8190869: C2: missing strength reduction of Math.pow(x, 2.0D) to x*x

Reviewed-by: kvn

src/hotspot/cpu/x86/macroAssembler_x86_pow.cpp

@@ -765,6 +765,11 @@ ALIGNED_(8) juint _log2_pow[] =
     0xfefa39efUL, 0x3fe62e42UL, 0xfefa39efUL, 0xbfe62e42UL
 };
 
+ALIGNED_(8) juint _DOUBLE2[] =
+{
+    0x00000000UL, 0x40000000UL
+};
+
 //registers,
 // input: xmm0, xmm1
 // scratch: xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7
@@ -803,13 +808,21 @@ void MacroAssembler::fast_pow(XMMRegister xmm0, XMMRegister xmm1, XMMRegister xm
   address HIGHMASK_LOG_X = (address)_HIGHMASK_LOG_X;
   address HALFMASK = (address)_HALFMASK;
   address log2 = (address)_log2_pow;
+  address DOUBLE2 = (address)_DOUBLE2;
 
 
   bind(start);
   subq(rsp, 40);
   movsd(Address(rsp, 8), xmm0);
   movsd(Address(rsp, 16), xmm1);
 
+  // Special case: pow(x, 2.0) => x * x
+  movdq(tmp1, xmm1);
+  cmp64(tmp1, ExternalAddress(DOUBLE2));
+  jccb(Assembler::notEqual, B1_2);
+  mulsd(xmm0, xmm0);
+  jmp(B1_5);
+
   bind(B1_2);
   pextrw(eax, xmm0, 3);
   xorpd(xmm2, xmm2);

src/hotspot/share/opto/library_call.cpp

@@ -1789,10 +1789,19 @@ bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
     return StubRoutines::dexp() != NULL ?
       runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(),  "dexp") :
       runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp),  "EXP");
-  case vmIntrinsics::_dpow:
-    return StubRoutines::dpow() != NULL ?
-      runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
+  case vmIntrinsics::_dpow: {
+    Node* exp = round_double_node(argument(2));
+    const TypeD* d = _gvn.type(exp)->isa_double_constant();
+    if (d != NULL && d->getd() == 2.0) {
+      // Special case: pow(x, 2.0) => x * x
+      Node* base = round_double_node(argument(0));
+      set_result(_gvn.transform(new MulDNode(base, base)));
+      return true;
+    }
+    return StubRoutines::dexp() != NULL ?
+      runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(),  "dpow") :
       runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow),  "POW");
+  }
 #undef FN_PTR
 
    // These intrinsics are not yet correctly implemented