[AVX512] Extended avx512_sqrt_packed (sqrt instructions) to VL subset.

Refactored through AVX512_maskable

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@220806 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index 04dc681432a88397304ca255c8ba201914fb2ce5..3fd18de6537ded160eda2284896f710a1b491676 100644 (file)
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -4273,32 +4273,24 @@ def : Pat <(v8f64 (int_x86_avx512_rcp28_pd (v8f64 VR512:$src),
               (bc_v8f64 (v16i32 immAllZerosV)), (i8 -1), FROUND_NO_EXC)),
            (VRCP28PDZrb VR512:$src)>;
 
               (bc_v8f64 (v16i32 immAllZerosV)), (i8 -1), FROUND_NO_EXC)),
            (VRCP28PDZrb VR512:$src)>;
 

-multiclass avx512_sqrt_packed<bits<8> opc, string OpcodeStr, SDNode OpNode,
-                              OpndItins itins_s, OpndItins itins_d> {
-  def PSZrr :AVX512PSI<opc, MRMSrcReg, (outs VR512:$dst), (ins VR512:$src),
-             !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
-             [(set VR512:$dst, (v16f32 (OpNode VR512:$src)))], itins_s.rr>,
-             EVEX, EVEX_V512;
-
-  let mayLoad = 1 in
-  def PSZrm : AVX512PSI<opc, MRMSrcMem, (outs VR512:$dst), (ins f512mem:$src),
-              !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
-              [(set VR512:$dst, 
-                (OpNode (v16f32 (bitconvert (memopv16f32 addr:$src)))))],
-              itins_s.rm>, EVEX, EVEX_V512, EVEX_CD8<32, CD8VF>;
-
-  def PDZrr : AVX512PDI<opc, MRMSrcReg, (outs VR512:$dst), (ins VR512:$src),
-              !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
-              [(set VR512:$dst, (v8f64 (OpNode VR512:$src)))], itins_d.rr>,
-              EVEX, EVEX_V512;
-
-  let mayLoad = 1 in
-    def PDZrm : AVX512PDI<opc, MRMSrcMem, (outs VR512:$dst), (ins f512mem:$src),
-                !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
-                [(set VR512:$dst, (OpNode
-                  (v8f64 (bitconvert (memopv16f32 addr:$src)))))],
-                itins_d.rm>, EVEX, EVEX_V512, EVEX_CD8<64, CD8VF>;
+multiclass avx512_sqrt_packed<bits<8> opc, string OpcodeStr,
+                              SDNode OpNode, X86VectorVTInfo _>{
+  defm rr: AVX512_maskable<opc, MRMSrcReg, _, (outs _.RC:$dst),
+                         (ins _.RC:$src), OpcodeStr, "$src", "$src",
+                         (_.FloatVT (OpNode _.RC:$src))>, EVEX;
+  let mayLoad = 1 in {
+    defm rm: AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst),
+                           (ins _.MemOp:$src), OpcodeStr, "$src", "$src",
+                           (OpNode (_.FloatVT
+                             (bitconvert (_.LdFrag addr:$src))))>, EVEX;

 
 

+    defm rmb: AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst),
+                            (ins _.ScalarMemOp:$src), OpcodeStr,
+                            "${src}"##_.BroadcastStr, "${src}"##_.BroadcastStr,
+                            (OpNode (_.FloatVT
+                              (X86VBroadcast (_.ScalarLdFrag addr:$src))))>,
+                            EVEX, EVEX_B;
+  }

 }
 
 multiclass avx512_sqrt_scalar<bits<8> opc, string OpcodeStr,
 }
 
 multiclass avx512_sqrt_scalar<bits<8> opc, string OpcodeStr,


@@ -4362,12 +4354,36 @@ multiclass avx512_sqrt_scalar<bits<8> opc, string OpcodeStr,
   }
 }
 
