AVX512: Implemented encoding and intrinsics for vpalignr

Added tests for intrinsics and encoding.

Differential Revision: http://reviews.llvm.org/D12270

llvm-svn: 246428

4 files changed, 92 insertions, 34 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index f9f79955828..1f591e0b51d 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -6902,7 +6902,7 @@ static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
     Hi = DAG.getBitcast(AlignVT, Hi);
 
     return DAG.getBitcast(
-        VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Hi, Lo,
+        VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Lo, Hi,
                         DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
   }
 
@@ -15695,12 +15695,16 @@ static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget
         Src1, Src2, Imm, Rnd),
         Mask, PassThru, Subtarget, DAG);
     }
+    case INTR_TYPE_3OP_IMM8_MASK:
     case INTR_TYPE_3OP_MASK: {
       SDValue Src1 = Op.getOperand(1);
       SDValue Src2 = Op.getOperand(2);
       SDValue Src3 = Op.getOperand(3);
       SDValue PassThru = Op.getOperand(4);
       SDValue Mask = Op.getOperand(5);
+
+      if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK)
+        Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3);
       // We specify 2 possible opcodes for intrinsics with rounding modes.
       // First, we check if the intrinsic may have non-default rounding mode,
       // (IntrData->Opc1 != 0), then we check the rounding mode operand.
diff --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td
index c0614ac03ab..b82818a2cb5 100644
--- a/llvm/lib/Target/X86/X86InstrAVX512.td
+++ b/llvm/lib/Target/X86/X86InstrAVX512.td
@@ -6443,22 +6443,33 @@ multiclass avx512_fp_packed_imm<bits<8> opc, string OpcodeStr, SDNode OpNode,
 
 //handle instruction  reg_vec1 = op(reg_vec2,reg_vec3,imm)
 //                               op(reg_vec2,mem_vec,imm)
+multiclass avx512_3Op_rm_imm8<bits<8> opc, string OpcodeStr, SDNode OpNode,
+                             X86VectorVTInfo DestInfo, X86VectorVTInfo SrcInfo>{
+
+  defm rri : AVX512_maskable<opc, MRMSrcReg, DestInfo, (outs DestInfo.RC:$dst),
+                  (ins SrcInfo.RC:$src1, SrcInfo.RC:$src2, u8imm:$src3),
+                  OpcodeStr, "$src3, $src2, $src1", "$src1, $src2, $src3",
+                  (DestInfo.VT (OpNode (SrcInfo.VT SrcInfo.RC:$src1),
+                               (SrcInfo.VT SrcInfo.RC:$src2),
+                               (i8 imm:$src3)))>;
+  let mayLoad = 1 in
+    defm rmi : AVX512_maskable<opc, MRMSrcMem, DestInfo, (outs DestInfo.RC:$dst),
+                  (ins SrcInfo.RC:$src1, SrcInfo.MemOp:$src2, u8imm:$src3),
+                  OpcodeStr, "$src3, $src2, $src1", "$src1, $src2, $src3",
+                  (DestInfo.VT (OpNode (SrcInfo.VT SrcInfo.RC:$src1),
+                               (SrcInfo.VT (bitconvert
+                                                  (SrcInfo.LdFrag addr:$src2))),
+                               (i8 imm:$src3)))>;
+}
+
+//handle instruction  reg_vec1 = op(reg_vec2,reg_vec3,imm)
+//                               op(reg_vec2,mem_vec,imm)
 //                               op(reg_vec2,broadcast(eltVt),imm)
 multiclass avx512_3Op_imm8<bits<8> opc, string OpcodeStr, SDNode OpNode,
-                                                            X86VectorVTInfo _>{
-  defm rri : AVX512_maskable<opc, MRMSrcReg, _, (outs _.RC:$dst),
-                      (ins _.RC:$src1, _.RC:$src2, u8imm:$src3),
-                      OpcodeStr, "$src3, $src2, $src1", "$src1, $src2, $src3",
-                      (OpNode (_.VT _.RC:$src1),
-                              (_.VT _.RC:$src2),
-                              (i8 imm:$src3))>;
-  let mayLoad = 1 in {
-    defm rmi : AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst),
-                      (ins _.RC:$src1, _.MemOp:$src2, u8imm:$src3),
-                      OpcodeStr, "$src3, $src2, $src1", "$src1, $src2, $src3",
-                      (OpNode (_.VT _.RC:$src1),
-                              (_.VT (bitconvert (_.LdFrag addr:$src2))),
-                              (i8 imm:$src3))>;
+                           X86VectorVTInfo _>:
+  avx512_3Op_rm_imm8<opc, OpcodeStr, OpNode, _, _>{
+
+  let mayLoad = 1 in
     defm rmbi : AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst),
                       (ins _.RC:$src1, _.ScalarMemOp:$src2, u8imm:$src3),
                       OpcodeStr, "$src3, ${src2}"##_.BroadcastStr##", $src1",
@@ -6466,7 +6477,6 @@ multiclass avx512_3Op_imm8<bits<8> opc, string OpcodeStr, SDNode OpNode,
                       (OpNode (_.VT _.RC:$src1),
                               (_.VT (X86VBroadcast(_.ScalarLdFrag addr:$src2))),
                               (i8 imm:$src3))>, EVEX_B;
-  }
 }
 
