llvm/lib/Target/ARM/ARMScheduleV6.td
Evan Cheng 9fe2009956 Sorry, several patches in one.
TargetInstrInfo:
Change produceSameValue() to take MachineRegisterInfo as an optional argument.
When in SSA form, targets can use it to make more aggressive equality analysis.

Machine LICM:
1. Eliminate isLoadFromConstantMemory, use MI.isInvariantLoad instead.
2. Fix a bug which prevent CSE of instructions which are not re-materializable.
3. Use improved form of produceSameValue.

ARM:
1. Teach ARM produceSameValue to look pass some PIC labels.
2. Look for operands from different loads of different constant pool entries
   which have same values.
3. Re-implement PIC GA materialization using movw + movt. Combine the pair with
   a "add pc" or "ldr [pc]" to form pseudo instructions. This makes it possible
   to re-materialize the instruction, allow machine LICM to hoist the set of
   instructions out of the loop and make it possible to CSE them. It's a bit
   hacky, but it significantly improve code quality.
4. Some minor bug fixes as well.

With the fixes, using movw + movt to materialize GAs significantly outperform the
load from constantpool method. 186.crafty and 255.vortex improved > 20%, 254.gap
and 176.gcc ~10%.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@123905 91177308-0d34-0410-b5e6-96231b3b80d8
2011-01-20 08:34:58 +00:00

295 lines
12 KiB
TableGen

//===- ARMScheduleV6.td - ARM v6 Scheduling Definitions ----*- tablegen -*-===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file defines the itinerary class data for the ARM v6 processors.
//
//===----------------------------------------------------------------------===//
// Model based on ARM1176
//
// Functional Units
def V6_Pipe : FuncUnit; // pipeline
// Scheduling information derived from "ARM1176JZF-S Technical Reference Manual"
//
def ARMV6Itineraries : ProcessorItineraries<
[V6_Pipe], [], [
//
// No operand cycles
InstrItinData<IIC_iALUx , [InstrStage<1, [V6_Pipe]>]>,
//
// Binary Instructions that produce a result
InstrItinData<IIC_iALUi , [InstrStage<1, [V6_Pipe]>], [2, 2]>,
InstrItinData<IIC_iALUr , [InstrStage<1, [V6_Pipe]>], [2, 2, 2]>,
InstrItinData<IIC_iALUsi , [InstrStage<1, [V6_Pipe]>], [2, 2, 1]>,
InstrItinData<IIC_iALUsr , [InstrStage<2, [V6_Pipe]>], [3, 3, 2, 1]>,
//
// Bitwise Instructions that produce a result
InstrItinData<IIC_iBITi , [InstrStage<1, [V6_Pipe]>], [2, 2]>,