   }
 }
 

+multiclass avx512_sqrt_packed_all<bits<8> opc, string OpcodeStr,
+                                  SDNode OpNode> {
+  defm PSZ : avx512_sqrt_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode,
+                                v16f32_info>,
+                                EVEX_V512, PS, EVEX_CD8<32, CD8VF>;
+  defm PDZ : avx512_sqrt_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode,
+                                v8f64_info>,
+                                EVEX_V512, VEX_W, PD, EVEX_CD8<64, CD8VF>;
+  // Define only if AVX512VL feature is present.
+  let Predicates = [HasVLX] in {
+    defm PSZ128 : avx512_sqrt_packed<opc, !strconcat(OpcodeStr, "ps"),
+                                     OpNode, v4f32x_info>,
+                                     EVEX_V128, PS, EVEX_CD8<32, CD8VF>;
+    defm PSZ256 : avx512_sqrt_packed<opc, !strconcat(OpcodeStr, "ps"),
+                                     OpNode, v8f32x_info>,
+                                     EVEX_V256, PS, EVEX_CD8<32, CD8VF>;
+    defm PDZ128 : avx512_sqrt_packed<opc, !strconcat(OpcodeStr, "pd"),
+                                     OpNode, v2f64x_info>,
+                                     EVEX_V128, VEX_W, PD, EVEX_CD8<64, CD8VF>;
+    defm PDZ256 : avx512_sqrt_packed<opc, !strconcat(OpcodeStr, "pd"),
+                                     OpNode, v4f64x_info>,
+                                     EVEX_V256, VEX_W, PD, EVEX_CD8<64, CD8VF>;
+  }
+}
+
+defm VSQRT : avx512_sqrt_packed_all<0x51, "vsqrt", fsqrt>;

 
 defm VSQRT  : avx512_sqrt_scalar<0x51, "sqrt", 
                 int_x86_avx512_sqrt_ss, int_x86_avx512_sqrt_sd, 
 
 defm VSQRT  : avx512_sqrt_scalar<0x51, "sqrt", 
                 int_x86_avx512_sqrt_ss, int_x86_avx512_sqrt_sd, 

-                SSE_SQRTSS, SSE_SQRTSD>,
-              avx512_sqrt_packed<0x51, "vsqrt", fsqrt,
-                SSE_SQRTPS, SSE_SQRTPD>;
+                SSE_SQRTSS, SSE_SQRTSD>;

 
 let Predicates = [HasAVX512] in {
   def : Pat<(v16f32 (int_x86_avx512_sqrt_ps_512 (v16f32 VR512:$src1),
 
 let Predicates = [HasAVX512] in {
   def : Pat<(v16f32 (int_x86_avx512_sqrt_ps_512 (v16f32 VR512:$src1),

diff --git a/test/MC/X86/avx512-encodings.s b/test/MC/X86/avx512-encodings.s
index 64e615377b16eee01dc7bf2b566b7f1b74a32984..231226f347ea9b5c8acc06660c0663a4a314de6e 100644 (file)
--- a/test/MC/X86/avx512-encodings.s
+++ b/test/MC/X86/avx512-encodings.s
@@ -3417,6 +3417,118 @@
 // CHECK:  encoding: [0x62,0xe2,0x7d,0x58,0x4e,0x82,0xfc,0xfd,0xff,0xff]
           vrsqrt14ps -516(%rdx){1to16}, %zmm16
 
 // CHECK:  encoding: [0x62,0xe2,0x7d,0x58,0x4e,0x82,0xfc,0xfd,0xff,0xff]
           vrsqrt14ps -516(%rdx){1to16}, %zmm16
 

