unknown>");
if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
FuncName = G->getSymbol();
else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
FuncName = G->getGlobal()->getName();
DiagnosticInfoUnsupported NoCalls(
Fn, Reason + FuncName, CLI.DL.getDebugLoc());
if (!CLI.IsTailCall) {
for (unsigned I = 0, E = CLI.Ins.size(); I != E; ++I)
return DAG.getEntryNode();
SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
SmallVectorImpl<SDValue> &InVals) const {
return lowerUnhandledCall(CLI, InVals, "unsupported call to function ");
SelectionDAG &DAG) const {
const Function &Fn = DAG.getMachineFunction().getFunction();
DiagnosticInfoUnsupported NoDynamicAlloca(Fn, "unsupported dynamic alloca",
auto Ops = {DAG.getConstant(0, SDLoc(), Op.getValueType()), Op.getOperand(0)};
return DAG.getMergeValues(Ops, SDLoc());
SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
SelectionDAG &DAG) const {
switch (Op.getOpcode()) {
Op->print(errs(), &DAG);
llvm_unreachable("Custom lowering code for this "
"instruction is not implemented yet!");
case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
case ISD::SDIVREM: return LowerSDIVREM(Op, DAG);
case ISD::FREM: return LowerFREM(Op, DAG);
case ISD::FCEIL: return LowerFCEIL(Op, DAG);
case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
case ISD::FRINT: return LowerFRINT(Op, DAG);
case ISD::FROUND: return LowerFROUND(Op, DAG);
case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
case ISD::FLOG:
return LowerFLOG(Op, DAG, numbers::ln2f);
case ISD::FLOG10:
return LowerFLOG(Op, DAG, numbers::ln2f / numbers::ln10f);
case ISD::FEXP:
return lowerFEXP(Op, DAG);
case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
case ISD::FP_TO_FP16: return LowerFP_TO_FP16(Op, DAG);
return LowerFP_TO_INT(Op, DAG);
case ISD::CTTZ:
case ISD::CTLZ:
return LowerCTLZ_CTTZ(Op, DAG);
return Op;
void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
SmallVectorImpl<SDValue> &Results,
SelectionDAG &DAG) const {
switch (N->getOpcode()) {
// Different parts of legalization seem to interpret which type of
// sign_extend_inreg is the one to check for custom lowering. The extended
// from type is what really matters, but some places check for custom
// lowering of the result type. This results in trying to use
// ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
// nothing here and let the illegal result integer be handled normally.
SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
SDValue Op,
SelectionDAG &DAG) const {
const DataLayout &DL = DAG.getDataLayout();
GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
const GlobalValue *GV = G->getGlobal();
if (G->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
G->getAddressSpace() == AMDGPUAS::REGION_ADDRESS) {
if (!MFI->isModuleEntryFunction() &&
!GV->getName().equals("")) {
SDLoc DL(Op);
const Function &Fn = DAG.getMachineFunction().getFunction();
DiagnosticInfoUnsupported BadLDSDecl(
Fn, "local memory global used by non-kernel function",
DL.getDebugLoc(), DS_Warning);
// We currently don't have a way to correctly allocate LDS objects that
// aren't directly associated with a kernel. We do force inlining of
// functions that use local objects. However, if these dead functions are
// not eliminated, we don't want a compile time error. Just emit a warning
// and a trap, since there should be no callable path here.
SDValue Trap = DAG.getNode(ISD::TRAP, DL, MVT::Other, DAG.getEntryNode());
SDValue OutputChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
Trap, DAG.getRoot());
return DAG.getUNDEF(Op.getValueType());
// XXX: What does the value of G->getOffset() mean?
assert(G->getOffset() == 0 &&
"Do not know what to do with an non-zero offset");
// TODO: We could emit code to handle the initialization somewhere.
// We ignore the initializer for now and legalize it to allow selection.
