8353686: Optimize Math.cbrt for x86 64 bit platforms

Reviewed-by: sviswanathan, sparasa, jbhateja

Author: missa-prime

src/hotspot/cpu/x86/assembler_x86.cpp

@@ -2867,6 +2867,17 @@ void Assembler::mov(Register dst, Register src) {
   movq(dst, src);
 }
 
+void Assembler::movapd(XMMRegister dst, Address src) {
+  NOT_LP64(assert(VM_Version::supports_sse2(), ""));
+  InstructionMark im(this);
+  InstructionAttr attributes(AVX_128bit, /* rex_w */ VM_Version::supports_evex(), /* legacy_mode */ false, /* no_mask_reg */ true, /* uses_vl */ true);
+  attributes.set_address_attributes(/* tuple_type */ EVEX_FVM, /* input_size_in_bits */ EVEX_NObit);
+  attributes.set_rex_vex_w_reverted();
+  simd_prefix(dst, xnoreg, src, VEX_SIMD_66, VEX_OPCODE_0F, &attributes);
+  emit_int8(0x28);
+  emit_operand(dst, src, 0);
+}
+
 void Assembler::movapd(XMMRegister dst, XMMRegister src) {
   int vector_len = VM_Version::supports_avx512novl() ? AVX_512bit : AVX_128bit;
   InstructionAttr attributes(vector_len, /* rex_w */ VM_Version::supports_evex(), /* legacy_mode */ false, /* no_mask_reg */ true, /* uses_vl */ true);
@@ -8072,6 +8083,14 @@ void Assembler::vandps(XMMRegister dst, XMMRegister nds, Address src, int vector
   emit_operand(dst, src, 0);
 }
 
+void Assembler::orpd(XMMRegister dst, XMMRegister src) {
+  NOT_LP64(assert(VM_Version::supports_sse2(), ""));
+  InstructionAttr attributes(AVX_128bit, /* rex_w */ !_legacy_mode_dq, /* legacy_mode */ _legacy_mode_dq, /* no_mask_reg */ true, /* uses_vl */ true);
+  attributes.set_rex_vex_w_reverted();
+  int encode = simd_prefix_and_encode(dst, dst, src, VEX_SIMD_66, VEX_OPCODE_0F, &attributes);
+  emit_int16(0x56, (0xC0 | encode));
+}
+
 void Assembler::unpckhpd(XMMRegister dst, XMMRegister src) {
   InstructionAttr attributes(AVX_128bit, /* rex_w */ VM_Version::supports_evex(), /* legacy_mode */ false, /* no_mask_reg */ true, /* uses_vl */ true);
   attributes.set_rex_vex_w_reverted();

src/hotspot/cpu/x86/assembler_x86.hpp

@@ -950,6 +950,7 @@ class Assembler : public AbstractAssembler  {
   // New cpus require use of movaps and movapd to avoid partial register stall
   // when moving between registers.
   void movaps(XMMRegister dst, XMMRegister src);
+  void movapd(XMMRegister dst, Address src);
   void movapd(XMMRegister dst, XMMRegister src);
 
   // End avoid using directly
@@ -2450,6 +2451,9 @@ class Assembler : public AbstractAssembler  {
   void vandpd(XMMRegister dst, XMMRegister nds, Address src, int vector_len);
   void vandps(XMMRegister dst, XMMRegister nds, Address src, int vector_len);
 
+  // Bitwise Logical OR of Packed Floating-Point Values
+  void orpd(XMMRegister dst, XMMRegister src);
+
   void unpckhpd(XMMRegister dst, XMMRegister src);
   void unpcklpd(XMMRegister dst, XMMRegister src);
 

src/hotspot/cpu/x86/c1_LIRGenerator_x86.cpp

@@ -720,7 +720,8 @@ void LIRGenerator::do_MathIntrinsic(Intrinsic* x) {
   if (x->id() == vmIntrinsics::_dexp || x->id() == vmIntrinsics::_dlog ||
       x->id() == vmIntrinsics::_dpow || x->id() == vmIntrinsics::_dcos ||
       x->id() == vmIntrinsics::_dsin || x->id() == vmIntrinsics::_dtan ||
-      x->id() == vmIntrinsics::_dlog10 || x->id() == vmIntrinsics::_dtanh
+      x->id() == vmIntrinsics::_dlog10 || x->id() == vmIntrinsics::_dtanh ||
+      x->id() == vmIntrinsics::_dcbrt
       ) {
     do_LibmIntrinsic(x);
     return;
@@ -807,7 +808,7 @@ void LIRGenerator::do_LibmIntrinsic(Intrinsic* x) {
       }
       break;
     case vmIntrinsics::_dpow:
-       if (StubRoutines::dpow() != nullptr) {
+      if (StubRoutines::dpow() != nullptr) {
         __ call_runtime_leaf(StubRoutines::dpow(), getThreadTemp(), result_reg, cc->args());
       } else {
         __ call_runtime_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dpow), getThreadTemp(), result_reg, cc->args());
@@ -828,18 +829,24 @@ void LIRGenerator::do_LibmIntrinsic(Intrinsic* x) {
       }
       break;
     case vmIntrinsics::_dtan:
-       if (StubRoutines::dtan() != nullptr) {
+      if (StubRoutines::dtan() != nullptr) {
         __ call_runtime_leaf(StubRoutines::dtan(), getThreadTemp(), result_reg, cc->args());
       } else {
         __ call_runtime_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtan), getThreadTemp(), result_reg, cc->args());
       }
       break;
     case vmIntrinsics::_dtanh:
-       assert(StubRoutines::dtanh() != nullptr, "tanh intrinsic not found");
-       if (StubRoutines::dtanh() != nullptr) {
+      assert(StubRoutines::dtanh() != nullptr, "tanh intrinsic not found");
+      if (StubRoutines::dtanh() != nullptr) {
         __ call_runtime_leaf(StubRoutines::dtanh(), getThreadTemp(), result_reg, cc->args());
       }
       break;
+    case vmIntrinsics::_dcbrt:
+      assert(StubRoutines::dcbrt() != nullptr, "cbrt intrinsic not found");
+      if (StubRoutines::dcbrt() != nullptr) {
+        __ call_runtime_leaf(StubRoutines::dcbrt(), getThreadTemp(), result_reg, cc->args());
+      }
+      break;
     default:  ShouldNotReachHere();
   }
 