+// CHECK: vsqrtpd %zmm19, %zmm19
+// CHECK:  encoding: [0x62,0xa1,0xfd,0x48,0x51,0xdb]
+          vsqrtpd %zmm19, %zmm19
+
+// CHECK: vsqrtpd %zmm19, %zmm19 {%k5}
+// CHECK:  encoding: [0x62,0xa1,0xfd,0x4d,0x51,0xdb]
+          vsqrtpd %zmm19, %zmm19 {%k5}
+
+// CHECK: vsqrtpd %zmm19, %zmm19 {%k5} {z}
+// CHECK:  encoding: [0x62,0xa1,0xfd,0xcd,0x51,0xdb]
+          vsqrtpd %zmm19, %zmm19 {%k5} {z}
+
+// CHECK: vsqrtpd (%rcx), %zmm19
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x48,0x51,0x19]
+          vsqrtpd (%rcx), %zmm19
+
+// CHECK: vsqrtpd 291(%rax,%r14,8), %zmm19
+// CHECK:  encoding: [0x62,0xa1,0xfd,0x48,0x51,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vsqrtpd 291(%rax,%r14,8), %zmm19
+
+// CHECK: vsqrtpd (%rcx){1to8}, %zmm19
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x58,0x51,0x19]
+          vsqrtpd (%rcx){1to8}, %zmm19
+
+// CHECK: vsqrtpd 8128(%rdx), %zmm19
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x48,0x51,0x5a,0x7f]
+          vsqrtpd 8128(%rdx), %zmm19
+
+// CHECK: vsqrtpd 8192(%rdx), %zmm19
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x48,0x51,0x9a,0x00,0x20,0x00,0x00]
+          vsqrtpd 8192(%rdx), %zmm19
+
+// CHECK: vsqrtpd -8192(%rdx), %zmm19
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x48,0x51,0x5a,0x80]
+          vsqrtpd -8192(%rdx), %zmm19
+
+// CHECK: vsqrtpd -8256(%rdx), %zmm19
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x48,0x51,0x9a,0xc0,0xdf,0xff,0xff]
+          vsqrtpd -8256(%rdx), %zmm19
+
+// CHECK: vsqrtpd 1016(%rdx){1to8}, %zmm19
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x58,0x51,0x5a,0x7f]
+          vsqrtpd 1016(%rdx){1to8}, %zmm19
+
+// CHECK: vsqrtpd 1024(%rdx){1to8}, %zmm19
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x58,0x51,0x9a,0x00,0x04,0x00,0x00]
+          vsqrtpd 1024(%rdx){1to8}, %zmm19
+
+// CHECK: vsqrtpd -1024(%rdx){1to8}, %zmm19
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x58,0x51,0x5a,0x80]
+          vsqrtpd -1024(%rdx){1to8}, %zmm19
+
+// CHECK: vsqrtpd -1032(%rdx){1to8}, %zmm19
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x58,0x51,0x9a,0xf8,0xfb,0xff,0xff]
+          vsqrtpd -1032(%rdx){1to8}, %zmm19
+
+// CHECK: vsqrtps %zmm29, %zmm28
+// CHECK:  encoding: [0x62,0x01,0x7c,0x48,0x51,0xe5]
+          vsqrtps %zmm29, %zmm28
+
+// CHECK: vsqrtps %zmm29, %zmm28 {%k3}
+// CHECK:  encoding: [0x62,0x01,0x7c,0x4b,0x51,0xe5]
+          vsqrtps %zmm29, %zmm28 {%k3}
+
+// CHECK: vsqrtps %zmm29, %zmm28 {%k3} {z}
+// CHECK:  encoding: [0x62,0x01,0x7c,0xcb,0x51,0xe5]
+          vsqrtps %zmm29, %zmm28 {%k3} {z}
+
+// CHECK: vsqrtps (%rcx), %zmm28
+// CHECK:  encoding: [0x62,0x61,0x7c,0x48,0x51,0x21]
+          vsqrtps (%rcx), %zmm28
+
+// CHECK: vsqrtps 291(%rax,%r14,8), %zmm28
+// CHECK:  encoding: [0x62,0x21,0x7c,0x48,0x51,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vsqrtps 291(%rax,%r14,8), %zmm28
+
+// CHECK: vsqrtps (%rcx){1to16}, %zmm28
+// CHECK:  encoding: [0x62,0x61,0x7c,0x58,0x51,0x21]
+          vsqrtps (%rcx){1to16}, %zmm28
+
+// CHECK: vsqrtps 8128(%rdx), %zmm28
+// CHECK:  encoding: [0x62,0x61,0x7c,0x48,0x51,0x62,0x7f]
+          vsqrtps 8128(%rdx), %zmm28
+
+// CHECK: vsqrtps 8192(%rdx), %zmm28
+// CHECK:  encoding: [0x62,0x61,0x7c,0x48,0x51,0xa2,0x00,0x20,0x00,0x00]
+          vsqrtps 8192(%rdx), %zmm28
+
+// CHECK: vsqrtps -8192(%rdx), %zmm28
+// CHECK:  encoding: [0x62,0x61,0x7c,0x48,0x51,0x62,0x80]
+          vsqrtps -8192(%rdx), %zmm28
+
+// CHECK: vsqrtps -8256(%rdx), %zmm28
+// CHECK:  encoding: [0x62,0x61,0x7c,0x48,0x51,0xa2,0xc0,0xdf,0xff,0xff]
+          vsqrtps -8256(%rdx), %zmm28
+
+// CHECK: vsqrtps 508(%rdx){1to16}, %zmm28
+// CHECK:  encoding: [0x62,0x61,0x7c,0x58,0x51,0x62,0x7f]
+          vsqrtps 508(%rdx){1to16}, %zmm28
+
+// CHECK: vsqrtps 512(%rdx){1to16}, %zmm28
+// CHECK:  encoding: [0x62,0x61,0x7c,0x58,0x51,0xa2,0x00,0x02,0x00,0x00]
+          vsqrtps 512(%rdx){1to16}, %zmm28
+
+// CHECK: vsqrtps -512(%rdx){1to16}, %zmm28
+// CHECK:  encoding: [0x62,0x61,0x7c,0x58,0x51,0x62,0x80]
+          vsqrtps -512(%rdx){1to16}, %zmm28
+
+// CHECK: vsqrtps -516(%rdx){1to16}, %zmm28
+// CHECK:  encoding: [0x62,0x61,0x7c,0x58,0x51,0xa2,0xfc,0xfd,0xff,0xff]
+          vsqrtps -516(%rdx){1to16}, %zmm28
+

