[AVX512] Implement EXTLOAD lowering with patterns to select existing VPMOVZX instruct...

llvm-svn: 275378

diff --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td
index 9bbc3cc..4c9cf38 100644 (file)
--- a/llvm/lib/Target/X86/X86InstrAVX512.td
+++ b/llvm/lib/Target/X86/X86InstrAVX512.td
@@ -6484,9 +6484,7 @@ def: Pat<(v16i8 (X86vtrunc (v16i16 VR256X:$src))),
 
 multiclass avx512_extend_common<bits<8> opc, string OpcodeStr,
               X86VectorVTInfo DestInfo, X86VectorVTInfo SrcInfo,
-              X86MemOperand x86memop, PatFrag LdFrag, SDPatternOperator OpNode,
-              bit IsCodeGenOnly>{
-  let isCodeGenOnly = IsCodeGenOnly in {
+              X86MemOperand x86memop, PatFrag LdFrag, SDPatternOperator OpNode>{
   defm rr   : AVX512_maskable<opc, MRMSrcReg, DestInfo, (outs DestInfo.RC:$dst),
                     (ins SrcInfo.RC:$src), OpcodeStr ,"$src", "$src",
                     (DestInfo.VT (OpNode (SrcInfo.VT SrcInfo.RC:$src)))>,
@@ -6496,145 +6494,177 @@ multiclass avx512_extend_common<bits<8> opc, string OpcodeStr,
                   (ins x86memop:$src), OpcodeStr ,"$src", "$src",
                   (DestInfo.VT (LdFrag addr:$src))>,
                 EVEX;
-  }//isCodeGenOnly
 }
 
 multiclass avx512_extend_BW<bits<8> opc, string OpcodeStr,
-          SDPatternOperator OpNode, bit IsCodeGenOnly,
+          SDPatternOperator OpNode,
           string ExtTy,PatFrag LdFrag = !cast<PatFrag>(ExtTy#"extloadvi8")> {
   let Predicates = [HasVLX, HasBWI] in {
     defm Z128:  avx512_extend_common<opc, OpcodeStr, v8i16x_info,
-                    v16i8x_info, i64mem, LdFrag, OpNode, IsCodeGenOnly>,
+                    v16i8x_info, i64mem, LdFrag, OpNode>,
                      EVEX_CD8<8, CD8VH>, T8PD, EVEX_V128;
 
     defm Z256:  avx512_extend_common<opc, OpcodeStr, v16i16x_info,
-                    v16i8x_info, i128mem, LdFrag, OpNode, IsCodeGenOnly>,
+                    v16i8x_info, i128mem, LdFrag, OpNode>,
                      EVEX_CD8<8, CD8VH>, T8PD, EVEX_V256;
   }
   let Predicates = [HasBWI] in {
     defm Z   :  avx512_extend_common<opc, OpcodeStr, v32i16_info,
-                    v32i8x_info, i256mem, LdFrag, OpNode, IsCodeGenOnly>,
+                    v32i8x_info, i256mem, LdFrag, OpNode>,
                      EVEX_CD8<8, CD8VH>, T8PD, EVEX_V512;
   }
 }
 
 multiclass avx512_extend_BD<bits<8> opc, string OpcodeStr,
-          SDPatternOperator OpNode, bit IsCodeGenOnly,
+          SDPatternOperator OpNode,
           string ExtTy,PatFrag LdFrag = !cast<PatFrag>(ExtTy#"extloadvi8")> {
   let Predicates = [HasVLX, HasAVX512] in {
     defm Z128:  avx512_extend_common<opc, OpcodeStr, v4i32x_info,
-                   v16i8x_info, i32mem, LdFrag, OpNode, IsCodeGenOnly>,
+                   v16i8x_info, i32mem, LdFrag, OpNode>,
                          EVEX_CD8<8, CD8VQ>, T8PD, EVEX_V128;
 
     defm Z256:  avx512_extend_common<opc, OpcodeStr, v8i32x_info,
-                   v16i8x_info, i64mem, LdFrag, OpNode, IsCodeGenOnly>,
+                   v16i8x_info, i64mem, LdFrag, OpNode>,
                          EVEX_CD8<8, CD8VQ>, T8PD, EVEX_V256;
   }
   let Predicates = [HasAVX512] in {
     defm Z   :  avx512_extend_common<opc, OpcodeStr, v16i32_info,
-                   v16i8x_info, i128mem, LdFrag, OpNode, IsCodeGenOnly>,
+                   v16i8x_info, i128mem, LdFrag, OpNode>,
                          EVEX_CD8<8, CD8VQ>, T8PD, EVEX_V512;
   }
 }
 
