Remove intrinsics for X86 BLSI, BLSMSK, and BLSR intrinsics and replace with custom...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@142642 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/llvm/IntrinsicsX86.td b/include/llvm/IntrinsicsX86.td
index db9c9ce00304be420cfebbb3a9d87d6224441a99..2bad822abeed045fa391bb649c2cbb32f8ab5944 100644 (file)
--- a/include/llvm/IntrinsicsX86.td
+++ b/include/llvm/IntrinsicsX86.td
@@ -1642,24 +1642,12 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
               Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
   def int_x86_bmi_bzhi_64 : GCCBuiltin<"__builtin_ia32_bzhi_di">,
               Intrinsic<[llvm_i64_ty], [llvm_i64_ty, llvm_i64_ty], [IntrNoMem]>;
-  def int_x86_bmi_blsi_32 :
-              Intrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoMem]>;
-  def int_x86_bmi_blsi_64 :
-              Intrinsic<[llvm_i64_ty], [llvm_i64_ty], [IntrNoMem]>;
-  def int_x86_bmi_blsmsk_32 :
-              Intrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoMem]>;
-  def int_x86_bmi_blsmsk_64 :
-              Intrinsic<[llvm_i64_ty], [llvm_i64_ty], [IntrNoMem]>;
-  def int_x86_bmi_blsr_32 :
-              Intrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoMem]>;
-  def int_x86_bmi_blsr_64 :
-              Intrinsic<[llvm_i64_ty], [llvm_i64_ty], [IntrNoMem]>;
-  def int_x86_bmi_pdep_32 :
+  def int_x86_bmi_pdep_32 : GCCBuiltin<"__builtin_ia32_pdep_si">,
               Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
-  def int_x86_bmi_pdep_64 :
+  def int_x86_bmi_pdep_64 : GCCBuiltin<"__builtin_ia32_pdep_di">,
               Intrinsic<[llvm_i64_ty], [llvm_i64_ty, llvm_i64_ty], [IntrNoMem]>;
-  def int_x86_bmi_pext_32 :
+  def int_x86_bmi_pext_32 : GCCBuiltin<"__builtin_ia32_pext_si">,
               Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
-  def int_x86_bmi_pext_64 :
+  def int_x86_bmi_pext_64 : GCCBuiltin<"__builtin_ia32_pext_di">,
               Intrinsic<[llvm_i64_ty], [llvm_i64_ty, llvm_i64_ty], [IntrNoMem]>;
 }

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index f1b160cdfae756248c1c6a1569e8a98e1be57fde..96f04e729810e3331608f22aae422eab9e654c70 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -1155,6 +1155,8 @@ X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
   setTargetDAGCombine(ISD::SINT_TO_FP);
   if (Subtarget->is64Bit())
     setTargetDAGCombine(ISD::MUL);
+  if (Subtarget->hasBMI())
+    setTargetDAGCombine(ISD::XOR);
 
   computeRegisterProperties();
 
@@ -13300,7 +13302,9 @@ static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
 
   EVT VT = N->getValueType(0);
 
-  // Create ANDN instructions
+  // Create ANDN, BLSI, and BLSR instructions
+  // BLSI is X & (-X)
+  // BLSR is X & (X-1)
   if (Subtarget->hasBMI() && (VT == MVT::i32 || VT == MVT::i64)) {
     SDValue N0 = N->getOperand(0);
     SDValue N1 = N->getOperand(1);
@@ -13313,6 +13317,26 @@ static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
     if (N1.getOpcode() == ISD::XOR && isAllOnes(N1.getOperand(1)))
       return DAG.getNode(X86ISD::ANDN, DL, VT, N1.getOperand(0), N0);
 