InstrItinData<IIC_iBITr , [InstrStage<1, [V6_Pipe]>], [2, 2, 2]>,
InstrItinData<IIC_iBITsi , [InstrStage<1, [V6_Pipe]>], [2, 2, 1]>,
InstrItinData<IIC_iBITsr , [InstrStage<2, [V6_Pipe]>], [3, 3, 2, 1]>,
//
// Unary Instructions that produce a result
InstrItinData<IIC_iUNAr , [InstrStage<1, [V6_Pipe]>], [2, 2]>,
InstrItinData<IIC_iUNAsi , [InstrStage<1, [V6_Pipe]>], [2, 1]>,
//
// Zero and sign extension instructions
InstrItinData<IIC_iEXTr , [InstrStage<1, [V6_Pipe]>], [1, 1]>,
InstrItinData<IIC_iEXTAr , [InstrStage<1, [V6_Pipe]>], [2, 2, 1]>,
InstrItinData<IIC_iEXTAsr , [InstrStage<2, [V6_Pipe]>], [3, 3, 2, 1]>,
//
// Compare instructions
InstrItinData<IIC_iCMPi , [InstrStage<1, [V6_Pipe]>], [2]>,
InstrItinData<IIC_iCMPr , [InstrStage<1, [V6_Pipe]>], [2, 2]>,
InstrItinData<IIC_iCMPsi , [InstrStage<1, [V6_Pipe]>], [2, 1]>,
InstrItinData<IIC_iCMPsr , [InstrStage<2, [V6_Pipe]>], [3, 2, 1]>,
//
// Test instructions
InstrItinData<IIC_iTSTi , [InstrStage<1, [V6_Pipe]>], [2]>,
InstrItinData<IIC_iTSTr , [InstrStage<1, [V6_Pipe]>], [2, 2]>,
InstrItinData<IIC_iTSTsi , [InstrStage<1, [V6_Pipe]>], [2, 1]>,
InstrItinData<IIC_iTSTsr , [InstrStage<2, [V6_Pipe]>], [3, 2, 1]>,
//
// Move instructions, unconditional
InstrItinData<IIC_iMOVi , [InstrStage<1, [V6_Pipe]>], [2]>,
InstrItinData<IIC_iMOVr , [InstrStage<1, [V6_Pipe]>], [2, 2]>,
InstrItinData<IIC_iMOVsi , [InstrStage<1, [V6_Pipe]>], [2, 1]>,
InstrItinData<IIC_iMOVsr , [InstrStage<2, [V6_Pipe]>], [3, 2, 1]>,
InstrItinData<IIC_iMOVix2 , [InstrStage<1, [V6_Pipe]>,
InstrStage<1, [V6_Pipe]>], [2]>,
InstrItinData<IIC_iMOVix2addpc,[InstrStage<1, [V6_Pipe]>,
InstrStage<1, [V6_Pipe]>,
InstrStage<1, [V6_Pipe]>], [3]>,
InstrItinData<IIC_iMOVix2ld , [InstrStage<1, [V6_Pipe]>,
InstrStage<1, [V6_Pipe]>,
InstrStage<1, [V6_Pipe]>], [5]>,
//
// Move instructions, conditional
InstrItinData<IIC_iCMOVi , [InstrStage<1, [V6_Pipe]>], [3]>,
InstrItinData<IIC_iCMOVr , [InstrStage<1, [V6_Pipe]>], [3, 2]>,
InstrItinData<IIC_iCMOVsi , [InstrStage<1, [V6_Pipe]>], [3, 1]>,
InstrItinData<IIC_iCMOVsr , [InstrStage<1, [V6_Pipe]>], [4, 2, 1]>,
InstrItinData<IIC_iCMOVix2 , [InstrStage<1, [V6_Pipe]>,
InstrStage<1, [V6_Pipe]>], [4]>,
//
// MVN instructions
InstrItinData<IIC_iMVNi , [InstrStage<1, [V6_Pipe]>], [2]>,
InstrItinData<IIC_iMVNr , [InstrStage<1, [V6_Pipe]>], [2, 2]>,
InstrItinData<IIC_iMVNsi , [InstrStage<1, [V6_Pipe]>], [2, 1]>,
InstrItinData<IIC_iMVNsr , [InstrStage<2, [V6_Pipe]>], [3, 2, 1]>,