// The initializer will anyway get errored out during assembly emission.
unsigned Offset = MFI->allocateLDSGlobal(DL, *cast<GlobalVariable>(GV));
return DAG.getConstant(Offset, SDLoc(Op), Op.getValueType());
return SDValue();
SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
SelectionDAG &DAG) const {
SmallVector<SDValue, 8> Args;
EVT VT = Op.getValueType();
if (VT == MVT::v4i16 || VT == MVT::v4f16) {
SDLoc SL(Op);
SDValue Lo = DAG.getNode(ISD::BITCAST, SL, MVT::i32, Op.getOperand(0));
SDValue Hi = DAG.getNode(ISD::BITCAST, SL, MVT::i32, Op.getOperand(1));
SDValue BV = DAG.getBuildVector(MVT::v2i32, SL, { Lo, Hi });
return DAG.getNode(ISD::BITCAST, SL, VT, BV);
for (const SDUse &U : Op->ops())
DAG.ExtractVectorElements(U.get(), Args);
return DAG.getBuildVector(Op.getValueType(), SDLoc(Op), Args);
SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
SelectionDAG &DAG) const {
SmallVector<SDValue, 8> Args;
unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
EVT VT = Op.getValueType();
EVT SrcVT = Op.getOperand(0).getValueType();
// For these types, we have some TableGen patterns except if the index is 1
if (((SrcVT == MVT::v4f16 && VT == MVT::v2f16) ||
(SrcVT == MVT::v4i16 && VT == MVT::v2i16)) &&
Start != 1)
return Op;
DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
return DAG.getBuildVector(Op.getValueType(), SDLoc(Op), Args);
/// Generate Min/Max node
SDValue AMDGPUTargetLowering::combineFMinMaxLegacy(const SDLoc &DL, EVT VT,
SDValue LHS, SDValue RHS,
SDValue True, SDValue False,
SDValue CC,
DAGCombinerInfo &DCI) const {
if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
return SDValue();
SelectionDAG &DAG = DCI.DAG;
ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
switch (CCOpcode) {
case ISD::SETNE:
case ISD::SETEQ:
case ISD::SETUO:
case ISD::SETO:
case ISD::SETULT: {
if (LHS == True)
case ISD::SETLE:
case ISD::SETLT: {
// Ordered. Assume ordered for undefined.
// Only do this after legalization to avoid interfering with other combines
// which might occur.
if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
return SDValue();
// We need to permute the operands to get the correct NaN behavior. The
// selected operand is the second one based on the failing compare with NaN,
// so permute it based on the compare type the hardware uses.
if (LHS == True)
case ISD::SETUGT: {
if (LHS == True)
case ISD::SETGT:
case ISD::SETGE:
case ISD::SETOGT: {
if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
return SDValue();
if (LHS == True)
llvm_unreachable("Invalid setcc condcode!");
return SDValue();
std::pair<SDValue, SDValue>
AMDGPUTargetLowering::split64BitValue(SDValue Op, SelectionDAG &DAG) const {
SDLoc SL(Op);
SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);
const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
const SDValue One = DAG.getConstant(1, SL, MVT::i32);
SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, Zero);
SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One);
return std::make_pair(Lo, Hi);
SDValue AMDGPUTargetLowering::getLoHalf64(SDValue Op, SelectionDAG &DAG) const {
SDLoc SL(Op);
SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);
const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, Zero);
SDValue AMDGPUTargetLowering::getHiHalf64(SDValue Op, SelectionDAG &DAG) const {
SDLoc SL(Op);
SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);
const SDValue One = DAG.getConstant(1, SL, MVT::i32);
return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One);
// Split a vector type into two parts. The first part is a power of two vector.
// The second part is whatever is left over, and is a scalar if it would
// otherwise be a 1-vector.
std::pair<EVT, EVT>
AMDGPUTargetLowering::getSplitDestVTs(const EVT &VT, SelectionDAG &DAG) const {
EVT EltVT = VT.getVectorElementType();
unsigned NumElts = VT.getVectorNumElements();
unsigned LoNumElts = PowerOf2Ceil((NumElts + 1) / 2);
LoVT = EVT::getVectorVT(*DAG.getContext(), EltVT, LoNumElts);
HiVT = NumElts - LoNumElts == 1
? EltVT
: EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts - LoNumElts);
return std::make_pair(LoVT, HiVT);
// Split a vector value into two parts of types LoVT and HiVT. HiVT could be
// scalar.
