[X86][AVX] Add support for 64-bit VZEXT_LOAD of 256-bit vectors to EltsFromConsecutiveLoads

[RKSimon]

Summary

This patch enables 64-bit VZEXT_LOAD of 256-bit vectors.

Follow up to D16217 and D16729

This change uncovered an odd pattern where VZEXT_LOAD v4i64 was being lowered to a load of the lower v2i64 (so the 2nd i64 destination element wasn't being zeroed), I can't find any use/reason for this and have removed the pattern and replaced it so only the 1st i64 element is loaded and the upper bits all zeroed. This matches the description for X86ISD::VZEXT_LOAD

[delena]

lib/Target/X86/X86ISelLowering.cpp

5644 ↗

(On Diff #46529)

Could you, please, add 512 bit vector here?

lib/Target/X86/X86InstrSSE.td

5061 ↗

(On Diff #46529)

The same pattern should be added to X86InstAVX512.td, right?

[RKSimon]

Added 512-bit vector support

[delena]

Thank you.
LGTM

lib/Target/X86/X86ISelLowering.cpp

5645 ↗

(On Diff #46630)

v8i64 is always legal for 512-bit vector.

[RKSimon]

Thanks Elena - as this can get run before legalization I've left the 8f64 legality test in.

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

if (IsConsecutiveLoad && FirstLoadedElt == 0 &&
 
     return NewLd;
   }
 
   int LoadSize =
       (1 + LastLoadedElt - FirstLoadedElt) * LDBaseVT.getStoreSizeInBits();
 
   // VZEXT_LOAD - consecutive load/undefs followed by zeros/undefs.
-  // TODO: The code below fires only for for loading the low 64-bits of a
-  // of a 128-bit vector. It's probably worth generalizing more.
   if (IsConsecutiveLoad && FirstLoadedElt == 0 && LoadSize == 64 &&
-      (VT.is128BitVector() && TLI.isTypeLegal(MVT::v2i64))) {
-    SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
+      ((VT.is128BitVector() && TLI.isTypeLegal(MVT::v2i64)) ||
+       (VT.is256BitVector() && TLI.isTypeLegal(MVT::v4i64)) ||
+       (VT.is512BitVector() && TLI.isTypeLegal(MVT::v8i64)))) {
+    MVT VecVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
+    SDVTList Tys = DAG.getVTList(VecVT, MVT::Other);
     SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
     SDValue ResNode =
         DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
                                 LDBase->getPointerInfo(),
                                 LDBase->getAlignment(),
                                 false/*isVolatile*/, true/*ReadMem*/,
                                 false/*WriteMem*/);
 

llvm/trunk/lib/Target/X86/X86InstrAVX512.td

defm Z128 : avx512_vector_fpclass<opc, OpcodeStr, OpNode, _.info128, "{x}",
                                       broadcast>, EVEX_V128;
     defm Z256 : avx512_vector_fpclass<opc, OpcodeStr, OpNode, _.info256, "{y}",
                                       broadcast>, EVEX_V256;
   }
 }
 
 multiclass avx512_fp_fpclass_all<string OpcodeStr, bits<8> opcVec,
              bits<8> opcScalar, SDNode VecOpNode, SDNode ScalarOpNode, Predicate prd>{
-  defm PS : avx512_vector_fpclass_all<OpcodeStr,  avx512vl_f32_info, opcVec, 
+  defm PS : avx512_vector_fpclass_all<OpcodeStr,  avx512vl_f32_info, opcVec,
                                       VecOpNode, prd, "{l}">, EVEX_CD8<32, CD8VF>;
-  defm PD : avx512_vector_fpclass_all<OpcodeStr,  avx512vl_f64_info, opcVec, 
+  defm PD : avx512_vector_fpclass_all<OpcodeStr,  avx512vl_f64_info, opcVec,
                                       VecOpNode, prd, "{q}">,EVEX_CD8<64, CD8VF> , VEX_W;
   defm SS : avx512_scalar_fpclass<opcScalar, OpcodeStr, ScalarOpNode,
                                       f32x_info, prd>, EVEX_CD8<32, CD8VT1>;
   defm SD : avx512_scalar_fpclass<opcScalar, OpcodeStr, ScalarOpNode,
                                       f64x_info, prd>, EVEX_CD8<64, CD8VT1>, VEX_W;
 }
 
 defm VFPCLASS : avx512_fp_fpclass_all<"vfpclass", 0x66, 0x67, X86Vfpclass,
def : Pat<(store VK1:$src, addr:$dst),
               sub_8bit))>;
   def : Pat<(store VK2:$src, addr:$dst),
             (MOV8mr addr:$dst,
              (EXTRACT_SUBREG (KMOVWrk (COPY_TO_REGCLASS VK2:$src, VK16)),
               sub_8bit))>;
   def : Pat<(store VK4:$src, addr:$dst),
             (MOV8mr addr:$dst,
              (EXTRACT_SUBREG (KMOVWrk (COPY_TO_REGCLASS VK4:$src, VK16)),
-              sub_8bit))>;  
+              sub_8bit))>;
   def : Pat<(store VK8:$src, addr:$dst),
             (MOV8mr addr:$dst,
              (EXTRACT_SUBREG (KMOVWrk (COPY_TO_REGCLASS VK8:$src, VK16)),
               sub_8bit))>;
-  
   def : Pat<(store (i8 (bitconvert (v8i1 VK8:$src))), addr:$dst),
             (KMOVWmk addr:$dst, (COPY_TO_REGCLASS VK8:$src, VK16))>;
   def : Pat<(v8i1 (bitconvert (i8 (load addr:$src)))),
             (COPY_TO_REGCLASS (KMOVWkm addr:$src), VK8)>;
 }
 let Predicates = [HasAVX512] in {
   def : Pat<(store (i16 (bitconvert (v16i1 VK16:$src))), addr:$dst),
             (KMOVWmk addr:$dst, VK16:$src)>;
multiclass avx512_load<bits<8> opc, string OpcodeStr, X86VectorVTInfo _,
                          bit IsReMaterializable = 1> {
   let hasSideEffects = 0 in {
   def rr : AVX512PI<opc, MRMSrcReg, (outs _.RC:$dst), (ins _.RC:$src),
                     !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"), [],
                     _.ExeDomain>, EVEX;
   def rrkz : AVX512PI<opc, MRMSrcReg, (outs _.RC:$dst),
                       (ins _.KRCWM:$mask,  _.RC:$src),
                       !strconcat(OpcodeStr, "\t{$src, ${dst} {${mask}} {z}|",
-                       "${dst} {${mask}} {z}, $src}"), 
+                       "${dst} {${mask}} {z}, $src}"),
                        [(set _.RC:$dst, (_.VT (vselect _.KRCWM:$mask,
                                            (_.VT _.RC:$src),
                                            _.ImmAllZerosV)))], _.ExeDomain>,
                        EVEX, EVEX_KZ;
 
   let canFoldAsLoad = 1, isReMaterializable = IsReMaterializable,
       SchedRW = [WriteLoad] in
   def rm : AVX512PI<opc, MRMSrcMem, (outs _.RC:$dst), (ins _.MemOp:$src),
def VMOVQI2PQIZrm : AVX512XSI<0x7E, MRMSrcMem, (outs VR128X:$dst),
                       [(set VR128X:$dst,
                         (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>,
                       EVEX, VEX_W, EVEX_CD8<8, CD8VT8>;
 
 //===----------------------------------------------------------------------===//
 // AVX-512  MOVSS, MOVSD
 //===----------------------------------------------------------------------===//
 
