Several big changes:
1. Use flags on the instructions in the .td file to indicate the PPC970 unit
type instead of a table in the .cpp file. Much cleaner.
2. Change the hazard recognizer to build d-groups according to the actual
algorithm used, not my flawed understanding of it.
3. Model "must be in the first slot" and "must be the only instr in a group"
accurately.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26719 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/PowerPC/PPCHazardRecognizers.cpp b/lib/Target/PowerPC/PPCHazardRecognizers.cpp
index a88cb12..b52e993 100644
--- a/lib/Target/PowerPC/PPCHazardRecognizers.cpp
+++ b/lib/Target/PowerPC/PPCHazardRecognizers.cpp
@@ -14,6 +14,7 @@
#define DEBUG_TYPE "sched"
#include "PPCHazardRecognizers.h"
#include "PPC.h"
+#include "PPCInstrInfo.h"
#include "llvm/Support/Debug.h"
#include <iostream>
using namespace llvm;
@@ -23,15 +24,14 @@
// PowerPC 970 Hazard Recognizer
//
// This models the dispatch group formation of the PPC970 processor. Dispatch
-// groups are bundles of up to five instructions that can contain up to two ALU
-// (aka FXU) ops, two FPU ops, two Load/Store ops, one CR op, one VALU op, one
-// VPERM op, and one BRANCH op. If the code contains more instructions in a
-// sequence than the dispatch group can contain (e.g. three loads in a row) the
-// processor terminates the dispatch group early, wasting execution resources.
+// groups are bundles of up to five instructions that can contain various mixes
+// of instructions. The PPC970 can dispatch a peak of 4 non-branch and one
+// branch instruction per-cycle.
//
-// In addition to these restrictions, there are a number of other restrictions:
-// some instructions, e.g. branches, are required to be the last instruction in
-// a group. Additionally, only branches can issue in the 5th (last) slot.
+// There are a number of restrictions to dispatch group formation: some
+// instructions can only be issued in the first slot of a dispatch group, & some
+// instructions fill an entire dispatch group. Additionally, only branches can
+// issue in the 5th (last) slot.
//
// Finally, there are a number of "structural" hazards on the PPC970. These
// conditions cause large performance penalties due to misprediction, recovery,
@@ -41,8 +41,6 @@
// conditions, we insert no-op instructions when appropriate.
//
// FIXME: This is missing some significant cases:
-// -1. Handle all of the instruction types in GetInstrType.
-// 0. Handling of instructions that must be the first/last in a group.
// 1. Modeling of microcoded instructions.
// 2. Handling of cracked instructions.
// 3. Handling of serialized operations.
@@ -50,103 +48,70 @@
// e.g. integer divides that only execute in the second slot.
//
-PPCHazardRecognizer970::PPCHazardRecognizer970() {
+PPCHazardRecognizer970::PPCHazardRecognizer970(const TargetInstrInfo &tii)
+ : TII(tii) {
EndDispatchGroup();
}
void PPCHazardRecognizer970::EndDispatchGroup() {
DEBUG(std::cerr << "=== Start of dispatch group\n");
- // Pipeline units.
- NumFXU = NumLSU = NumFPU = 0;
- HasCR = HasSPR = HasVALU = HasVPERM = false;
NumIssued = 0;
// Structural hazard info.
HasCTRSet = false;
- StorePtr1 = StorePtr2 = SDOperand();
- StoreSize = 0;
+ NumStores = 0;
}
-PPCHazardRecognizer970::PPC970InstrType
-PPCHazardRecognizer970::GetInstrType(unsigned Opcode) {
- if (Opcode < ISD::BUILTIN_OP_END)
- return PseudoInst;
+PPCII::PPC970_Unit
+PPCHazardRecognizer970::GetInstrType(unsigned Opcode,
+ bool &isFirst, bool &isSingle,
+ bool &isLoad, bool &isStore){
+ if (Opcode < ISD::BUILTIN_OP_END) {
+ isFirst = isSingle = isLoad = isStore = false;
+ return PPCII::PPC970_Pseudo;
+ }
Opcode -= ISD::BUILTIN_OP_END;
- switch (Opcode) {
- case PPC::FMRSD: return PseudoInst; // Usually coallesced away.
