diff options
Diffstat (limited to 'llvm/lib')
| -rw-r--r-- | llvm/lib/Target/X86/X86ISelLowering.cpp | 2 | ||||
| -rw-r--r-- | llvm/lib/Target/X86/X86InstrAVX512.td | 27 |
2 files changed, 28 insertions, 1 deletions
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp index 7d16dd9b623..161cf17bfd1 100644 --- a/llvm/lib/Target/X86/X86ISelLowering.cpp +++ b/llvm/lib/Target/X86/X86ISelLowering.cpp @@ -21387,7 +21387,7 @@ static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget &Subtarget, bool LShift = VT.is128BitVector() || (VT.is256BitVector() && Subtarget.hasInt256()); - bool AShift = LShift && (Subtarget.hasVLX() || + bool AShift = LShift && (Subtarget.hasAVX512() || (VT != MVT::v2i64 && VT != MVT::v4i64)); return (Opcode == ISD::SRA) ? AShift : LShift; } diff --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td index 5fda485dc8e..20708afe2dc 100644 --- a/llvm/lib/Target/X86/X86InstrAVX512.td +++ b/llvm/lib/Target/X86/X86InstrAVX512.td @@ -4798,6 +4798,33 @@ defm VPSLL : avx512_shift_types<0xF2, 0xF3, 0xF1, "vpsll", X86vshl>; defm VPSRA : avx512_shift_types<0xE2, 0xE2, 0xE1, "vpsra", X86vsra>; defm VPSRL : avx512_shift_types<0xD2, 0xD3, 0xD1, "vpsrl", X86vsrl>; +// Use 512bit VPSRA/VPSRAI version to implement v2i64/v4i64 in case NoVLX. +let Predicates = [HasAVX512, NoVLX] in { + def : Pat<(v4i64 (X86vsra (v4i64 VR256X:$src1), (v2i64 VR128X:$src2))), + (EXTRACT_SUBREG (v8i64 + (VPSRAQZrr + (v8i64 (INSERT_SUBREG (IMPLICIT_DEF), VR256X:$src1, sub_ymm)), + VR128X:$src2)), sub_ymm)>; + + def : Pat<(v2i64 (X86vsra (v2i64 VR128X:$src1), (v2i64 VR128X:$src2))), + (EXTRACT_SUBREG (v8i64 + (VPSRAQZrr + (v8i64 (INSERT_SUBREG (IMPLICIT_DEF), VR128X:$src1, sub_xmm)), + VR128X:$src2)), sub_xmm)>; + + def : Pat<(v4i64 (X86vsrai (v4i64 VR256X:$src1), (i8 imm:$src2))), + (EXTRACT_SUBREG (v8i64 + (VPSRAQZri + (v8i64 (INSERT_SUBREG (IMPLICIT_DEF), VR256X:$src1, sub_ymm)), + imm:$src2)), sub_ymm)>; + + def : Pat<(v2i64 (X86vsrai (v2i64 VR128X:$src1), (i8 imm:$src2))), + (EXTRACT_SUBREG (v8i64 + (VPSRAQZri + (v8i64 (INSERT_SUBREG (IMPLICIT_DEF), VR128X:$src1, sub_xmm)), + imm:$src2)), sub_xmm)>; +} + //===-------------------------------------------------------------------===// // Variable Bit Shifts //===-------------------------------------------------------------------===// |
