diff --git a/lib/Target/ARM/ARMScheduleV7.td b/lib/Target/ARM/ARMScheduleV7.td index 677b4e91a83..ffec43e6b2c 100644 --- a/lib/Target/ARM/ARMScheduleV7.td +++ b/lib/Target/ARM/ARMScheduleV7.td @@ -1018,6 +1018,72 @@ def CortexA9Itineraries : ProcessorItineraries<[ InstrStage<1, [FU_Pipe0, FU_Pipe1]>, InstrStage<4, [FU_NPipe]>], [9, 3, 2, 1]>, // + // Double-register FP Unary + InstrItinData, + // Extra 3 latency cycle since wbck is 6 cycles + InstrStage2<7, [FU_DRegsVFP], 0, Reserved>, + InstrStage<1, [FU_Pipe0, FU_Pipe1]>, + InstrStage<1, [FU_NPipe]>], [5, 2]>, + // + // Quad-register FP Unary + // Result written in N5, but that is relative to the last cycle of multicycle, + // so we use 6 for those cases + InstrItinData, + // Extra 3 latency cycle since wbck is 7 cycles + InstrStage2<8, [FU_DRegsVFP], 0, Reserved>, + InstrStage<1, [FU_Pipe0, FU_Pipe1]>, + InstrStage<2, [FU_NPipe]>], [6, 2]>, + // + // Double-register FP Binary + // FIXME: We're using this itin for many instructions and [2, 2] here is too + // optimistic. + InstrItinData, + // Extra 3 latency cycle since wbck is 7 cycles + InstrStage2<7, [FU_DRegsVFP], 0, Reserved>, + InstrStage<1, [FU_Pipe0, FU_Pipe1]>, + InstrStage<1, [FU_NPipe]>], [5, 2, 2]>, + // + // Quad-register FP Binary + // Result written in N5, but that is relative to the last cycle of multicycle, + // so we use 6 for those cases + // FIXME: We're using this itin for many instructions and [2, 2] here is too + // optimistic. + InstrItinData, + // Extra 3 latency cycle since wbck is 8 cycles + InstrStage2<8, [FU_DRegsVFP], 0, Reserved>, + InstrStage<1, [FU_Pipe0, FU_Pipe1]>, + InstrStage<2, [FU_NPipe]>], [6, 2, 2]>, + // + // Double-register FP Multiple-Accumulate + InstrItinData, + // Extra 3 latency cycle since wbck is 7 cycles + InstrStage2<8, [FU_DRegsVFP], 0, Reserved>, + InstrStage<1, [FU_Pipe0, FU_Pipe1]>, + InstrStage<2, [FU_NPipe]>], [6, 3, 2, 1]>, + // + // Quad-register FP Multiple-Accumulate + // Result written in N9, but that is relative to the last cycle of multicycle, + // so we use 10 for those cases + InstrItinData, + // Extra 3 latency cycle since wbck is 9 cycles + InstrStage2<10, [FU_DRegsVFP], 0, Reserved>, + InstrStage<1, [FU_Pipe0, FU_Pipe1]>, + InstrStage<4, [FU_NPipe]>], [8, 4, 2, 1]>, + // + // Double-register Reciprical Step + InstrItinData, + // Extra 3 latency cycle since wbck is 7 cycles + InstrStage2<8, [FU_DRegsVFP], 0, Reserved>, + InstrStage<1, [FU_Pipe0, FU_Pipe1]>, + InstrStage<2, [FU_NPipe]>], [6, 2, 2]>, + // + // Quad-register Reciprical Step + InstrItinData, + // Extra 3 latency cycle since wbck is 9 cycles + InstrStage2<10, [FU_DRegsVFP], 0, Reserved>, + InstrStage<1, [FU_Pipe0, FU_Pipe1]>, + InstrStage<4, [FU_NPipe]>], [8, 2, 2]>, + // // Double-register Permute InstrItinData, // Extra 3 latency cycle since wbck is 6 cycles