[X86][AVX512VLBW] add support in byte shift and SAD

add byte shift left/right
add SAD - compute sum of absolute differences

Differential Revision: http://reviews.llvm.org/D12479

llvm-svn: 246654

4 files changed, 83 insertions, 7 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 560e8934ce6..b715cd5931c 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -15576,6 +15576,9 @@ static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget
     case INTR_TYPE_2OP:
       return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
         Op.getOperand(2));
+    case INTR_TYPE_2OP_IMM8:
+      return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
+                         DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(2)));
     case INTR_TYPE_3OP:
       return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
         Op.getOperand(2), Op.getOperand(3));
diff --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td
index ebc11dc4330..d93deaec55d 100644
--- a/llvm/lib/Target/X86/X86InstrAVX512.td
+++ b/llvm/lib/Target/X86/X86InstrAVX512.td
@@ -6822,3 +6822,71 @@ multiclass avx512_shufp<string OpcodeStr, AVX512VLVectorVTInfo VTInfo_I,
 
 defm VSHUFPS: avx512_shufp<"vshufps", avx512vl_i32_info, avx512vl_f32_info>, PS;
 defm VSHUFPD: avx512_shufp<"vshufpd", avx512vl_i64_info, avx512vl_f64_info>, PD, VEX_W;
+//===----------------------------------------------------------------------===//
+// AVX-512 - Byte shift Left/Right
+//===----------------------------------------------------------------------===//
+
+multiclass avx512_shift_packed<bits<8> opc, SDNode OpNode, Format MRMr,
+                             Format MRMm, string OpcodeStr, X86VectorVTInfo _>{
+  def rr : AVX512<opc, MRMr,
+             (outs _.RC:$dst), (ins _.RC:$src1, u8imm:$src2),
+             !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
+             [(set _.RC:$dst,(_.VT (OpNode _.RC:$src1, (i8 imm:$src2))))]>;
+  let mayLoad = 1 in
+    def rm : AVX512<opc, MRMm,
+             (outs _.RC:$dst), (ins _.MemOp:$src1, u8imm:$src2),
+             !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
+             [(set _.RC:$dst,(_.VT (OpNode 
+                                   (_.LdFrag addr:$src1), (i8 imm:$src2))))]>;
+}
+
+multiclass avx512_shift_packed_all<bits<8> opc, SDNode OpNode, Format MRMr, 
+                                 Format MRMm, string OpcodeStr, Predicate prd>{
+  let Predicates = [prd] in
+    defm Z512 : avx512_shift_packed<opc, OpNode, MRMr, MRMm, 
+                                    OpcodeStr, v8i64_info>, EVEX_V512;
+  let Predicates = [prd, HasVLX] in {
+    defm Z256 : avx512_shift_packed<opc, OpNode, MRMr, MRMm, 
+                                    OpcodeStr, v4i64x_info>, EVEX_V256;
+    defm Z128 : avx512_shift_packed<opc, OpNode, MRMr, MRMm, 
+                                    OpcodeStr, v2i64x_info>, EVEX_V128;
+  }
+}
+defm VPSLLDQ : avx512_shift_packed_all<0x73, X86vshldq, MRM7r, MRM7m, "vpslldq", 
+                                       HasBWI>, AVX512PDIi8Base, EVEX_4V;
+defm VPSRLDQ : avx512_shift_packed_all<0x73, X86vshrdq, MRM3r, MRM3m, "vpsrldq", 
+                                       HasBWI>, AVX512PDIi8Base, EVEX_4V;
+
+
+multiclass avx512_psadbw_packed<bits<8> opc, SDNode OpNode, 
+                                string OpcodeStr, X86VectorVTInfo _src>{
+  def rr : AVX512BI<opc, MRMSrcReg,
+             (outs _src.RC:$dst), (ins _src.RC:$src1, _src.RC:$src2),
+             !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
+             [(set _src.RC:$dst,(_src.VT 
+                                (OpNode _src.RC:$src1, _src.RC:$src2)))]>;
+  let mayLoad = 1 in
+    def rm : AVX512BI<opc, MRMSrcMem,
+             (outs _src.RC:$dst), (ins _src.RC:$src1, _src.MemOp:$src2),
+             !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
+             [(set _src.RC:$dst,(_src.VT 
+                                (OpNode _src.RC:$src1, 
+                                (_src.VT (bitconvert 
+                                          (_src.LdFrag addr:$src2))))))]>;
+}
+
+multiclass avx512_psadbw_packed_all<bits<8> opc, SDNode OpNode, 
+                                    string OpcodeStr, Predicate prd> {
+  let Predicates = [prd] in
+    defm Z512 : avx512_psadbw_packed<opc, OpNode, OpcodeStr, v64i8_info>, 
+                                    EVEX_V512;
+  let Predicates = [prd, HasVLX] in {
+    defm Z256 : avx512_psadbw_packed<opc, OpNode, OpcodeStr, v32i8x_info>, 
+                                    EVEX_V256;
+    defm Z128 : avx512_psadbw_packed<opc, OpNode, OpcodeStr, v16i8x_info>,
+                                    EVEX_V128;
+  }
+}
+
+defm VPSADBW : avx512_psadbw_packed_all<0xf6, X86psadbw, "vpsadbw", 
+                                       HasBWI>, EVEX_4V;
diff --git a/llvm/lib/Target/X86/X86InstrSSE.td b/llvm/lib/Target/X86/X86InstrSSE.td
index b4907074605..c42d789c203 100644
--- a/llvm/lib/Target/X86/X86InstrSSE.td
+++ b/llvm/lib/Target/X86/X86InstrSSE.td
@@ -4137,8 +4137,10 @@ defm VPSRAW : PDI_binop_rmi<0xE1, 0x71, MRM4r, "vpsraw", X86vsra, X86vsrai,
 defm VPSRAD : PDI_binop_rmi<0xE2, 0x72, MRM4r, "vpsrad", X86vsra, X86vsrai,
                             VR128, v4i32, v4i32, bc_v4i32, loadv2i64,
                             SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
+} // Predicates = [HasAVX]
 