+    // Check LHS for neg
+    if (N0.getOpcode() == ISD::SUB && N0.getOperand(1) == N1 &&
+        isZero(N0.getOperand(0)))
+      return DAG.getNode(X86ISD::BLSI, DL, VT, N1);
+
+    // Check RHS for neg
+    if (N1.getOpcode() == ISD::SUB && N1.getOperand(1) == N0 &&
+        isZero(N1.getOperand(0)))
+      return DAG.getNode(X86ISD::BLSI, DL, VT, N0);
+
+    // Check LHS for X-1
+    if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1 &&
+        isAllOnes(N0.getOperand(1)))
+      return DAG.getNode(X86ISD::BLSR, DL, VT, N1);
+
+    // Check RHS for X-1
+    if (N1.getOpcode() == ISD::ADD && N1.getOperand(0) == N0 &&
+        isAllOnes(N1.getOperand(1)))
+      return DAG.getNode(X86ISD::BLSR, DL, VT, N0);
+
     return SDValue();
   }
 
@@ -13500,6 +13524,33 @@ static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
+static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
+                                 TargetLowering::DAGCombinerInfo &DCI,
+                                 const X86Subtarget *Subtarget) {
+  if (DCI.isBeforeLegalizeOps())
+    return SDValue();
+
+  EVT VT = N->getValueType(0);
+
+  if (VT != MVT::i32 && VT != MVT::i64)
+    return SDValue();
+
+  // Create BLSMSK instructions by finding X ^ (X-1)
+  SDValue N0 = N->getOperand(0);
+  SDValue N1 = N->getOperand(1);
+  DebugLoc DL = N->getDebugLoc();
+
+  if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1 &&
+      isAllOnes(N0.getOperand(1)))
+    return DAG.getNode(X86ISD::BLSMSK, DL, VT, N1);
+
+  if (N1.getOpcode() == ISD::ADD && N1.getOperand(0) == N0 &&
+      isAllOnes(N1.getOperand(1)))
+    return DAG.getNode(X86ISD::BLSMSK, DL, VT, N0);
+
+  return SDValue();
+}
+
 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
                                    const X86Subtarget *Subtarget) {
@@ -14180,6 +14231,7 @@ SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case ISD::SRL:            return PerformShiftCombine(N, DAG, Subtarget);
   case ISD::AND:            return PerformAndCombine(N, DAG, DCI, Subtarget);
   case ISD::OR:             return PerformOrCombine(N, DAG, DCI, Subtarget);
+  case ISD::XOR:            return PerformXorCombine(N, DAG, DCI, Subtarget);
   case ISD::LOAD:           return PerformLOADCombine(N, DAG, Subtarget);
   case ISD::STORE:          return PerformSTORECombine(N, DAG, Subtarget);
   case ISD::SINT_TO_FP:     return PerformSINT_TO_FPCombine(N, DAG, this);

diff --git a/lib/Target/X86/X86ISelLowering.h b/lib/Target/X86/X86ISelLowering.h
index 342a5e61754510ac46daf8d4bfe41650f743c1c9..0903b9f7ab0de3e1138acd3418a2f75241438ccd 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.h
+++ b/lib/Target/X86/X86ISelLowering.h
@@ -230,6 +230,10 @@ namespace llvm {
 
       ANDN, // ANDN - Bitwise AND NOT with FLAGS results.
 
+      BLSI,   // BLSI - Extract lowest set isolated bit
+      BLSMSK, // BLSMSK - Get mask up to lowest set bit
+      BLSR,   // BLSR - Reset lowest set bit
+
       UMUL, // LOW, HI, FLAGS = umul LHS, RHS
 
       // MUL_IMM - X86 specific multiply by immediate.

diff --git a/lib/Target/X86/X86InstrInfo.td b/lib/Target/X86/X86InstrInfo.td
index 0994ab94767077fc6f60866022bfa1227468e6db..aa35cf0f0097410f55f7b2c85ea4f5c5fdb7a138 100644 (file)
--- a/lib/Target/X86/X86InstrInfo.td
+++ b/lib/Target/X86/X86InstrInfo.td
@@ -226,6 +226,10 @@ def X86and_flag  : SDNode<"X86ISD::AND",  SDTBinaryArithWithFlags,
                           [SDNPCommutative]>;
 def X86andn_flag : SDNode<"X86ISD::ANDN", SDTBinaryArithWithFlags>;
 
