[X86] Custom legalize vXi1 extract_subvector with KSHIFTR.

This allows us to remove some isel patterns.

This is mostly NFC, but we now use KSHIFTB instead of KSHIFTW with DQI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@321576 91177308-0d34-0410-b5e6-96231b3b80d8

5 files changed, 57 insertions, 60 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 8bc25c9f080..833305ae0bd 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -1186,9 +1186,8 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
     setOperationAction(ISD::CONCAT_VECTORS,     MVT::v16i1, Custom);
     setOperationAction(ISD::INSERT_SUBVECTOR,   MVT::v8i1,  Custom);
     setOperationAction(ISD::INSERT_SUBVECTOR,   MVT::v16i1, Custom);
-    for (auto VT : { MVT::v1i1, MVT::v2i1, MVT::v4i1, MVT::v8i1,
-                     MVT::v16i1, MVT::v32i1, MVT::v64i1 })
-      setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
+    for (auto VT : { MVT::v1i1, MVT::v8i1 })
+      setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
 
     for (MVT VT : MVT::fp_vector_valuetypes())
       setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
@@ -1428,6 +1427,8 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
     setOperationAction(ISD::CONCAT_VECTORS,     MVT::v64i1, Custom);
     setOperationAction(ISD::INSERT_SUBVECTOR,   MVT::v32i1, Custom);
     setOperationAction(ISD::INSERT_SUBVECTOR,   MVT::v64i1, Custom);
+    for (auto VT : { MVT::v16i1, MVT::v32i1 })
+      setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
 
     // Extends from v32i1 masks to 256-bit vectors.
     setOperationAction(ISD::SIGN_EXTEND,        MVT::v32i8, Custom);
@@ -1540,6 +1541,8 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
     setOperationAction(ISD::CONCAT_VECTORS,     MVT::v8i1, Custom);
     setOperationAction(ISD::CONCAT_VECTORS,     MVT::v4i1, Custom);
     setOperationAction(ISD::INSERT_SUBVECTOR,   MVT::v4i1, Custom);
+    for (auto VT : { MVT::v2i1, MVT::v4i1 })
+      setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
 
     // Extends from v2i1/v4i1 masks to 128-bit vectors.
     setOperationAction(ISD::ZERO_EXTEND,      MVT::v4i32, Custom);
@@ -15070,6 +15073,42 @@ static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget &Subtarget,
   return insert1BitVector(Op, DAG, Subtarget);
 }
 
+static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget &Subtarget,
+                                      SelectionDAG &DAG) {
+  assert(Op.getSimpleValueType().getVectorElementType() == MVT::i1 &&
+         "Only vXi1 extract_subvectors need custom lowering");
+
+  SDLoc dl(Op);
+  SDValue Vec = Op.getOperand(0);
+  SDValue Idx = Op.getOperand(1);
+
+  if (!isa<ConstantSDNode>(Idx))
+    return SDValue();
+
+  unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
+  if (IdxVal == 0) // the operation is legal
+    return Op;
+
+  MVT VecVT = Vec.getSimpleValueType();
+  unsigned NumElems = VecVT.getVectorNumElements();
+
+  // Extend to natively supported kshift.
+  MVT WideVecVT = VecVT;
+  if ((!Subtarget.hasDQI() && NumElems == 8) || NumElems < 8) {
+    WideVecVT = Subtarget.hasDQI() ? MVT::v8i1 : MVT::v16i1;
+    Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideVecVT,
+                      DAG.getUNDEF(WideVecVT), Vec,
+                      DAG.getIntPtrConstant(0, dl));
+  }
+
+  // Shift to the LSB.
+  Vec = DAG.getNode(X86ISD::KSHIFTR, dl, WideVecVT, Vec,
+                    DAG.getConstant(IdxVal, dl, MVT::i8));
+
+  return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, Op.getValueType(), Vec,
+                     DAG.getIntPtrConstant(0, dl));
+}
+
 // Returns the appropriate wrapper opcode for a global reference.
 unsigned X86TargetLowering::getGlobalWrapperKind(const GlobalValue *GV) const {
   // References to absolute symbols are never PC-relative.
@@ -24595,6 +24634,7 @@ SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
   case ISD::INSERT_VECTOR_ELT:  return LowerINSERT_VECTOR_ELT(Op, DAG);
   case ISD::INSERT_SUBVECTOR:   return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
+  case ISD::EXTRACT_SUBVECTOR:  return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
   case ISD::SCALAR_TO_VECTOR:   return LowerSCALAR_TO_VECTOR(Op, Subtarget,DAG);
   case ISD::ConstantPool:       return LowerConstantPool(Op, DAG);
   case ISD::GlobalAddress:      return LowerGlobalAddress(Op, DAG);
diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index dcd84930741..5b7b9baf003 100644
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -3087,49 +3087,6 @@ defm : operation_subvector_mask_lowering<VK16, v16i1, VK64, v64i1>;
 
