// Instructions with folded loads need to read the memory operand immediately,
// but other register operands don't have to be read until the load is ready.
-// These are marked with ReadAfterLd.
+// These operands are marked with ReadAfterLd.
def ReadAfterLd : SchedRead;
// Instructions with both a load and a store folded are modeled as a folded
// load + WriteRMW.
def WriteRMW : SchedWrite;
-// Most instructions can fold loads, so every SchedWrite comes in two variants:
-// With and without a folded load.
+// Most instructions can fold loads, so almost every SchedWrite comes in two
+// variants: With and without a folded load.
+// An X86FoldableSchedWrite holds a reference to the corresponding SchedWrite
+// with a folded load.
+class X86FoldableSchedWrite : SchedWrite {
+ // The SchedWrite to use when a load is folded into the instruction.
+ SchedWrite Folded;
+}
-// Arithmetic.
-def WriteALU : SchedWrite; // Simple integer ALU op.
-def WriteALULd : SchedWrite; // ALU op with folded load.
-def WriteIMul : SchedWrite; // Integer multiplication.
-def WriteIMulLd : SchedWrite;
-def WriteIDiv : SchedWrite; // Integer division.
-def WriteIDivLd : SchedWrite;
-def WriteLEA : SchedWrite; // LEA instructions can't fold loads.
+// Multiclass that produces a linked pair of SchedWrites.
+multiclass X86SchedWritePair {
+ // Register-Memory operation.
+ def Ld : SchedWrite;
+ // Register-Register operation.
+ def NAME : X86FoldableSchedWrite {
+ let Folded = !cast<SchedWrite>(NAME#"Ld");
+ }
+}
+// Arithmetic.
+defm WriteALU : X86SchedWritePair; // Simple integer ALU op.
+defm WriteIMul : X86SchedWritePair; // Integer multiplication.
+def WriteIMulH : SchedWrite; // Integer multiplication, high part.
+defm WriteIDiv : X86SchedWritePair; // Integer division.
+def WriteLEA : SchedWrite; // LEA instructions can't fold loads.
+
+// Integer shifts and rotates.
+defm WriteShift : X86SchedWritePair;
+
+// Loads, stores, and moves, not folded with other operations.
+def WriteLoad : SchedWrite;
+def WriteStore : SchedWrite;
+def WriteMove : SchedWrite;
+
+// Idioms that clear a register, like xorps %xmm0, %xmm0.
+// These can often bypass execution ports completely.
+def WriteZero : SchedWrite;
+
+// Branches don't produce values, so they have no latency, but they still
+// consume resources. Indirect branches can fold loads.
+defm WriteJump : X86SchedWritePair;
+
+// Floating point. This covers both scalar and vector operations.
+defm WriteFAdd : X86SchedWritePair; // Floating point add/sub/compare.
+defm WriteFMul : X86SchedWritePair; // Floating point multiplication.
+defm WriteFDiv : X86SchedWritePair; // Floating point division.
+defm WriteFSqrt : X86SchedWritePair; // Floating point square root.
+defm WriteFRcp : X86SchedWritePair; // Floating point reciprocal.
+defm WriteFMA : X86SchedWritePair; // Fused Multiply Add.
+
+// FMA Scheduling helper class.
+class FMASC { X86FoldableSchedWrite Sched = WriteFAdd; }
+
+// Vector integer operations.
+defm WriteVecALU : X86SchedWritePair; // Vector integer ALU op, no logicals.
+defm WriteVecShift : X86SchedWritePair; // Vector integer shifts.
+defm WriteVecIMul : X86SchedWritePair; // Vector integer multiply.
+
+// Vector bitwise operations.
+// These are often used on both floating point and integer vectors.
+defm WriteVecLogic : X86SchedWritePair; // Vector and/or/xor.
+defm WriteShuffle : X86SchedWritePair; // Vector shuffles and blends.
+
+// Conversion between integer and float.
+defm WriteCvtF2I : X86SchedWritePair; // Float -> Integer.
+defm WriteCvtI2F : X86SchedWritePair; // Integer -> Float.
+defm WriteCvtF2F : X86SchedWritePair; // Float -> Float size conversion.
+
+// Catch-all for expensive system instructions.
+def WriteSystem : SchedWrite;
+
+// Old microcoded instructions that nobody use.
+def WriteMicrocoded : SchedWrite;
//===----------------------------------------------------------------------===//
// Instruction Itinerary classes used for X86
-def IIC_DEFAULT : InstrItinClass;
def IIC_ALU_MEM : InstrItinClass;
def IIC_ALU_NONMEM : InstrItinClass;
def IIC_LEA : InstrItinClass;
def IIC_SSE_PABS_RR : InstrItinClass;
def IIC_SSE_PABS_RM : InstrItinClass;
-def IIC_SSE_SQRTP_RR : InstrItinClass;
-def IIC_SSE_SQRTP_RM : InstrItinClass;
-def IIC_SSE_SQRTS_RR : InstrItinClass;
-def IIC_SSE_SQRTS_RM : InstrItinClass;
+def IIC_SSE_SQRTPS_RR : InstrItinClass;
+def IIC_SSE_SQRTPS_RM : InstrItinClass;
+def IIC_SSE_SQRTSS_RR : InstrItinClass;
+def IIC_SSE_SQRTSS_RM : InstrItinClass;
+def IIC_SSE_SQRTPD_RR : InstrItinClass;
+def IIC_SSE_SQRTPD_RM : InstrItinClass;
+def IIC_SSE_SQRTSD_RR : InstrItinClass;
+def IIC_SSE_SQRTSD_RM : InstrItinClass;
def IIC_SSE_RCPP_RR : InstrItinClass;
def IIC_SSE_RCPP_RM : InstrItinClass;
// Resources beyond the decoder operate on micro-ops and are bufferred
// so adjacent micro-ops don't directly compete.
//
-// MinLatency=0 indicates that RAW dependencies can be decoded in the
-// same cycle.
+// MicroOpBufferSize > 1 indicates that RAW dependencies can be
+// decoded in the same cycle. The value 32 is a reasonably arbitrary
+// number of in-flight instructions.
//
// HighLatency=10 is optimistic. X86InstrInfo::isHighLatencyDef
// indicates high latency opcodes. Alternatively, InstrItinData
// latencies. Since these latencies are not used for pipeline hazards,
// they do not need to be exact.
//
-// ILPWindow=10 is an arbitrary threshold that approximates cycles of
-// latency hidden by instruction buffers. The actual value is not very
-// important but should be zero for inorder and nonzero for OOO processors.
-//
// The GenericModel contains no instruciton itineraries.
def GenericModel : SchedMachineModel {
let IssueWidth = 4;
- let MinLatency = 0;
+ let MicroOpBufferSize = 32;
let LoadLatency = 4;
let HighLatency = 10;
- let ILPWindow = 10;
}
include "X86ScheduleAtom.td"
+include "X86SchedSandyBridge.td"
+include "X86SchedHaswell.td"