std::pair<SDValue, SDValue>
AMDGPUTargetLowering::splitVector(const SDValue &N, const SDLoc &DL,
const EVT &LoVT, const EVT &HiVT,
SelectionDAG &DAG) const {
assert(LoVT.getVectorNumElements() +
(HiVT.isVector() ? HiVT.getVectorNumElements() : 1) <=
N.getValueType().getVectorNumElements() &&
"More vector elements requested than available!");
DAG.getVectorIdxConstant(0, DL));
SDValue Hi = DAG.getNode(
HiVT, N, DAG.getVectorIdxConstant(LoVT.getVectorNumElements(), DL));
return std::make_pair(Lo, Hi);
SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue Op,
SelectionDAG &DAG) const {
LoadSDNode *Load = cast<LoadSDNode>(Op);
EVT VT = Op.getValueType();
SDLoc SL(Op);
// If this is a 2 element vector, we really want to scalarize and not create
// weird 1 element vectors.
if (VT.getVectorNumElements() == 2) {
SDValue Ops[2];
std::tie(Ops[0], Ops[1]) = scalarizeVectorLoad(Load, DAG);
return DAG.getMergeValues(Ops, SL);
SDValue BasePtr = Load->getBasePtr();
EVT MemVT = Load->getMemoryVT();
const MachinePointerInfo &SrcValue = Load->getMemOperand()->getPointerInfo();
SDValue Lo, Hi;
std::tie(LoVT, HiVT) = getSplitDestVTs(VT, DAG);
std::tie(LoMemVT, HiMemVT) = getSplitDestVTs(MemVT, DAG);
std::tie(Lo, Hi) = splitVector(Op, SL, LoVT, HiVT, DAG);
unsigned Size = LoMemVT.getStoreSize();
unsigned BaseAlign = Load->getAlignment();
unsigned HiAlign = MinAlign(BaseAlign, Size);
SDValue LoLoad = DAG.getExtLoad(Load->getExtensionType(), SL, LoVT,
Load->getChain(), BasePtr, SrcValue, LoMemVT,
BaseAlign, Load->getMemOperand()->getFlags());
SDValue HiPtr = DAG.getObjectPtrOffset(SL, BasePtr, TypeSize::Fixed(Size));
SDValue HiLoad =
DAG.getExtLoad(Load->getExtensionType(), SL, HiVT, Load->getChain(),
HiPtr, SrcValue.getWithOffset(LoMemVT.getStoreSize()),
HiMemVT, HiAlign, Load->getMemOperand()->getFlags());
SDValue Join;
if (LoVT == HiVT) {
// This is the case that the vector is power of two so was evenly split.
Join = DAG.getNode(ISD::CONCAT_VECTORS, SL, VT, LoLoad, HiLoad);
} else {
DAG.getVectorIdxConstant(0, SL));
Join = DAG.getNode(
VT, Join, HiLoad,
DAG.getVectorIdxConstant(LoVT.getVectorNumElements(), SL));
SDValue Ops[] = {Join, DAG.getNode(ISD::TokenFactor, SL, MVT::Other,
LoLoad.getValue(1), HiLoad.getValue(1))};
return DAG.getMergeValues(Ops, SL);
SDValue AMDGPUTargetLowering::WidenOrSplitVectorLoad(SDValue Op,
SelectionDAG &DAG) const {
LoadSDNode *Load = cast<LoadSDNode>(Op);
EVT VT = Op.getValueType();
SDValue BasePtr = Load->getBasePtr();
EVT MemVT = Load->getMemoryVT();
SDLoc SL(Op);
const MachinePointerInfo &SrcValue = Load->getMemOperand()->getPointerInfo();
unsigned BaseAlign = Load->getAlignment();
unsigned NumElements = MemVT.getVectorNumElements();
// Widen from vec3 to vec4 when the load is at least 8-byte aligned
// or 16-byte fully dereferenceable. Otherwise, split the vector load.
if (NumElements != 3 ||
(BaseAlign < 8 &&
!SrcValue.isDereferenceable(16, *DAG.getContext(), DAG.getDataLayout())))
return SplitVectorLoad(Op, DAG);
assert(NumElements == 3);
EVT WideVT =
EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), 4);
EVT WideMemVT =
EVT::getVectorVT(*DAG.getContext(), MemVT.getVectorElementType(), 4);
SDValue WideLoad = DAG.getExtLoad(
Load->getExtensionType(), SL, WideVT, Load->getChain(), BasePtr, SrcValue,
WideMemVT, BaseAlign, Load->getMemOperand()->getFlags());
return DAG.getMergeValues(
DAG.getVectorIdxConstant(0, SL)),
SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
SelectionDAG &DAG) const {
StoreSDNode *Store = cast<StoreSDNode>(Op);
SDValue Val = Store->getValue();
EVT VT = Val.getValueType();
// If this is a 2 element vector, we really want to scalarize and not create
// weird 1 element vectors.