 defm : operation_subvector_mask_lowering<VK32, v32i1, VK64, v64i1>;
 
-
-multiclass vextract_for_mask_to_mask<string InstrStr, X86KVectorVTInfo From,
-                                     X86KVectorVTInfo To, Predicate prd> {
-let Predicates = [prd] in
-  def :
-    Pat<(To.KVT(extract_subvector(From.KVT From.KRC:$src), (iPTR imm:$imm8))),
-        (To.KVT(COPY_TO_REGCLASS
-                  (!cast<Instruction>(InstrStr#"ri") From.KVT:$src,
-                      (i8 imm:$imm8)), To.KRC))>;
-}
-
-multiclass vextract_for_mask_to_mask_legal_w<X86KVectorVTInfo From,
-                                             X86KVectorVTInfo To> {
-def :
-  Pat<(To.KVT(extract_subvector(From.KVT From.KRC:$src), (iPTR imm:$imm8))),
-      (To.KVT(COPY_TO_REGCLASS
-               (KSHIFTRWri(COPY_TO_REGCLASS From.KRC:$src, VK16),
-                   (i8 imm:$imm8)), To.KRC))>;
-}
-
-defm : vextract_for_mask_to_mask_legal_w<v2i1_info, v1i1_info>;
-defm : vextract_for_mask_to_mask_legal_w<v4i1_info, v1i1_info>;
-defm : vextract_for_mask_to_mask_legal_w<v8i1_info, v1i1_info>;
-defm : vextract_for_mask_to_mask_legal_w<v4i1_info, v2i1_info>;
-defm : vextract_for_mask_to_mask_legal_w<v8i1_info, v2i1_info>;
-defm : vextract_for_mask_to_mask_legal_w<v8i1_info, v4i1_info>;
-
-defm : vextract_for_mask_to_mask<"KSHIFTRW", v16i1_info, v1i1_info, HasAVX512>;
-defm : vextract_for_mask_to_mask<"KSHIFTRD", v32i1_info, v1i1_info, HasBWI>;
-defm : vextract_for_mask_to_mask<"KSHIFTRQ", v64i1_info, v1i1_info, HasBWI>;
-defm : vextract_for_mask_to_mask<"KSHIFTRW", v16i1_info, v2i1_info, HasAVX512>;
-defm : vextract_for_mask_to_mask<"KSHIFTRD", v32i1_info, v2i1_info, HasBWI>;
-defm : vextract_for_mask_to_mask<"KSHIFTRQ", v64i1_info, v2i1_info, HasBWI>;
-defm : vextract_for_mask_to_mask<"KSHIFTRW", v16i1_info, v4i1_info, HasAVX512>;
-defm : vextract_for_mask_to_mask<"KSHIFTRD", v32i1_info, v4i1_info, HasBWI>;
-defm : vextract_for_mask_to_mask<"KSHIFTRQ", v64i1_info, v4i1_info, HasBWI>;
-defm : vextract_for_mask_to_mask<"KSHIFTRW", v16i1_info, v8i1_info, HasAVX512>;
-defm : vextract_for_mask_to_mask<"KSHIFTRD", v32i1_info, v8i1_info, HasBWI>;
-defm : vextract_for_mask_to_mask<"KSHIFTRQ", v64i1_info, v8i1_info, HasBWI>;
-defm : vextract_for_mask_to_mask<"KSHIFTRD", v32i1_info, v16i1_info, HasBWI>;
-defm : vextract_for_mask_to_mask<"KSHIFTRQ", v64i1_info, v16i1_info, HasBWI>;
-defm : vextract_for_mask_to_mask<"KSHIFTRQ", v64i1_info, v32i1_info, HasBWI>;
-
 // Patterns for kmask shift
 multiclass mask_shift_lowering<RegisterClass RC, ValueType VT> {
   def : Pat<(VT (X86kshiftl RC:$src, (i8 imm:$imm))),
diff --git a/test/CodeGen/X86/avx512-extract-subvector-load-store.ll b/test/CodeGen/X86/avx512-extract-subvector-load-store.ll
index e1ed8ea98a1..1f2e77dbc62 100644
--- a/test/CodeGen/X86/avx512-extract-subvector-load-store.ll
+++ b/test/CodeGen/X86/avx512-extract-subvector-load-store.ll
