8312425: [vectorapi] AArch64: Optimize vector math operations with SLEEF

make/autoconf/flags-cflags.m4

@@ -920,6 +920,38 @@ AC_DEFUN([FLAGS_SETUP_CFLAGS_CPU_DEP],
         IF_FALSE: [$2FDLIBM_CFLAGS=""])
   fi
   AC_SUBST($2FDLIBM_CFLAGS)
+
+  # Check whether the compiler supports the Arm C Language Extensions (ACLE)
+  # for SVE. Set SVE_CFLAGS to -march=armv8-a+sve if it does.
+  # ACLE and this flag are required to build the aarch64 SVE related functions in
+  # libvectormath.
+  if test "x${OPENJDK_TARGET_CPU}" = "xaarch64"; then
+    if test "x$TOOLCHAIN_TYPE" = xgcc || test "x$TOOLCHAIN_TYPE" = xclang; then
+      AC_LANG_PUSH(C)
+      OLD_CFLAGS="$CFLAGS"
+      CFLAGS="$CFLAGS -march=armv8-a+sve"
+
+      AC_MSG_CHECKING([if Arm SVE ACLE is supported])
+      AC_COMPILE_IFELSE([AC_LANG_PROGRAM([#include <arm_sve.h>],
+          [
+            svint32_t r = svdup_n_s32(1);
+            return 0;
+          ])],
+          [
+            AC_MSG_RESULT([yes])
+            $2SVE_CFLAGS="-march=armv8-a+sve"
+          ],
+          [
+            AC_MSG_RESULT([no])
+            $2SVE_CFLAGS=""
+          ]
+      )
+
+      CFLAGS="$OLD_CFLAGS"
+      AC_LANG_POP(C)
+    fi
+  fi
+  AC_SUBST($2SVE_CFLAGS)
 ])
 
 # FLAGS_SETUP_GCC6_COMPILER_FLAGS([PREFIX])

make/autoconf/spec.gmk.template

@@ -839,6 +839,9 @@ OS_VERSION_MAJOR := @OS_VERSION_MAJOR@
 OS_VERSION_MINOR := @OS_VERSION_MINOR@
 OS_VERSION_MICRO := @OS_VERSION_MICRO@
 
+# Arm SVE
+SVE_CFLAGS := @SVE_CFLAGS@
+
 # Images directory definitions
 JDK_IMAGE_SUBDIR := jdk
 JRE_IMAGE_SUBDIR := jre

make/modules/jdk.incubator.vector/Lib.gmk

@@ -37,3 +37,16 @@ ifeq ($(call isTargetOs, linux windows)+$(call isTargetCpu, x86_64)+$(INCLUDE_CO
 
   TARGETS += $(BUILD_LIBJSVML)
 endif
+
+ifeq ($(call isTargetOs, linux)+$(call isTargetCpu, aarch64)+$(INCLUDE_COMPILER2), true+true+true)
+  $(eval $(call SetupJdkLibrary, BUILD_LIBVECTORMATH, \
+      NAME := vectormath, \
+      CFLAGS := $(CFLAGS_JDKLIB) -Wno-error=unused-function, \
+      vector_math_sve.c_CFLAGS := $(SVE_CFLAGS), \
+      LDFLAGS := $(LDFLAGS_JDKLIB) \
+          $(call SET_SHARED_LIBRARY_ORIGIN), \
+      LIBS := $(JDKLIB_LIBS) \
+  ))
+
+  TARGETS += $(BUILD_LIBVECTORMATH)
+endif

src/hotspot/cpu/aarch64/aarch64.ad

@@ -2319,14 +2319,18 @@ const TypeVectMask* Matcher::predicate_reg_type(const Type* elemTy, int length)
   return new TypeVectMask(elemTy, length);
 }
 
-// Vector calling convention not yet implemented.
 bool Matcher::supports_vector_calling_convention(void) {
-  return false;
+  return EnableVectorSupport && UseVectorStubs;
 }
 
 OptoRegPair Matcher::vector_return_value(uint ideal_reg) {
-  Unimplemented();
-  return OptoRegPair(0, 0);
+  assert(EnableVectorSupport && UseVectorStubs, "sanity");
+  int lo = V0_num;
+  int hi = V0_H_num;
+  if (ideal_reg == Op_VecX || ideal_reg == Op_VecA) {
+    hi = V0_K_num;
+  }
+  return OptoRegPair(hi, lo);
 }
 
