[AVX512] Add DQ subvector inserts

In AVX512f we support 64x2 and 32x8 inserts via matching them to 32x4 and 64x4
respectively.  These are matched by "Alt" Pat<>'s (Alt stands for alternative
VTs).

Since DQ has native support for these intructions, I peeled off the non-"Alt"
part of the baseclass into vinsert_for_size_no_alt. The DQ instructions are
derived from this multiclass.  The "Alt" Pat<>'s are disabled with DQ.

Fixes <rdar://problem/18426089>

llvm-svn: 219874

diff --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td
index 788fcab..f20c882 100644 (file)
--- a/llvm/lib/Target/X86/X86InstrAVX512.td
+++ b/llvm/lib/Target/X86/X86InstrAVX512.td
@@ -347,11 +347,10 @@ def : Pat<(v8f64 immAllZerosV), (AVX512_512_SET0)>;
 // AVX-512 - VECTOR INSERT
 //
 
-multiclass vinsert_for_size<int Opcode,
-                             X86VectorVTInfo From, X86VectorVTInfo To,
-                             X86VectorVTInfo AltFrom, X86VectorVTInfo AltTo,
-                             PatFrag vinsert_insert,
-                             SDNodeXForm INSERT_get_vinsert_imm> {
+multiclass vinsert_for_size_no_alt<int Opcode,
+                                   X86VectorVTInfo From, X86VectorVTInfo To,
+                                   PatFrag vinsert_insert,
+                                   SDNodeXForm INSERT_get_vinsert_imm> {
   let hasSideEffects = 0, ExeDomain = To.ExeDomain in {
     def rr : AVX512AIi8<Opcode, MRMSrcReg, (outs VR512:$dst),
                (ins VR512:$src1, From.RC:$src2, i8imm:$src3),
@@ -372,14 +371,24 @@ multiclass vinsert_for_size<int Opcode,
                []>,
              EVEX_4V, EVEX_V512, EVEX_CD8<From.EltSize, From.CD8TupleForm>;
   }
+}
 
+multiclass vinsert_for_size<int Opcode,
+                            X86VectorVTInfo From, X86VectorVTInfo To,
+                            X86VectorVTInfo AltFrom, X86VectorVTInfo AltTo,
+                            PatFrag vinsert_insert,
+                            SDNodeXForm INSERT_get_vinsert_imm> :
+  vinsert_for_size_no_alt<Opcode, From, To,
+                          vinsert_insert, INSERT_get_vinsert_imm> {
   // Codegen pattern with the alternative types, e.g. v2i64 -> v8i64 for
-  // vinserti32x4
-  def : Pat<(vinsert_insert:$ins
-               (AltTo.VT VR512:$src1), (AltFrom.VT From.RC:$src2), (iPTR imm)),
-            (AltTo.VT (!cast<Instruction>(NAME # From.EltSize # "x4rr")
-                          VR512:$src1, From.RC:$src2,
-                          (INSERT_get_vinsert_imm VR512:$ins)))>;
+  // vinserti32x4.  Only add this if 64x2 and friends are not supported
+  // natively via AVX512DQ.
+  let Predicates = [NoDQI] in
+    def : Pat<(vinsert_insert:$ins
+                 (AltTo.VT VR512:$src1), (AltFrom.VT From.RC:$src2), (iPTR imm)),
+              (AltTo.VT (!cast<Instruction>(NAME # From.EltSize # "x4rr")
+                            VR512:$src1, From.RC:$src2,
+                            (INSERT_get_vinsert_imm VR512:$ins)))>;
 }
 
 multiclass vinsert_for_type<ValueType EltVT32, int Opcode128,
@@ -391,6 +400,12 @@ multiclass vinsert_for_type<ValueType EltVT32, int Opcode128,
                                  X86VectorVTInfo< 8, EltVT64, VR512>,
                                  vinsert128_insert,
                                  INSERT_get_vinsert128_imm>;
+  let Predicates = [HasDQI] in
+    defm NAME # "64x2" : vinsert_for_size_no_alt<Opcode128,
+                                 X86VectorVTInfo< 2, EltVT64, VR128X>,
+                                 X86VectorVTInfo< 8, EltVT64, VR512>,
+                                 vinsert128_insert,
+                                 INSERT_get_vinsert128_imm>, VEX_W;
   defm NAME # "64x4" : vinsert_for_size<Opcode256,
                                  X86VectorVTInfo< 4, EltVT64, VR256X>,
                                  X86VectorVTInfo< 8, EltVT64, VR512>,
@@ -398,6 +413,12 @@ multiclass vinsert_for_type<ValueType EltVT32, int Opcode128,
                                  X86VectorVTInfo<16, EltVT32, VR512>,
                                  vinsert256_insert,
                                  INSERT_get_vinsert256_imm>, VEX_W;
+  let Predicates = [HasDQI] in
+    defm NAME # "32x8" : vinsert_for_size_no_alt<Opcode256,
+                                 X86VectorVTInfo< 8, EltVT32, VR256X>,
+                                 X86VectorVTInfo<16, EltVT32, VR512>,
+                                 vinsert256_insert,
+                                 INSERT_get_vinsert256_imm>;
 }
 