 // CHECK: vsubpd %zmm9, %zmm12, %zmm9
 // CHECK:  encoding: [0x62,0x51,0x9d,0x48,0x5c,0xc9]
           vsubpd %zmm9, %zmm12, %zmm9
 // CHECK: vsubpd %zmm9, %zmm12, %zmm9
 // CHECK:  encoding: [0x62,0x51,0x9d,0x48,0x5c,0xc9]
           vsubpd %zmm9, %zmm12, %zmm9

diff --git a/test/MC/X86/x86-64-avx512f_vl.s b/test/MC/X86/x86-64-avx512f_vl.s
index 039b4afab55f128610a5e77117e4efe3e86da051..1d5c4c4e9c02c9da36fd55399231be715c219ca5 100644 (file)
--- a/test/MC/X86/x86-64-avx512f_vl.s
+++ b/test/MC/X86/x86-64-avx512f_vl.s
@@ -4432,6 +4432,230 @@
 // CHECK:  encoding: [0x62,0x62,0x7d,0x38,0x4e,0x9a,0xfc,0xfd,0xff,0xff]
           vrsqrt14ps -516(%rdx){1to8}, %ymm27
 
 // CHECK:  encoding: [0x62,0x62,0x7d,0x38,0x4e,0x9a,0xfc,0xfd,0xff,0xff]
           vrsqrt14ps -516(%rdx){1to8}, %ymm27
 