 // Is this branch offset short enough that a short branch can be used?
@@ -16146,6 +16150,22 @@ instruct CallLeafDirect(method meth)
   ins_pipe(pipe_class_call);
 %}
 
+// Call Runtime Instruction without safepoint and with vector arguments
+instruct CallLeafDirectVector(method meth)
+%{
+  match(CallLeafVector);
+
+  effect(USE meth);
+
+  ins_cost(CALL_COST);
+
+  format %{ "CALL, runtime leaf vector $meth" %}
+
+  ins_encode(aarch64_enc_java_to_runtime(meth));
+
+  ins_pipe(pipe_class_call);
+%}
+
 // Call Runtime Instruction
 
 instruct CallLeafNoFPDirect(method meth)

src/hotspot/cpu/aarch64/sharedRuntime_aarch64.cpp

@@ -884,7 +884,20 @@ static int c_calling_convention_priv(const BasicType *sig_bt,
 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
                                              uint num_bits,
                                              uint total_args_passed) {
-  Unimplemented();
+  // More than 8 argument inputs are not supported now.
+  assert(total_args_passed <= Argument::n_float_register_parameters_c, "unsupported");
+  assert(num_bits >= 64 && num_bits <= 2048 && is_power_of_2(num_bits), "unsupported");
+
+  static const FloatRegister VEC_ArgReg[Argument::n_float_register_parameters_c] = {
+    v0, v1, v2, v3, v4, v5, v6, v7
+  };
+
+  // On SVE, we use the same vector registers with 128-bit vector registers on NEON.
+  int next_reg_val = num_bits == 64 ? 1 : 3;
+  for (uint i = 0; i < total_args_passed; i++) {
+    VMReg vmreg = VEC_ArgReg[i]->as_VMReg();
+    regs[i].set_pair(vmreg->next(next_reg_val), vmreg);
+  }
   return 0;
 }
 

src/hotspot/cpu/aarch64/stubGenerator_aarch64.cpp

@@ -42,6 +42,7 @@
 #include "oops/oop.inline.hpp"
 #include "prims/methodHandles.hpp"
 #include "prims/upcallLinker.hpp"
+#include "runtime/arguments.hpp"
 #include "runtime/atomic.hpp"
 #include "runtime/continuation.hpp"
 #include "runtime/continuationEntry.inline.hpp"
@@ -8535,6 +8536,72 @@ class StubGenerator: public StubCodeGenerator {
     if (UseAdler32Intrinsics) {
       StubRoutines::_updateBytesAdler32 = generate_updateBytesAdler32();
     }
+
+#ifdef COMPILER2
+    // Get native vector math stub routine addresses
+    void* libvectormath = nullptr;
+    char ebuf[1024];
+    char dll_name[JVM_MAXPATHLEN];
+    if (os::dll_locate_lib(dll_name, sizeof(dll_name), Arguments::get_dll_dir(), "vectormath")) {
+      libvectormath = os::dll_load(dll_name, ebuf, sizeof ebuf);
+    }
+    if (libvectormath != nullptr) {
+      // Method naming convention
+      //   All the methods are named as <OP><T><N>_<U><suffix>
+      //   Where:
+      //     <OP>     is the operation name, e.g. sin
+      //     <T>      is optional to indicate float/double
+      //              "f/d" for vector float/double operation
+      //     <N>      is the number of elements in the vector
+      //              "2/4" for neon, and "x" for sve
+      //     <U>      is the precision level
+      //              "u10/u05" represents 1.0/0.5 ULP error bounds
+      //               We use "u10" for all operations by default
+      //               But for those functions do not have u10 support, we use "u05" instead
+      //     <suffix> indicates neon/sve
+      //              "sve/advsimd" for sve/neon implementations
+      //     e.g. sinfx_u10sve is the method for computing vector float sin using SVE instructions
+      //          cosd2_u10advsimd is the method for computing 2 elements vector double cos using NEON instructions
+      //
+      log_info(library)("Loaded library %s, handle " INTPTR_FORMAT, JNI_LIB_PREFIX "vectormath" JNI_LIB_SUFFIX, p2i(libvectormath));
+
+      // Math vector stubs implemented with SVE for scalable vector size.
+      if (UseSVE > 0) {
+        for (int op = 0; op < VectorSupport::NUM_VECTOR_OP_MATH; op++) {
+          int vop = VectorSupport::VECTOR_OP_MATH_START + op;
+
+          // The native library does not support u10 level of "hypot".
+          const char* ulf = (vop == VectorSupport::VECTOR_OP_HYPOT) ? "u05" : "u10";
+
+          snprintf(ebuf, sizeof(ebuf), "%sfx_%ssve", VectorSupport::mathname[op], ulf);
+          StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_SCALABLE][op] = (address)os::dll_lookup(libvectormath, ebuf);
+
+          snprintf(ebuf, sizeof(ebuf), "%sdx_%ssve", VectorSupport::mathname[op], ulf);
+          StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_SCALABLE][op] = (address)os::dll_lookup(libvectormath, ebuf);
+        }
+      }
+
+      // Math vector stubs implemented with NEON for 64/128 bits vector size.
+      for (int op = 0; op < VectorSupport::NUM_VECTOR_OP_MATH; op++) {
+        int vop = VectorSupport::VECTOR_OP_MATH_START + op;
+
+        // The native library does not support u10 level of "hypot".
+        const char* ulf = (vop == VectorSupport::VECTOR_OP_HYPOT) ? "u05" : "u10";
+
+        snprintf(ebuf, sizeof(ebuf), "%sf4_%sadvsimd", VectorSupport::mathname[op], ulf);
+        StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_64][op] = (address)os::dll_lookup(libvectormath, ebuf);
+
+        snprintf(ebuf, sizeof(ebuf), "%sf4_%sadvsimd", VectorSupport::mathname[op], ulf);
+        StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_128][op] = (address)os::dll_lookup(libvectormath, ebuf);
+
+        snprintf(ebuf, sizeof(ebuf), "%sd2_%sadvsimd", VectorSupport::mathname[op], ulf);
+        StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_128][op] = (address)os::dll_lookup(libvectormath, ebuf);
+      }
+    } else {
+      log_info(library)("Failed to load native vector math library!");
+    }
+#endif // COMPILER2
+
 #endif // COMPILER2_OR_JVMCI
   }
 