// Integer multiply pipeline
//
InstrItinData<IIC_iMUL16 , [InstrStage<1, [V6_Pipe]>], [4, 1, 1]>,
InstrItinData<IIC_iMAC16 , [InstrStage<1, [V6_Pipe]>], [4, 1, 1, 2]>,
InstrItinData<IIC_iMUL32 , [InstrStage<2, [V6_Pipe]>], [5, 1, 1]>,
InstrItinData<IIC_iMAC32 , [InstrStage<2, [V6_Pipe]>], [5, 1, 1, 2]>,
InstrItinData<IIC_iMUL64 , [InstrStage<3, [V6_Pipe]>], [6, 1, 1]>,
InstrItinData<IIC_iMAC64 , [InstrStage<3, [V6_Pipe]>], [6, 1, 1, 2]>,
// Integer load pipeline
//
// Immediate offset
InstrItinData<IIC_iLoad_i , [InstrStage<1, [V6_Pipe]>], [4, 1]>,
InstrItinData<IIC_iLoad_bh_i, [InstrStage<1, [V6_Pipe]>], [4, 1]>,
InstrItinData<IIC_iLoad_d_i , [InstrStage<1, [V6_Pipe]>], [4, 1]>,
//
// Register offset
InstrItinData<IIC_iLoad_r , [InstrStage<1, [V6_Pipe]>], [4, 1, 1]>,
InstrItinData<IIC_iLoad_bh_r, [InstrStage<1, [V6_Pipe]>], [4, 1, 1]>,
InstrItinData<IIC_iLoad_d_r , [InstrStage<1, [V6_Pipe]>], [4, 1, 1]>,
//
// Scaled register offset, issues over 2 cycles
InstrItinData<IIC_iLoad_si , [InstrStage<2, [V6_Pipe]>], [5, 2, 1]>,
InstrItinData<IIC_iLoad_bh_si, [InstrStage<2, [V6_Pipe]>], [5, 2, 1]>,
//
// Immediate offset with update
InstrItinData<IIC_iLoad_iu , [InstrStage<1, [V6_Pipe]>], [4, 2, 1]>,
InstrItinData<IIC_iLoad_bh_iu, [InstrStage<1, [V6_Pipe]>], [4, 2, 1]>,
//
// Register offset with update
InstrItinData<IIC_iLoad_ru , [InstrStage<1, [V6_Pipe]>], [4, 2, 1, 1]>,
InstrItinData<IIC_iLoad_bh_ru, [InstrStage<1, [V6_Pipe]>], [4, 2, 1, 1]>,
InstrItinData<IIC_iLoad_d_ru , [InstrStage<1, [V6_Pipe]>], [4, 2, 1, 1]>,
//
// Scaled register offset with update, issues over 2 cycles
InstrItinData<IIC_iLoad_siu, [InstrStage<2, [V6_Pipe]>], [5, 2, 2, 1]>,
InstrItinData<IIC_iLoad_bh_siu,[InstrStage<2, [V6_Pipe]>], [5, 2, 2, 1]>,
//
// Load multiple, def is the 5th operand.
InstrItinData<IIC_iLoad_m , [InstrStage<3, [V6_Pipe]>], [1, 1, 1, 1, 4]>,
//
// Load multiple + update, defs are the 1st and 5th operands.
InstrItinData<IIC_iLoad_mu , [InstrStage<3, [V6_Pipe]>], [2, 1, 1, 1, 4]>,
//
// Load multiple plus branch
InstrItinData<IIC_iLoad_mBr, [InstrStage<3, [V6_Pipe]>,
InstrStage<1, [V6_Pipe]>], [1, 2, 1, 1, 4]>,
//
// iLoadi + iALUr for t2LDRpci_pic.
InstrItinData<IIC_iLoadiALU, [InstrStage<1, [V6_Pipe]>,
InstrStage<1, [V6_Pipe]>], [3, 1]>,
//
// Pop, def is the 3rd operand.
InstrItinData<IIC_iPop , [InstrStage<3, [V6_Pipe]>], [1, 1, 4]>,
//
// Pop + branch, def is the 3rd operand.