+// CHECK: vsqrtpd %xmm26, %xmm29
+// CHECK:  encoding: [0x62,0x01,0xfd,0x08,0x51,0xea]
+          vsqrtpd %xmm26, %xmm29
+
+// CHECK: vsqrtpd %xmm26, %xmm29 {%k3}
+// CHECK:  encoding: [0x62,0x01,0xfd,0x0b,0x51,0xea]
+          vsqrtpd %xmm26, %xmm29 {%k3}
+
+// CHECK: vsqrtpd %xmm26, %xmm29 {%k3} {z}
+// CHECK:  encoding: [0x62,0x01,0xfd,0x8b,0x51,0xea]
+          vsqrtpd %xmm26, %xmm29 {%k3} {z}
+
+// CHECK: vsqrtpd (%rcx), %xmm29
+// CHECK:  encoding: [0x62,0x61,0xfd,0x08,0x51,0x29]
+          vsqrtpd (%rcx), %xmm29
+
+// CHECK: vsqrtpd 291(%rax,%r14,8), %xmm29
+// CHECK:  encoding: [0x62,0x21,0xfd,0x08,0x51,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vsqrtpd 291(%rax,%r14,8), %xmm29
+
+// CHECK: vsqrtpd (%rcx){1to2}, %xmm29
+// CHECK:  encoding: [0x62,0x61,0xfd,0x18,0x51,0x29]
+          vsqrtpd (%rcx){1to2}, %xmm29
+
+// CHECK: vsqrtpd 2032(%rdx), %xmm29
+// CHECK:  encoding: [0x62,0x61,0xfd,0x08,0x51,0x6a,0x7f]
+          vsqrtpd 2032(%rdx), %xmm29
+
+// CHECK: vsqrtpd 2048(%rdx), %xmm29
+// CHECK:  encoding: [0x62,0x61,0xfd,0x08,0x51,0xaa,0x00,0x08,0x00,0x00]
+          vsqrtpd 2048(%rdx), %xmm29
+
+// CHECK: vsqrtpd -2048(%rdx), %xmm29
+// CHECK:  encoding: [0x62,0x61,0xfd,0x08,0x51,0x6a,0x80]
+          vsqrtpd -2048(%rdx), %xmm29
+
+// CHECK: vsqrtpd -2064(%rdx), %xmm29
+// CHECK:  encoding: [0x62,0x61,0xfd,0x08,0x51,0xaa,0xf0,0xf7,0xff,0xff]
+          vsqrtpd -2064(%rdx), %xmm29
+
+// CHECK: vsqrtpd 1016(%rdx){1to2}, %xmm29
+// CHECK:  encoding: [0x62,0x61,0xfd,0x18,0x51,0x6a,0x7f]
+          vsqrtpd 1016(%rdx){1to2}, %xmm29
+
+// CHECK: vsqrtpd 1024(%rdx){1to2}, %xmm29
+// CHECK:  encoding: [0x62,0x61,0xfd,0x18,0x51,0xaa,0x00,0x04,0x00,0x00]
+          vsqrtpd 1024(%rdx){1to2}, %xmm29
+
+// CHECK: vsqrtpd -1024(%rdx){1to2}, %xmm29
+// CHECK:  encoding: [0x62,0x61,0xfd,0x18,0x51,0x6a,0x80]
+          vsqrtpd -1024(%rdx){1to2}, %xmm29
+
+// CHECK: vsqrtpd -1032(%rdx){1to2}, %xmm29
+// CHECK:  encoding: [0x62,0x61,0xfd,0x18,0x51,0xaa,0xf8,0xfb,0xff,0xff]
+          vsqrtpd -1032(%rdx){1to2}, %xmm29
+
+// CHECK: vsqrtpd %ymm20, %ymm18
+// CHECK:  encoding: [0x62,0xa1,0xfd,0x28,0x51,0xd4]
+          vsqrtpd %ymm20, %ymm18
+
+// CHECK: vsqrtpd %ymm20, %ymm18 {%k3}
+// CHECK:  encoding: [0x62,0xa1,0xfd,0x2b,0x51,0xd4]
+          vsqrtpd %ymm20, %ymm18 {%k3}
+
+// CHECK: vsqrtpd %ymm20, %ymm18 {%k3} {z}
+// CHECK:  encoding: [0x62,0xa1,0xfd,0xab,0x51,0xd4]
+          vsqrtpd %ymm20, %ymm18 {%k3} {z}
+
+// CHECK: vsqrtpd (%rcx), %ymm18
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x28,0x51,0x11]
+          vsqrtpd (%rcx), %ymm18
+
+// CHECK: vsqrtpd 291(%rax,%r14,8), %ymm18
+// CHECK:  encoding: [0x62,0xa1,0xfd,0x28,0x51,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vsqrtpd 291(%rax,%r14,8), %ymm18
+
+// CHECK: vsqrtpd (%rcx){1to4}, %ymm18
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x38,0x51,0x11]
+          vsqrtpd (%rcx){1to4}, %ymm18
+
+// CHECK: vsqrtpd 4064(%rdx), %ymm18
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x28,0x51,0x52,0x7f]
+          vsqrtpd 4064(%rdx), %ymm18
+
+// CHECK: vsqrtpd 4096(%rdx), %ymm18
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x28,0x51,0x92,0x00,0x10,0x00,0x00]
+          vsqrtpd 4096(%rdx), %ymm18
+
+// CHECK: vsqrtpd -4096(%rdx), %ymm18
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x28,0x51,0x52,0x80]
+          vsqrtpd -4096(%rdx), %ymm18
+
+// CHECK: vsqrtpd -4128(%rdx), %ymm18
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x28,0x51,0x92,0xe0,0xef,0xff,0xff]
+          vsqrtpd -4128(%rdx), %ymm18
+
+// CHECK: vsqrtpd 1016(%rdx){1to4}, %ymm18
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x38,0x51,0x52,0x7f]
+          vsqrtpd 1016(%rdx){1to4}, %ymm18
+
+// CHECK: vsqrtpd 1024(%rdx){1to4}, %ymm18
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x38,0x51,0x92,0x00,0x04,0x00,0x00]
+          vsqrtpd 1024(%rdx){1to4}, %ymm18
+
+// CHECK: vsqrtpd -1024(%rdx){1to4}, %ymm18
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x38,0x51,0x52,0x80]
+          vsqrtpd -1024(%rdx){1to4}, %ymm18
+
+// CHECK: vsqrtpd -1032(%rdx){1to4}, %ymm18
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x38,0x51,0x92,0xf8,0xfb,0xff,0xff]
+          vsqrtpd -1032(%rdx){1to4}, %ymm18
+