if (VT.getVectorNumElements() == 2)
return scalarizeVectorStore(Store, DAG);
EVT MemVT = Store->getMemoryVT();
SDValue Chain = Store->getChain();
SDValue BasePtr = Store->getBasePtr();
SDLoc SL(Op);
SDValue Lo, Hi;
std::tie(LoVT, HiVT) = getSplitDestVTs(VT, DAG);
std::tie(LoMemVT, HiMemVT) = getSplitDestVTs(MemVT, DAG);
std::tie(Lo, Hi) = splitVector(Val, SL, LoVT, HiVT, DAG);
SDValue HiPtr = DAG.getObjectPtrOffset(SL, BasePtr, LoMemVT.getStoreSize());
const MachinePointerInfo &SrcValue = Store->getMemOperand()->getPointerInfo();
unsigned BaseAlign = Store->getAlignment();
unsigned Size = LoMemVT.getStoreSize();
unsigned HiAlign = MinAlign(BaseAlign, Size);
SDValue LoStore =
DAG.getTruncStore(Chain, SL, Lo, BasePtr, SrcValue, LoMemVT, BaseAlign,
SDValue HiStore =
DAG.getTruncStore(Chain, SL, Hi, HiPtr, SrcValue.getWithOffset(Size),
HiMemVT, HiAlign, Store->getMemOperand()->getFlags());
return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, LoStore, HiStore);
// This is a shortcut for integer division because we have fast i32<->f32
// conversions, and fast f32 reciprocal instructions. The fractional part of a
// float is enough to accurately represent up to a 24-bit signed integer.
SDValue AMDGPUTargetLowering::LowerDIVREM24(SDValue Op, SelectionDAG &DAG,
bool Sign) const {
SDLoc DL(Op);
EVT VT = Op.getValueType();
SDValue LHS = Op.getOperand(0);
SDValue RHS = Op.getOperand(1);
MVT IntVT = MVT::i32;
MVT FltVT = MVT::f32;
unsigned LHSSignBits = DAG.ComputeNumSignBits(LHS);
if (LHSSignBits < 9)
return SDValue();
unsigned RHSSignBits = DAG.ComputeNumSignBits(RHS);
if (RHSSignBits < 9)
return SDValue();
unsigned BitSize = VT.getSizeInBits();
unsigned SignBits = std::min(LHSSignBits, RHSSignBits);
unsigned DivBits = BitSize - SignBits;
if (Sign)
ISD::NodeType ToFp = Sign ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
ISD::NodeType ToInt = Sign ? ISD::FP_TO_SINT : ISD::FP_TO_UINT;
SDValue jq = DAG.getConstant(1, DL, IntVT);
if (Sign) {
// char|short jq = ia ^ ib;
jq = DAG.getNode(ISD::XOR, DL, VT, LHS, RHS);
// jq = jq >> (bitsize - 2)
jq = DAG.getNode(ISD::SRA, DL, VT, jq,
DAG.getConstant(BitSize - 2, DL, VT));
// jq = jq | 0x1
jq = DAG.getNode(ISD::OR, DL, VT, jq, DAG.getConstant(1, DL, VT));
// int ia = (int)LHS;
SDValue ia = LHS;
// int ib, (int)RHS;
SDValue ib = RHS;
// float fa = (float)ia;
SDValue fa = DAG.getNode(ToFp, DL, FltVT, ia);
// float fb = (float)ib;
SDValue fb = DAG.getNode(ToFp, DL, FltVT, ib);
SDValue fq = DAG.getNode(ISD::FMUL, DL, FltVT,
fa, DAG.getNode(AMDGPUISD::RCP, DL, FltVT, fb));
// fq = trunc(fq);
fq = DAG.getNode(ISD::FTRUNC, DL, FltVT, fq);
// float fqneg = -fq;
SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FltVT, fq);
MachineFunction &MF = DAG.getMachineFunction();
const AMDGPUMachineFunction *MFI = MF.getInfo<AMDGPUMachineFunction>();
// float fr = mad(fqneg, fb, fa);
unsigned OpCode = !Subtarget->hasMadMacF32Insts() ?
(unsigned)ISD::FMA :
!MFI->getMode().allFP32Denormals() ?