+def X86blsi_flag : SDNode<"X86ISD::BLSI",  SDTUnaryArithWithFlags>;
+def X86blsmsk_flag : SDNode<"X86ISD::BLSMSK",  SDTUnaryArithWithFlags>;
+def X86blsr_flag : SDNode<"X86ISD::BLSR",  SDTUnaryArithWithFlags>;
+
 def X86mul_imm : SDNode<"X86ISD::MUL_IMM", SDTIntBinOp>;
 
 def X86WinAlloca : SDNode<"X86ISD::WIN_ALLOCA", SDTX86Void,
@@ -1401,30 +1405,30 @@ let Predicates = [HasBMI], Defs = [EFLAGS] in {
 }
 
 multiclass bmi_bls<string mnemonic, Format RegMRM, Format MemMRM,
-                   RegisterClass RC, X86MemOperand x86memop, Intrinsic Int,
-                   PatFrag ld_frag> {
+                  RegisterClass RC, X86MemOperand x86memop, SDNode OpNode,
+                  PatFrag ld_frag> {
   def rr : I<0xF3, RegMRM, (outs RC:$dst), (ins RC:$src),
              !strconcat(mnemonic, "\t{$src, $dst|$dst, $src}"),
-             [(set RC:$dst, (Int RC:$src)), (implicit EFLAGS)]>, T8, VEX_4V;
+             [(set RC:$dst, EFLAGS, (OpNode RC:$src))]>, T8, VEX_4V;
   def rm : I<0xF3, MemMRM, (outs RC:$dst), (ins x86memop:$src),
              !strconcat(mnemonic, "\t{$src, $dst|$dst, $src}"),
-             [(set RC:$dst, (Int (ld_frag addr:$src))), (implicit EFLAGS)]>,
+             [(set RC:$dst, EFLAGS, (OpNode (ld_frag addr:$src)))]>,
              T8, VEX_4V;
 }
 
 let Predicates = [HasBMI], Defs = [EFLAGS] in {
   defm BLSR32 : bmi_bls<"blsr{l}", MRM1r, MRM1m, GR32, i32mem,
-                        int_x86_bmi_blsr_32, loadi32>;
+                        X86blsr_flag, loadi32>;
   defm BLSR64 : bmi_bls<"blsr{q}", MRM1r, MRM1m, GR64, i64mem,
-                        int_x86_bmi_blsr_64, loadi64>, VEX_W;
+                        X86blsr_flag, loadi64>, VEX_W;
   defm BLSMSK32 : bmi_bls<"blsmsk{l}", MRM2r, MRM2m, GR32, i32mem,
-                          int_x86_bmi_blsmsk_32, loadi32>;
+                          X86blsmsk_flag, loadi32>;
   defm BLSMSK64 : bmi_bls<"blsmsk{q}", MRM2r, MRM2m, GR64, i64mem,
-                          int_x86_bmi_blsmsk_64, loadi64>, VEX_W;
+                          X86blsmsk_flag, loadi64>, VEX_W;
   defm BLSI32 : bmi_bls<"blsi{l}", MRM3r, MRM3m, GR32, i32mem,
-                        int_x86_bmi_blsi_32, loadi32>;
+                        X86blsi_flag, loadi32>;
   defm BLSI64 : bmi_bls<"blsi{q}", MRM3r, MRM3m, GR64, i64mem,
-                        int_x86_bmi_blsi_64, loadi64>, VEX_W;
+                        X86blsi_flag, loadi64>, VEX_W;
 }
 