@@ -6,7 +6,7 @@ define void @load_v8i1_broadcast_4_v2i1(<8 x i1>* %a0,<2 x double> %a1,<2 x doub
 ; AVX512-LABEL: load_v8i1_broadcast_4_v2i1:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    kmovb (%rdi), %k0
-; AVX512-NEXT:    kshiftrw $4, %k0, %k0
+; AVX512-NEXT:    kshiftrb $4, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm2
 ; AVX512-NEXT:    vpbroadcastq %xmm2, %xmm2
 ; AVX512-NEXT:    vpmovq2m %xmm2, %k1
@@ -37,7 +37,7 @@ define void @load_v8i1_broadcast_7_v2i1(<8 x i1>* %a0,<2 x double> %a1,<2 x doub
 ; AVX512-LABEL: load_v8i1_broadcast_7_v2i1:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    kmovb (%rdi), %k0
-; AVX512-NEXT:    kshiftrw $6, %k0, %k0
+; AVX512-NEXT:    kshiftrb $6, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm2
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
 ; AVX512-NEXT:    vpmovq2m %xmm2, %k1
@@ -624,7 +624,7 @@ define void @load_v2i1_broadcast_1_v1i1_store(<2 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v2i1_broadcast_1_v1i1_store:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    kmovb (%rdi), %k0
-; AVX512-NEXT:    kshiftrw $1, %k0, %k0
+; AVX512-NEXT:    kshiftrb $1, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
@@ -645,7 +645,7 @@ define void @load_v3i1_broadcast_1_v1i1_store(<3 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v3i1_broadcast_1_v1i1_store:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    kmovb (%rdi), %k0
-; AVX512-NEXT:    kshiftrw $1, %k0, %k0
+; AVX512-NEXT:    kshiftrb $1, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
@@ -666,7 +666,7 @@ define void @load_v3i1_broadcast_2_v1i1_store(<3 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v3i1_broadcast_2_v1i1_store:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    kmovb (%rdi), %k0
-; AVX512-NEXT:    kshiftrw $2, %k0, %k0
+; AVX512-NEXT:    kshiftrb $2, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
@@ -687,7 +687,7 @@ define void @load_v4i1_broadcast_2_v1i1_store(<4 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v4i1_broadcast_2_v1i1_store:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    kmovb (%rdi), %k0
-; AVX512-NEXT:    kshiftrw $2, %k0, %k0
+; AVX512-NEXT:    kshiftrb $2, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
@@ -708,7 +708,7 @@ define void @load_v4i1_broadcast_3_v1i1_store(<4 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v4i1_broadcast_3_v1i1_store:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    kmovb (%rdi), %k0
-; AVX512-NEXT:    kshiftrw $3, %k0, %k0
+; AVX512-NEXT:    kshiftrb $3, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
@@ -729,7 +729,7 @@ define void @load_v8i1_broadcast_4_v1i1_store(<8 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v8i1_broadcast_4_v1i1_store:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    kmovb (%rdi), %k0
-; AVX512-NEXT:    kshiftrw $4, %k0, %k0
