[AVX512] Bring back vector-shuffle lowering support through broadcasts

[oota-llvm.git] / lib / Target / X86 / X86InstrAVX512.td

diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index 07fe310d0395be2683ef45b4d38927c54f639073..3dbc3d2abd8fdd62775ee5af007321afb70f48bb 100644 (file)
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -711,6 +711,16 @@ def : Pat<(v16f32 (X86VBroadcast (v4f32 VR128X:$src))),
 def : Pat<(v8f64 (X86VBroadcast (v2f64 VR128X:$src))),
           (VBROADCASTSDZrr VR128X:$src)>;
 
 def : Pat<(v8f64 (X86VBroadcast (v2f64 VR128X:$src))),
           (VBROADCASTSDZrr VR128X:$src)>;
 

+def : Pat<(v16f32 (X86VBroadcast (v16f32 VR512:$src))),
+          (VBROADCASTSSZrr (EXTRACT_SUBREG (v16f32 VR512:$src), sub_xmm))>;
+def : Pat<(v8f64 (X86VBroadcast (v8f64 VR512:$src))),
+          (VBROADCASTSDZrr (EXTRACT_SUBREG (v8f64 VR512:$src), sub_xmm))>;
+
+def : Pat<(v16i32 (X86VBroadcast (v16i32 VR512:$src))),
+          (VPBROADCASTDZrr (EXTRACT_SUBREG (v16i32 VR512:$src), sub_xmm))>;
+def : Pat<(v8i64 (X86VBroadcast (v8i64 VR512:$src))),
+          (VPBROADCASTQZrr (EXTRACT_SUBREG (v8i64 VR512:$src), sub_xmm))>;
+

 def : Pat<(v16f32 (int_x86_avx512_vbroadcast_ss_ps_512 (v4f32 VR128X:$src))),
           (VBROADCASTSSZrr VR128X:$src)>;
 def : Pat<(v8f64 (int_x86_avx512_vbroadcast_sd_pd_512 (v2f64 VR128X:$src))),
 def : Pat<(v16f32 (int_x86_avx512_vbroadcast_ss_ps_512 (v4f32 VR128X:$src))),
           (VBROADCASTSSZrr VR128X:$src)>;
 def : Pat<(v8f64 (int_x86_avx512_vbroadcast_sd_pd_512 (v2f64 VR128X:$src))),


@@ -783,14 +793,37 @@ multiclass avx512_perm_imm<bits<8> opc, string OpcodeStr, SDNode OpNode,
 }
 }
 
 }
 }
 

+multiclass avx512_permil<bits<8> OpcImm, bits<8> OpcVar, X86VectorVTInfo _,
+                         X86VectorVTInfo Ctrl> :
+     avx512_perm_imm<OpcImm, "vpermil" # _.Suffix, X86VPermilpi, _> {
+  let ExeDomain = _.ExeDomain in {
+    def rr : AVX5128I<OpcVar, MRMSrcReg, (outs _.RC:$dst),
+                     (ins _.RC:$src1, _.RC:$src2),
+                     !strconcat("vpermil" # _.Suffix,
+                         " \t{$src2, $src1, $dst|$dst, $src1, $src2}"),
+                     [(set _.RC:$dst,
+                         (_.VT (X86VPermilpv _.RC:$src1,
+                                  (Ctrl.VT Ctrl.RC:$src2))))]>,
+             EVEX_4V;
+    def rm : AVX5128I<OpcVar, MRMSrcMem, (outs _.RC:$dst),
+                     (ins _.RC:$src1, Ctrl.MemOp:$src2),
+                     !strconcat("vpermil" # _.Suffix,
+                         " \t{$src2, $src1, $dst|$dst, $src1, $src2}"),
+                     [(set _.RC:$dst,
+                         (_.VT (X86VPermilpv _.RC:$src1,
+                                  (Ctrl.VT (Ctrl.MemOpFrag addr:$src2)))))]>,
+             EVEX_4V;
+  }
+}
+

 defm VPERMQZ :    avx512_perm_imm<0x00, "vpermq", X86VPermi, v8i64_info>,
                   EVEX_V512, VEX_W;
 defm VPERMPDZ :   avx512_perm_imm<0x01, "vpermpd", X86VPermi, v8f64_info>,
                   EVEX_V512, VEX_W;
 
 defm VPERMQZ :    avx512_perm_imm<0x00, "vpermq", X86VPermi, v8i64_info>,
                   EVEX_V512, VEX_W;
 defm VPERMPDZ :   avx512_perm_imm<0x01, "vpermpd", X86VPermi, v8f64_info>,
                   EVEX_V512, VEX_W;
 

-defm VPERMILPSZ : avx512_perm_imm<0x04, "vpermilps", X86VPermilpi, v16f32_info>,
+defm VPERMILPSZ : avx512_permil<0x04, 0x0C, v16f32_info, v16i32_info>,

                   EVEX_V512;
                   EVEX_V512;

-defm VPERMILPDZ : avx512_perm_imm<0x05, "vpermilpd", X86VPermilpi, v8f64_info>,
+defm VPERMILPDZ : avx512_permil<0x05, 0x0D, v8f64_info, v8i64_info>,

                   EVEX_V512, VEX_W;
 
 def : Pat<(v16i32 (X86VPermilpi VR512:$src1, (i8 imm:$imm))),
                   EVEX_V512, VEX_W;
 
 def : Pat<(v16i32 (X86VPermilpi VR512:$src1, (i8 imm:$imm))),