-multiclass avx512_move_scalar <string asm, SDNode OpNode, 
+multiclass avx512_move_scalar <string asm, SDNode OpNode,
                               X86VectorVTInfo _> {
-  defm rr_Int : AVX512_maskable_scalar<0x10, MRMSrcReg, _, (outs _.RC:$dst), 
+  defm rr_Int : AVX512_maskable_scalar<0x10, MRMSrcReg, _, (outs _.RC:$dst),
                     (ins _.RC:$src1, _.RC:$src2),
-                    asm, "$src2, $src1","$src1, $src2", 
+                    asm, "$src2, $src1","$src1, $src2",
                     (_.VT (OpNode (_.VT _.RC:$src1),
                                    (_.VT _.RC:$src2))),
                                    IIC_SSE_MOV_S_RR>, EVEX_4V;
   let Constraints = "$src1 = $dst" , mayLoad = 1 in
     defm rm_Int : AVX512_maskable_3src_scalar<0x10, MRMSrcMem, _,
-                    (outs _.RC:$dst), 
+                    (outs _.RC:$dst),
                     (ins _.ScalarMemOp:$src),
                     asm,"$src","$src",
-                    (_.VT (OpNode (_.VT _.RC:$src1), 
-                               (_.VT (scalar_to_vector 
+                    (_.VT (OpNode (_.VT _.RC:$src1),
+                               (_.VT (scalar_to_vector
                                      (_.ScalarLdFrag addr:$src)))))>, EVEX;
   let isCodeGenOnly = 1 in {
-    def rr : AVX512PI<0x10, MRMSrcReg, (outs _.RC:$dst), 
+    def rr : AVX512PI<0x10, MRMSrcReg, (outs _.RC:$dst),
                (ins _.RC:$src1, _.FRC:$src2),
                !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
                [(set _.RC:$dst, (_.VT (OpNode _.RC:$src1,
                                       (scalar_to_vector _.FRC:$src2))))],
                _.ExeDomain,IIC_SSE_MOV_S_RR>, EVEX_4V;
   let mayLoad = 1 in
     def rm : AVX512PI<0x10, MRMSrcMem, (outs _.FRC:$dst), (ins _.ScalarMemOp:$src),
                !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
                [(set _.FRC:$dst, (_.ScalarLdFrag addr:$src))],
                _.ExeDomain, IIC_SSE_MOV_S_RM>, EVEX;
   }
   let mayStore = 1 in {
     def mr: AVX512PI<0x11, MRMDestMem, (outs), (ins _.ScalarMemOp:$dst, _.FRC:$src),
                !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
                [(store _.FRC:$src, addr:$dst)],  _.ExeDomain, IIC_SSE_MOV_S_MR>,
                EVEX;
-    def mrk: AVX512PI<0x11, MRMDestMem, (outs), 
+    def mrk: AVX512PI<0x11, MRMDestMem, (outs),
                 (ins _.ScalarMemOp:$dst, VK1WM:$mask, _.FRC:$src),
                 !strconcat(asm, "\t{$src, $dst {${mask}}|$dst {${mask}}, $src}"),
                 [], _.ExeDomain, IIC_SSE_MOV_S_MR>, EVEX, EVEX_K;
   } // mayStore
 }
 
 defm VMOVSSZ : avx512_move_scalar<"vmovss", X86Movss, f32x_info>,
                                   VEX_LIG, XS, EVEX_CD8<32, CD8VT1>;
let Predicates = [HasAVX512] in {
 
   // Use regular 128-bit instructions to match 256-bit scalar_to_vec+zext.
   def : Pat<(v8i32 (X86vzmovl (insert_subvector undef,
                                (v4i32 (scalar_to_vector GR32:$src)),(iPTR 0)))),
             (SUBREG_TO_REG (i32 0), (VMOVDI2PDIZrr GR32:$src), sub_xmm)>;
   def : Pat<(v4i64 (X86vzmovl (insert_subvector undef,
                                (v2i64 (scalar_to_vector GR64:$src)),(iPTR 0)))),
             (SUBREG_TO_REG (i64 0), (VMOV64toPQIZrr GR64:$src), sub_xmm)>;
+  def : Pat<(v4i64 (X86vzload addr:$src)),
+            (SUBREG_TO_REG (i64 0), (VMOVZPQILo2PQIZrm addr:$src), sub_xmm)>;
+
+  // Use regular 128-bit instructions to match 512-bit scalar_to_vec+zext.
+  def : Pat<(v8i64 (X86vzload addr:$src)),
+            (SUBREG_TO_REG (i64 0), (VMOVZPQILo2PQIZrm addr:$src), sub_xmm)>;
 }
 
 def : Pat<(v16i32 (X86Vinsert (v16i32 immAllZerosV), GR32:$src2, (iPTR 0))),
         (SUBREG_TO_REG (i32 0), (VMOVDI2PDIZrr GR32:$src2), sub_xmm)>;
 
 def : Pat<(v8i64 (X86Vinsert (bc_v8i64 (v16i32 immAllZerosV)), GR64:$src2, (iPTR 0))),
         (SUBREG_TO_REG (i32 0), (VMOV64toPQIZrr GR64:$src2), sub_xmm)>;
 
defm VPMULHW : avx512_binop_rm_vl_w<0xE5, "vpmulhw", mulhs, SSE_INTALU_ITINS_P,
                                     HasBWI, 1>;
 defm VPMULHUW : avx512_binop_rm_vl_w<0xE4, "vpmulhuw", mulhu, SSE_INTMUL_ITINS_P,
                                      HasBWI, 1>;
 defm VPMULHRSW : avx512_binop_rm_vl_w<0x0B, "vpmulhrsw", X86mulhrs, SSE_INTMUL_ITINS_P,
                                       HasBWI, 1>, T8PD;
 defm VPAVG : avx512_binop_rm_vl_bw<0xE0, 0xE3, "vpavg", X86avg,
                                    SSE_INTALU_ITINS_P, HasBWI, 1>;
 
-multiclass avx512_binop_all<bits<8> opc, string OpcodeStr, OpndItins itins, 
+multiclass avx512_binop_all<bits<8> opc, string OpcodeStr, OpndItins itins,
                             AVX512VLVectorVTInfo _SrcVTInfo, AVX512VLVectorVTInfo _DstVTInfo,
                             SDNode OpNode, Predicate prd,  bit IsCommutable = 0> {
   let Predicates = [prd] in
     defm NAME#Z : avx512_binop_rm2<opc, OpcodeStr, itins, OpNode,
                                  _SrcVTInfo.info512, _DstVTInfo.info512,
                                  v8i64_info, IsCommutable>,
                                   EVEX_V512, EVEX_CD8<64, CD8VF>, VEX_W;
   let Predicates = [HasVLX, prd] in {
     defm NAME#Z256 : avx512_binop_rm2<opc, OpcodeStr, itins, OpNode,
-                                      _SrcVTInfo.info256, _DstVTInfo.info256, 
+                                      _SrcVTInfo.info256, _DstVTInfo.info256,
                                       v4i64x_info, IsCommutable>,
                                       EVEX_V256, EVEX_CD8<64, CD8VF>, VEX_W;
     defm NAME#Z128 : avx512_binop_rm2<opc, OpcodeStr, itins, OpNode,
-                                      _SrcVTInfo.info128, _DstVTInfo.info128, 
+                                      _SrcVTInfo.info128, _DstVTInfo.info128,
                                       v2i64x_info, IsCommutable>,
                                      EVEX_V128, EVEX_CD8<64, CD8VF>, VEX_W;
   }
 }
 
 defm VPMULDQ : avx512_binop_all<0x28, "vpmuldq", SSE_INTALU_ITINS_P,
                                 avx512vl_i32_info, avx512vl_i64_info,
                                 X86pmuldq, HasAVX512, 1>,T8PD;
-defm VPMULUDQ : avx512_binop_all<0xF4, "vpmuludq", SSE_INTMUL_ITINS_P, 
+defm VPMULUDQ : avx512_binop_all<0xF4, "vpmuludq", SSE_INTMUL_ITINS_P,
                                 avx512vl_i32_info, avx512vl_i64_info,
                                 X86pmuludq, HasAVX512, 1>;
 defm VPMULTISHIFTQB : avx512_binop_all<0x83, "vpmultishiftqb", SSE_INTALU_ITINS_P,
                                 avx512vl_i8_info, avx512vl_i8_info,
                                 X86multishift, HasVBMI, 0>, T8PD;
 
 multiclass avx512_packs_rmb<bits<8> opc, string OpcodeStr, SDNode OpNode,
                             X86VectorVTInfo _Src, X86VectorVTInfo _Dst> {
defm rmb : AVX512_maskable_cmp<opc, MRMSrcMem, _, (outs _.KRC:$dst),
                     (ins _.RC:$src1, _.ScalarMemOp:$src2), OpcodeStr,
                     "${src2}"##_.BroadcastStr##", $src1",
                     "$src1, ${src2}"##_.BroadcastStr,
                     (OpNode (_.VT _.RC:$src1), (_.VT (X86VBroadcast
                                                 (_.ScalarLdFrag addr:$src2))))>,
                     EVEX_B, EVEX_4V, EVEX_CD8<_.EltSize, CD8VF>;
 }
 