 //handle scalar instruction  reg_vec1 = op(reg_vec2,reg_vec3,imm)
@@ -6542,6 +6552,20 @@ multiclass avx512_common_fp_sae_packed_imm<string OpcodeStr,
   }
 }
 
+multiclass avx512_common_3Op_rm_imm8<bits<8> opc, SDNode OpNode, string OpStr,
+                   AVX512VLVectorVTInfo DestInfo, AVX512VLVectorVTInfo SrcInfo>{
+  let Predicates = [HasBWI] in {
+    defm Z    : avx512_3Op_rm_imm8<opc, OpStr, OpNode, DestInfo.info512,
+                           SrcInfo.info512>, EVEX_V512, AVX512AIi8Base, EVEX_4V;
+  }
+  let Predicates = [HasBWI, HasVLX] in {
+    defm Z128 : avx512_3Op_rm_imm8<opc, OpStr, OpNode, DestInfo.info128,
+                           SrcInfo.info128>, EVEX_V128, AVX512AIi8Base, EVEX_4V;
+    defm Z256 : avx512_3Op_rm_imm8<opc, OpStr, OpNode,  DestInfo.info256,
+                           SrcInfo.info256>, EVEX_V256, AVX512AIi8Base, EVEX_4V;
+  }
+}
+
 multiclass avx512_common_3Op_imm8<string OpcodeStr, AVX512VLVectorVTInfo _,
                                 bits<8> opc, SDNode OpNode>{
   let Predicates = [HasAVX512] in {
@@ -6665,6 +6689,28 @@ defm VALIGND: avx512_valign<"valignd", avx512vl_i32_info, avx512vl_f32_info>,
 defm VALIGNQ: avx512_valign<"valignq", avx512vl_i64_info, avx512vl_f64_info>,
                                                   EVEX_CD8<64, CD8VF>, VEX_W;
 
+multiclass avx512_vpalign_lowering<X86VectorVTInfo _ , list<Predicate> p>{
+  let Predicates = p in
+    def NAME#_.VTName#rri:
+          Pat<(_.VT (X86PAlignr _.RC:$src1, _.RC:$src2, (i8 imm:$imm))),
+              (!cast<Instruction>(NAME#_.ZSuffix#rri)
+                    _.RC:$src1, _.RC:$src2, imm:$imm)>;
+}
+
+multiclass avx512_vpalign_lowering_common<AVX512VLVectorVTInfo _>:
+      avx512_vpalign_lowering<_.info512, [HasBWI]>,
+      avx512_vpalign_lowering<_.info128, [HasBWI, HasVLX]>,
+      avx512_vpalign_lowering<_.info256, [HasBWI, HasVLX]>;
+
+defm VPALIGN:   avx512_common_3Op_rm_imm8<0x0F, X86PAlignr, "vpalignr" ,
+                                          avx512vl_i8_info, avx512vl_i8_info>,
+                avx512_vpalign_lowering_common<avx512vl_i16_info>,
+                avx512_vpalign_lowering_common<avx512vl_i32_info>,
+                avx512_vpalign_lowering_common<avx512vl_f32_info>,
+                avx512_vpalign_lowering_common<avx512vl_i64_info>,
+                avx512_vpalign_lowering_common<avx512vl_f64_info>,
+                EVEX_CD8<8, CD8VF>;
+
 multiclass avx512_unary_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
                            X86VectorVTInfo _> {
   defm rr : AVX512_maskable<opc, MRMSrcReg, _, (outs _.RC:$dst),
diff --git a/llvm/lib/Target/X86/X86InstrSSE.td b/llvm/lib/Target/X86/X86InstrSSE.td
index 3e072bc4bbc..0342ac2d48f 100644
--- a/llvm/lib/Target/X86/X86InstrSSE.td
+++ b/llvm/lib/Target/X86/X86InstrSSE.td
@@ -5799,37 +5799,37 @@ let Predicates = [HasAVX2] in
 let Constraints = "$src1 = $dst", Predicates = [UseSSSE3] in
   defm PALIGN : ssse3_palignr<"palignr">;
 
-let Predicates = [HasAVX2] in {
+let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
 def : Pat<(v8i32 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
-          (VPALIGNR256rr VR256:$src2, VR256:$src1, imm:$imm)>;
+          (VPALIGNR256rr VR256:$src1, VR256:$src2, imm:$imm)>;
 def : Pat<(v8f32 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
-          (VPALIGNR256rr VR256:$src2, VR256:$src1, imm:$imm)>;
+          (VPALIGNR256rr VR256:$src1, VR256:$src2, imm:$imm)>;
 def : Pat<(v16i16 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
-          (VPALIGNR256rr VR256:$src2, VR256:$src1, imm:$imm)>;
+          (VPALIGNR256rr VR256:$src1, VR256:$src2, imm:$imm)>;
 def : Pat<(v32i8 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
-          (VPALIGNR256rr VR256:$src2, VR256:$src1, imm:$imm)>;
+          (VPALIGNR256rr VR256:$src1, VR256:$src2, imm:$imm)>;
 }
 
-let Predicates = [HasAVX] in {
+let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
 def : Pat<(v4i32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
-          (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
+          (VPALIGNR128rr VR128:$src1, VR128:$src2, imm:$imm)>;
 def : Pat<(v4f32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
-          (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
+          (VPALIGNR128rr VR128:$src1, VR128:$src2, imm:$imm)>;
 def : Pat<(v8i16 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
-          (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
+          (VPALIGNR128rr VR128:$src1, VR128:$src2, imm:$imm)>;
 def : Pat<(v16i8 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
-          (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
+          (VPALIGNR128rr VR128:$src1, VR128:$src2, imm:$imm)>;
 }
 
 let Predicates = [UseSSSE3] in {
 def : Pat<(v4i32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
-          (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
+          (PALIGNR128rr VR128:$src1, VR128:$src2, imm:$imm)>;
 def : Pat<(v4f32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
-          (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
+          (PALIGNR128rr VR128:$src1, VR128:$src2, imm:$imm)>;
 def : Pat<(v8i16 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
-          (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
+          (PALIGNR128rr VR128:$src1, VR128:$src2, imm:$imm)>;
 def : Pat<(v16i8 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
-          (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
+          (PALIGNR128rr VR128:$src1, VR128:$src2, imm:$imm)>;
 }
 