+; AVX512-NEXT:    kshiftrb $4, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
@@ -750,7 +750,7 @@ define void @load_v8i1_broadcast_4_v2i1_store(<8 x i1>* %a0,<2 x i1>* %a1) {
 ; AVX512-LABEL: load_v8i1_broadcast_4_v2i1_store:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    kmovb (%rdi), %k0
-; AVX512-NEXT:    kshiftrw $4, %k0, %k0
+; AVX512-NEXT:    kshiftrb $4, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm0
 ; AVX512-NEXT:    vpbroadcastq %xmm0, %xmm0
 ; AVX512-NEXT:    vpmovq2m %xmm0, %k0
@@ -779,7 +779,7 @@ define void @load_v8i1_broadcast_7_v1i1_store(<8 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v8i1_broadcast_7_v1i1_store:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    kmovb (%rdi), %k0
-; AVX512-NEXT:    kshiftrw $7, %k0, %k0
+; AVX512-NEXT:    kshiftrb $7, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
@@ -800,7 +800,7 @@ define void @load_v8i1_broadcast_7_v2i1_store(<8 x i1>* %a0,<2 x i1>* %a1) {
 ; AVX512-LABEL: load_v8i1_broadcast_7_v2i1_store:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    kmovb (%rdi), %k0
-; AVX512-NEXT:    kshiftrw $6, %k0, %k0
+; AVX512-NEXT:    kshiftrb $6, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm0
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
 ; AVX512-NEXT:    vpmovq2m %xmm0, %k0
diff --git a/test/CodeGen/X86/avx512-skx-insert-subvec.ll b/test/CodeGen/X86/avx512-skx-insert-subvec.ll
index 6bee0de181a..f6cb093d521 100644
--- a/test/CodeGen/X86/avx512-skx-insert-subvec.ll
+++ b/test/CodeGen/X86/avx512-skx-insert-subvec.ll
@@ -136,7 +136,7 @@ define <4 x i1> @test9(<8 x i1> %a, <8 x i1> %b) {
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vpsllw $15, %xmm0, %xmm0
 ; CHECK-NEXT:    vpmovw2m %xmm0, %k0
-; CHECK-NEXT:    kshiftrw $4, %k0, %k0
+; CHECK-NEXT:    kshiftrb $4, %k0, %k0
 ; CHECK-NEXT:    vpmovm2d %k0, %xmm0
 ; CHECK-NEXT:    retq
   %res = shufflevector <8 x i1> %a, <8 x i1> %b, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
@@ -148,7 +148,7 @@ define <2 x i1> @test10(<4 x i1> %a, <4 x i1> %b) {
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vpslld $31, %xmm0, %xmm0
 ; CHECK-NEXT:    vptestmd %xmm0, %xmm0, %k0
-; CHECK-NEXT:    kshiftrw $2, %k0, %k0
+; CHECK-NEXT:    kshiftrb $2, %k0, %k0
 ; CHECK-NEXT:    vpmovm2q %k0, %xmm0
 ; CHECK-NEXT:    retq
   %res = shufflevector <4 x i1> %a, <4 x i1> %b, <2 x i32> <i32 2, i32 3>
diff --git a/test/CodeGen/X86/pr33349.ll b/test/CodeGen/X86/pr33349.ll
index b1428ba6667..8f9c861d9ec 100644
--- a/test/CodeGen/X86/pr33349.ll
+++ b/test/CodeGen/X86/pr33349.ll
@@ -40,7 +40,7 @@ target triple = "x86_64-unknown-linux-gnu"
 ; SKX:       # %bb.0: # %bb
 ; SKX-NEXT:    vpslld $31, %xmm0, %xmm0
 ; SKX-NEXT:    vptestmd %xmm0, %xmm0, %k0
-; SKX-NEXT:    kshiftrw $2, %k0, %k1
+; SKX-NEXT:    kshiftrb $2, %k0, %k1
 ; SKX-NEXT:    kshiftrw $1, %k1, %k2
 ; SKX-NEXT:    kmovd %k2, %eax
 ; SKX-NEXT:    testb $1, %al