- case PPC::BCTRL:
- case PPC::BL:
- case PPC::BLA:
- case PPC::BLR:
- return BR;
- case PPC::MCRF:
- case PPC::MFCR:
- case PPC::MFOCRF:
- return CR;
- case PPC::MFLR:
- case PPC::MFCTR:
- case PPC::MTLR:
- case PPC::MTCTR:
- return SPR;
- case PPC::LFS:
- case PPC::LFD:
- case PPC::LWZ:
- case PPC::LFSX:
- case PPC::LWZX:
- case PPC::LBZ:
- case PPC::LHA:
- case PPC::LHZ:
- case PPC::LWZU:
- return LSU_LD;
- case PPC::STFS:
- case PPC::STFD:
- case PPC::STW:
- case PPC::STB:
- case PPC::STH:
- case PPC::STWU:
- return LSU_ST;
- case PPC::DIVW:
- case PPC::DIVWU:
- case PPC::DIVD:
- case PPC::DIVDU:
- return FXU_FIRST;
- case PPC::FADDS:
- case PPC::FCTIWZ:
- case PPC::FRSP:
- case PPC::FSUB:
- return FPU;
- }
+ const TargetInstrDescriptor &TID = TII.get(Opcode);
- return FXU;
+ isLoad = TID.Flags & M_LOAD_FLAG;
+ isStore = TID.Flags & M_STORE_FLAG;
+
+ unsigned TSFlags = TID.TSFlags;
+
+ isFirst = TSFlags & PPCII::PPC970_First;
+ isSingle = TSFlags & PPCII::PPC970_Single;
+ return (PPCII::PPC970_Unit)(TSFlags & PPCII::PPC970_Mask);
}
/// isLoadOfStoredAddress - If we have a load from the previously stored pointer
/// as indicated by StorePtr1/StorePtr2/StoreSize, return true.
bool PPCHazardRecognizer970::
isLoadOfStoredAddress(unsigned LoadSize, SDOperand Ptr1, SDOperand Ptr2) const {
- // Handle exact and commuted addresses.
- if (Ptr1 == StorePtr1 && Ptr2 == StorePtr2)
- return true;
- if (Ptr2 == StorePtr1 && Ptr1 == StorePtr2)
- return true;
-
- // Okay, we don't have an exact match, if this is an indexed offset, see if we
- // have overlap (which happens during fp->int conversion for example).
- if (StorePtr2 == Ptr2) {
- if (ConstantSDNode *StoreOffset = dyn_cast<ConstantSDNode>(StorePtr1))
- if (ConstantSDNode *LoadOffset = dyn_cast<ConstantSDNode>(Ptr1)) {
- // Okay the base pointers match, so we have [c1+r] vs [c2+r]. Check to
- // see if the load and store actually overlap.
- int StoreOffs = StoreOffset->getValue();
- int LoadOffs = LoadOffset->getValue();
- if (StoreOffs < LoadOffs) {
- if (int(StoreOffs+StoreSize) > LoadOffs) return true;
- } else {
- if (int(LoadOffs+LoadSize) > StoreOffs) return true;
+ for (unsigned i = 0, e = NumStores; i != e; ++i) {
+ // Handle exact and commuted addresses.
+ if (Ptr1 == StorePtr1[i] && Ptr2 == StorePtr2[i])
+ return true;
+ if (Ptr2 == StorePtr1[i] && Ptr1 == StorePtr2[i])
+ return true;
+
+ // Okay, we don't have an exact match, if this is an indexed offset, see if
+ // we have overlap (which happens during fp->int conversion for example).