 multiclass avx512_extend_BQ<bits<8> opc, string OpcodeStr,
-          SDPatternOperator OpNode, bit IsCodeGenOnly,
+          SDPatternOperator OpNode,
           string ExtTy,PatFrag LdFrag = !cast<PatFrag>(ExtTy#"extloadvi8")> {
   let Predicates = [HasVLX, HasAVX512] in {
     defm Z128:  avx512_extend_common<opc, OpcodeStr, v2i64x_info,
-                   v16i8x_info, i16mem, LdFrag, OpNode, IsCodeGenOnly>,
+                   v16i8x_info, i16mem, LdFrag, OpNode>,
                      EVEX_CD8<8, CD8VO>, T8PD, EVEX_V128;
 
     defm Z256:  avx512_extend_common<opc, OpcodeStr, v4i64x_info,
-                   v16i8x_info, i32mem, LdFrag, OpNode, IsCodeGenOnly>,
+                   v16i8x_info, i32mem, LdFrag, OpNode>,
                      EVEX_CD8<8, CD8VO>, T8PD, EVEX_V256;
   }
   let Predicates = [HasAVX512] in {
     defm Z   :  avx512_extend_common<opc, OpcodeStr, v8i64_info,
-                   v16i8x_info, i64mem, LdFrag, OpNode, IsCodeGenOnly>,
+                   v16i8x_info, i64mem, LdFrag, OpNode>,
                      EVEX_CD8<8, CD8VO>, T8PD, EVEX_V512;
   }
 }
 
 multiclass avx512_extend_WD<bits<8> opc, string OpcodeStr,
-         SDPatternOperator OpNode, bit IsCodeGenOnly,
+         SDPatternOperator OpNode,
          string ExtTy,PatFrag LdFrag = !cast<PatFrag>(ExtTy#"extloadvi16")> {
   let Predicates = [HasVLX, HasAVX512] in {
     defm Z128:  avx512_extend_common<opc, OpcodeStr, v4i32x_info,
-                   v8i16x_info, i64mem, LdFrag, OpNode, IsCodeGenOnly>,
+                   v8i16x_info, i64mem, LdFrag, OpNode>,
                      EVEX_CD8<16, CD8VH>, T8PD, EVEX_V128;
 
     defm Z256:  avx512_extend_common<opc, OpcodeStr, v8i32x_info,
-                   v8i16x_info, i128mem, LdFrag, OpNode, IsCodeGenOnly>,
+                   v8i16x_info, i128mem, LdFrag, OpNode>,
                      EVEX_CD8<16, CD8VH>, T8PD, EVEX_V256;
   }
   let Predicates = [HasAVX512] in {
     defm Z   :  avx512_extend_common<opc, OpcodeStr, v16i32_info,
-                   v16i16x_info, i256mem, LdFrag, OpNode, IsCodeGenOnly>,
+                   v16i16x_info, i256mem, LdFrag, OpNode>,
                      EVEX_CD8<16, CD8VH>, T8PD, EVEX_V512;
   }
 }
 
 multiclass avx512_extend_WQ<bits<8> opc, string OpcodeStr,
-         SDPatternOperator OpNode, bit IsCodeGenOnly,
+         SDPatternOperator OpNode,
          string ExtTy,PatFrag LdFrag = !cast<PatFrag>(ExtTy#"extloadvi16")> {
   let Predicates = [HasVLX, HasAVX512] in {
     defm Z128:  avx512_extend_common<opc, OpcodeStr, v2i64x_info,
-                   v8i16x_info, i32mem, LdFrag, OpNode, IsCodeGenOnly>,
+                   v8i16x_info, i32mem, LdFrag, OpNode>,
                      EVEX_CD8<16, CD8VQ>, T8PD, EVEX_V128;
 
     defm Z256:  avx512_extend_common<opc, OpcodeStr, v4i64x_info,
-                   v8i16x_info, i64mem, LdFrag, OpNode, IsCodeGenOnly>,
+                   v8i16x_info, i64mem, LdFrag, OpNode>,
                      EVEX_CD8<16, CD8VQ>, T8PD, EVEX_V256;
   }
   let Predicates = [HasAVX512] in {
     defm Z   :  avx512_extend_common<opc, OpcodeStr, v8i64_info,
-                   v8i16x_info, i128mem, LdFrag, OpNode, IsCodeGenOnly>,
+                   v8i16x_info, i128mem, LdFrag, OpNode>,
                      EVEX_CD8<16, CD8VQ>, T8PD, EVEX_V512;
   }
 }
 