InstrItinData<IIC_iPop_Br, [InstrStage<3, [V6_Pipe]>,
InstrStage<1, [V6_Pipe]>], [1, 2, 4]>,
// Integer store pipeline
//
// Immediate offset
InstrItinData<IIC_iStore_i , [InstrStage<1, [V6_Pipe]>], [2, 1]>,
InstrItinData<IIC_iStore_bh_i, [InstrStage<1, [V6_Pipe]>], [2, 1]>,
InstrItinData<IIC_iStore_d_i , [InstrStage<1, [V6_Pipe]>], [2, 1]>,
//
// Register offset
InstrItinData<IIC_iStore_r , [InstrStage<1, [V6_Pipe]>], [2, 1, 1]>,
InstrItinData<IIC_iStore_bh_r, [InstrStage<1, [V6_Pipe]>], [2, 1, 1]>,
InstrItinData<IIC_iStore_d_r , [InstrStage<1, [V6_Pipe]>], [2, 1, 1]>,
//
// Scaled register offset, issues over 2 cycles
InstrItinData<IIC_iStore_si , [InstrStage<2, [V6_Pipe]>], [2, 2, 1]>,
InstrItinData<IIC_iStore_bh_si, [InstrStage<2, [V6_Pipe]>], [2, 2, 1]>,
//
// Immediate offset with update
InstrItinData<IIC_iStore_iu , [InstrStage<1, [V6_Pipe]>], [2, 2, 1]>,
InstrItinData<IIC_iStore_bh_iu, [InstrStage<1, [V6_Pipe]>], [2, 2, 1]>,
//
// Register offset with update
InstrItinData<IIC_iStore_ru, [InstrStage<1, [V6_Pipe]>], [2, 2, 1, 1]>,
InstrItinData<IIC_iStore_bh_ru,[InstrStage<1, [V6_Pipe]>], [2, 2, 1, 1]>,
InstrItinData<IIC_iStore_d_ru, [InstrStage<1, [V6_Pipe]>], [2, 2, 1, 1]>,
//
// Scaled register offset with update, issues over 2 cycles
InstrItinData<IIC_iStore_siu, [InstrStage<2, [V6_Pipe]>], [2, 2, 2, 1]>,
InstrItinData<IIC_iStore_bh_siu,[InstrStage<2, [V6_Pipe]>], [2, 2, 2, 1]>,
//
// Store multiple
InstrItinData<IIC_iStore_m , [InstrStage<3, [V6_Pipe]>]>,
//
// Store multiple + update
InstrItinData<IIC_iStore_mu , [InstrStage<3, [V6_Pipe]>], [2]>,
// Branch
//
// no delay slots, so the latency of a branch is unimportant
InstrItinData<IIC_Br , [InstrStage<1, [V6_Pipe]>]>,
// VFP
// Issue through integer pipeline, and execute in NEON unit. We assume
// RunFast mode so that NFP pipeline is used for single-precision when
// possible.
//
// FP Special Register to Integer Register File Move
InstrItinData<IIC_fpSTAT , [InstrStage<1, [V6_Pipe]>], [3]>,
//
// Single-precision FP Unary
InstrItinData<IIC_fpUNA32 , [InstrStage<1, [V6_Pipe]>], [5, 2]>,
//
// Double-precision FP Unary
InstrItinData<IIC_fpUNA64 , [InstrStage<1, [V6_Pipe]>], [5, 2]>,
//
// Single-precision FP Compare
InstrItinData<IIC_fpCMP32 , [InstrStage<1, [V6_Pipe]>], [2, 2]>,
//
// Double-precision FP Compare
InstrItinData<IIC_fpCMP64 , [InstrStage<1, [V6_Pipe]>], [2, 2]>,
//
// Single to Double FP Convert
InstrItinData<IIC_fpCVTSD , [InstrStage<1, [V6_Pipe]>], [5, 2]>,
//
// Double to Single FP Convert
InstrItinData<IIC_fpCVTDS , [InstrStage<1, [V6_Pipe]>], [5, 2]>,
//
// Single-Precision FP to Integer Convert
InstrItinData<IIC_fpCVTSI , [InstrStage<1, [V6_Pipe]>], [9, 2]>,