-// Use 512bit version to implement 128/256 bit in case NoVLX.  
+// Use 512bit version to implement 128/256 bit in case NoVLX.
 multiclass avx512_vptest_lowering<SDNode OpNode, X86VectorVTInfo ExtendInfo,
                                   X86VectorVTInfo _, string Suffix> {
     def : Pat<(_.KVT (OpNode (_.VT _.RC:$src1), (_.VT _.RC:$src2))),
               (_.KVT (COPY_TO_REGCLASS
                        (!cast<Instruction>(NAME # Suffix # "Zrr")
-                         (INSERT_SUBREG (ExtendInfo.VT (IMPLICIT_DEF)), 
+                         (INSERT_SUBREG (ExtendInfo.VT (IMPLICIT_DEF)),
                                         _.RC:$src1, _.SubRegIdx),
-                         (INSERT_SUBREG (ExtendInfo.VT (IMPLICIT_DEF)), 
+                         (INSERT_SUBREG (ExtendInfo.VT (IMPLICIT_DEF)),
                                         _.RC:$src2, _.SubRegIdx)),
                      _.KRC))>;
 }
 
 multiclass avx512_vptest_dq_sizes<bits<8> opc, string OpcodeStr, SDNode OpNode,
                                   AVX512VLVectorVTInfo _, string Suffix> {
   let Predicates  = [HasAVX512] in
   defm Z : avx512_vptest<opc, OpcodeStr, OpNode, _.info512>,
            avx512_vptest_mb<opc, OpcodeStr, OpNode, _.info512>, EVEX_V512;
 
   let Predicates = [HasAVX512, HasVLX] in {
   defm Z256 : avx512_vptest<opc, OpcodeStr, OpNode, _.info256>,
               avx512_vptest_mb<opc, OpcodeStr, OpNode, _.info256>, EVEX_V256;
   defm Z128 : avx512_vptest<opc, OpcodeStr, OpNode, _.info128>,
               avx512_vptest_mb<opc, OpcodeStr, OpNode, _.info128>, EVEX_V128;
   }
   let Predicates = [HasAVX512, NoVLX] in {
   defm Z256_Alt : avx512_vptest_lowering< OpNode, _.info512, _.info256, Suffix>;
   defm Z128_Alt : avx512_vptest_lowering< OpNode, _.info512, _.info128, Suffix>;
-  } 
+  }
 }
 
 multiclass avx512_vptest_dq<bits<8> opc, string OpcodeStr, SDNode OpNode> {
   defm D : avx512_vptest_dq_sizes<opc, OpcodeStr#"d", OpNode,
                                  avx512vl_i32_info, "D">;
   defm Q : avx512_vptest_dq_sizes<opc, OpcodeStr#"q", OpNode,
                                  avx512vl_i64_info, "Q">, VEX_W;
 }
defm WZ256: avx512_vptest<opc, OpcodeStr#"w", OpNode, v16i16x_info>,
               EVEX_V256, VEX_W;
   defm WZ128: avx512_vptest<opc, OpcodeStr#"w", OpNode, v8i16x_info>,
               EVEX_V128, VEX_W;
   defm BZ256: avx512_vptest<opc, OpcodeStr#"b", OpNode, v32i8x_info>,
               EVEX_V256;
   defm BZ128: avx512_vptest<opc, OpcodeStr#"b", OpNode, v16i8x_info>,
               EVEX_V128;
   }
-  
   let Predicates = [HasAVX512, NoVLX] in {
   defm BZ256_Alt : avx512_vptest_lowering< OpNode, v64i8_info, v32i8x_info, "B">;
   defm BZ128_Alt : avx512_vptest_lowering< OpNode, v64i8_info, v16i8x_info, "B">;
   defm WZ256_Alt : avx512_vptest_lowering< OpNode, v32i16_info, v16i16x_info, "W">;
   defm WZ128_Alt : avx512_vptest_lowering< OpNode, v32i16_info, v8i16x_info, "W">;
-  } 
+  }
 
 }
 
 multiclass avx512_vptest_all_forms<bits<8> opc_wb, bits<8> opc_dq, string OpcodeStr,
                                    SDNode OpNode> :
   avx512_vptest_wb <opc_wb, OpcodeStr, OpNode>,
   avx512_vptest_dq<opc_dq, OpcodeStr, OpNode>;
 

 multiclass avx512_var_shift_types<bits<8> opc, string OpcodeStr,
                                  SDNode OpNode> {
   defm D : avx512_var_shift_sizes<opc, OpcodeStr#"d", OpNode,
                                  avx512vl_i32_info>;
   defm Q : avx512_var_shift_sizes<opc, OpcodeStr#"q", OpNode,
                                  avx512vl_i64_info>, VEX_W;
 }
 
-// Use 512bit version to implement 128/256 bit in case NoVLX.  
+// Use 512bit version to implement 128/256 bit in case NoVLX.
 multiclass avx512_var_shift_w_lowering<AVX512VLVectorVTInfo _, SDNode OpNode> {
   let Predicates = [HasBWI, NoVLX] in {
-  def : Pat<(_.info256.VT (OpNode (_.info256.VT _.info256.RC:$src1), 
+  def : Pat<(_.info256.VT (OpNode (_.info256.VT _.info256.RC:$src1),
                                   (_.info256.VT _.info256.RC:$src2))),
-            (EXTRACT_SUBREG                
+            (EXTRACT_SUBREG
                 (!cast<Instruction>(NAME#"WZrr")
                     (INSERT_SUBREG (_.info512.VT (IMPLICIT_DEF)), VR256X:$src1, sub_ymm),
                     (INSERT_SUBREG (_.info512.VT (IMPLICIT_DEF)), VR256X:$src2, sub_ymm)),
              sub_ymm)>;
 
-  def : Pat<(_.info128.VT (OpNode (_.info128.VT _.info128.RC:$src1), 
+  def : Pat<(_.info128.VT (OpNode (_.info128.VT _.info128.RC:$src1),
                                   (_.info128.VT _.info128.RC:$src2))),
-            (EXTRACT_SUBREG                
+            (EXTRACT_SUBREG
                 (!cast<Instruction>(NAME#"WZrr")
                     (INSERT_SUBREG (_.info512.VT (IMPLICIT_DEF)), VR128X:$src1, sub_xmm),
                     (INSERT_SUBREG (_.info512.VT (IMPLICIT_DEF)), VR128X:$src2, sub_xmm)),
              sub_xmm)>;
   }
 }
 
 multiclass avx512_var_shift_w<bits<8> opc, string OpcodeStr,

 
 defm VPERMQ : avx512_vpermi_dq_sizes<0x00, MRMSrcReg, MRMSrcMem, "vpermq",
                              X86VPermi, avx512vl_i64_info>,
                              EVEX, AVX512AIi8Base, EVEX_CD8<64, CD8VF>, VEX_W;
 defm VPERMPD : avx512_vpermi_dq_sizes<0x01, MRMSrcReg, MRMSrcMem, "vpermpd",
                              X86VPermi, avx512vl_f64_info>,
                              EVEX, AVX512AIi8Base, EVEX_CD8<64, CD8VF>, VEX_W;
 //===----------------------------------------------------------------------===//
-// AVX-512 - VPERMIL 
+// AVX-512 - VPERMIL
 //===----------------------------------------------------------------------===//
 
 multiclass avx512_permil_vec<bits<8> OpcVar, string OpcodeStr,  SDNode OpNode,
                              X86VectorVTInfo _, X86VectorVTInfo Ctrl> {
   defm rr: AVX512_maskable<OpcVar, MRMSrcReg, _, (outs _.RC:$dst),
                   (ins _.RC:$src1, Ctrl.RC:$src2), OpcodeStr,
                   "$src2, $src1", "$src1, $src2",
                   (_.VT (OpNode _.RC:$src1,
def : Pat<(f64 (uint_to_fp GR32:$src)),
           (VCVTUSI2SDZrr (f64 (IMPLICIT_DEF)), GR32:$src)>;
 def : Pat<(f64 (uint_to_fp GR64:$src)),
           (VCVTUSI642SDZrr (f64 (IMPLICIT_DEF)), GR64:$src)>;
 }
 
 //===----------------------------------------------------------------------===//
 // AVX-512  Scalar convert from float/double to integer
 //===----------------------------------------------------------------------===//
-multiclass avx512_cvt_s_int_round<bits<8> opc, RegisterClass SrcRC, 
+multiclass avx512_cvt_s_int_round<bits<8> opc, RegisterClass SrcRC,
                                   RegisterClass DstRC, Intrinsic Int,
                            Operand memop, ComplexPattern mem_cpat, string asm> {
   let hasSideEffects = 0, Predicates = [HasAVX512] in {
     def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src),
                 !strconcat(asm,"\t{$src, $dst|$dst, $src}"),
                 [(set DstRC:$dst, (Int SrcRC:$src))]>, EVEX, VEX_LIG;
     def rb : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src, AVX512RC:$rc),
-                !strconcat(asm,"\t{$rc, $src, $dst|$dst, $src, $rc}"), []>, 
+                !strconcat(asm,"\t{$rc, $src, $dst|$dst, $src, $rc}"), []>,
                 EVEX, VEX_LIG, EVEX_B, EVEX_RC;
     let mayLoad = 1 in
     def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins memop:$src),
                 !strconcat(asm,"\t{$src, $dst|$dst, $src}"), []>, EVEX, VEX_LIG;
-  } // hasSideEffects = 0, Predicates = [HasAVX512] 
+  } // hasSideEffects = 0, Predicates = [HasAVX512]
 }
 