 multiclass avx512_extend_DQ<bits<8> opc, string OpcodeStr,
-         SDPatternOperator OpNode, bit IsCodeGenOnly,
+         SDPatternOperator OpNode,
          string ExtTy,PatFrag LdFrag = !cast<PatFrag>(ExtTy#"extloadvi32")> {
 
   let Predicates = [HasVLX, HasAVX512] in {
     defm Z128:  avx512_extend_common<opc, OpcodeStr, v2i64x_info,
-                   v4i32x_info, i64mem, LdFrag, OpNode, IsCodeGenOnly>,
+                   v4i32x_info, i64mem, LdFrag, OpNode>,
                      EVEX_CD8<32, CD8VH>, T8PD, EVEX_V128;
 
     defm Z256:  avx512_extend_common<opc, OpcodeStr, v4i64x_info,
-                   v4i32x_info, i128mem, LdFrag, OpNode, IsCodeGenOnly>,
+                   v4i32x_info, i128mem, LdFrag, OpNode>,
                      EVEX_CD8<32, CD8VH>, T8PD, EVEX_V256;
   }
   let Predicates = [HasAVX512] in {
     defm Z   :  avx512_extend_common<opc, OpcodeStr, v8i64_info,
-                   v8i32x_info, i256mem, LdFrag, OpNode, IsCodeGenOnly>,
+                   v8i32x_info, i256mem, LdFrag, OpNode>,
                      EVEX_CD8<32, CD8VH>, T8PD, EVEX_V512;
   }
 }
 
-defm VPMOVZXBW : avx512_extend_BW<0x30, "vpmovzxbw", X86vzext, 0, "z">;
-defm VPMOVZXBD : avx512_extend_BD<0x31, "vpmovzxbd", X86vzext, 0, "z">;
-defm VPMOVZXBQ : avx512_extend_BQ<0x32, "vpmovzxbq", X86vzext, 0, "z">;
-defm VPMOVZXWD : avx512_extend_WD<0x33, "vpmovzxwd", X86vzext, 0, "z">;
-defm VPMOVZXWQ : avx512_extend_WQ<0x34, "vpmovzxwq", X86vzext, 0, "z">;
-defm VPMOVZXDQ : avx512_extend_DQ<0x35, "vpmovzxdq", X86vzext, 0, "z">;
+defm VPMOVZXBW : avx512_extend_BW<0x30, "vpmovzxbw", X86vzext, "z">;
+defm VPMOVZXBD : avx512_extend_BD<0x31, "vpmovzxbd", X86vzext, "z">;
+defm VPMOVZXBQ : avx512_extend_BQ<0x32, "vpmovzxbq", X86vzext, "z">;
+defm VPMOVZXWD : avx512_extend_WD<0x33, "vpmovzxwd", X86vzext, "z">;
+defm VPMOVZXWQ : avx512_extend_WQ<0x34, "vpmovzxwq", X86vzext, "z">;
+defm VPMOVZXDQ : avx512_extend_DQ<0x35, "vpmovzxdq", X86vzext, "z">;
 
-defm VPMOVSXBW: avx512_extend_BW<0x20, "vpmovsxbw", X86vsext, 0, "s">;
-defm VPMOVSXBD: avx512_extend_BD<0x21, "vpmovsxbd", X86vsext, 0, "s">;
-defm VPMOVSXBQ: avx512_extend_BQ<0x22, "vpmovsxbq", X86vsext, 0, "s">;
-defm VPMOVSXWD: avx512_extend_WD<0x23, "vpmovsxwd", X86vsext, 0, "s">;
-defm VPMOVSXWQ: avx512_extend_WQ<0x24, "vpmovsxwq", X86vsext, 0, "s">;
-defm VPMOVSXDQ: avx512_extend_DQ<0x25, "vpmovsxdq", X86vsext, 0, "s">;
+defm VPMOVSXBW: avx512_extend_BW<0x20, "vpmovsxbw", X86vsext, "s">;
+defm VPMOVSXBD: avx512_extend_BD<0x21, "vpmovsxbd", X86vsext, "s">;
+defm VPMOVSXBQ: avx512_extend_BQ<0x22, "vpmovsxbq", X86vsext, "s">;
+defm VPMOVSXWD: avx512_extend_WD<0x23, "vpmovsxwd", X86vsext, "s">;
+defm VPMOVSXWQ: avx512_extend_WQ<0x24, "vpmovsxwq", X86vsext, "s">;
+defm VPMOVSXDQ: avx512_extend_DQ<0x25, "vpmovsxdq", X86vsext, "s">;
 
 // EXTLOAD patterns, implemented using vpmovz
-defm VPMOVAXBW : avx512_extend_BW<0x30, "vpmovzxbw", null_frag, 1, "">;
-defm VPMOVAXBD : avx512_extend_BD<0x31, "vpmovzxbd", null_frag, 1, "">;
-defm VPMOVAXBQ : avx512_extend_BQ<0x32, "vpmovzxbq", null_frag, 1, "">;
-defm VPMOVAXWD : avx512_extend_WD<0x33, "vpmovzxwd", null_frag, 1, "">;
-defm VPMOVAXWQ : avx512_extend_WQ<0x34, "vpmovzxwq", null_frag, 1, "">;
-defm VPMOVAXDQ : avx512_extend_DQ<0x35, "vpmovzxdq", null_frag, 1, "">;
+multiclass avx512_ext_lowering<string InstrStr, X86VectorVTInfo To,
+                               X86VectorVTInfo From, PatFrag LdFrag> {
+  def : Pat<(To.VT (LdFrag addr:$src)),
+            (!cast<Instruction>("VPMOVZX"#InstrStr#"rm") addr:$src)>;
+  def : Pat<(To.VT (vselect To.KRCWM:$mask, (LdFrag addr:$src), To.RC:$src0)),
+            (!cast<Instruction>("VPMOVZX"#InstrStr#"rmk") To.RC:$src0,
+             To.KRC:$mask, addr:$src)>;
+  def : Pat<(To.VT (vselect To.KRCWM:$mask, (LdFrag addr:$src),
+                    To.ImmAllZerosV)),
+            (!cast<Instruction>("VPMOVZX"#InstrStr#"rmkz") To.KRC:$mask,
+             addr:$src)>;
+}
+
+let Predicates = [HasVLX, HasBWI] in {
+  defm : avx512_ext_lowering<"BWZ128", v8i16x_info,  v16i8x_info,  extloadvi8>;
+  defm : avx512_ext_lowering<"BWZ256", v16i16x_info, v16i8x_info,  extloadvi8>;
+}
+let Predicates = [HasBWI] in {
+  defm : avx512_ext_lowering<"BWZ",    v32i16_info,  v32i8x_info,  extloadvi8>;
+}
+let Predicates = [HasVLX, HasAVX512] in {
+  defm : avx512_ext_lowering<"BDZ128", v4i32x_info,  v16i8x_info,  extloadvi8>;
+  defm : avx512_ext_lowering<"BDZ256", v8i32x_info,  v16i8x_info,  extloadvi8>;
+  defm : avx512_ext_lowering<"BQZ128", v2i64x_info,  v16i8x_info,  extloadvi8>;
+  defm : avx512_ext_lowering<"BQZ256", v4i64x_info,  v16i8x_info,  extloadvi8>;
+  defm : avx512_ext_lowering<"WDZ128", v4i32x_info,  v8i16x_info,  extloadvi16>;
+  defm : avx512_ext_lowering<"WDZ256", v8i32x_info,  v8i16x_info,  extloadvi16>;
+  defm : avx512_ext_lowering<"WQZ128", v2i64x_info,  v8i16x_info,  extloadvi16>;
+  defm : avx512_ext_lowering<"WQZ256", v4i64x_info,  v8i16x_info,  extloadvi16>;
+  defm : avx512_ext_lowering<"DQZ128", v2i64x_info,  v4i32x_info,  extloadvi32>;
+  defm : avx512_ext_lowering<"DQZ256", v4i64x_info,  v4i32x_info,  extloadvi32>;
+}
+let Predicates = [HasAVX512] in {
+  defm : avx512_ext_lowering<"BDZ",    v16i32_info,  v16i8x_info,  extloadvi8>;
+  defm : avx512_ext_lowering<"BQZ",    v8i64_info,   v16i8x_info,  extloadvi8>;
+  defm : avx512_ext_lowering<"WDZ",    v16i32_info,  v16i16x_info, extloadvi16>;
+  defm : avx512_ext_lowering<"WQZ",    v8i64_info,   v8i16x_info,  extloadvi16>;
+  defm : avx512_ext_lowering<"DQZ",    v8i64_info,   v8i32x_info,  extloadvi32>;
+}
 
 //===----------------------------------------------------------------------===//
 // GATHER - SCATTER Operations