+// CHECK: vsqrtps %xmm28, %xmm19
+// CHECK:  encoding: [0x62,0x81,0x7c,0x08,0x51,0xdc]
+          vsqrtps %xmm28, %xmm19
+
+// CHECK: vsqrtps %xmm28, %xmm19 {%k7}
+// CHECK:  encoding: [0x62,0x81,0x7c,0x0f,0x51,0xdc]
+          vsqrtps %xmm28, %xmm19 {%k7}
+
+// CHECK: vsqrtps %xmm28, %xmm19 {%k7} {z}
+// CHECK:  encoding: [0x62,0x81,0x7c,0x8f,0x51,0xdc]
+          vsqrtps %xmm28, %xmm19 {%k7} {z}
+
+// CHECK: vsqrtps (%rcx), %xmm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x08,0x51,0x19]
+          vsqrtps (%rcx), %xmm19
+
+// CHECK: vsqrtps 291(%rax,%r14,8), %xmm19
+// CHECK:  encoding: [0x62,0xa1,0x7c,0x08,0x51,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vsqrtps 291(%rax,%r14,8), %xmm19
+
+// CHECK: vsqrtps (%rcx){1to4}, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x18,0x51,0x19]
+          vsqrtps (%rcx){1to4}, %xmm19
+
+// CHECK: vsqrtps 2032(%rdx), %xmm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x08,0x51,0x5a,0x7f]
+          vsqrtps 2032(%rdx), %xmm19
+
+// CHECK: vsqrtps 2048(%rdx), %xmm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x08,0x51,0x9a,0x00,0x08,0x00,0x00]
+          vsqrtps 2048(%rdx), %xmm19
+
+// CHECK: vsqrtps -2048(%rdx), %xmm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x08,0x51,0x5a,0x80]
+          vsqrtps -2048(%rdx), %xmm19
+
+// CHECK: vsqrtps -2064(%rdx), %xmm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x08,0x51,0x9a,0xf0,0xf7,0xff,0xff]
+          vsqrtps -2064(%rdx), %xmm19
+
+// CHECK: vsqrtps 508(%rdx){1to4}, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x18,0x51,0x5a,0x7f]
+          vsqrtps 508(%rdx){1to4}, %xmm19
+
+// CHECK: vsqrtps 512(%rdx){1to4}, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x18,0x51,0x9a,0x00,0x02,0x00,0x00]
+          vsqrtps 512(%rdx){1to4}, %xmm19
+
+// CHECK: vsqrtps -512(%rdx){1to4}, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x18,0x51,0x5a,0x80]
+          vsqrtps -512(%rdx){1to4}, %xmm19
+
+// CHECK: vsqrtps -516(%rdx){1to4}, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x18,0x51,0x9a,0xfc,0xfd,0xff,0xff]
+          vsqrtps -516(%rdx){1to4}, %xmm19
+
+// CHECK: vsqrtps %ymm25, %ymm19
+// CHECK:  encoding: [0x62,0x81,0x7c,0x28,0x51,0xd9]
+          vsqrtps %ymm25, %ymm19
+
+// CHECK: vsqrtps %ymm25, %ymm19 {%k2}
+// CHECK:  encoding: [0x62,0x81,0x7c,0x2a,0x51,0xd9]
+          vsqrtps %ymm25, %ymm19 {%k2}
+
+// CHECK: vsqrtps %ymm25, %ymm19 {%k2} {z}
+// CHECK:  encoding: [0x62,0x81,0x7c,0xaa,0x51,0xd9]
+          vsqrtps %ymm25, %ymm19 {%k2} {z}
+
+// CHECK: vsqrtps (%rcx), %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x28,0x51,0x19]
+          vsqrtps (%rcx), %ymm19
+
+// CHECK: vsqrtps 291(%rax,%r14,8), %ymm19
+// CHECK:  encoding: [0x62,0xa1,0x7c,0x28,0x51,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vsqrtps 291(%rax,%r14,8), %ymm19
+
+// CHECK: vsqrtps (%rcx){1to8}, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x38,0x51,0x19]
+          vsqrtps (%rcx){1to8}, %ymm19
+
+// CHECK: vsqrtps 4064(%rdx), %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x28,0x51,0x5a,0x7f]
+          vsqrtps 4064(%rdx), %ymm19
+
+// CHECK: vsqrtps 4096(%rdx), %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x28,0x51,0x9a,0x00,0x10,0x00,0x00]
+          vsqrtps 4096(%rdx), %ymm19
+
+// CHECK: vsqrtps -4096(%rdx), %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x28,0x51,0x5a,0x80]
+          vsqrtps -4096(%rdx), %ymm19
+
+// CHECK: vsqrtps -4128(%rdx), %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x28,0x51,0x9a,0xe0,0xef,0xff,0xff]
+          vsqrtps -4128(%rdx), %ymm19
+
+// CHECK: vsqrtps 508(%rdx){1to8}, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x38,0x51,0x5a,0x7f]
+          vsqrtps 508(%rdx){1to8}, %ymm19
+
+// CHECK: vsqrtps 512(%rdx){1to8}, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x38,0x51,0x9a,0x00,0x02,0x00,0x00]
+          vsqrtps 512(%rdx){1to8}, %ymm19
+
+// CHECK: vsqrtps -512(%rdx){1to8}, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x38,0x51,0x5a,0x80]
+          vsqrtps -512(%rdx){1to8}, %ymm19
+
+// CHECK: vsqrtps -516(%rdx){1to8}, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x38,0x51,0x9a,0xfc,0xfd,0xff,0xff]
+          vsqrtps -516(%rdx){1to8}, %ymm19
+