 // Convert float/double to signed/unsigned int 32/64
 defm VCVTSS2SIZ: avx512_cvt_s_int_round<0x2D, VR128X, GR32, int_x86_sse_cvtss2si,
                                    ssmem, sse_load_f32, "cvtss2si">,
                                    XS, EVEX_CD8<32, CD8VT1>;
-defm VCVTSS2SI64Z: avx512_cvt_s_int_round<0x2D, VR128X, GR64, 
+defm VCVTSS2SI64Z: avx512_cvt_s_int_round<0x2D, VR128X, GR64,
                                   int_x86_sse_cvtss2si64,
                                    ssmem, sse_load_f32, "cvtss2si">,
                                    XS, VEX_W, EVEX_CD8<32, CD8VT1>;
-defm VCVTSS2USIZ: avx512_cvt_s_int_round<0x79, VR128X, GR32, 
+defm VCVTSS2USIZ: avx512_cvt_s_int_round<0x79, VR128X, GR32,
                                   int_x86_avx512_cvtss2usi,
                                    ssmem, sse_load_f32, "cvtss2usi">,
                                    XS, EVEX_CD8<32, CD8VT1>;
 defm VCVTSS2USI64Z: avx512_cvt_s_int_round<0x79, VR128X, GR64,
                                    int_x86_avx512_cvtss2usi64, ssmem,
                                    sse_load_f32, "cvtss2usi">, XS, VEX_W,
                                    EVEX_CD8<32, CD8VT1>;
 defm VCVTSD2SIZ: avx512_cvt_s_int_round<0x2D, VR128X, GR32, int_x86_sse2_cvtsd2si,
                                    sdmem, sse_load_f64, "cvtsd2si">,
                                    XD, EVEX_CD8<64, CD8VT1>;
-defm VCVTSD2SI64Z: avx512_cvt_s_int_round<0x2D, VR128X, GR64, 
+defm VCVTSD2SI64Z: avx512_cvt_s_int_round<0x2D, VR128X, GR64,
                                    int_x86_sse2_cvtsd2si64,
                                    sdmem, sse_load_f64, "cvtsd2si">,
                                    XD, VEX_W, EVEX_CD8<64, CD8VT1>;
-defm VCVTSD2USIZ:   avx512_cvt_s_int_round<0x79, VR128X, GR32, 
+defm VCVTSD2USIZ:   avx512_cvt_s_int_round<0x79, VR128X, GR32,
                                    int_x86_avx512_cvtsd2usi,
                                    sdmem, sse_load_f64, "cvtsd2usi">,
                                    XD, EVEX_CD8<64, CD8VT1>;
 defm VCVTSD2USI64Z: avx512_cvt_s_int_round<0x79, VR128X, GR64,
                                    int_x86_avx512_cvtsd2usi64, sdmem,
                                    sse_load_f64, "cvtsd2usi">, XD, VEX_W,
                                    EVEX_CD8<64, CD8VT1>;
 
defm Int_VCVTSI2SD64Z : sse12_cvt_sint_3addr<0x2A, GR64, VR128X,
             SSE_CVT_Scalar, 0>, XD, EVEX_4V, VEX_W;
 
   defm Int_VCVTUSI2SDZ : sse12_cvt_sint_3addr<0x2A, GR32, VR128X,
             int_x86_avx512_cvtusi2sd, i32mem, loadi32, "cvtusi2sd{l}",
             SSE_CVT_Scalar, 0>, XD, EVEX_4V;
 } // isCodeGenOnly = 1, Predicates = [HasAVX512]
 
 // Convert float/double to signed/unsigned int 32/64 with truncation
-multiclass avx512_cvt_s_all<bits<8> opc, string asm, X86VectorVTInfo _SrcRC, 
-                            X86VectorVTInfo _DstRC, SDNode OpNode, 
+multiclass avx512_cvt_s_all<bits<8> opc, string asm, X86VectorVTInfo _SrcRC,
+                            X86VectorVTInfo _DstRC, SDNode OpNode,
                             SDNode OpNodeRnd>{
 let Predicates = [HasAVX512] in {
   def rr : SI<opc, MRMSrcReg, (outs _DstRC.RC:$dst), (ins _SrcRC.FRC:$src),
               !strconcat(asm,"\t{$src, $dst|$dst, $src}"),
               [(set _DstRC.RC:$dst, (OpNode _SrcRC.FRC:$src))]>, EVEX;
   def rb : SI<opc, MRMSrcReg, (outs _DstRC.RC:$dst), (ins _SrcRC.FRC:$src),
                 !strconcat(asm,"\t{{sae}, $src, $dst|$dst, $src, {sae}}"),
                 []>, EVEX, EVEX_B;
   def rm : SI<opc, MRMSrcMem, (outs _DstRC.RC:$dst), (ins _SrcRC.MemOp:$src),
               !strconcat(asm,"\t{$src, $dst|$dst, $src}"),
-              [(set _DstRC.RC:$dst, (OpNode (_SrcRC.ScalarLdFrag addr:$src)))]>, 
+              [(set _DstRC.RC:$dst, (OpNode (_SrcRC.ScalarLdFrag addr:$src)))]>,
               EVEX;
 
   let isCodeGenOnly = 1,hasSideEffects = 0 in {
       def rr_Int : SI<opc, MRMSrcReg, (outs _DstRC.RC:$dst), (ins _SrcRC.RC:$src),
                 !strconcat(asm,"\t{$src, $dst|$dst, $src}"),
                [(set _DstRC.RC:$dst, (OpNodeRnd _SrcRC.RC:$src,
                                      (i32 FROUND_CURRENT)))]>, EVEX, VEX_LIG;
       def rb_Int : SI<opc, MRMSrcReg, (outs _DstRC.RC:$dst), (ins _SrcRC.RC:$src),
                 !strconcat(asm,"\t{{sae}, $src, $dst|$dst, $src, {sae}}"),
-                [(set _DstRC.RC:$dst, (OpNodeRnd _SrcRC.RC:$src, 
-                                      (i32 FROUND_NO_EXC)))]>, 
+                [(set _DstRC.RC:$dst, (OpNodeRnd _SrcRC.RC:$src,
+                                      (i32 FROUND_NO_EXC)))]>,
                                       EVEX,VEX_LIG , EVEX_B;
       let mayLoad = 1 in
-        def rm_Int : SI<opc, MRMSrcMem, (outs _DstRC.RC:$dst), 
+        def rm_Int : SI<opc, MRMSrcMem, (outs _DstRC.RC:$dst),
                     (ins _SrcRC.MemOp:$src),
                     !strconcat(asm,"\t{$src, $dst|$dst, $src}"),
                     []>, EVEX, VEX_LIG;
 
   } // isCodeGenOnly = 1, hasSideEffects = 0
 } //HasAVX512
 }
 
 
-defm VCVTTSS2SIZ: avx512_cvt_s_all<0x2C, "cvttss2si", f32x_info, i32x_info, 
-                        fp_to_sint,X86cvttss2IntRnd>, 
+defm VCVTTSS2SIZ: avx512_cvt_s_all<0x2C, "cvttss2si", f32x_info, i32x_info,
+                        fp_to_sint,X86cvttss2IntRnd>,
                         XS, EVEX_CD8<32, CD8VT1>;
-defm VCVTTSS2SI64Z: avx512_cvt_s_all<0x2C, "cvttss2si", f32x_info, i64x_info, 
-                        fp_to_sint,X86cvttss2IntRnd>, 
+defm VCVTTSS2SI64Z: avx512_cvt_s_all<0x2C, "cvttss2si", f32x_info, i64x_info,
+                        fp_to_sint,X86cvttss2IntRnd>,
                         VEX_W, XS, EVEX_CD8<32, CD8VT1>;
-defm VCVTTSD2SIZ: avx512_cvt_s_all<0x2C, "cvttsd2si", f64x_info, i32x_info, 
+defm VCVTTSD2SIZ: avx512_cvt_s_all<0x2C, "cvttsd2si", f64x_info, i32x_info,
                         fp_to_sint,X86cvttsd2IntRnd>,
                         XD, EVEX_CD8<64, CD8VT1>;
-defm VCVTTSD2SI64Z: avx512_cvt_s_all<0x2C, "cvttsd2si", f64x_info, i64x_info, 
-                        fp_to_sint,X86cvttsd2IntRnd>, 
+defm VCVTTSD2SI64Z: avx512_cvt_s_all<0x2C, "cvttsd2si", f64x_info, i64x_info,
+                        fp_to_sint,X86cvttsd2IntRnd>,
                         VEX_W, XD, EVEX_CD8<64, CD8VT1>;
 
-defm VCVTTSS2USIZ: avx512_cvt_s_all<0x78, "cvttss2usi", f32x_info, i32x_info, 
-                        fp_to_uint,X86cvttss2UIntRnd>, 
+defm VCVTTSS2USIZ: avx512_cvt_s_all<0x78, "cvttss2usi", f32x_info, i32x_info,
+                        fp_to_uint,X86cvttss2UIntRnd>,
                         XS, EVEX_CD8<32, CD8VT1>;
-defm VCVTTSS2USI64Z: avx512_cvt_s_all<0x78, "cvttss2usi", f32x_info, i64x_info, 
-                        fp_to_uint,X86cvttss2UIntRnd>, 
+defm VCVTTSS2USI64Z: avx512_cvt_s_all<0x78, "cvttss2usi", f32x_info, i64x_info,
+                        fp_to_uint,X86cvttss2UIntRnd>,
                         XS,VEX_W, EVEX_CD8<32, CD8VT1>;
-defm VCVTTSD2USIZ: avx512_cvt_s_all<0x78, "cvttsd2usi", f64x_info, i32x_info, 
-                        fp_to_uint,X86cvttsd2UIntRnd>, 
+defm VCVTTSD2USIZ: avx512_cvt_s_all<0x78, "cvttsd2usi", f64x_info, i32x_info,
+                        fp_to_uint,X86cvttsd2UIntRnd>,
                         XD, EVEX_CD8<64, CD8VT1>;
-defm VCVTTSD2USI64Z: avx512_cvt_s_all<0x78, "cvttsd2usi", f64x_info, i64x_info, 
-                        fp_to_uint,X86cvttsd2UIntRnd>, 
+defm VCVTTSD2USI64Z: avx512_cvt_s_all<0x78, "cvttsd2usi", f64x_info, i64x_info,
+                        fp_to_uint,X86cvttsd2UIntRnd>,
                         XD, VEX_W, EVEX_CD8<64, CD8VT1>;
 let Predicates = [HasAVX512] in {
   def : Pat<(i32 (int_x86_sse_cvttss2si (v4f32 VR128X:$src))),
             (VCVTTSS2SIZrr_Int (COPY_TO_REGCLASS VR128X:$src, FR32X))>;
   def : Pat<(i64 (int_x86_sse_cvttss2si64 (v4f32 VR128X:$src))),
             (VCVTTSS2SI64Zrr_Int (COPY_TO_REGCLASS VR128X:$src, FR32X))>;
   def : Pat<(i32 (int_x86_sse2_cvttsd2si (v2f64 VR128X:$src))),
             (VCVTTSD2SIZrr_Int (COPY_TO_REGCLASS VR128X:$src, FR64X))>;
   def : Pat<(i64 (int_x86_sse2_cvttsd2si64 (v2f64 VR128X:$src))),
             (VCVTTSD2SI64Zrr_Int (COPY_TO_REGCLASS VR128X:$src, FR64X))>;
 
 } // HasAVX512
 //===----------------------------------------------------------------------===//
 // AVX-512  Convert form float to double and back
 //===----------------------------------------------------------------------===//
 multiclass avx512_cvt_fp_scalar<bits<8> opc, string OpcodeStr, X86VectorVTInfo _,
                          X86VectorVTInfo _Src, SDNode OpNode> {
   defm rr : AVX512_maskable_scalar<opc, MRMSrcReg, _, (outs _.RC:$dst),
-                         (ins _Src.RC:$src1, _Src.RC:$src2), OpcodeStr, 
+                         (ins _Src.RC:$src1, _Src.RC:$src2), OpcodeStr,
                          "$src2, $src1", "$src1, $src2",
                          (_.VT (OpNode (_Src.VT _Src.RC:$src1),
-                                       (_Src.VT _Src.RC:$src2)))>, 
+                                       (_Src.VT _Src.RC:$src2)))>,
                          EVEX_4V, VEX_LIG, Sched<[WriteCvtF2F]>;
   defm rm : AVX512_maskable_scalar<opc, MRMSrcMem, _, (outs _.RC:$dst),
-                         (ins _Src.RC:$src1, _Src.MemOp:$src2), OpcodeStr, 
+                         (ins _Src.RC:$src1, _Src.MemOp:$src2), OpcodeStr,
                          "$src2, $src1", "$src1, $src2",
-                         (_.VT (OpNode (_Src.VT _Src.RC:$src1), 
-                                  (_Src.VT (scalar_to_vector 
-                                            (_Src.ScalarLdFrag addr:$src2)))))>, 
+                         (_.VT (OpNode (_Src.VT _Src.RC:$src1),
+                                  (_Src.VT (scalar_to_vector
+                                            (_Src.ScalarLdFrag addr:$src2)))))>,
                          EVEX_4V, VEX_LIG, Sched<[WriteCvtF2FLd, ReadAfterLd]>;
 }
 
 // Scalar Coversion with SAE - suppress all exceptions
 multiclass avx512_cvt_fp_sae_scalar<bits<8> opc, string OpcodeStr, X86VectorVTInfo _,
                          X86VectorVTInfo _Src, SDNode OpNodeRnd> {
   defm rrb : AVX512_maskable_scalar<opc, MRMSrcReg, _, (outs _.RC:$dst),
                         (ins _Src.RC:$src1, _Src.RC:$src2), OpcodeStr,
                         "{sae}, $src2, $src1", "$src1, $src2, {sae}",
-                        (_.VT (OpNodeRnd (_Src.VT _Src.RC:$src1), 
+                        (_.VT (OpNodeRnd (_Src.VT _Src.RC:$src1),
                                          (_Src.VT _Src.RC:$src2),
                                          (i32 FROUND_NO_EXC)))>,
                         EVEX_4V, VEX_LIG, EVEX_B;
 }
 
 // Scalar Conversion with rounding control (RC)
 multiclass avx512_cvt_fp_rc_scalar<bits<8> opc, string OpcodeStr, X86VectorVTInfo _,
                          X86VectorVTInfo _Src, SDNode OpNodeRnd> {
   defm rrb : AVX512_maskable_scalar<opc, MRMSrcReg, _, (outs _.RC:$dst),
                         (ins _Src.RC:$src1, _Src.RC:$src2, AVX512RC:$rc), OpcodeStr,
                         "$rc, $src2, $src1", "$src1, $src2, $rc",
-                        (_.VT (OpNodeRnd (_Src.VT _Src.RC:$src1), 
+                        (_.VT (OpNodeRnd (_Src.VT _Src.RC:$src1),
                                          (_Src.VT _Src.RC:$src2), (i32 imm:$rc)))>,
                         EVEX_4V, VEX_LIG, Sched<[WriteCvtF2FLd, ReadAfterLd]>,
                         EVEX_B, EVEX_RC;
 }
-multiclass avx512_cvt_fp_scalar_sd2ss<bits<8> opc, string OpcodeStr, SDNode OpNode, 
-                                  SDNode OpNodeRnd, X86VectorVTInfo _src, 
+multiclass avx512_cvt_fp_scalar_sd2ss<bits<8> opc, string OpcodeStr, SDNode OpNode,
+                                  SDNode OpNodeRnd, X86VectorVTInfo _src,
                                                         X86VectorVTInfo _dst> {
   let Predicates = [HasAVX512] in {
     defm Z : avx512_cvt_fp_scalar<opc, OpcodeStr, _dst, _src, OpNode>,
              avx512_cvt_fp_rc_scalar<opc, OpcodeStr, _dst, _src,
                                OpNodeRnd>, VEX_W, EVEX_CD8<64, CD8VT1>,
                                EVEX_V512, XD;
   }
 }
 
-multiclass avx512_cvt_fp_scalar_ss2sd<bits<8> opc, string OpcodeStr, SDNode OpNode, 
-                                    SDNode OpNodeRnd, X86VectorVTInfo _src, 
+multiclass avx512_cvt_fp_scalar_ss2sd<bits<8> opc, string OpcodeStr, SDNode OpNode,
+                                    SDNode OpNodeRnd, X86VectorVTInfo _src,
                                                           X86VectorVTInfo _dst> {
   let Predicates = [HasAVX512] in {
     defm Z : avx512_cvt_fp_scalar<opc, OpcodeStr, _dst, _src, OpNode>,
-             avx512_cvt_fp_sae_scalar<opc, OpcodeStr, _dst, _src, OpNodeRnd>, 
+             avx512_cvt_fp_sae_scalar<opc, OpcodeStr, _dst, _src, OpNodeRnd>,
              EVEX_CD8<32, CD8VT1>, XS, EVEX_V512;
   }
 }
 defm VCVTSD2SS : avx512_cvt_fp_scalar_sd2ss<0x5A, "vcvtsd2ss", X86fround,
                                          X86froundRnd, f64x_info, f32x_info>;
-defm VCVTSS2SD : avx512_cvt_fp_scalar_ss2sd<0x5A, "vcvtss2sd", X86fpext, 
+defm VCVTSS2SD : avx512_cvt_fp_scalar_ss2sd<0x5A, "vcvtss2sd", X86fpext,
                                           X86fpextRnd,f32x_info, f64x_info >;
 
-def : Pat<(f64 (fextend FR32X:$src)), 
-          (COPY_TO_REGCLASS (VCVTSS2SDZrr (COPY_TO_REGCLASS FR32X:$src, VR128X), 
+def : Pat<(f64 (fextend FR32X:$src)),
+          (COPY_TO_REGCLASS (VCVTSS2SDZrr (COPY_TO_REGCLASS FR32X:$src, VR128X),
                                (COPY_TO_REGCLASS FR32X:$src, VR128X)), VR128X)>,
           Requires<[HasAVX512]>;
 def : Pat<(f64 (fextend (loadf32 addr:$src))),
           (COPY_TO_REGCLASS (VCVTSS2SDZrm (v4f32 (IMPLICIT_DEF)), addr:$src), VR128X)>,
           Requires<[HasAVX512]>;
 
 def : Pat<(f64 (extloadf32 addr:$src)),
       (COPY_TO_REGCLASS (VCVTSS2SDZrm (v4f32 (IMPLICIT_DEF)), addr:$src), VR128X)>,
       Requires<[HasAVX512, OptForSize]>;
 
 def : Pat<(f64 (extloadf32 addr:$src)),
-          (COPY_TO_REGCLASS (VCVTSS2SDZrr (v4f32 (IMPLICIT_DEF)), 
+          (COPY_TO_REGCLASS (VCVTSS2SDZrr (v4f32 (IMPLICIT_DEF)),
                     (COPY_TO_REGCLASS (VMOVSSZrm addr:$src), VR128X)), VR128X)>,
           Requires<[HasAVX512, OptForSpeed]>;
 
-def : Pat<(f32 (fround FR64X:$src)), 
-          (COPY_TO_REGCLASS (VCVTSD2SSZrr (COPY_TO_REGCLASS FR64X:$src, VR128X), 
+def : Pat<(f32 (fround FR64X:$src)),
+          (COPY_TO_REGCLASS (VCVTSD2SSZrr (COPY_TO_REGCLASS FR64X:$src, VR128X),
                     (COPY_TO_REGCLASS FR64X:$src, VR128X)), VR128X)>,
            Requires<[HasAVX512]>;
 //===----------------------------------------------------------------------===//
 // AVX-512  Vector convert from signed/unsigned integer to float/double
 //          and from float/double to signed/unsigned integer
 //===----------------------------------------------------------------------===//
 
 multiclass avx512_vcvt_fp<bits<8> opc, string OpcodeStr, X86VectorVTInfo _,
def : Pat<(v8f32 (fround (loadv8f64 addr:$src))),
             (VCVTPD2PSZrm addr:$src)>;
   def : Pat<(v8f64 (extloadv8f32 addr:$src)),
             (VCVTPS2PDZrm addr:$src)>;
 }
 
 //===----------------------------------------------------------------------===//
 // Half precision conversion instructions
 //===----------------------------------------------------------------------===//
-multiclass avx512_cvtph2ps<X86VectorVTInfo _dest, X86VectorVTInfo _src, 
+multiclass avx512_cvtph2ps<X86VectorVTInfo _dest, X86VectorVTInfo _src,
                            X86MemOperand x86memop, PatFrag ld_frag> {
   defm rr : AVX512_maskable<0x13, MRMSrcReg, _dest ,(outs _dest.RC:$dst), (ins _src.RC:$src),
                     "vcvtph2ps", "$src", "$src",
                    (X86cvtph2ps (_src.VT _src.RC:$src),
                                                 (i32 FROUND_CURRENT))>, T8PD;
   let hasSideEffects = 0, mayLoad = 1 in {
     defm rm : AVX512_maskable<0x13, MRMSrcMem, _dest, (outs _dest.RC:$dst), (ins x86memop:$src),
-                      "vcvtph2ps", "$src", "$src", 
+                      "vcvtph2ps", "$src", "$src",
                       (X86cvtph2ps (_src.VT (bitconvert (ld_frag addr:$src))),
                                        (i32 FROUND_CURRENT))>, T8PD;
   }
 }
 
 multiclass avx512_cvtph2ps_sae<X86VectorVTInfo _dest, X86VectorVTInfo _src> {
   defm rb : AVX512_maskable<0x13, MRMSrcReg, _dest ,(outs _dest.RC:$dst), (ins _src.RC:$src),
                     "vcvtph2ps", "{sae}, $src", "$src, {sae}",
                    (X86cvtph2ps (_src.VT _src.RC:$src),
                                                 (i32 FROUND_NO_EXC))>, T8PD, EVEX_B;
 
 }
 
 let Predicates = [HasAVX512] in {
   defm VCVTPH2PSZ : avx512_cvtph2ps<v16f32_info, v16i16x_info, f256mem, loadv4i64>,
-                    avx512_cvtph2ps_sae<v16f32_info, v16i16x_info>, 
+                    avx512_cvtph2ps_sae<v16f32_info, v16i16x_info>,
                     EVEX, EVEX_V512, EVEX_CD8<32, CD8VH>;
   let Predicates = [HasVLX] in {
-    defm VCVTPH2PSZ256 : avx512_cvtph2ps<v8f32x_info, v8i16x_info, f128mem, 
+    defm VCVTPH2PSZ256 : avx512_cvtph2ps<v8f32x_info, v8i16x_info, f128mem,
                          loadv2i64>,EVEX, EVEX_V256, EVEX_CD8<32, CD8VH>;
     defm VCVTPH2PSZ128 : avx512_cvtph2ps<v4f32x_info, v8i16x_info, f64mem,
                          loadv2i64>, EVEX, EVEX_V128, EVEX_CD8<32, CD8VH>;
   }
 }
 
-multiclass avx512_cvtps2ph<X86VectorVTInfo _dest, X86VectorVTInfo _src, 
+multiclass avx512_cvtps2ph<X86VectorVTInfo _dest, X86VectorVTInfo _src,
                            X86MemOperand x86memop> {
   defm rr : AVX512_maskable<0x1D, MRMDestReg, _dest ,(outs _dest.RC:$dst),
                (ins _src.RC:$src1, i32u8imm:$src2),
-                    "vcvtps2ph", "$src2, $src1", "$src1, $src2", 
+                    "vcvtps2ph", "$src2, $src1", "$src1, $src2",
                    (X86cvtps2ph (_src.VT _src.RC:$src1),
-                                (i32 imm:$src2), 
+                                (i32 imm:$src2),
                                 (i32 FROUND_CURRENT))>, AVX512AIi8Base;
   let hasSideEffects = 0, mayStore = 1 in {
     def mr : AVX512AIi8<0x1D, MRMDestMem, (outs),
                (ins x86memop:$dst, _src.RC:$src1, i32u8imm:$src2),
-               "vcvtps2ph\t{$src2, $src1, $dst|$dst, $src1, $src2}", 
+               "vcvtps2ph\t{$src2, $src1, $dst|$dst, $src1, $src2}",
                [(store (_dest.VT (X86cvtps2ph (_src.VT _src.RC:$src1),
                                        (i32 imm:$src2), (i32 FROUND_CURRENT) )),
                                        addr:$dst)]>;
     def mrk : AVX512AIi8<0x1D, MRMDestMem, (outs),
                (ins x86memop:$dst, _dest.KRCWM:$mask, _src.RC:$src1, i32u8imm:$src2),
-               "vcvtps2ph\t{$src2, $src1, $dst {${mask}}|$dst {${mask}}, $src1, $src2}", 
+               "vcvtps2ph\t{$src2, $src1, $dst {${mask}}|$dst {${mask}}, $src1, $src2}",
                 []>, EVEX_K;
   }
 }
 multiclass avx512_cvtps2ph_sae<X86VectorVTInfo _dest, X86VectorVTInfo _src> {
   defm rb : AVX512_maskable<0x1D, MRMDestReg, _dest ,(outs _dest.RC:$dst),
                (ins _src.RC:$src1, i32u8imm:$src2),
-                    "vcvtps2ph", "$src2, {sae}, $src1", "$src1, $src2, {sae}", 
+                    "vcvtps2ph", "$src2, {sae}, $src1", "$src1, $src2, {sae}",
                    (X86cvtps2ph (_src.VT _src.RC:$src1),
-                                (i32 imm:$src2), 
+                                (i32 imm:$src2),
                                 (i32 FROUND_NO_EXC))>, EVEX_B, AVX512AIi8Base;
 }
 let Predicates = [HasAVX512] in {
   defm VCVTPS2PHZ : avx512_cvtps2ph<v16i16x_info, v16f32_info, f256mem>,
                     avx512_cvtps2ph_sae<v16i16x_info, v16f32_info>,
                       EVEX, EVEX_V512, EVEX_CD8<32, CD8VH>;
   let Predicates = [HasVLX] in {
     defm VCVTPS2PHZ256 : avx512_cvtps2ph<v8i16x_info, v8f32x_info, f128mem>,
                         EVEX, EVEX_V256, EVEX_CD8<32, CD8VH>;
     defm VCVTPS2PHZ128 : avx512_cvtps2ph<v8i16x_info, v4f32x_info, f128mem>,
                         EVEX, EVEX_V128, EVEX_CD8<32, CD8VH>;
   }
 }
 
 //  Unordered/Ordered scalar fp compare with Sea and set EFLAGS
 multiclass avx512_ord_cmp_sae<bits<8> opc, X86VectorVTInfo _, SDNode OpNode,
                             string OpcodeStr> {
   def rb: AVX512<opc, MRMSrcReg, (outs), (ins _.RC:$src1, _.RC:$src2),
                  !strconcat(OpcodeStr, "\t{{sae}, $src2, $src1|$src1, $src2, {sae}}"),
-                 [(set EFLAGS, (OpNode (_.VT _.RC:$src1), _.RC:$src2, 
+                 [(set EFLAGS, (OpNode (_.VT _.RC:$src1), _.RC:$src2,
                                                         (i32 FROUND_NO_EXC)))],
                  IIC_SSE_COMIS_RR>, EVEX, EVEX_B, VEX_LIG, EVEX_V128,
                  Sched<[WriteFAdd]>;
 }
 
 let Defs = [EFLAGS], Predicates = [HasAVX512] in {
   defm VUCOMISSZ : avx512_ord_cmp_sae<0x2E, v4f32x_info, X86ucomiSae, "vucomiss">,
                                    AVX512PSIi8Base, EVEX_CD8<32, CD8VT1>;

 defm VPMOVM2 : avx512_convert_mask_to_vector<"vpmovm2">;
 
 multiclass convert_vector_to_mask_common<bits<8> opc, X86VectorVTInfo _, string OpcodeStr > {
     def rr : AVX512XS8I<opc, MRMSrcReg, (outs _.KRC:$dst), (ins _.RC:$src),
                         !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
                         [(set _.KRC:$dst, (X86cvt2mask (_.VT _.RC:$src)))]>, EVEX;
 }
 
-// Use 512bit version to implement 128/256 bit in case NoVLX.  
-multiclass convert_vector_to_mask_lowering<X86VectorVTInfo ExtendInfo, 
+// Use 512bit version to implement 128/256 bit in case NoVLX.
+multiclass convert_vector_to_mask_lowering<X86VectorVTInfo ExtendInfo,
                                                             X86VectorVTInfo _> {
 
   def : Pat<(_.KVT (X86cvt2mask (_.VT _.RC:$src))),
             (_.KVT (COPY_TO_REGCLASS
                      (!cast<Instruction>(NAME#"Zrr")
-                       (INSERT_SUBREG (ExtendInfo.VT (IMPLICIT_DEF)), 
+                       (INSERT_SUBREG (ExtendInfo.VT (IMPLICIT_DEF)),
                                       _.RC:$src, _.SubRegIdx)),
                    _.KRC))>;
 }
 
 multiclass avx512_convert_vector_to_mask<bits<8> opc, string OpcodeStr,
                                    AVX512VLVectorVTInfo VTInfo, Predicate prd> {
   let Predicates = [prd] in
     defm Z : convert_vector_to_mask_common <opc, VTInfo.info512, OpcodeStr>,
multiclass avx512_shift_packed<bits<8> opc, SDNode OpNode, Format MRMr,
   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 
+             [(set _.RC:$dst,(_.VT (OpNode
                                    (_.LdFrag addr:$src1), (i8 imm:$src2))))]>;
 }
 
-multiclass avx512_shift_packed_all<bits<8> opc, SDNode OpNode, Format MRMr, 
+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, 
+    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, 
+    defm Z256 : avx512_shift_packed<opc, OpNode, MRMr, MRMm,
                                     OpcodeStr, v4i64x_info>, EVEX_V256;
-    defm Z128 : avx512_shift_packed<opc, OpNode, MRMr, MRMm, 
+    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", 
+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", 
+defm VPSRLDQ : avx512_shift_packed_all<0x73, X86vshrdq, MRM3r, MRM3m, "vpsrldq",
                                        HasBWI>, AVX512PDIi8Base, EVEX_4V;
 
 
-multiclass avx512_psadbw_packed<bits<8> opc, SDNode OpNode, 
+multiclass avx512_psadbw_packed<bits<8> opc, SDNode OpNode,
                                 string OpcodeStr, X86VectorVTInfo _dst,
                                 X86VectorVTInfo _src>{
   def rr : AVX512BI<opc, MRMSrcReg,
              (outs _dst.RC:$dst), (ins _src.RC:$src1, _src.RC:$src2),
              !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
              [(set _dst.RC:$dst,(_dst.VT
                                 (OpNode (_src.VT _src.RC:$src1),
                                         (_src.VT _src.RC:$src2))))]>;
   let mayLoad = 1 in
     def rm : AVX512BI<opc, MRMSrcMem,
              (outs _dst.RC:$dst), (ins _src.RC:$src1, _src.MemOp:$src2),
              !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
              [(set _dst.RC:$dst,(_dst.VT
                                 (OpNode (_src.VT _src.RC:$src1),
                                 (_src.VT (bitconvert
                                           (_src.LdFrag addr:$src2))))))]>;
 }
 
-multiclass avx512_psadbw_packed_all<bits<8> opc, SDNode OpNode, 
+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, v8i64_info,
                                     v64i8_info>, EVEX_V512;
   let Predicates = [prd, HasVLX] in {
     defm Z256 : avx512_psadbw_packed<opc, OpNode, OpcodeStr, v4i64x_info,
                                     v32i8x_info>, EVEX_V256;
     defm Z128 : avx512_psadbw_packed<opc, OpNode, OpcodeStr, v2i64x_info,
                                     v16i8x_info>, EVEX_V128;
   }
 }
 
-defm VPSADBW : avx512_psadbw_packed_all<0xf6, X86psadbw, "vpsadbw", 
+defm VPSADBW : avx512_psadbw_packed_all<0xf6, X86psadbw, "vpsadbw",
                                        HasBWI>, EVEX_4V;
 
 multiclass avx512_ternlog<bits<8> opc, string OpcodeStr, SDNode OpNode,
                                                             X86VectorVTInfo _>{
   let Constraints = "$src1 = $dst" in {
   defm rri : AVX512_maskable_3src<opc, MRMSrcReg, _, (outs _.RC:$dst),
                       (ins _.RC:$src2, _.RC:$src3, u8imm:$src4),
                       OpcodeStr, "$src4, $src3, $src2", "$src2, $src3, $src3",
multiclass avx512_fixupimm_packed<bits<8> opc, string OpcodeStr, SDNode OpNode,
   } // Constraints = "$src1 = $dst"
 }
 
 multiclass avx512_fixupimm_packed_sae<bits<8> opc, string OpcodeStr,
                                              SDNode OpNode, X86VectorVTInfo _>{
 let Constraints = "$src1 = $dst" in {
   defm rrib : AVX512_maskable_3src<opc, MRMSrcReg, _, (outs _.RC:$dst),
                       (ins _.RC:$src2, _.RC:$src3, i32u8imm:$src4),
-                      OpcodeStr##_.Suffix, "$src4, {sae}, $src3, $src2", 
+                      OpcodeStr##_.Suffix, "$src4, {sae}, $src3, $src2",
                       "$src2, $src3, {sae}, $src4",
                       (OpNode (_.VT _.RC:$src1),
                                 (_.VT _.RC:$src2),
                                 (_.IntVT _.RC:$src3),
                                 (i32 imm:$src4),
                                 (i32 FROUND_NO_EXC))>, EVEX_B;
   }
 }
multiclass avx512_fixupimm_packed_all<AVX512VLVectorVTInfo _Vec>{
   let Predicates = [HasAVX512, HasVLX] in {
     defm Z128 : avx512_fixupimm_packed<0x54, "vfixupimm", X86VFixupimm, _Vec.info128>,
                                   AVX512AIi8Base, EVEX_4V, EVEX_V128;
     defm Z256 : avx512_fixupimm_packed<0x54, "vfixupimm", X86VFixupimm, _Vec.info256>,
                                   AVX512AIi8Base, EVEX_4V, EVEX_V256;
   }
 }
 
-defm VFIXUPIMMSS : avx512_fixupimm_scalar<0x55, "vfixupimm", X86VFixupimmScalar, 
-                                          f32x_info, v4i32x_info>, 
+defm VFIXUPIMMSS : avx512_fixupimm_scalar<0x55, "vfixupimm", X86VFixupimmScalar,
+                                          f32x_info, v4i32x_info>,
                          AVX512AIi8Base, VEX_LIG, EVEX_4V, EVEX_CD8<32, CD8VT1>;
-defm VFIXUPIMMSD : avx512_fixupimm_scalar<0x55, "vfixupimm", X86VFixupimmScalar, 
-                                          f64x_info, v2i64x_info>, 
+defm VFIXUPIMMSD : avx512_fixupimm_scalar<0x55, "vfixupimm", X86VFixupimmScalar,
+                                          f64x_info, v2i64x_info>,
                          AVX512AIi8Base, VEX_LIG, EVEX_4V, EVEX_CD8<64, CD8VT1>, VEX_W;
-defm VFIXUPIMMPS : avx512_fixupimm_packed_all<avx512vl_f32_info>, 
+defm VFIXUPIMMPS : avx512_fixupimm_packed_all<avx512vl_f32_info>,
                          EVEX_CD8<32, CD8VF>;
-defm VFIXUPIMMPD : avx512_fixupimm_packed_all<avx512vl_f64_info>, 
+defm VFIXUPIMMPD : avx512_fixupimm_packed_all<avx512vl_f64_info>,
                          EVEX_CD8<64, CD8VF>, VEX_W;

llvm/trunk/lib/Target/X86/X86InstrFragmentsSIMD.td

def alignedstore512 : PatFrag<(ops node:$val, node:$ptr),
   return cast<StoreSDNode>(N)->getAlignment() >= 64;
 }]>;
 
