8291893: riscv: remove fence.i used in user space

Reviewed-by: fyang, vkempik

Author: Yadong Wang

src/hotspot/cpu/riscv/assembler_riscv.cpp

@@ -311,13 +311,6 @@ void Assembler::movptr(Register Rd, address addr) {
   addi(Rd, Rd, offset);
 }
 
-void Assembler::ifence() {
-  fence_i();
-  if (UseConservativeFence) {
-    fence(ir, ir);
-  }
-}
-
 #define INSN(NAME, NEG_INSN)                                                         \
   void Assembler::NAME(Register Rs, Register Rt, const address &dest) {              \
     NEG_INSN(Rt, Rs, dest);                                                          \

src/hotspot/cpu/riscv/assembler_riscv.hpp

@@ -339,7 +339,6 @@ class Assembler : public AbstractAssembler {
   void movptr(Register Rd, address addr);
   void movptr_with_offset(Register Rd, address addr, int32_t &offset);
   void movptr(Register Rd, uintptr_t imm64);
-  void ifence();
   void j(const address &dest, Register temp = t0);
   void j(const Address &adr, Register temp = t0);
   void j(Label &l, Register temp = t0);
@@ -961,7 +960,6 @@ class Assembler : public AbstractAssembler {
     emit(insn);                                             \
   }
 
-  INSN(fence_i, 0b0001111, 0b001, 0b000000000000);
   INSN(ecall,   0b1110011, 0b000, 0b000000000000);
   INSN(_ebreak, 0b1110011, 0b000, 0b000000000001);
 

src/hotspot/cpu/riscv/compiledIC_riscv.cpp

@@ -69,8 +69,8 @@ address CompiledStaticCall::emit_to_interp_stub(CodeBuffer &cbuf, address mark)
 #undef __
 
 int CompiledStaticCall::to_interp_stub_size() {
-  // fence_i + fence* + (lui, addi, slli, addi, slli, addi) + (lui, addi, slli, addi, slli) + jalr
-  return NativeFenceI::instruction_size() + 12 * NativeInstruction::instruction_size;
+  // (lui, addi, slli, addi, slli, addi) + (lui, addi, slli, addi, slli) + jalr
+  return 12 * NativeInstruction::instruction_size;
 }
 
 int CompiledStaticCall::to_trampoline_stub_size() {
@@ -98,7 +98,7 @@ void CompiledDirectStaticCall::set_to_interpreted(const methodHandle& callee, ad
 
   // Creation also verifies the object.
   NativeMovConstReg* method_holder
-    = nativeMovConstReg_at(stub + NativeFenceI::instruction_size());
+    = nativeMovConstReg_at(stub);
 #ifdef ASSERT
   NativeGeneralJump* jump = nativeGeneralJump_at(method_holder->next_instruction_address());
 
@@ -119,7 +119,7 @@ void CompiledDirectStaticCall::set_stub_to_clean(static_stub_Relocation* static_
   assert(CompiledICLocker::is_safe(stub), "mt unsafe call");
   // Creation also verifies the object.
   NativeMovConstReg* method_holder
-    = nativeMovConstReg_at(stub + NativeFenceI::instruction_size());
+    = nativeMovConstReg_at(stub);
   method_holder->set_data(0);
   NativeJump* jump = nativeJump_at(method_holder->next_instruction_address());
   jump->set_jump_destination((address)-1);
@@ -139,7 +139,7 @@ void CompiledDirectStaticCall::verify() {
   assert(stub != NULL, "no stub found for static call");
   // Creation also verifies the object.
   NativeMovConstReg* method_holder
-    = nativeMovConstReg_at(stub + NativeFenceI::instruction_size());
+    = nativeMovConstReg_at(stub);
   NativeJump* jump = nativeJump_at(method_holder->next_instruction_address());
 
   // Verify state.

src/hotspot/cpu/riscv/globals_riscv.hpp

@@ -88,8 +88,7 @@ define_pd_global(intx, InlineSmallCode,          1000);
   product(bool, TraceTraps, false, "Trace all traps the signal handler")         \
   /* For now we're going to be safe and add the I/O bits to userspace fences. */ \
   product(bool, UseConservativeFence, true,                                      \
-          "Extend i for r and o for w in the pred/succ flags of fence;"          \
-          "Extend fence.i to fence.i + fence.")                                  \
+          "Extend i for r and o for w in the pred/succ flags of fence")          \
   product(bool, AvoidUnalignedAccesses, true,                                    \
           "Avoid generating unaligned memory accesses")                          \
   product(bool, UseRVV, false, EXPERIMENTAL, "Use RVV instructions")             \

src/hotspot/cpu/riscv/icache_riscv.cpp

@@ -30,6 +30,14 @@
 #define __ _masm->
 
 static int icache_flush(address addr, int lines, int magic) {
+  // To make a store to instruction memory visible to all RISC-V harts,
+  // the writing hart has to execute a data FENCE before requesting that
+  // all remote RISC-V harts execute a FENCE.I.
+  //
+  // No sush assurance is defined at the interface level of the builtin
+  // method, and so we should make sure it works.
+  __asm__ volatile("fence rw, rw" : : : "memory");
+
   __builtin___clear_cache(addr, addr + (lines << ICache::log2_line_size));
   return magic;
 }

src/hotspot/cpu/riscv/macroAssembler_riscv.cpp

@@ -555,7 +555,6 @@ void MacroAssembler::emit_static_call_stub() {
   // CompiledDirectStaticCall::set_to_interpreted knows the
   // exact layout of this stub.
 