src/hotspot/cpu/x86/macroAssembler_x86.cpp

@@ -2250,6 +2250,16 @@ void MacroAssembler::evmovdqaq(XMMRegister dst, AddressLiteral src, int vector_l
   }
 }
 
+void MacroAssembler::movapd(XMMRegister dst, AddressLiteral src, Register rscratch) {
+  assert(rscratch != noreg || always_reachable(src), "missing");
+
+  if (reachable(src)) {
+    Assembler::movapd(dst, as_Address(src));
+  } else {
+    lea(rscratch, src);
+    Assembler::movapd(dst, Address(rscratch, 0));
+  }
+}
 
 void MacroAssembler::movdqa(XMMRegister dst, AddressLiteral src, Register rscratch) {
   assert(rscratch != noreg || always_reachable(src), "missing");

src/hotspot/cpu/x86/macroAssembler_x86.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -995,6 +995,8 @@ class MacroAssembler: public Assembler {
   void andpd(XMMRegister dst, Address        src) { Assembler::andpd(dst, src); }
   void andpd(XMMRegister dst, AddressLiteral src, Register rscratch = noreg);
 
+  void andnpd(XMMRegister dst, XMMRegister src) { Assembler::andnpd(dst, src); }
+
   void andps(XMMRegister dst, XMMRegister    src) { Assembler::andps(dst, src); }
   void andps(XMMRegister dst, Address        src) { Assembler::andps(dst, src); }
   void andps(XMMRegister dst, AddressLiteral src, Register rscratch = noreg);
@@ -1007,6 +1009,8 @@ class MacroAssembler: public Assembler {
   void comisd(XMMRegister dst, Address        src) { Assembler::comisd(dst, src); }
   void comisd(XMMRegister dst, AddressLiteral src, Register rscratch = noreg);
 
+  void orpd(XMMRegister dst, XMMRegister src) { Assembler::orpd(dst, src); }
+
   void cmp32_mxcsr_std(Address mxcsr_save, Register tmp, Register rscratch = noreg);
   void ldmxcsr(Address src) { Assembler::ldmxcsr(src); }
   void ldmxcsr(AddressLiteral src, Register rscratch = noreg);
@@ -1241,6 +1245,9 @@ class MacroAssembler: public Assembler {
   void evmovdquq(XMMRegister dst, KRegister mask, AddressLiteral src, bool merge, int vector_len, Register rscratch = noreg);
   void evmovdqaq(XMMRegister dst, KRegister mask, AddressLiteral src, bool merge, int vector_len, Register rscratch = noreg);
 
+  using Assembler::movapd;
+  void movapd(XMMRegister dst, AddressLiteral src, Register rscratch = noreg);
+
   // Move Aligned Double Quadword
   void movdqa(XMMRegister dst, XMMRegister    src) { Assembler::movdqa(dst, src); }
   void movdqa(XMMRegister dst, Address        src) { Assembler::movdqa(dst, src); }

src/hotspot/cpu/x86/stubGenerator_x86_64.cpp

@@ -3692,6 +3692,9 @@ void StubGenerator::generate_libm_stubs() {
     if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dtanh)) {
       StubRoutines::_dtanh = generate_libmTanh(); // from stubGenerator_x86_64_tanh.cpp
     }
+    if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dcbrt)) {
+      StubRoutines::_dcbrt = generate_libmCbrt(); // from stubGenerator_x86_64_cbrt.cpp
+    }
     if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dexp)) {
       StubRoutines::_dexp = generate_libmExp(); // from stubGenerator_x86_64_exp.cpp
     }

src/hotspot/cpu/x86/stubGenerator_x86_64.hpp

@@ -556,6 +556,7 @@ class StubGenerator: public StubCodeGenerator {
   address generate_libmCos();
   address generate_libmTan();
   address generate_libmTanh();
+  address generate_libmCbrt();
   address generate_libmExp();
   address generate_libmPow();
   address generate_libmLog();