(unsigned)ISD::FMAD :
SDValue fr = DAG.getNode(OpCode, DL, FltVT, fqneg, fb, fa);
// int iq = (int)fq;
SDValue iq = DAG.getNode(ToInt, DL, IntVT, fq);
// fr = fabs(fr);
fr = DAG.getNode(ISD::FABS, DL, FltVT, fr);
// fb = fabs(fb);
fb = DAG.getNode(ISD::FABS, DL, FltVT, fb);
EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
// int cv = fr >= fb;
SDValue cv = DAG.getSetCC(DL, SetCCVT, fr, fb, ISD::SETOGE);
// jq = (cv ? jq : 0);
jq = DAG.getNode(ISD::SELECT, DL, VT, cv, jq, DAG.getConstant(0, DL, VT));
// dst = iq + jq;
SDValue Div = DAG.getNode(ISD::ADD, DL, VT, iq, jq);
// Rem needs compensation, it's easier to recompute it
SDValue Rem = DAG.getNode(ISD::MUL, DL, VT, Div, RHS);
Rem = DAG.getNode(ISD::SUB, DL, VT, LHS, Rem);
// Truncate to number of bits this divide really is.
if (Sign) {
SDValue InRegSize
= DAG.getValueType(EVT::getIntegerVT(*DAG.getContext(), DivBits));
Div = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, Div, InRegSize);
Rem = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, Rem, InRegSize);
} else {
SDValue TruncMask = DAG.getConstant((UINT64_C(1) << DivBits) - 1, DL, VT);
Div = DAG.getNode(ISD::AND, DL, VT, Div, TruncMask);
Rem = DAG.getNode(ISD::AND, DL, VT, Rem, TruncMask);
return DAG.getMergeValues({ Div, Rem }, DL);
void AMDGPUTargetLowering::LowerUDIVREM64(SDValue Op,
SelectionDAG &DAG,
SmallVectorImpl<SDValue> &Results) const {
SDLoc DL(Op);
EVT VT = Op.getValueType();
assert(VT == MVT::i64 && "LowerUDIVREM64 expects an i64");
EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
SDValue One = DAG.getConstant(1, DL, HalfVT);
SDValue Zero = DAG.getConstant(0, DL, HalfVT);
//HiLo split
SDValue LHS = Op.getOperand(0);
SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, Zero);
SDValue RHS = Op.getOperand(1);
SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, Zero);
if (DAG.MaskedValueIsZero(RHS, APInt::getHighBitsSet(64, 32)) &&
DAG.MaskedValueIsZero(LHS, APInt::getHighBitsSet(64, 32))) {
SDValue Res = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
LHS_Lo, RHS_Lo);
SDValue DIV = DAG.getBuildVector(MVT::v2i32, DL, {Res.getValue(0), Zero});
SDValue REM = DAG.getBuildVector(MVT::v2i32, DL, {Res.getValue(1), Zero});
Results.push_back(DAG.getNode(ISD::BITCAST, DL, MVT::i64, DIV));
Results.push_back(DAG.getNode(ISD::BITCAST, DL, MVT::i64, REM));
if (isTypeLegal(MVT::i64)) {
// The algorithm here is based on ideas from "Software Integer Division",
// Tom Rodeheffer, August 2008.
MachineFunction &MF = DAG.getMachineFunction();
const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
// Compute denominator reciprocal.
unsigned FMAD = !Subtarget->hasMadMacF32Insts() ?
(unsigned)ISD::FMA :
!MFI->getMode().allFP32Denormals() ?
(unsigned)ISD::FMAD :
SDValue Cvt_Lo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, RHS_Lo);
SDValue Cvt_Hi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, RHS_Hi);
SDValue Mad1 = DAG.getNode(FMAD, DL, MVT::f32, Cvt_Hi,
DAG.getConstantFP(APInt(32, 0x4f800000).bitsToFloat(), DL, MVT::f32),
SDValue Rcp = DAG.getNode(AMDGPUISD::RCP, DL, MVT::f32, Mad1);
SDValue Mul1 = DAG.getNode(ISD::FMUL, DL, MVT::f32, Rcp,
DAG.getConstantFP(APInt(32, 0x5f7ffffc).bitsToFloat(), DL, MVT::f32));
SDValue Mul2 = DAG.getNode(ISD::FMUL, DL, MVT::f32, Mul1,
DAG.getConstantFP(APInt(32, 0x2f800000).bitsToFloat(), DL, MVT::f32));
SDValue Trunc = DAG.getNode(ISD::FTRUNC, DL, MVT::f32, Mul2);
SDValue Mad2 = DAG.getNode(FMAD, DL, MVT::f32, Trunc,
DAG.getConstantFP(APInt(32, 0xcf800000).bitsToFloat(), DL, MVT::f32),
SDValue Rcp_Lo = DAG.getNode(ISD::FP_TO_UINT, DL, HalfVT, Mad2);
SDValue Rcp_Hi = DAG.getNode(ISD::FP_TO_UINT, DL, HalfVT, Trunc);
SDValue Rcp64 = DAG.getBitcast(VT,
DAG.getBuildVector(MVT::v2i32, DL, {Rcp_Lo, Rcp_Hi}));
SDValue Zero64 = DAG.getConstant(0, DL, VT);
SDValue One64 = DAG.getConstant(1, DL, VT);
SDValue Zero1 = DAG.getConstant(0, DL, MVT::i1);
SDVTList HalfCarryVT = DAG.getVTList(HalfVT, MVT::i1);
// First round of UNR (Unsigned integer Newton-Raphson).
SDValue Neg_RHS = DAG.getNode(ISD::SUB, DL, VT, Zero64, RHS);
SDValue Mullo1 = DAG.getNode(ISD::MUL, DL, VT, Neg_RHS, Rcp64);
SDValue Mulhi1 = DAG.getNode(ISD::MULHU, DL, VT, Rcp64, Mullo1);
SDValue Mulhi1_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi1,
SDValue Mulhi1_Hi =
DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi1, One);
SDValue Add1_Lo = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Rcp_Lo,
Mulhi1_Lo, Zero1);
SDValue Add1_Hi = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Rcp_Hi,
Mulhi1_Hi, Add1_Lo.getValue(1));
SDValue Add1 = DAG.getBitcast(VT,
DAG.getBuildVector(MVT::v2i32, DL, {Add1_Lo, Add1_Hi}));
// Second round of UNR.
SDValue Mullo2 = DAG.getNode(ISD::MUL, DL, VT, Neg_RHS, Add1);
SDValue Mulhi2 = DAG.getNode(ISD::MULHU, DL, VT, Add1, Mullo2);
SDValue Mulhi2_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi2,
SDValue Mulhi2_Hi =
DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mulhi2, One);
SDValue Add2_Lo = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Add1_Lo,
Mulhi2_Lo, Zero1);
SDValue Add2_Hi = DAG.getNode(ISD::ADDCARRY, DL, HalfCarryVT, Add1_Hi,
Mulhi2_Hi, Add2_Lo.getValue(1));
SDValue Add2 = DAG.getBitcast(VT,
DAG.getBuildVector(MVT::v2i32, DL, {Add2_Lo, Add2_Hi}));
SDValue Mulhi3 = DAG.getNode(ISD::MULHU, DL, VT, LHS, Add2);
SDValue Mul3 = DAG.getNode(ISD::MUL, DL, VT, RHS, Mulhi3);
SDValue Mul3_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mul3, Zero);
SDValue Mul3_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, Mul3, One);
SDValue Sub1_Lo = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, LHS_Lo,
Mul3_Lo, Zero1);
SDValue Sub1_Hi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, LHS_Hi,
Mul3_Hi, Sub1_Lo.getValue(1));
SDValue Sub1_Mi = DAG.getNode(ISD::SUB, DL, HalfVT, LHS_Hi, Mul3_Hi);
SDValue Sub1 = DAG.getBitcast(VT,
DAG.getBuildVector(MVT::v2i32, DL, {Sub1_Lo, Sub1_Hi}));
SDValue MinusOne = DAG.getConstant(0xffffffffu, DL, HalfVT);
SDValue C1 = DAG.getSelectCC(DL, Sub1_Hi, RHS_Hi, MinusOne, Zero,
SDValue C2 = DAG.getSelectCC(DL, Sub1_Lo, RHS_Lo, MinusOne, Zero,
SDValue C3 = DAG.getSelectCC(DL, Sub1_Hi, RHS_Hi, C2, C1, ISD::SETEQ);
// TODO: Here and below portions of the code can be enclosed into if/endif.
// Currently control flow is unconditional and we have 4 selects after
// potential endif to substitute PHIs.
// if C3 != 0 ...