-let ExeDomain = SSEPackedInt, SchedRW = [WriteVecShift] in {
+let ExeDomain = SSEPackedInt, SchedRW = [WriteVecShift] ,
+                                    Predicates = [HasAVX, NoVLX_Or_NoBWI]in {
   // 128-bit logical shifts.
   def VPSLLDQri : PDIi8<0x73, MRM7r,
                     (outs VR128:$dst), (ins VR128:$src1, u8imm:$src2),
@@ -4153,8 +4155,7 @@ let ExeDomain = SSEPackedInt, SchedRW = [WriteVecShift] in {
                       (v2i64 (X86vshrdq VR128:$src1, (i8 imm:$src2))))]>,
                     VEX_4V;
   // PSRADQri doesn't exist in SSE[1-3].
-}
-} // Predicates = [HasAVX]
+} // Predicates = [HasAVX, NoVLX_Or_NoBWI]
 
 let Predicates = [HasAVX2, NoVLX] in {
 defm VPSLLWY : PDI_binop_rmi<0xF1, 0x71, MRM6r, "vpsllw", X86vshl, X86vshli,
@@ -4183,8 +4184,10 @@ defm VPSRAWY : PDI_binop_rmi<0xE1, 0x71, MRM4r, "vpsraw", X86vsra, X86vsrai,
 defm VPSRADY : PDI_binop_rmi<0xE2, 0x72, MRM4r, "vpsrad", X86vsra, X86vsrai,
                              VR256, v8i32, v4i32, bc_v4i32, loadv2i64,
                              SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
+}// Predicates = [HasAVX2]
 
-let ExeDomain = SSEPackedInt, SchedRW = [WriteVecShift], hasSideEffects = 0 in {
+let ExeDomain = SSEPackedInt, SchedRW = [WriteVecShift], hasSideEffects = 0 , 
+                                    Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
   // 256-bit logical shifts.
   def VPSLLDQYri : PDIi8<0x73, MRM7r,
                     (outs VR256:$dst), (ins VR256:$src1, u8imm:$src2),
@@ -4199,8 +4202,7 @@ let ExeDomain = SSEPackedInt, SchedRW = [WriteVecShift], hasSideEffects = 0 in {
                       (v4i64 (X86vshrdq VR256:$src1, (i8 imm:$src2))))]>,
                     VEX_4V, VEX_L;
   // PSRADQYri doesn't exist in SSE[1-3].
-}
-} // Predicates = [HasAVX2]
+} // Predicates = [HasAVX2, NoVLX_Or_NoBWI]
 
 let Constraints = "$src1 = $dst" in {
 defm PSLLW : PDI_binop_rmi<0xF1, 0x71, MRM6r, "psllw", X86vshl, X86vshli,
diff --git a/llvm/lib/Target/X86/X86IntrinsicsInfo.h b/llvm/lib/Target/X86/X86IntrinsicsInfo.h
index 49219926719..409ba59b1dc 100644
--- a/llvm/lib/Target/X86/X86IntrinsicsInfo.h
+++ b/llvm/lib/Target/X86/X86IntrinsicsInfo.h
@@ -19,7 +19,7 @@ namespace llvm {
 enum IntrinsicType {
   INTR_NO_TYPE,
   GATHER, SCATTER, PREFETCH, RDSEED, RDRAND, RDPMC, RDTSC, XTEST, ADX,
-  INTR_TYPE_1OP, INTR_TYPE_2OP, INTR_TYPE_3OP, INTR_TYPE_4OP,
+  INTR_TYPE_1OP, INTR_TYPE_2OP, INTR_TYPE_2OP_IMM8, INTR_TYPE_3OP, INTR_TYPE_4OP,
   CMP_MASK, CMP_MASK_CC, VSHIFT, VSHIFT_MASK, COMI,
   INTR_TYPE_1OP_MASK, INTR_TYPE_1OP_MASK_RM,
   INTR_TYPE_2OP_MASK, INTR_TYPE_2OP_MASK_RM,
@@ -1426,6 +1426,9 @@ static const IntrinsicData  IntrinsicsWithoutChain[] = {
                      X86ISD::VPERMV3, 0),
   X86_INTRINSIC_DATA(avx512_maskz_vpermt2var_q_512, VPERM_3OP_MASKZ,
                      X86ISD::VPERMV3, 0),
+  X86_INTRINSIC_DATA(avx512_psad_bw_512, INTR_TYPE_2OP, X86ISD::PSADBW, 0),
+  X86_INTRINSIC_DATA(avx512_psll_dq_512, INTR_TYPE_2OP_IMM8, X86ISD::VSHLDQ, 0),
+  X86_INTRINSIC_DATA(avx512_psrl_dq_512, INTR_TYPE_2OP_IMM8, X86ISD::VSRLDQ, 0),
   X86_INTRINSIC_DATA(avx512_rcp28_pd,   INTR_TYPE_1OP_MASK_RM,X86ISD::RCP28, 0),
   X86_INTRINSIC_DATA(avx512_rcp28_ps,   INTR_TYPE_1OP_MASK_RM,X86ISD::RCP28, 0),
   X86_INTRINSIC_DATA(avx512_rcp28_sd,   INTR_TYPE_SCALAR_MASK_RM, X86ISD::RCP28, 0),