+ if (StorePtr2[i] == Ptr2) {
+ if (ConstantSDNode *StoreOffset = dyn_cast<ConstantSDNode>(StorePtr1[i]))
+ if (ConstantSDNode *LoadOffset = dyn_cast<ConstantSDNode>(Ptr1)) {
+ // Okay the base pointers match, so we have [c1+r] vs [c2+r]. Check
+ // to see if the load and store actually overlap.
+ int StoreOffs = StoreOffset->getValue();
+ int LoadOffs = LoadOffset->getValue();
+ if (StoreOffs < LoadOffs) {
+ if (int(StoreOffs+StoreSize) > LoadOffs) return true;
+ } else {
+ if (int(LoadOffs+LoadSize) > StoreOffs) return true;
+ }
}
- }
+ }
}
return false;
}
@@ -157,53 +122,76 @@
/// pipeline flush.
HazardRecognizer::HazardType PPCHazardRecognizer970::
getHazardType(SDNode *Node) {
- PPC970InstrType InstrType = GetInstrType(Node->getOpcode());
- if (InstrType == PseudoInst) return NoHazard;
+ bool isFirst, isSingle, isLoad, isStore;
+ PPCII::PPC970_Unit InstrType =
+ GetInstrType(Node->getOpcode(), isFirst, isSingle, isLoad, isStore);
+ if (InstrType == PPCII::PPC970_Pseudo) return NoHazard;
unsigned Opcode = Node->getOpcode()-ISD::BUILTIN_OP_END;
+ // We can only issue a PPC970_First/PPC970_Single instruction (such as
+ // crand/mtspr/etc) if this is the first cycle of the dispatch group.
+ if (NumIssued != 0 && (isFirst || isSingle) )
+ return Hazard;
+
switch (InstrType) {
default: assert(0 && "Unknown instruction type!");
- case FXU:
- case FXU_FIRST: if (NumFXU == 2) return Hazard;
- case LSU_ST:
- case LSU_LD: if (NumLSU == 2) return Hazard;
- case FPU: if (NumFPU == 2) return Hazard;
- case CR: if (HasCR) return Hazard;
- case SPR: if (HasSPR) return Hazard;
- case VALU: if (HasVALU) return Hazard;
- case VPERM: if (HasVPERM) return Hazard;
- case BR: break;
+ case PPCII::PPC970_FXU:
+ case PPCII::PPC970_LSU:
+ case PPCII::PPC970_FPU:
+ case PPCII::PPC970_VALU:
+ case PPCII::PPC970_VPERM:
+ // We can only issue a branch as the last instruction in a group.
+ if (NumIssued == 4) return Hazard;
+ break;
+ case PPCII::PPC970_CRU:
+ // We can only issue a CR instruction in the first two slots.
+ if (NumIssued >= 2) return Hazard;
+ break;
+ case PPCII::PPC970_BRU:
+ break;
}
-
- // We can only issue a CR or SPR instruction, or an FXU instruction that needs
- // to lead a dispatch group as the first instruction in the group.
- if (NumIssued != 0 &&
- (InstrType == CR || InstrType == SPR || InstrType == FXU_FIRST))
- return Hazard;
-
- // We can only issue a branch as the last instruction in a group.
- if (NumIssued == 4 && InstrType != BR)
- return Hazard;
-
+
// Do not allow MTCTR and BCTRL to be in the same dispatch group.
if (HasCTRSet && Opcode == PPC::BCTRL)
return NoopHazard;
// If this is a load following a store, make sure it's not to the same or
// overlapping address.