SDValue Sub2_Lo = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub1_Lo,
RHS_Lo, Zero1);
SDValue Sub2_Mi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub1_Mi,
RHS_Hi, Sub1_Lo.getValue(1));
SDValue Sub2_Hi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub2_Mi,
Zero, Sub2_Lo.getValue(1));
SDValue Sub2 = DAG.getBitcast(VT,
DAG.getBuildVector(MVT::v2i32, DL, {Sub2_Lo, Sub2_Hi}));
SDValue Add3 = DAG.getNode(ISD::ADD, DL, VT, Mulhi3, One64);
SDValue C4 = DAG.getSelectCC(DL, Sub2_Hi, RHS_Hi, MinusOne, Zero,
SDValue C5 = DAG.getSelectCC(DL, Sub2_Lo, RHS_Lo, MinusOne, Zero,
SDValue C6 = DAG.getSelectCC(DL, Sub2_Hi, RHS_Hi, C5, C4, ISD::SETEQ);
// if (C6 != 0)
SDValue Add4 = DAG.getNode(ISD::ADD, DL, VT, Add3, One64);
SDValue Sub3_Lo = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub2_Lo,
RHS_Lo, Zero1);
SDValue Sub3_Mi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub2_Mi,
RHS_Hi, Sub2_Lo.getValue(1));
SDValue Sub3_Hi = DAG.getNode(ISD::SUBCARRY, DL, HalfCarryVT, Sub3_Mi,
Zero, Sub3_Lo.getValue(1));
SDValue Sub3 = DAG.getBitcast(VT,
DAG.getBuildVector(MVT::v2i32, DL, {Sub3_Lo, Sub3_Hi}));
// endif C6
// endif C3
SDValue Sel1 = DAG.getSelectCC(DL, C6, Zero, Add4, Add3, ISD::SETNE);
SDValue Div = DAG.getSelectCC(DL, C3, Zero, Sel1, Mulhi3, ISD::SETNE);
SDValue Sel2 = DAG.getSelectCC(DL, C6, Zero, Sub3, Sub2, ISD::SETNE);
SDValue Rem = DAG.getSelectCC(DL, C3, Zero, Sel2, Sub1, ISD::SETNE);
// r600 expandion.
// Get Speculative values
SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, Zero, REM_Part, LHS_Hi, ISD::SETEQ);
SDValue REM = DAG.getBuildVector(MVT::v2i32, DL, {REM_Lo, Zero});
REM = DAG.getNode(ISD::BITCAST, DL, MVT::i64, REM);
SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, Zero, DIV_Part, Zero, ISD::SETEQ);
SDValue DIV_Lo = Zero;
const unsigned halfBitWidth = HalfVT.getSizeInBits();
for (unsigned i = 0; i < halfBitWidth; ++i) {
const unsigned bitPos = halfBitWidth - i - 1;
SDValue POS = DAG.getConstant(bitPos, DL, HalfVT);
// Get value of high bit
SDValue HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, One);
HBit = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, HBit);
// Shift
REM = DAG.getNode(ISD::SHL, DL, VT, REM, DAG.getConstant(1, DL, VT));
// Add LHS high bit
REM = DAG.getNode(ISD::OR, DL, VT, REM, HBit);
SDValue BIT = DAG.getConstant(1ULL << bitPos, DL, HalfVT);
SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, Zero, ISD::SETUGE);
DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
// Update REM
SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
SDValue DIV = DAG.getBuildVector(MVT::v2i32, DL, {DIV_Lo, DIV_Hi});
DIV = DAG.getNode(ISD::BITCAST, DL, MVT::i64, DIV);
SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
EVT VT = Op.getValueType();
if (VT == MVT::i64) {
SmallVector<SDValue, 2> Results;
LowerUDIVREM64(Op, DAG, Results);
return DAG.getMergeValues(Results, DL);
if (VT == MVT::i32) {
if (SDValue Res = LowerDIVREM24(Op, DAG, false))
return Res;
SDValue X = Op.getOperand(0);
SDValue Y = Op.getOperand(1);
// See AMDGPUCodeGenPrepare::expandDivRem32 for a description of the
// algorithm used here.
// Initial estimate of inv(y).
// One round of UNR.
SDValue NegY = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Y);
SDValue NegYZ = DAG.getNode(ISD::MUL, DL, VT, NegY, Z);
Z = DAG.getNode(ISD::ADD, DL, VT, Z,
DAG.getNode(ISD::MULHU, DL, VT, Z, NegYZ));
// Quotient/remainder estimate.
SDValue Q = DAG.getNode(ISD::MULHU, DL, VT, X, Z);
SDValue R =
DAG.getNode(ISD::SUB, DL, VT, X, DAG.getNode(ISD::MUL, DL, VT, Q, Y));
// First quotient/remainder refinement.
EVT CCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
SDValue One = DAG.getConstant(1, DL, VT);
SDValue Cond = DAG.getSetCC(DL, CCVT, R, Y, ISD::SETUGE);
Q = DAG.getNode(ISD::SELECT, DL, VT, Cond,
DAG.getNode(ISD::ADD, DL, VT, Q, One), Q);
R = DAG.getNode(ISD::SELECT, DL, VT, Cond,
DAG.getNode(ISD::SUB, DL, VT, R, Y), R);
// Second quotient/remainder refinement.
Cond = DAG.getSetCC(DL, CCVT, R, Y, ISD::SETUGE);
Q = DAG.getNode(ISD::SELECT, DL, VT, Cond,
DAG.getNode(ISD::ADD, DL, VT, Q, One), Q);
R = DAG.getNode(ISD::SELECT, DL, VT, Cond,
DAG.getNode(ISD::SUB, DL, VT, R, Y), R);
return DAG.getMergeValues({Q, R}, DL);
SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
EVT VT = Op.getValueType();
SDValue LHS = Op.getOperand(0);
SDValue RHS = Op.getOperand(1);
SDValue Zero = DAG.getConstant(0, DL, VT);
SDValue NegOne = DAG.getConstant(-1, DL, VT);
if (VT == MVT::i32) {
if (SDValue Res = LowerDIVREM24(Op, DAG, true))
return Res;
if (VT == MVT::i64 &&
DAG.ComputeNumSignBits(LHS) > 32 &&
DAG.ComputeNumSignBits(RHS) > 32) {
EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
//HiLo split
SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, Zero);
SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, Zero);
SDValue DIVREM = DAG.getNode(ISD::SDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
LHS_Lo, RHS_Lo);
SDValue Res[2] = {
DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(0)),
DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(1))
return DAG.getMergeValues(Res, DL);
SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
SDValue RSign = LHSign; // Remainder sign is the same as LHS
LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign);
RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign);
LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign);
RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign);
SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS);
SDValue Rem = Div.getValue(1);
Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign);
Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign);
Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign);
Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign);
SDValue Res[2] = {
return DAG.getMergeValues(Res, DL);
// (frem x, y) -> (fma (fneg (ftrunc (fdiv x, y))), y, x)
SDValue AMDGPUTargetLowering::LowerFREM(SDValue Op, SelectionDAG &DAG) const {
SDLoc SL(Op);
EVT VT = Op.getValueType();
auto Flags = Op->getFlags();
SDValue X = Op.getOperand(0);
SDValue Y = Op.getOperand(1);
SDValue Div = DAG.getNode(ISD::FDIV, SL, VT, X, Y, Flags);
SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, VT, Div, Flags);
SDValue Neg = DAG.getNode(ISD::FNEG, SL, VT, Trunc, Flags);
// TODO: For f32 use FMAD instead if !hasFastFMA32?
return DAG.getNode(ISD::FMA, SL, VT, Neg, Y, X, Flags);
SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
SDLoc SL(Op);
SDValue Src = Op.getOperand(0);
// result = trunc(src)
// if (src > 0.0 && src != result)
// result += 1.0
SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64);
const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f64);
getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
// TODO: Should this propagate fast-math-flags?
return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
static SDValue extractF64Exponent(SDValue Hi, const SDLoc &SL,
SelectionDAG &DAG) {
const unsigned FractBits = 52;
const unsigned ExpBits = 11;
SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_U32, SL, MVT::i32,
DAG.getConstant(FractBits - 32, SL, MVT::i32),
DAG.getConstant(ExpBits, SL, MVT::i32));
SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
DAG.getConstant(1023, SL, MVT::i32));
return Exp;
SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
SDLoc SL(Op);
SDValue Src = Op.getOperand(0);
assert(Op.getValueType() == MVT::f64);
const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
// Extract the upper half, since this is where we will find the sign and
// exponent.
SDValue Hi = getHiHalf64(Src, DAG);