 // Like 'load', but always requires 128-bit vector alignment.
 def alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
   return cast<LoadSDNode>(N)->getAlignment() >= 16;
 }]>;
 
-// Like 'X86vzload', but always requires 128-bit vector alignment.
-def alignedX86vzload : PatFrag<(ops node:$ptr), (X86vzload node:$ptr), [{
-  return cast<MemSDNode>(N)->getAlignment() >= 16;
-}]>;
-
 // Like 'load', but always requires 256-bit vector alignment.
 def alignedload256 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
   return cast<LoadSDNode>(N)->getAlignment() >= 32;
 }]>;
 
 // Like 'load', but always requires 512-bit vector alignment.
 def alignedload512 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
   return cast<LoadSDNode>(N)->getAlignment() >= 64;

llvm/trunk/lib/Target/X86/X86InstrSSE.td

 let Predicates = [UseAVX], AddedComplexity = 20 in {
   def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
             (VMOVZQI2PQIrm addr:$src)>;
   def : Pat<(v2i64 (X86vzload addr:$src)),
             (VMOVZQI2PQIrm addr:$src)>;
   def : Pat<(v4i64 (X86vzmovl (insert_subvector undef,
               (v2i64 (scalar_to_vector (loadi64 addr:$src))), (iPTR 0)))),
             (SUBREG_TO_REG (i64 0), (VMOVZQI2PQIrm addr:$src), sub_xmm)>;
+  def : Pat<(v4i64 (X86vzload addr:$src)),
+            (SUBREG_TO_REG (i64 0), (VMOVZQI2PQIrm addr:$src), sub_xmm)>;
 }
 
 let Predicates = [UseSSE2], AddedComplexity = 20 in {
   def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
             (MOVZQI2PQIrm addr:$src)>;
   def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
 }
 
-let Predicates = [HasAVX] in {
-def : Pat<(v4i64 (alignedX86vzload addr:$src)),
-          (SUBREG_TO_REG (i32 0), (VMOVAPSrm addr:$src), sub_xmm)>;
-def : Pat<(v4i64 (X86vzload addr:$src)),
-          (SUBREG_TO_REG (i32 0), (VMOVUPSrm addr:$src), sub_xmm)>;
-}
-
 //===---------------------------------------------------------------------===//
 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
 // IA32 document. movq xmm1, xmm2 does clear the high bits.
 //
 let ExeDomain = SSEPackedInt, SchedRW = [WriteVecLogic] in {
 let AddedComplexity = 15 in
 def VMOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
                         "vmovq\t{$src, $dst|$dst, $src}",

llvm/trunk/test/CodeGen/X86/merge-consecutive-loads-256.ll

 ; AVX-NEXT:    retq
   %ptr1 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 2
   %val1 = load <4 x float>, <4 x float>* %ptr1
   %res = shufflevector <4 x float> zeroinitializer, <4 x float> %val1, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
   ret <8 x float> %res
 }
 
 define <8 x float> @merge_8f32_f32_12zzuuzz(float* %ptr) nounwind uwtable noinline ssp {
-; AVX1-LABEL: merge_8f32_f32_12zzuuzz:
-; AVX1:       # BB#0:
-; AVX1-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
-; AVX1-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: merge_8f32_f32_12zzuuzz:
-; AVX2:       # BB#0:
-; AVX2-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX2-NEXT:    retq
-;
-; AVX512F-LABEL: merge_8f32_f32_12zzuuzz:
-; AVX512F:       # BB#0:
-; AVX512F-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
-; AVX512F-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; AVX512F-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX512F-NEXT:    retq
+; AVX-LABEL: merge_8f32_f32_12zzuuzz:
+; AVX:       # BB#0:
+; AVX-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
+; AVX-NEXT:    retq
   %ptr0 = getelementptr inbounds float, float* %ptr, i64 1
   %ptr1 = getelementptr inbounds float, float* %ptr, i64 2
   %val0 = load float, float* %ptr0
   %val1 = load float, float* %ptr1
   %res0 = insertelement <8 x float> undef, float %val0, i32 0
   %res1 = insertelement <8 x float> %res0, float %val1, i32 1
   %res2 = insertelement <8 x float> %res1, float   0.0, i32 2
   %res3 = insertelement <8 x float> %res2, float   0.0, i32 3

llvm/trunk/test/CodeGen/X86/merge-consecutive-loads-512.ll

; ALL-NEXT:    retq
   %res7 = insertelement <8 x i64> %res5, i64 %val7, i32 7
   ret <8 x i64> %res7
 }
 
 define <16 x float> @merge_16f32_f32_89zzzuuuuuuuuuuuz(float* %ptr) nounwind uwtable noinline ssp {
 ; ALL-LABEL: merge_16f32_f32_89zzzuuuuuuuuuuuz:
 ; ALL:       # BB#0:
 ; ALL-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
-; ALL-NEXT:    vxorpd %xmm1, %xmm1, %xmm1
-; ALL-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; ALL-NEXT:    vxorpd %ymm1, %ymm1, %ymm1
-; ALL-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
 ; ALL-NEXT:    retq
   %ptr0 = getelementptr inbounds float, float* %ptr, i64 8
   %ptr1 = getelementptr inbounds float, float* %ptr, i64 9
   %val0 = load float, float* %ptr0
   %val1 = load float, float* %ptr1
   %res0 = insertelement <16 x float> undef, float %val0, i32 0
   %res1 = insertelement <16 x float> %res0, float %val1, i32 1
   %res2 = insertelement <16 x float> %res1, float   0.0, i32 2
; ALL-NEXT:    retq
   %resF = insertelement <16 x float> %resE, float %valF, i32 15
   ret <16 x float> %resF
 }
 
 define <16 x i32> @merge_16i32_i32_12zzzuuuuuuuuuuuz(i32* %ptr) nounwind uwtable noinline ssp {
 ; ALL-LABEL: merge_16i32_i32_12zzzuuuuuuuuuuuz:
 ; ALL:       # BB#0:
 ; ALL-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
-; ALL-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; ALL-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; ALL-NEXT:    vpxor %ymm1, %ymm1, %ymm1
-; ALL-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; ALL-NEXT:    retq
   %ptr0 = getelementptr inbounds i32, i32* %ptr, i64 1
   %ptr1 = getelementptr inbounds i32, i32* %ptr, i64 2
   %val0 = load i32, i32* %ptr0
   %val1 = load i32, i32* %ptr1
   %res0 = insertelement <16 x i32> undef, i32 %val0, i32 0
   %res1 = insertelement <16 x i32> %res0, i32 %val1, i32 1
   %res2 = insertelement <16 x i32> %res1, i32     0, i32 2

 ; AVX512F:       # BB#0:
 ; AVX512F-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
 ; AVX512F-NEXT:    vxorps %ymm1, %ymm1, %ymm1
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: merge_32i16_i16_12u4uuuuuuuuuuuuuuuuuuuuuuuuuuzz:
 ; AVX512BW:       # BB#0:
 ; AVX512BW-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
-; AVX512BW-NEXT:    vpxor %ymm1, %ymm1, %ymm1
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512BW-NEXT:    retq
   %ptr0 = getelementptr inbounds i16, i16* %ptr, i64 1
   %ptr1 = getelementptr inbounds i16, i16* %ptr, i64 2
   %ptr3 = getelementptr inbounds i16, i16* %ptr, i64 4
   %val0 = load i16, i16* %ptr0
   %val1 = load i16, i16* %ptr1
   %val3 = load i16, i16* %ptr3
   %res0 = insertelement <32 x i16> undef, i16 %val0, i16 0
; AVX512BW-NEXT:    retq
   %resH = insertelement <32 x i16> %resG, i16     0, i16 17
   ret <32 x i16> %resH
 }
 
 define <64 x i8> @merge_64i8_i8_12u4uuu8uuuuuuzzzzuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuz(i8* %ptr) nounwind uwtable noinline ssp {
 ; AVX512F-LABEL: merge_64i8_i8_12u4uuu8uuuuuuzzzzuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuz:
 ; AVX512F:       # BB#0:
 ; AVX512F-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
-; AVX512F-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; AVX512F-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX512F-NEXT:    vxorps %ymm1, %ymm1, %ymm1
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: merge_64i8_i8_12u4uuu8uuuuuuzzzzuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuz:
 ; AVX512BW:       # BB#0:
 ; AVX512BW-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512BW-NEXT:    vpxor %ymm1, %ymm1, %ymm1
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512BW-NEXT:    retq
   %ptr0 = getelementptr inbounds i8, i8* %ptr, i64 1
   %ptr1 = getelementptr inbounds i8, i8* %ptr, i64 2
   %ptr3 = getelementptr inbounds i8, i8* %ptr, i64 4
   %ptr7 = getelementptr inbounds i8, i8* %ptr, i64 8
   %val0 = load i8, i8* %ptr0
   %val1 = load i8, i8* %ptr1
   %val3 = load i8, i8* %ptr3