 //===---------------------------------------------------------------------===//
diff --git a/llvm/lib/Target/X86/X86IntrinsicsInfo.h b/llvm/lib/Target/X86/X86IntrinsicsInfo.h
index c065f3d86c5..ed516294139 100644
--- a/llvm/lib/Target/X86/X86IntrinsicsInfo.h
+++ b/llvm/lib/Target/X86/X86IntrinsicsInfo.h
@@ -23,7 +23,8 @@ enum IntrinsicType {
   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,
-  INTR_TYPE_3OP_MASK, INTR_TYPE_3OP_MASK_RM, FMA_OP_MASK, FMA_OP_MASKZ, FMA_OP_MASK3, VPERM_3OP_MASK,
+  INTR_TYPE_3OP_MASK, INTR_TYPE_3OP_MASK_RM, INTR_TYPE_3OP_IMM8_MASK,
+  FMA_OP_MASK, FMA_OP_MASKZ, FMA_OP_MASK3, VPERM_3OP_MASK,
   VPERM_3OP_MASKZ,
   INTR_TYPE_SCALAR_MASK_RM, COMPRESS_EXPAND_IN_REG, COMPRESS_TO_MEM,
   TRUNCATE_TO_MEM_VI8, TRUNCATE_TO_MEM_VI16, TRUNCATE_TO_MEM_VI32,
@@ -753,6 +754,12 @@ static const IntrinsicData  IntrinsicsWithoutChain[] = {
   X86_INTRINSIC_DATA(avx512_mask_paddus_w_128, INTR_TYPE_2OP_MASK, X86ISD::ADDUS, 0),
   X86_INTRINSIC_DATA(avx512_mask_paddus_w_256, INTR_TYPE_2OP_MASK, X86ISD::ADDUS, 0),
   X86_INTRINSIC_DATA(avx512_mask_paddus_w_512, INTR_TYPE_2OP_MASK, X86ISD::ADDUS, 0),
+  X86_INTRINSIC_DATA(avx512_mask_palignr_128, INTR_TYPE_3OP_IMM8_MASK,
+                     X86ISD::PALIGNR, 0),
+  X86_INTRINSIC_DATA(avx512_mask_palignr_256, INTR_TYPE_3OP_IMM8_MASK,
+                     X86ISD::PALIGNR, 0),
+  X86_INTRINSIC_DATA(avx512_mask_palignr_512, INTR_TYPE_3OP_IMM8_MASK,
+                     X86ISD::PALIGNR, 0),
   X86_INTRINSIC_DATA(avx512_mask_pand_d_128, INTR_TYPE_2OP_MASK, ISD::AND, 0),
   X86_INTRINSIC_DATA(avx512_mask_pand_d_256, INTR_TYPE_2OP_MASK, ISD::AND, 0),
   X86_INTRINSIC_DATA(avx512_mask_pand_d_512, INTR_TYPE_2OP_MASK, ISD::AND, 0),
@@ -1199,9 +1206,10 @@ static const IntrinsicData  IntrinsicsWithoutChain[] = {
                      X86ISD::UNPCKL, 0),
   X86_INTRINSIC_DATA(avx512_mask_unpckl_ps_512, INTR_TYPE_2OP_MASK,
                      X86ISD::UNPCKL, 0),
-  X86_INTRINSIC_DATA(avx512_mask_valign_d_512, INTR_TYPE_3OP_MASK, X86ISD::VALIGN, 0),
-  X86_INTRINSIC_DATA(avx512_mask_valign_q_512, INTR_TYPE_3OP_MASK, X86ISD::VALIGN, 0),
-
+  X86_INTRINSIC_DATA(avx512_mask_valign_d_512, INTR_TYPE_3OP_IMM8_MASK,
+                     X86ISD::VALIGN, 0),
+  X86_INTRINSIC_DATA(avx512_mask_valign_q_512, INTR_TYPE_3OP_IMM8_MASK,
+                     X86ISD::VALIGN, 0),
   X86_INTRINSIC_DATA(avx512_mask_vfmadd_pd_128, FMA_OP_MASK, X86ISD::FMADD, 0),
   X86_INTRINSIC_DATA(avx512_mask_vfmadd_pd_256, FMA_OP_MASK, X86ISD::FMADD, 0),
   X86_INTRINSIC_DATA(avx512_mask_vfmadd_pd_512, FMA_OP_MASK, X86ISD::FMADD,