diff options
Diffstat (limited to 'lib/Target/PowerPC/PPCInstrVSX.td')
| -rw-r--r-- | lib/Target/PowerPC/PPCInstrVSX.td | 86 |
1 files changed, 86 insertions, 0 deletions
diff --git a/lib/Target/PowerPC/PPCInstrVSX.td b/lib/Target/PowerPC/PPCInstrVSX.td index ffba0e5aadb5..183512acaf9e 100644 --- a/lib/Target/PowerPC/PPCInstrVSX.td +++ b/lib/Target/PowerPC/PPCInstrVSX.td @@ -3494,6 +3494,17 @@ def DblToFlt { dag B1 = (f32 (fpround (f64 (extractelt v2f64:$B, 1)))); } +def ExtDbl { + dag A0S = (i32 (PPCmfvsr (f64 (PPCfctiwz (f64 (extractelt v2f64:$A, 0)))))); + dag A1S = (i32 (PPCmfvsr (f64 (PPCfctiwz (f64 (extractelt v2f64:$A, 1)))))); + dag B0S = (i32 (PPCmfvsr (f64 (PPCfctiwz (f64 (extractelt v2f64:$B, 0)))))); + dag B1S = (i32 (PPCmfvsr (f64 (PPCfctiwz (f64 (extractelt v2f64:$B, 1)))))); + dag A0U = (i32 (PPCmfvsr (f64 (PPCfctiwuz (f64 (extractelt v2f64:$A, 0)))))); + dag A1U = (i32 (PPCmfvsr (f64 (PPCfctiwuz (f64 (extractelt v2f64:$A, 1)))))); + dag B0U = (i32 (PPCmfvsr (f64 (PPCfctiwuz (f64 (extractelt v2f64:$B, 0)))))); + dag B1U = (i32 (PPCmfvsr (f64 (PPCfctiwuz (f64 (extractelt v2f64:$B, 1)))))); +} + def ByteToWord { dag LE_A0 = (i32 (sext_inreg (i32 (vector_extract v16i8:$A, 0)), i8)); dag LE_A1 = (i32 (sext_inreg (i32 (vector_extract v16i8:$A, 4)), i8)); @@ -3571,9 +3582,15 @@ def FltToULong { } def DblToInt { dag A = (i32 (PPCmfvsr (f64 (PPCfctiwz f64:$A)))); + dag B = (i32 (PPCmfvsr (f64 (PPCfctiwz f64:$B)))); + dag C = (i32 (PPCmfvsr (f64 (PPCfctiwz f64:$C)))); + dag D = (i32 (PPCmfvsr (f64 (PPCfctiwz f64:$D)))); } def DblToUInt { dag A = (i32 (PPCmfvsr (f64 (PPCfctiwuz f64:$A)))); + dag B = (i32 (PPCmfvsr (f64 (PPCfctiwuz f64:$B)))); + dag C = (i32 (PPCmfvsr (f64 (PPCfctiwuz f64:$C)))); + dag D = (i32 (PPCmfvsr (f64 (PPCfctiwuz f64:$D)))); } def DblToLong { dag A = (i64 (PPCmfvsr (f64 (PPCfctidz f64:$A)))); @@ -3612,6 +3629,47 @@ def MrgFP { dag BAlToFlt = (XVCVDPSP (XXPERMDI $B, $A, 3)); } +// Word-element merge dags - conversions from f64 to i32 merged into vectors. +def MrgWords { + // For big endian, we merge low and hi doublewords (A, B). + dag A0B0 = (v2f64 (XXPERMDI v2f64:$A, v2f64:$B, 0)); + dag A1B1 = (v2f64 (XXPERMDI v2f64:$A, v2f64:$B, 3)); + dag CVA1B1S = (v4i32 (XVCVDPSXWS A1B1)); + dag CVA0B0S = (v4i32 (XVCVDPSXWS A0B0)); + dag CVA1B1U = (v4i32 (XVCVDPUXWS A1B1)); + dag CVA0B0U = (v4i32 (XVCVDPUXWS A0B0)); + + // For little endian, we merge low and hi doublewords (B, A). + dag B1A1 = (v2f64 (XXPERMDI v2f64:$B, v2f64:$A, 0)); + dag B0A0 = (v2f64 (XXPERMDI v2f64:$B, v2f64:$A, 3)); + dag CVB1A1S = (v4i32 (XVCVDPSXWS B1A1)); + dag CVB0A0S = (v4i32 (XVCVDPSXWS B0A0)); + dag CVB1A1U = (v4i32 (XVCVDPUXWS B1A1)); + dag CVB0A0U = (v4i32 (XVCVDPUXWS B0A0)); + + // For big endian, we merge hi doublewords of (A, C) and (B, D), convert + // then merge. + dag AC = (v2f64 (XXPERMDI (COPY_TO_REGCLASS f64:$A, VSRC), + (COPY_TO_REGCLASS f64:$C, VSRC), 0)); + dag BD = (v2f64 (XXPERMDI (COPY_TO_REGCLASS f64:$B, VSRC), + (COPY_TO_REGCLASS f64:$D, VSRC), 0)); + dag CVACS = (v4i32 (XVCVDPSXWS AC)); + dag CVBDS = (v4i32 (XVCVDPSXWS BD)); + dag CVACU = (v4i32 (XVCVDPUXWS AC)); + dag CVBDU = (v4i32 (XVCVDPUXWS BD)); + + // For little endian, we merge hi doublewords of (D, B) and (C, A), convert + // then merge. + dag DB = (v2f64 (XXPERMDI (COPY_TO_REGCLASS f64:$D, VSRC), + (COPY_TO_REGCLASS f64:$B, VSRC), 0)); + dag CA = (v2f64 (XXPERMDI (COPY_TO_REGCLASS f64:$C, VSRC), + (COPY_TO_REGCLASS f64:$A, VSRC), 0)); + dag CVDBS = (v4i32 (XVCVDPSXWS DB)); + dag CVCAS = (v4i32 (XVCVDPSXWS CA)); + dag CVDBU = (v4i32 (XVCVDPUXWS DB)); + dag CVCAU = (v4i32 (XVCVDPUXWS CA)); +} + // Patterns for BUILD_VECTOR nodes. let AddedComplexity = 400 in { @@ -3679,6 +3737,20 @@ let AddedComplexity = 400 in { def : Pat<(v4f32 (build_vector DblToFlt.A0, DblToFlt.A1, DblToFlt.B0, DblToFlt.B1)), (v4f32 (VMRGEW MrgFP.ABhToFlt, MrgFP.ABlToFlt))>; + + // Convert 4 doubles to a vector of ints. + def : Pat<(v4i32 (build_vector DblToInt.A, DblToInt.B, + DblToInt.C, DblToInt.D)), + (v4i32 (VMRGEW MrgWords.CVACS, MrgWords.CVBDS))>; + def : Pat<(v4i32 (build_vector DblToUInt.A, DblToUInt.B, + DblToUInt.C, DblToUInt.D)), + (v4i32 (VMRGEW MrgWords.CVACU, MrgWords.CVBDU))>; + def : Pat<(v4i32 (build_vector ExtDbl.A0S, ExtDbl.A1S, + ExtDbl.B0S, ExtDbl.B1S)), + (v4i32 (VMRGEW MrgWords.CVA0B0S, MrgWords.CVA1B1S))>; + def : Pat<(v4i32 (build_vector ExtDbl.A0U, ExtDbl.A1U, + ExtDbl.B0U, ExtDbl.B1U)), + (v4i32 (VMRGEW MrgWords.CVA0B0U, MrgWords.CVA1B1U))>; } let Predicates = [IsLittleEndian, HasVSX] in { @@ -3693,6 +3765,20 @@ let AddedComplexity = 400 in { def : Pat<(v4f32 (build_vector DblToFlt.A0, DblToFlt.A1, DblToFlt.B0, DblToFlt.B1)), (v4f32 (VMRGEW MrgFP.BAhToFlt, MrgFP.BAlToFlt))>; + + // Convert 4 doubles to a vector of ints. + def : Pat<(v4i32 (build_vector DblToInt.A, DblToInt.B, + DblToInt.C, DblToInt.D)), + (v4i32 (VMRGEW MrgWords.CVDBS, MrgWords.CVCAS))>; + def : Pat<(v4i32 (build_vector DblToUInt.A, DblToUInt.B, + DblToUInt.C, DblToUInt.D)), + (v4i32 (VMRGEW MrgWords.CVDBU, MrgWords.CVCAU))>; + def : Pat<(v4i32 (build_vector ExtDbl.A0S, ExtDbl.A1S, + ExtDbl.B0S, ExtDbl.B1S)), + (v4i32 (VMRGEW MrgWords.CVB1A1S, MrgWords.CVB0A0S))>; + def : Pat<(v4i32 (build_vector ExtDbl.A0U, ExtDbl.A1U, + ExtDbl.B0U, ExtDbl.B1U)), + (v4i32 (VMRGEW MrgWords.CVB1A1U, MrgWords.CVB0A0U))>; } let Predicates = [HasDirectMove] in { |