-  ifence();
   mov_metadata(xmethod, (Metadata*)NULL);
 
   // Jump to the entry point of the i2c stub.
@@ -1668,7 +1667,6 @@ void MacroAssembler::movoop(Register dst, jobject obj, bool immediate) {
 
   // nmethod entry barrier necessitate using the constant pool. They have to be
   // ordered with respected to oop access.
-  // Using immediate literals would necessitate fence.i.
   if (BarrierSet::barrier_set()->barrier_set_nmethod() != NULL || !immediate) {
     address dummy = address(uintptr_t(pc()) & -wordSize); // A nearby aligned address
     ld_constant(dst, Address(dummy, rspec));
@@ -2746,7 +2744,6 @@ void MacroAssembler::build_frame(int framesize) {
   sd(fp, Address(sp, framesize - 2 * wordSize));
   sd(ra, Address(sp, framesize - wordSize));
   if (PreserveFramePointer) { add(fp, sp, framesize); }
-  verify_cross_modify_fence_not_required();
 }
 
 void MacroAssembler::remove_frame(int framesize) {
@@ -2797,7 +2794,6 @@ address MacroAssembler::read_polling_page(Register r, int32_t offset, relocInfo:
     lwu(zr, Address(r, offset));
     mark = inst_mark();
   }
-  verify_cross_modify_fence_not_required();
   return mark;
 }
 
@@ -3981,29 +3977,3 @@ void MacroAssembler::cmp_l2i(Register dst, Register src1, Register src2, Registe
   neg(dst, dst);
   bind(done);
 }
-
-void MacroAssembler::safepoint_ifence() {
-  ifence();
-#ifndef PRODUCT
-  if (VerifyCrossModifyFence) {
-    // Clear the thread state.
-    sb(zr, Address(xthread, in_bytes(JavaThread::requires_cross_modify_fence_offset())));
-  }
-#endif
-}
-
-#ifndef PRODUCT
-void MacroAssembler::verify_cross_modify_fence_not_required() {
-  if (VerifyCrossModifyFence) {
-    // Check if thread needs a cross modify fence.
-    lbu(t0, Address(xthread, in_bytes(JavaThread::requires_cross_modify_fence_offset())));
-    Label fence_not_required;
-    beqz(t0, fence_not_required);
-    // If it does then fail.
-    la(t0, RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::verify_cross_modify_fence_failure)));
-    mv(c_rarg0, xthread);
-    jalr(t0);
-    bind(fence_not_required);
-  }
-}
-#endif

src/hotspot/cpu/riscv/macroAssembler_riscv.hpp

@@ -46,9 +46,6 @@ class MacroAssembler: public Assembler {
 
   void safepoint_poll(Label& slow_path, bool at_return, bool acquire, bool in_nmethod);
 
-  // Place a fence.i after code may have been modified due to a safepoint.
-  void safepoint_ifence();
-
   // Alignment
   void align(int modulus, int extra_offset = 0);
 
@@ -836,9 +833,6 @@ class MacroAssembler: public Assembler {
 
   void load_reserved(Register addr, enum operand_size size, Assembler::Aqrl acquire);
   void store_conditional(Register addr, Register new_val, enum operand_size size, Assembler::Aqrl release);
-
-  // Check the current thread doesn't need a cross modify fence.
-  void verify_cross_modify_fence_not_required() PRODUCT_RETURN;
 };
 
 #ifdef ASSERT

src/hotspot/cpu/riscv/nativeInst_riscv.hpp

@@ -42,7 +42,6 @@
 // - - NativeIllegalInstruction
 // - - NativeCallTrampolineStub
 // - - NativeMembar
-// - - NativeFenceI
 
 // The base class for different kinds of native instruction abstractions.
 // Provides the primitive operations to manipulate code relative to this.
@@ -554,14 +553,6 @@ inline NativeMembar *NativeMembar_at(address addr) {
   return (NativeMembar*)addr;
 }
 
-class NativeFenceI : public NativeInstruction {
-public:
-  static inline int instruction_size() {
-    // 2 for fence.i + fence
-    return (UseConservativeFence ? 2 : 1) * NativeInstruction::instruction_size;
-  }
-};
-
 class NativePostCallNop: public NativeInstruction {
 public:
   bool check() const { return is_nop(); }

src/hotspot/cpu/riscv/sharedRuntime_riscv.cpp

@@ -354,10 +354,6 @@ static void patch_callers_callsite(MacroAssembler *masm) {
   __ la_patchable(t0, RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite)), offset);
   __ jalr(x1, t0, offset);
 
-  // Explicit fence.i required because fixup_callers_callsite may change the code
-  // stream.
-  __ safepoint_ifence();
-
   __ pop_CPU_state();
   // restore sp
   __ leave();

src/hotspot/os_cpu/linux_riscv/orderAccess_linux_riscv.hpp

@@ -54,10 +54,6 @@ inline void OrderAccess::fence() {
 }
 
 inline void OrderAccess::cross_modify_fence_impl() {
-  asm volatile("fence.i" : : : "memory");
-  if (UseConservativeFence) {
-    asm volatile("fence ir, ir" : : : "memory");
-  }
 }
 
 #endif // OS_CPU_LINUX_RISCV_ORDERACCESS_LINUX_RISCV_HPP