 multiclass bmi_bextr_bzhi<bits<8> opc, string mnemonic, RegisterClass RC,

diff --git a/test/CodeGen/X86/bmi.ll b/test/CodeGen/X86/bmi.ll
index 4b40d9033804a9e138c465ed5b8c26187c86bfe7..69cf7365c5073a83f6356528ed81d7b7e0eecf4a 100644 (file)
--- a/test/CodeGen/X86/bmi.ll
+++ b/test/CodeGen/X86/bmi.ll
@@ -89,59 +89,53 @@ define i64 @bzhi64(i64 %x, i64 %y) nounwind readnone {
 declare i64 @llvm.x86.bmi.bzhi.64(i64, i64) nounwind readnone
 
 define i32 @blsi32(i32 %x) nounwind readnone {
-  %tmp = tail call i32 @llvm.x86.bmi.blsi.32(i32 %x)
-  ret i32 %tmp
+  %tmp = sub i32 0, %x
+  %tmp2 = and i32 %x, %tmp
+  ret i32 %tmp2
 ; CHECK: blsi32:
 ; CHECK: blsil
 }
 
-declare i32 @llvm.x86.bmi.blsi.32(i32) nounwind readnone
-
 define i64 @blsi64(i64 %x) nounwind readnone {
-  %tmp = tail call i64 @llvm.x86.bmi.blsi.64(i64 %x)
-  ret i64 %tmp
+  %tmp = sub i64 0, %x
+  %tmp2 = and i64 %tmp, %x
+  ret i64 %tmp2
 ; CHECK: blsi64:
 ; CHECK: blsiq
 }
 
-declare i64 @llvm.x86.bmi.blsi.64(i64) nounwind readnone
-
 define i32 @blsmsk32(i32 %x) nounwind readnone {
-  %tmp = tail call i32 @llvm.x86.bmi.blsmsk.32(i32 %x)
-  ret i32 %tmp
+  %tmp = sub i32 %x, 1
+  %tmp2 = xor i32 %x, %tmp
+  ret i32 %tmp2
 ; CHECK: blsmsk32:
 ; CHECK: blsmskl
 }
 
-declare i32 @llvm.x86.bmi.blsmsk.32(i32) nounwind readnone
-
 define i64 @blsmsk64(i64 %x) nounwind readnone {
-  %tmp = tail call i64 @llvm.x86.bmi.blsmsk.64(i64 %x)
-  ret i64 %tmp
+  %tmp = sub i64 %x, 1
+  %tmp2 = xor i64 %tmp, %x
+  ret i64 %tmp2
 ; CHECK: blsmsk64:
 ; CHECK: blsmskq
 }
 
-declare i64 @llvm.x86.bmi.blsmsk.64(i64) nounwind readnone
-
 define i32 @blsr32(i32 %x) nounwind readnone {
-  %tmp = tail call i32 @llvm.x86.bmi.blsr.32(i32 %x)
-  ret i32 %tmp
+  %tmp = sub i32 %x, 1
+  %tmp2 = and i32 %x, %tmp
+  ret i32 %tmp2
 ; CHECK: blsr32:
 ; CHECK: blsrl
 }
 
-declare i32 @llvm.x86.bmi.blsr.32(i32) nounwind readnone
-
 define i64 @blsr64(i64 %x) nounwind readnone {
-  %tmp = tail call i64 @llvm.x86.bmi.blsr.64(i64 %x)
-  ret i64 %tmp
+  %tmp = sub i64 %x, 1
+  %tmp2 = and i64 %tmp, %x
+  ret i64 %tmp2
 ; CHECK: blsr64:
 ; CHECK: blsrq
 }
 
-declare i64 @llvm.x86.bmi.blsr.64(i64) nounwind readnone
-
 define i32 @pdep32(i32 %x, i32 %y) nounwind readnone {
   %tmp = tail call i32 @llvm.x86.bmi.pdep.32(i32 %x, i32 %y)
   ret i32 %tmp