- if (InstrType == LSU_LD && StoreSize) {
+ if (isLoad && StoreSize) {
unsigned LoadSize;
switch (Opcode) {
default: assert(0 && "Unknown load!");
- case PPC::LBZ: LoadSize = 1; break;
+ case PPC::LBZ:
+ case PPC::LBZX:
+ case PPC::LVEBX:
+ LoadSize = 1;
+ break;
case PPC::LHA:
- case PPC::LHZ: LoadSize = 2; break;
- case PPC::LWZU:
- case PPC::LFSX:
+ case PPC::LHAX:
+ case PPC::LHZ:
+ case PPC::LHZX:
+ case PPC::LVEHX:
+ LoadSize = 2;
+ break;
case PPC::LFS:
+ case PPC::LFSX:
+ case PPC::LWZ:
case PPC::LWZX:
- case PPC::LWZ: LoadSize = 4; break;
- case PPC::LFD: LoadSize = 8; break;
+ case PPC::LWZU:
+ case PPC::LWA:
+ case PPC::LWAX:
+ case PPC::LVEWX:
+ LoadSize = 4;
+ break;
+ case PPC::LFD:
+ case PPC::LFDX:
+ case PPC::LD:
+ case PPC::LDX:
+ LoadSize = 8;
+ break;
+ case PPC::LVX:
+ LoadSize = 16;
+ break;
}
if (isLoadOfStoredAddress(LoadSize,
@@ -215,41 +203,61 @@
}
void PPCHazardRecognizer970::EmitInstruction(SDNode *Node) {
- PPC970InstrType InstrType = GetInstrType(Node->getOpcode());
- if (InstrType == PseudoInst) return;
+ bool isFirst, isSingle, isLoad, isStore;
+ PPCII::PPC970_Unit InstrType =
+ GetInstrType(Node->getOpcode(), isFirst, isSingle, isLoad, isStore);
+ if (InstrType == PPCII::PPC970_Pseudo) return;
unsigned Opcode = Node->getOpcode()-ISD::BUILTIN_OP_END;
// Update structural hazard information.
if (Opcode == PPC::MTCTR) HasCTRSet = true;
// Track the address stored to.
- if (InstrType == LSU_ST) {
- StorePtr1 = Node->getOperand(1);
- StorePtr2 = Node->getOperand(2);
+ if (isStore) {
+ unsigned ThisStoreSize;
switch (Opcode) {
default: assert(0 && "Unknown store instruction!");
- case PPC::STB: StoreSize = 1; break;
- case PPC::STH: StoreSize = 2; break;
+ case PPC::STBX:
+ case PPC::STB:
+ case PPC::STVEBX:
+ ThisStoreSize = 1;
+ break;
+ case PPC::STHX:
+ case PPC::STH:
+ case PPC::STVEHX:
+ ThisStoreSize = 2;
+ break;
case PPC::STFS:
+ case PPC::STFSX:
case PPC::STWU:
- case PPC::STW: StoreSize = 4; break;
- case PPC::STFD: StoreSize = 8; break;
+ case PPC::STWX:
+ case PPC::STWUX:
+ case PPC::STW:
+ case PPC::STVEWX:
+ case PPC::STFIWX:
+ ThisStoreSize = 4;
+ break;
+ case PPC::STD:
+ case PPC::STDU:
+ case PPC::STFD:
+ case PPC::STFDX:
+ case PPC::STDX:
+ case PPC::STDUX:
+ ThisStoreSize = 8;
+ break;
+ case PPC::STVX:
+ ThisStoreSize = 16;
+ break;
}
+
+ StoreSize[NumStores] = ThisStoreSize;
+ StorePtr1[NumStores] = Node->getOperand(1);
+ StorePtr2[NumStores] = Node->getOperand(2);
+ ++NumStores;
}
- switch (InstrType) {
- default: assert(0 && "Unknown instruction type!");
- case FXU:
- case FXU_FIRST: ++NumFXU; break;
- case LSU_LD:
- case LSU_ST: ++NumLSU; break;
- case FPU: ++NumFPU; break;
- case CR: HasCR = true; break;
- case SPR: HasSPR = true; break;
- case VALU: HasVALU = true; break;
- case VPERM: HasVPERM = true; break;
- case BR: NumIssued = 4; return; // ends a d-group.
- }
+ if (InstrType == PPCII::PPC970_BRU || isSingle)
+ NumIssued = 4; // Terminate a d-group.
++NumIssued;
if (NumIssued == 5)