src/hotspot/cpu/x86/stubGenerator_x86_64.cpp

@@ -4334,41 +4334,41 @@ void StubGenerator::generate_compiler_stubs() {
 
     log_info(library)("Loaded library %s, handle " INTPTR_FORMAT, JNI_LIB_PREFIX "jsvml" JNI_LIB_SUFFIX, p2i(libjsvml));
     if (UseAVX > 2) {
-      for (int op = 0; op < VectorSupport::NUM_SVML_OP; op++) {
-        int vop = VectorSupport::VECTOR_OP_SVML_START + op;
+      for (int op = 0; op < VectorSupport::NUM_VECTOR_OP_MATH; op++) {
+        int vop = VectorSupport::VECTOR_OP_MATH_START + op;
         if ((!VM_Version::supports_avx512dq()) &&
             (vop == VectorSupport::VECTOR_OP_LOG || vop == VectorSupport::VECTOR_OP_LOG10 || vop == VectorSupport::VECTOR_OP_POW)) {
           continue;
         }
-        snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf16_ha_z0", VectorSupport::svmlname[op]);
+        snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf16_ha_z0", VectorSupport::mathname[op]);
         StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_512][op] = (address)os::dll_lookup(libjsvml, ebuf);
 
-        snprintf(ebuf, sizeof(ebuf), "__jsvml_%s8_ha_z0", VectorSupport::svmlname[op]);
+        snprintf(ebuf, sizeof(ebuf), "__jsvml_%s8_ha_z0", VectorSupport::mathname[op]);
         StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_512][op] = (address)os::dll_lookup(libjsvml, ebuf);
       }
     }
     const char* avx_sse_str = (UseAVX >= 2) ? "l9" : ((UseAVX == 1) ? "e9" : "ex");
-    for (int op = 0; op < VectorSupport::NUM_SVML_OP; op++) {
-      int vop = VectorSupport::VECTOR_OP_SVML_START + op;
+    for (int op = 0; op < VectorSupport::NUM_VECTOR_OP_MATH; op++) {
+      int vop = VectorSupport::VECTOR_OP_MATH_START + op;
       if (vop == VectorSupport::VECTOR_OP_POW) {
         continue;
       }
-      snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf4_ha_%s", VectorSupport::svmlname[op], avx_sse_str);
+      snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf4_ha_%s", VectorSupport::mathname[op], avx_sse_str);
       StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_64][op] = (address)os::dll_lookup(libjsvml, ebuf);
 
-      snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf4_ha_%s", VectorSupport::svmlname[op], avx_sse_str);
+      snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf4_ha_%s", VectorSupport::mathname[op], avx_sse_str);
       StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_128][op] = (address)os::dll_lookup(libjsvml, ebuf);
 
-      snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf8_ha_%s", VectorSupport::svmlname[op], avx_sse_str);
+      snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf8_ha_%s", VectorSupport::mathname[op], avx_sse_str);
       StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_256][op] = (address)os::dll_lookup(libjsvml, ebuf);
 
-      snprintf(ebuf, sizeof(ebuf), "__jsvml_%s1_ha_%s", VectorSupport::svmlname[op], avx_sse_str);
+      snprintf(ebuf, sizeof(ebuf), "__jsvml_%s1_ha_%s", VectorSupport::mathname[op], avx_sse_str);
       StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_64][op] = (address)os::dll_lookup(libjsvml, ebuf);
 
-      snprintf(ebuf, sizeof(ebuf), "__jsvml_%s2_ha_%s", VectorSupport::svmlname[op], avx_sse_str);
+      snprintf(ebuf, sizeof(ebuf), "__jsvml_%s2_ha_%s", VectorSupport::mathname[op], avx_sse_str);
       StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_128][op] = (address)os::dll_lookup(libjsvml, ebuf);
 
-      snprintf(ebuf, sizeof(ebuf), "__jsvml_%s4_ha_%s", VectorSupport::svmlname[op], avx_sse_str);
+      snprintf(ebuf, sizeof(ebuf), "__jsvml_%s4_ha_%s", VectorSupport::mathname[op], avx_sse_str);
       StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_256][op] = (address)os::dll_lookup(libjsvml, ebuf);
     }
   }

src/hotspot/share/opto/callnode.cpp

@@ -755,7 +755,7 @@ Node *CallNode::match( const ProjNode *proj, const Matcher *match ) {
 
     if (Opcode() == Op_CallLeafVector) {
       // If the return is in vector, compute appropriate regmask taking into account the whole range
-      if(ideal_reg >= Op_VecS && ideal_reg <= Op_VecZ) {
+      if(ideal_reg >= Op_VecA && ideal_reg <= Op_VecZ) {
         if(OptoReg::is_valid(regs.second())) {
           for (OptoReg::Name r = regs.first(); r <= regs.second(); r = OptoReg::add(r, 1)) {
             rm.Insert(r);

src/hotspot/share/opto/library_call.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020, 2023, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2020, 2024, 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
@@ -364,7 +364,7 @@ class LibraryCallKit : public GraphKit {
   bool inline_index_vector();
   bool inline_index_partially_in_upper_range();
 
-  Node* gen_call_to_svml(int vector_api_op_id, BasicType bt, int num_elem, Node* opd1, Node* opd2);
+  Node* gen_call_to_vector_math(int vector_api_op_id, BasicType bt, int num_elem, Node* opd1, Node* opd2);
 
   enum VectorMaskUseType {
     VecMaskUseLoad  = 1 << 0,

src/hotspot/share/opto/vectorIntrinsics.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020, 2023, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2020, 2024, 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
@@ -536,12 +536,12 @@ bool LibraryCallKit::inline_vector_nary_operation(int n) {
   Node* operation = nullptr;
   if (opc == Op_CallLeafVector) {
     assert(UseVectorStubs, "sanity");
-    operation = gen_call_to_svml(opr->get_con(), elem_bt, num_elem, opd1, opd2);
+    operation = gen_call_to_vector_math(opr->get_con(), elem_bt, num_elem, opd1, opd2);
     if (operation == nullptr) {
       if (C->print_intrinsics()) {
-        tty->print_cr("  ** svml call failed for %s_%s_%d",
-                         (elem_bt == T_FLOAT)?"float":"double",
-                         VectorSupport::svmlname[opr->get_con() - VectorSupport::VECTOR_OP_SVML_START],
+        tty->print_cr("  ** Vector math call failed for %s_%s_%d",
+                         (elem_bt == T_FLOAT) ? "float" : "double",
+                         VectorSupport::mathname[opr->get_con() - VectorSupport::VECTOR_OP_MATH_START],
                          num_elem * type2aelembytes(elem_bt));
       }
       return false;
@@ -2174,24 +2174,24 @@ bool LibraryCallKit::inline_vector_rearrange() {
   return true;
 }
 
