From 7f155d7d2bb8c702ed0f64393e38ec1cc26e2f08 Mon Sep 17 00:00:00 2001 From: Arnold Schwaighofer Date: Fri, 7 Jun 2013 00:04:28 +0000 Subject: [PATCH] ARM sched model: Add SIMD/VFP load/store instructions on Swift Reapply 183270 again (because three is a magic number). This should now no longer seg fault after r183459. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@183464 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/Target/ARM/ARMScheduleSwift.td | 364 +++++++++++++++++++++++++++++ 1 file changed, 364 insertions(+) diff --git a/lib/Target/ARM/ARMScheduleSwift.td b/lib/Target/ARM/ARMScheduleSwift.td index 79034cb5218..1c9058c2502 100644 --- a/lib/Target/ARM/ARMScheduleSwift.td +++ b/lib/Target/ARM/ARMScheduleSwift.td @@ -1679,6 +1679,370 @@ let SchedModel = SwiftModel in { // Not serializing. def : InstRW<[SwiftWriteP0TwoCycle], (instregex "FMSTAT")>; + // 4.2.39 Advanced SIMD and VFP, Load Single Element + def : InstRW<[SwiftWriteLM4Cy], (instregex "VLDRD$", "VLDRS$")>; + + // 4.2.40 Advanced SIMD and VFP, Store Single Element + def : InstRW<[SwiftWriteLM4Cy], (instregex "VSTRD$", "VSTRS$")>; + + // 4.2.41 Advanced SIMD and VFP, Load Multiple + // 4.2.42 Advanced SIMD and VFP, Store Multiple + + // Resource requirement for permuting, just reserves the resources. + foreach Num = 1-28 in { + def SwiftVLDMPerm#Num : SchedWriteRes<[SwiftUnitP1]> { + let Latency = 0; + let NumMicroOps = Num; + let ResourceCycles = [Num]; + } + } + + // Pre RA pseudos - load/store to a Q register as a D register pair. + def : InstRW<[SwiftWriteLM4Cy], (instregex "VLDMQIA$", "VSTMQIA$")>; + + // Post RA not modelled accurately. We assume that register use of width 64 + // bit maps to a D register, 128 maps to a Q register. Not all different kinds + // are accurately represented. + def SwiftWriteVLDM : SchedWriteVariant<[ + // Load of one S register. + SchedVar, + // Load of one D register. + SchedVar, + // Load of 3 S register. + SchedVar, + // Load of a Q register (not neccessarily true). We should not be mapping to + // 4 S registers, either. + SchedVar, + // Load of 5 S registers. + SchedVar, + // Load of 3 D registers. (Must also be able to handle s register list - + // though, not accurate) + SchedVar, + // Load of 7 S registers. + SchedVar, + // Load of two Q registers. + SchedVar, + // Load of 9 S registers. + SchedVar, + // Load of 5 D registers. + SchedVar, + // Inaccurate: reuse describtion from 9 S registers. + SchedVar, + // Load of three Q registers. + SchedVar, + // Inaccurate: reuse describtion from 9 S registers. + SchedVar, + // Load of 7 D registers inaccurate. + SchedVar, + SchedVar, + // Load of 4 Q registers. + SchedVar, + // Unknow number of registers, just use resources for two registers. + SchedVar + ]> { let Variadic = 1; } + + def : InstRW<[SwiftWriteVLDM], (instregex "VLDM[SD](IA|DB)$")>; + + def : InstRW<[SwiftWriteP01OneCycle2x, SwiftWriteVLDM], + (instregex "VLDM[SD](IA|DB)_UPD$")>; + + def SwiftWriteVSTM : SchedWriteVariant<[ + // One S register. + SchedVar, + // One D register. + SchedVar, + // Three S registers. + SchedVar, + // Assume one Q register. + SchedVar, + SchedVar, + // Assume three D registers. + SchedVar, + SchedVar, + // Assume two Q registers. + SchedVar, + SchedVar, + // Assume 5 D registers. + SchedVar, + SchedVar, + // Asume three Q registers. + SchedVar, + SchedVar, + // Assume 7 D registers. + SchedVar, + SchedVar, + // Assume four Q registers. + SchedVar, + // Asumme two Q registers. + SchedVar + ]> { let Variadic = 1; } + + def : InstRW<[SwiftWriteVSTM], (instregex "VSTM[SD](IA|DB)$")>; + + def : InstRW<[SwiftWriteP01OneCycle2x, SwiftWriteVSTM], + (instregex "VSTM[SD](IA|DB)_UPD")>; + + // 4.2.43 Advanced SIMD, Element or Structure Load and Store + def SwiftWrite2xP2FourCy : SchedWriteRes<[SwiftUnitP2]> { + let Latency = 4; + let ResourceCycles = [2]; + } + def SwiftWrite3xP2FourCy : SchedWriteRes<[SwiftUnitP2]> { + let Latency = 4; + let ResourceCycles = [3]; + } + foreach Num = 1-2 in { + def SwiftExt#Num#xP0 : SchedWriteRes<[SwiftUnitP0]> { + let Latency = 0; + let NumMicroOps = Num; + let ResourceCycles = [Num]; + } + } + // VLDx + // Multiple structures. + // Single element structure loads. + // We assume aligned. + // Single/two register. + def : InstRW<[SwiftWriteLM4Cy], (instregex "VLD1(d|q)(8|16|32|64)$")>; + def : InstRW<[SwiftWriteLM4Cy, SwiftWriteP01OneCycle], + (instregex "VLD1(d|q)(8|16|32|64)wb")>; + // Three register. + def : InstRW<[SwiftWrite3xP2FourCy], + (instregex "VLD1(d|q)(8|16|32|64)T$", "VLD1d64TPseudo")>; + def : InstRW<[SwiftWrite3xP2FourCy, SwiftWriteP01OneCycle], + (instregex "VLD1(d|q)(8|16|32|64)Twb")>; + /// Four Register. + def : InstRW<[SwiftWrite2xP2FourCy], + (instregex "VLD1(d|q)(8|16|32|64)Q$", "VLD1d64QPseudo")>; + def : InstRW<[SwiftWrite2xP2FourCy, SwiftWriteP01OneCycle], + (instregex "VLD1(d|q)(8|16|32|64)Qwb")>; + // Two element structure loads. + // Two/four register. + def : InstRW<[SwiftWriteLM9Cy, SwiftExt2xP0, SwiftVLDMPerm2], + (instregex "VLD2(d|q|b)(8|16|32)$", "VLD2q(8|16|32)Pseudo$")>; + def : InstRW<[SwiftWriteLM9Cy, SwiftWriteP01OneCycle, SwiftExt2xP0, + SwiftVLDMPerm2], + (instregex "VLD2(d|q|b)(8|16|32)wb", "VLD2q(8|16|32)PseudoWB")>; + // Three element structure. + def : InstRW<[SwiftWriteLM9Cy, SwiftWriteLM9CyNo, SwiftWriteLM9CyNo, + SwiftVLDMPerm3, SwiftWrite3xP2FourCy], + (instregex "VLD3(d|q)(8|16|32)$")>; + def : InstRW<[SwiftWriteLM9Cy, SwiftVLDMPerm3, SwiftWrite3xP2FourCy], + (instregex "VLD3(d|q)(8|16|32)(oddP|P)seudo$")>; + + def : InstRW<[SwiftWriteLM9Cy, SwiftWriteLM9CyNo, SwiftWriteLM9CyNo, + SwiftWriteP01OneCycle, SwiftVLDMPerm3, SwiftWrite3xP2FourCy], + (instregex "VLD3(d|q)(8|16|32)_UPD$")>; + def : InstRW<[SwiftWriteLM9Cy, SwiftWriteP01OneCycle, SwiftVLDMPerm3, + SwiftWrite3xP2FourCy], + (instregex "VLD3(d|q)(8|16|32)(oddP|P)seudo_UPD")>; + // Four element structure loads. + def : InstRW<[SwiftWriteLM11Cy, SwiftWriteLM11Cy, SwiftWriteLM11Cy, + SwiftWriteLM11Cy, SwiftExt2xP0, SwiftVLDMPerm4, + SwiftWrite3xP2FourCy], + (instregex "VLD4(d|q)(8|16|32)$")>; + def : InstRW<[SwiftWriteLM11Cy, SwiftExt2xP0, SwiftVLDMPerm4, + SwiftWrite3xP2FourCy], + (instregex "VLD4(d|q)(8|16|32)(oddP|P)seudo$")>; + def : InstRW<[SwiftWriteLM11Cy, SwiftWriteLM11Cy, SwiftWriteLM11Cy, + SwiftWriteLM11Cy, SwiftWriteP01OneCycle, SwiftExt2xP0, + SwiftVLDMPerm4, SwiftWrite3xP2FourCy], + (instregex "VLD4(d|q)(8|16|32)_UPD")>; + def : InstRW<[SwiftWriteLM11Cy, SwiftWriteP01OneCycle, SwiftExt2xP0, + SwiftVLDMPerm4, SwiftWrite3xP2FourCy], + (instregex "VLD4(d|q)(8|16|32)(oddP|P)seudo_UPD")>; + + // Single all/lane loads. + // One element structure. + def : InstRW<[SwiftWriteLM6Cy, SwiftVLDMPerm2], + (instregex "VLD1(LN|DUP)(d|q)(8|16|32)$", "VLD1(LN|DUP)(d|q)(8|16|32)Pseudo$")>; + def : InstRW<[SwiftWriteLM6Cy, SwiftWriteP01OneCycle, SwiftVLDMPerm2], + (instregex "VLD1(LN|DUP)(d|q)(8|16|32)(wb|_UPD)", + "VLD1LNq(8|16|32)Pseudo_UPD")>; + // Two element structure. + def : InstRW<[SwiftWriteLM6Cy, SwiftWriteLM6Cy, SwiftExt1xP0, SwiftVLDMPerm2], + (instregex "VLD2(DUP|LN)(d|q)(8|16|32|8x2|16x2|32x2)$", + "VLD2LN(d|q)(8|16|32)Pseudo$")>; + def : InstRW<[SwiftWriteLM6Cy, SwiftWriteLM6Cy, SwiftWriteP01OneCycle, + SwiftExt1xP0, SwiftVLDMPerm2], + (instregex "VLD2LN(d|q)(8|16|32)_UPD$")>; + def : InstRW<[SwiftWriteLM6Cy, SwiftWriteP01OneCycle, SwiftWriteLM6Cy, + SwiftExt1xP0, SwiftVLDMPerm2], + (instregex "VLD2DUPd(8|16|32|8x2|16x2|32x2)wb")>; + def : InstRW<[SwiftWriteLM6Cy, SwiftWriteP01OneCycle, SwiftWriteLM6Cy, + SwiftExt1xP0, SwiftVLDMPerm2], + (instregex "VLD2LN(d|q)(8|16|32)Pseudo_UPD")>; + // Three element structure. + def : InstRW<[SwiftWriteLM7Cy, SwiftWriteLM8Cy, SwiftWriteLM8Cy, SwiftExt1xP0, + SwiftVLDMPerm3], + (instregex "VLD3(DUP|LN)(d|q)(8|16|32)$", + "VLD3(LN|DUP)(d|q)(8|16|32)Pseudo$")>; + def : InstRW<[SwiftWriteLM7Cy, SwiftWriteLM8Cy, SwiftWriteLM8Cy, + SwiftWriteP01OneCycle, SwiftExt1xP0, SwiftVLDMPerm3], + (instregex "VLD3(LN|DUP)(d|q)(8|16|32)_UPD")>; + def : InstRW<[SwiftWriteLM7Cy, SwiftWriteP01OneCycle, SwiftWriteLM8Cy, + SwiftWriteLM8Cy, SwiftExt1xP0, SwiftVLDMPerm3], + (instregex "VLD3(LN|DUP)(d|q)(8|16|32)Pseudo_UPD")>; + // Four element struture. + def : InstRW<[SwiftWriteLM8Cy, SwiftWriteLM9Cy, SwiftWriteLM10CyNo, + SwiftWriteLM10CyNo, SwiftExt1xP0, SwiftVLDMPerm5], + (instregex "VLD4(LN|DUP)(d|q)(8|16|32)$", + "VLD4(LN|DUP)(d|q)(8|16|32)Pseudo$")>; + def : InstRW<[SwiftWriteLM8Cy, SwiftWriteLM9Cy, SwiftWriteLM10CyNo, + SwiftWriteLM10CyNo, SwiftWriteP01OneCycle, SwiftExt1xP0, + SwiftVLDMPerm5], + (instregex "VLD4(DUP|LN)(d|q)(8|16|32)_UPD")>; + def : InstRW<[SwiftWriteLM8Cy, SwiftWriteP01OneCycle, SwiftWriteLM9Cy, + SwiftWriteLM10CyNo, SwiftWriteLM10CyNo, SwiftExt1xP0, + SwiftVLDMPerm5], + (instregex "VLD4(DUP|LN)(d|q)(8|16|32)Pseudo_UPD")>; + // VSTx + // Multiple structures. + // Single element structure store. + def : InstRW<[SwiftWrite1xP2], (instregex "VST1d(8|16|32|64)$")>; + def : InstRW<[SwiftWrite2xP2], (instregex "VST1q(8|16|32|64)$")>; + def : InstRW<[SwiftWriteP01OneCycle, SwiftWrite1xP2], + (instregex "VST1d(8|16|32|64)wb")>; + def : InstRW<[SwiftWriteP01OneCycle, SwiftWrite2xP2], + (instregex "VST1q(8|16|32|64)wb")>; + def : InstRW<[SwiftWrite3xP2], + (instregex "VST1d(8|16|32|64)T$", "VST1d64TPseudo$")>; + def : InstRW<[SwiftWriteP01OneCycle, SwiftWrite3xP2], + (instregex "VST1d(8|16|32|64)Twb", "VST1d64TPseudoWB")>; + def : InstRW<[SwiftWrite4xP2], + (instregex "VST1d(8|16|32|64)(Q|QPseudo)$")>; + def : InstRW<[SwiftWriteP01OneCycle, SwiftWrite4xP2], + (instregex "VST1d(8|16|32|64)(Qwb|QPseudoWB)")>; + // Two element structure store. + def : InstRW<[SwiftWrite1xP2, SwiftVLDMPerm1], + (instregex "VST2(d|b)(8|16|32)$")>; + def : InstRW<[SwiftWriteP01OneCycle, SwiftWrite1xP2, SwiftVLDMPerm1], + (instregex "VST2(b|d)(8|16|32)wb")>; + def : InstRW<[SwiftWrite2xP2, SwiftVLDMPerm2], + (instregex "VST2q(8|16|32)$", "VST2q(8|16|32)Pseudo$")>; + def : InstRW<[SwiftWrite2xP2, SwiftVLDMPerm2], + (instregex "VST2q(8|16|32)wb", "VST2q(8|16|32)PseudoWB")>; + // Three element structure store. + def : InstRW<[SwiftWrite4xP2, SwiftVLDMPerm2], + (instregex "VST3(d|q)(8|16|32)$", "VST3(d|q)(8|16|32)(oddP|P)seudo$")>; + def : InstRW<[SwiftWriteP01OneCycle, SwiftWrite4xP2, SwiftVLDMPerm2], + (instregex "VST3(d|q)(8|16|32)_UPD", + "VST3(d|q)(8|16|32)(oddP|P)seudo_UPD$")>; + // Four element structure store. + def : InstRW<[SwiftWrite4xP2, SwiftVLDMPerm2], + (instregex "VST4(d|q)(8|16|32)$", "VST4(d|q)(8|16|32)(oddP|P)seudo$")>; + def : InstRW<[SwiftWriteP01OneCycle, SwiftWrite4xP2, SwiftVLDMPerm4], + (instregex "VST4(d|q)(8|16|32)_UPD", + "VST4(d|q)(8|16|32)(oddP|P)seudo_UPD$")>; + // Single/all lane store. + // One element structure. + def : InstRW<[SwiftWrite1xP2, SwiftVLDMPerm1], + (instregex "VST1LNd(8|16|32)$", "VST1LNq(8|16|32)Pseudo$")>; + def : InstRW<[SwiftWriteP01OneCycle, SwiftWrite1xP2, SwiftVLDMPerm1], + (instregex "VST1LNd(8|16|32)_UPD", "VST1LNq(8|16|32)Pseudo_UPD")>; + // Two element structure. + def : InstRW<[SwiftWrite1xP2, SwiftVLDMPerm2], + (instregex "VST2LN(d|q)(8|16|32)$", "VST2LN(d|q)(8|16|32)Pseudo$")>; + def : InstRW<[SwiftWriteP01OneCycle, SwiftWrite1xP2, SwiftVLDMPerm2], + (instregex "VST2LN(d|q)(8|16|32)_UPD", + "VST2LN(d|q)(8|16|32)Pseudo_UPD")>; + // Three element structure. + def : InstRW<[SwiftWrite4xP2, SwiftVLDMPerm2], + (instregex "VST3LN(d|q)(8|16|32)$", "VST3LN(d|q)(8|16|32)Pseudo$")>; + def : InstRW<[SwiftWriteP01OneCycle, SwiftWrite4xP2, SwiftVLDMPerm2], + (instregex "VST3LN(d|q)(8|16|32)_UPD", + "VST3LN(d|q)(8|16|32)Pseudo_UPD")>; + // Four element structure. + def : InstRW<[SwiftWrite2xP2, SwiftVLDMPerm2], + (instregex "VST4LN(d|q)(8|16|32)$", "VST4LN(d|q)(8|16|32)Pseudo$")>; + def : InstRW<[SwiftWriteP01OneCycle, SwiftWrite2xP2, SwiftVLDMPerm2], + (instregex "VST4LN(d|q)(8|16|32)_UPD", + "VST4LN(d|q)(8|16|32)Pseudo_UPD")>; + // Preload. def : WriteRes { let Latency = 0; let ResourceCycles = [0];