//
// Double-Precision FP to Integer Convert
InstrItinData<IIC_fpCVTDI , [InstrStage<1, [V6_Pipe]>], [9, 2]>,
//
// Integer to Single-Precision FP Convert
InstrItinData<IIC_fpCVTIS , [InstrStage<1, [V6_Pipe]>], [9, 2]>,
//
// Integer to Double-Precision FP Convert
InstrItinData<IIC_fpCVTID , [InstrStage<1, [V6_Pipe]>], [9, 2]>,
//
// Single-precision FP ALU
InstrItinData<IIC_fpALU32 , [InstrStage<1, [V6_Pipe]>], [9, 2, 2]>,
//
// Double-precision FP ALU
InstrItinData<IIC_fpALU64 , [InstrStage<1, [V6_Pipe]>], [9, 2, 2]>,
//
// Single-precision FP Multiply
InstrItinData<IIC_fpMUL32 , [InstrStage<1, [V6_Pipe]>], [9, 2, 2]>,
//
// Double-precision FP Multiply
InstrItinData<IIC_fpMUL64 , [InstrStage<2, [V6_Pipe]>], [9, 2, 2]>,
//
// Single-precision FP MAC
InstrItinData<IIC_fpMAC32 , [InstrStage<1, [V6_Pipe]>], [9, 2, 2, 2]>,
//
// Double-precision FP MAC
InstrItinData<IIC_fpMAC64 , [InstrStage<2, [V6_Pipe]>], [9, 2, 2, 2]>,
//
// Single-precision FP DIV
InstrItinData<IIC_fpDIV32 , [InstrStage<15, [V6_Pipe]>], [20, 2, 2]>,
//
// Double-precision FP DIV
InstrItinData<IIC_fpDIV64 , [InstrStage<29, [V6_Pipe]>], [34, 2, 2]>,
//
// Single-precision FP SQRT
InstrItinData<IIC_fpSQRT32 , [InstrStage<15, [V6_Pipe]>], [20, 2, 2]>,
//
// Double-precision FP SQRT
InstrItinData<IIC_fpSQRT64 , [InstrStage<29, [V6_Pipe]>], [34, 2, 2]>,
//
// Integer to Single-precision Move
InstrItinData<IIC_fpMOVIS, [InstrStage<1, [V6_Pipe]>], [10, 1]>,
//
// Integer to Double-precision Move
InstrItinData<IIC_fpMOVID, [InstrStage<1, [V6_Pipe]>], [10, 1, 1]>,
//
// Single-precision to Integer Move
InstrItinData<IIC_fpMOVSI, [InstrStage<1, [V6_Pipe]>], [10, 1]>,
//
// Double-precision to Integer Move
InstrItinData<IIC_fpMOVDI, [InstrStage<1, [V6_Pipe]>], [10, 10, 1]>,
//
// Single-precision FP Load
InstrItinData<IIC_fpLoad32 , [InstrStage<1, [V6_Pipe]>], [5, 2, 2]>,
//
// Double-precision FP Load
InstrItinData<IIC_fpLoad64 , [InstrStage<1, [V6_Pipe]>], [5, 2, 2]>,
//
// FP Load Multiple
InstrItinData<IIC_fpLoad_m , [InstrStage<3, [V6_Pipe]>], [2, 1, 1, 5]>,
//
// FP Load Multiple + update
InstrItinData<IIC_fpLoad_mu, [InstrStage<3, [V6_Pipe]>], [3, 2, 1, 1, 5]>,
//
// Single-precision FP Store
InstrItinData<IIC_fpStore32 , [InstrStage<1, [V6_Pipe]>], [2, 2, 2]>,
//
// Double-precision FP Store
// use FU_Issue to enforce the 1 load/store per cycle limit
InstrItinData<IIC_fpStore64 , [InstrStage<1, [V6_Pipe]>], [2, 2, 2]>,
//
// FP Store Multiple
InstrItinData<IIC_fpStore_m, [InstrStage<3, [V6_Pipe]>], [2, 2, 2, 2]>,
//
// FP Store Multiple + update
InstrItinData<IIC_fpStore_mu,[InstrStage<3, [V6_Pipe]>], [3, 2, 2, 2, 2]>
]>;