 defm VINSERTF : vinsert_for_type<f32, 0x18, f64, 0x1a>;

diff --git a/llvm/lib/Target/X86/X86InstrInfo.td b/llvm/lib/Target/X86/X86InstrInfo.td
index 4c55f7c..3dbf819 100644 (file)
--- a/llvm/lib/Target/X86/X86InstrInfo.td
+++ b/llvm/lib/Target/X86/X86InstrInfo.td
@@ -710,6 +710,7 @@ def HasCDI       : Predicate<"Subtarget->hasCDI()">;
 def HasPFI       : Predicate<"Subtarget->hasPFI()">;
 def HasERI       : Predicate<"Subtarget->hasERI()">;
 def HasDQI       : Predicate<"Subtarget->hasDQI()">;
+def NoDQI        : Predicate<"!Subtarget->hasDQI()">;
 def HasBWI       : Predicate<"Subtarget->hasBWI()">;
 def HasVLX       : Predicate<"Subtarget->hasVLX()">,
                      AssemblerPredicate<"FeatureVLX", "AVX-512 VLX ISA">;

diff --git a/llvm/test/CodeGen/X86/avx512-insert-extract.ll b/llvm/test/CodeGen/X86/avx512-insert-extract.ll
index fa60476..eba895e 100644 (file)
--- a/llvm/test/CodeGen/X86/avx512-insert-extract.ll
+++ b/llvm/test/CodeGen/X86/avx512-insert-extract.ll
@@ -13,9 +13,11 @@ define <16 x float> @test1(<16 x float> %x, float* %br, float %y) nounwind {
 }
 
 ;CHECK-LABEL: test2:
-;CHECK: vinsertf32x4 $0
+;KNL: vinsertf32x4 $0
+;SKX: vinsertf64x2 $0
 ;CHECK: vextractf32x4 $3
-;CHECK: vinsertf32x4 $3
+;KNL: vinsertf32x4 $3
+;SKX: vinsertf64x2 $3
 ;CHECK: ret
 define <8 x double> @test2(<8 x double> %x, double* %br, double %y) nounwind {
   %rrr = load double* %br
@@ -36,7 +38,8 @@ define <16 x float> @test3(<16 x float> %x) nounwind {
 
 ;CHECK-LABEL: test4:
 ;CHECK: vextracti32x4 $2
-;CHECK: vinserti32x4 $0
+;KNL: vinserti32x4 $0
+;SKX: vinserti64x2 $0
 ;CHECK: ret
 define <8 x i64> @test4(<8 x i64> %x) nounwind {
   %eee = extractelement <8 x i64> %x, i32 4

diff --git a/llvm/test/MC/X86/avx512-encodings.s b/llvm/test/MC/X86/avx512-encodings.s
index 8326a7c..8dc89da 100644 (file)
--- a/llvm/test/MC/X86/avx512-encodings.s
+++ b/llvm/test/MC/X86/avx512-encodings.s
@@ -4205,6 +4205,14 @@ vinserti32x4  $1, %xmm21, %zmm5, %zmm17
 // CHECK: encoding: [0x62,0xe3,0x1d,0x40,0x38,0x4f,0x10,0x01]
 vinserti32x4  $1, 256(%rdi), %zmm28, %zmm17
 
+// CHECK: vinserti32x8
+// CHECK: encoding: [0x62,0xd3,0x4d,0x40,0x3a,0xdb,0x01]
+vinserti32x8  $1, %ymm11, %zmm22, %zmm3
+
+// CHECK: vinsertf64x2
+// CHECK: encoding: [0x62,0xf3,0xed,0x48,0x18,0x4f,0x10,0x01]
+vinsertf64x2  $1, 256(%rdi), %zmm2, %zmm1
+
 // CHECK: vextracti32x4
 // CHECK: encoding: [0x62,0x33,0x7d,0x48,0x39,0xc9,0x01]
 vextracti32x4  $1, %zmm9, %xmm17