 // CHECK: vmovapd %xmm22, (%rcx)
 // CHECK:  encoding: [0x62,0xe1,0xfd,0x08,0x29,0x31]
           vmovapd %xmm22, (%rcx)
 // CHECK: vmovapd %xmm22, (%rcx)
 // CHECK:  encoding: [0x62,0xe1,0xfd,0x08,0x29,0x31]
           vmovapd %xmm22, (%rcx)

diff --git a/utils/TableGen/X86DisassemblerTables.cpp b/utils/TableGen/X86DisassemblerTables.cpp
index dc1764cdbb19cd2ff04533d9d13025e99f68d714..d7e981c8fdfea43b4fdfabb8e53acb3d2e8024fe 100644 (file)
--- a/utils/TableGen/X86DisassemblerTables.cpp
+++ b/utils/TableGen/X86DisassemblerTables.cpp
@@ -171,12 +171,17 @@ static inline bool inheritsFrom(InstructionContext child,
   case IC_EVEX_OPSIZE:
     return inheritsFrom(child, IC_EVEX_W_OPSIZE) ||
            inheritsFrom(child, IC_EVEX_L_W_OPSIZE);
   case IC_EVEX_OPSIZE:
     return inheritsFrom(child, IC_EVEX_W_OPSIZE) ||
            inheritsFrom(child, IC_EVEX_L_W_OPSIZE);

+  case IC_EVEX_B:
+    return false;

   case IC_EVEX_W:
   case IC_EVEX_W_XS:
   case IC_EVEX_W_XD:
   case IC_EVEX_W_OPSIZE:
     return false;
   case IC_EVEX_L:
   case IC_EVEX_W:
   case IC_EVEX_W_XS:
   case IC_EVEX_W_XD:
   case IC_EVEX_W_OPSIZE:
     return false;
   case IC_EVEX_L:

+  case IC_EVEX_L_K_B:
+  case IC_EVEX_L_KZ_B:
+  case IC_EVEX_L_B:

   case IC_EVEX_L_XS:
   case IC_EVEX_L_XD:
   case IC_EVEX_L_OPSIZE:
   case IC_EVEX_L_XS:
   case IC_EVEX_L_XD:
   case IC_EVEX_L_OPSIZE: