[X86] AVX512: Add writemask variants for vperm*2*

This includes assembler and codegen support (see the new tests in
avx512-encodings.s and avx512-shuffle.ll).

<rdar://problem/17492620>

llvm-svn: 212221

1 files changed, 68 insertions, 14 deletions

diff --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td
index 02df80e5666..daff30bea45 100644
--- a/llvm/lib/Target/X86/X86InstrAVX512.td
+++ b/llvm/lib/Target/X86/X86InstrAVX512.td
@@ -609,7 +609,7 @@ defm VPERMPDZ  : avx512_perm<0x16, "vpermpd", VR512,  memopv8f64, f512mem,
 // -- VPERM2I - 3 source operands form --
 multiclass avx512_perm_3src<bits<8> opc, string OpcodeStr, RegisterClass RC,
                           PatFrag mem_frag, X86MemOperand x86memop,
-                          SDNode OpNode, ValueType OpVT> {
+                          SDNode OpNode, ValueType OpVT, RegisterClass KRC> {
 let Constraints = "$src1 = $dst" in {
   def rr : AVX5128I<opc, MRMSrcReg, (outs RC:$dst),
                    (ins RC:$src1, RC:$src2, RC:$src3),
@@ -619,32 +619,86 @@ let Constraints = "$src1 = $dst" in {
                      (OpVT (OpNode RC:$src1, RC:$src2, RC:$src3)))]>,
                     EVEX_4V;
 
+  def rrk : AVX5128I<opc, MRMSrcReg, (outs RC:$dst),
+                   (ins RC:$src1, KRC:$mask, RC:$src2, RC:$src3),
+                   !strconcat(OpcodeStr,
+                       " \t{$src3, $src2, $dst {${mask}}|"
+                       "$dst {${mask}}, $src2, $src3}"),
+                   [(set RC:$dst, (OpVT (vselect KRC:$mask,
+                                           (OpNode RC:$src1, RC:$src2,
+                                              RC:$src3),
+                                           RC:$src1)))]>,
+                    EVEX_4V, EVEX_K;
+
+  let AddedComplexity = 30 in // Prefer over VMOV*rrkz Pat<>
+    def rrkz : AVX5128I<opc, MRMSrcReg, (outs RC:$dst),
+                   (ins RC:$src1, KRC:$mask, RC:$src2, RC:$src3),
+                   !strconcat(OpcodeStr,
+                       " \t{$src3, $src2, $dst {${mask}} {z} |",
+                       "$dst {${mask}} {z}, $src2, $src3}"),
+                   [(set RC:$dst, (OpVT (vselect KRC:$mask,
+                                           (OpNode RC:$src1, RC:$src2,
+                                              RC:$src3),
+                                           (OpVT (bitconvert
+                                              (v16i32 immAllZerosV))))))]>,
+                    EVEX_4V, EVEX_KZ;
+
   def rm : AVX5128I<opc, MRMSrcMem, (outs RC:$dst),
                    (ins RC:$src1, RC:$src2, x86memop:$src3),
                    !strconcat(OpcodeStr,
                     " \t{$src3, $src2, $dst|$dst, $src2, $src3}"),
                    [(set RC:$dst,
-                     (OpVT (OpNode RC:$src1, RC:$src2, 
+                     (OpVT (OpNode RC:$src1, RC:$src2,
                       (mem_frag addr:$src3))))]>, EVEX_4V;
+
+  def rmk : AVX5128I<opc, MRMSrcMem, (outs RC:$dst),
+                   (ins RC:$src1, KRC:$mask, RC:$src2, x86memop:$src3),
+                   !strconcat(OpcodeStr,
+                    " \t{$src3, $src2, $dst {${mask}}|"
+                    "$dst {${mask}}, $src2, $src3}"),
+                   [(set RC:$dst,
+                       (OpVT (vselect KRC:$mask,
+                                      (OpNode RC:$src1, RC:$src2,
+                                         (mem_frag addr:$src3)),
+                                      RC:$src1)))]>,
+                    EVEX_4V, EVEX_K;
+
+  let AddedComplexity = 10 in // Prefer over the rrkz variant
+    def rmkz : AVX5128I<opc, MRMSrcMem, (outs RC:$dst),
+                   (ins RC:$src1, KRC:$mask, RC:$src2, x86memop:$src3),
+                   !strconcat(OpcodeStr,
+                    " \t{$src3, $src2, $dst {${mask}} {z}|"
+                    "$dst {${mask}} {z}, $src2, $src3}"),
+                   [(set RC:$dst,
+                     (OpVT (vselect KRC:$mask,
+                                    (OpNode RC:$src1, RC:$src2,
+                                            (mem_frag addr:$src3)),
+                                    (OpVT (bitconvert
+                                       (v16i32 immAllZerosV))))))]>,
+                    EVEX_4V, EVEX_KZ;
   }
 }
-defm VPERMI2D  : avx512_perm_3src<0x76, "vpermi2d",  VR512, memopv16i32, i512mem, 
-                               X86VPermiv3, v16i32>, EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VPERMI2Q  : avx512_perm_3src<0x76, "vpermi2q",  VR512, memopv8i64, i512mem, 
-                               X86VPermiv3, v8i64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
-defm VPERMI2PS : avx512_perm_3src<0x77, "vpermi2ps",  VR512, memopv16f32, i512mem, 
-                               X86VPermiv3, v16f32>, EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VPERMI2PD : avx512_perm_3src<0x77, "vpermi2pd",  VR512, memopv8f64, i512mem, 
-                               X86VPermiv3, v8f64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+defm VPERMI2D  : avx512_perm_3src<0x76, "vpermi2d",  VR512, memopv16i32,
+                                  i512mem, X86VPermiv3, v16i32, VK16WM>,
+                 EVEX_V512, EVEX_CD8<32, CD8VF>;
+defm VPERMI2Q  : avx512_perm_3src<0x76, "vpermi2q",  VR512, memopv8i64,
+                                  i512mem, X86VPermiv3, v8i64, VK8WM>,
+                 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+defm VPERMI2PS : avx512_perm_3src<0x77, "vpermi2ps",  VR512, memopv16f32,
+                                  i512mem, X86VPermiv3, v16f32, VK16WM>,
+                 EVEX_V512, EVEX_CD8<32, CD8VF>;
+defm VPERMI2PD : avx512_perm_3src<0x77, "vpermi2pd",  VR512, memopv8f64,
+                                  i512mem, X86VPermiv3, v8f64, VK8WM>,
+                  EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
 
 defm VPERMT2D  : avx512_perm_3src<0x7E, "vpermt2d",  VR512, memopv16i32, i512mem, 
-                               X86VPermv3, v16i32>, EVEX_V512, EVEX_CD8<32, CD8VF>;
+                               X86VPermv3, v16i32, VK16WM>, EVEX_V512, EVEX_CD8<32, CD8VF>;
 defm VPERMT2Q  : avx512_perm_3src<0x7E, "vpermt2q",  VR512, memopv8i64, i512mem, 
-                               X86VPermv3, v8i64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+                               X86VPermv3, v8i64, VK8WM>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
 defm VPERMT2PS : avx512_perm_3src<0x7F, "vpermt2ps",  VR512, memopv16f32, i512mem, 
-                               X86VPermv3, v16f32>, EVEX_V512, EVEX_CD8<32, CD8VF>;
+                               X86VPermv3, v16f32, VK16WM>, EVEX_V512, EVEX_CD8<32, CD8VF>;
 defm VPERMT2PD : avx512_perm_3src<0x7F, "vpermt2pd",  VR512, memopv8f64, i512mem, 
-                               X86VPermv3, v8f64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+                               X86VPermv3, v8f64, VK8WM>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
 
 def : Pat<(v16f32 (int_x86_avx512_mask_vpermt_ps_512 (v16i32 VR512:$idx),
                    (v16f32 VR512:$src1), (v16f32 VR512:$src2), (i16 -1))),