-static address get_svml_address(int vop, int bits, BasicType bt, char* name_ptr, int name_len) {
+static address get_vector_math_address(int vop, int bits, BasicType bt, char* name_ptr, int name_len) {
   address addr = nullptr;
   assert(UseVectorStubs, "sanity");
   assert(name_ptr != nullptr, "unexpected");
-  assert((vop >= VectorSupport::VECTOR_OP_SVML_START) && (vop <= VectorSupport::VECTOR_OP_SVML_END), "unexpected");
-  int op = vop - VectorSupport::VECTOR_OP_SVML_START;
+  assert((vop >= VectorSupport::VECTOR_OP_MATH_START) && (vop <= VectorSupport::VECTOR_OP_MATH_END), "unexpected");
+  int op = vop - VectorSupport::VECTOR_OP_MATH_START;
 
   switch(bits) {
     case 64:  //fallthough
     case 128: //fallthough
     case 256: //fallthough
     case 512:
       if (bt == T_FLOAT) {
-        snprintf(name_ptr, name_len, "vector_%s_float%d", VectorSupport::svmlname[op], bits);
+        snprintf(name_ptr, name_len, "vector_%s_float_%dbits_fixed", VectorSupport::mathname[op], bits);
         addr = StubRoutines::_vector_f_math[exact_log2(bits/64)][op];
       } else {
         assert(bt == T_DOUBLE, "must be FP type only");
-        snprintf(name_ptr, name_len, "vector_%s_double%d", VectorSupport::svmlname[op], bits);
+        snprintf(name_ptr, name_len, "vector_%s_double_%dbits_fixed", VectorSupport::mathname[op], bits);
         addr = StubRoutines::_vector_d_math[exact_log2(bits/64)][op];
       }
       break;
@@ -2202,19 +2202,29 @@ static address get_svml_address(int vop, int bits, BasicType bt, char* name_ptr,
       break;
   }
 
+  if (addr == nullptr && Matcher::supports_scalable_vector()) {
+    if (bt == T_FLOAT) {
+      snprintf(name_ptr, name_len, "vector_%s_float_%dbits_scalable", VectorSupport::mathname[op], bits);
+      addr = StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_SCALABLE][op];
+    } else {
+      assert(bt == T_DOUBLE, "must be FP type only");
+      snprintf(name_ptr, name_len, "vector_%s_double_%dbits_scalable", VectorSupport::mathname[op], bits);
+      addr = StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_SCALABLE][op];
+    }
+  }
   return addr;
 }
 
-Node* LibraryCallKit::gen_call_to_svml(int vector_api_op_id, BasicType bt, int num_elem, Node* opd1, Node* opd2) {
+Node* LibraryCallKit::gen_call_to_vector_math(int vector_api_op_id, BasicType bt, int num_elem, Node* opd1, Node* opd2) {
   assert(UseVectorStubs, "sanity");
-  assert(vector_api_op_id >= VectorSupport::VECTOR_OP_SVML_START && vector_api_op_id <= VectorSupport::VECTOR_OP_SVML_END, "need valid op id");
+  assert(vector_api_op_id >= VectorSupport::VECTOR_OP_MATH_START && vector_api_op_id <= VectorSupport::VECTOR_OP_MATH_END, "need valid op id");
   assert(opd1 != nullptr, "must not be null");
   const TypeVect* vt = TypeVect::make(bt, num_elem);
   const TypeFunc* call_type = OptoRuntime::Math_Vector_Vector_Type(opd2 != nullptr ? 2 : 1, vt, vt);
   char name[100] = "";
 
-  // Get address for svml method.
-  address addr = get_svml_address(vector_api_op_id, vt->length_in_bytes() * BitsPerByte, bt, name, 100);
+  // Get address for vector math method.
+  address addr = get_vector_math_address(vector_api_op_id, vt->length_in_bytes() * BitsPerByte, bt, name, 100);
 
   if (addr == nullptr) {
     return nullptr;

src/hotspot/share/prims/vectorSupport.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020, 2023, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2020, 2024, 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
@@ -43,7 +43,7 @@
 #endif // COMPILER2
 
 #ifdef COMPILER2
-const char* VectorSupport::svmlname[VectorSupport::NUM_SVML_OP] = {
+const char* VectorSupport::mathname[VectorSupport::NUM_VECTOR_OP_MATH] = {
     "tan",
     "tanh",
     "sin",