[vm/ffi] Outlining state transitions in AOT

For single argument FFI calls:
- reduces trampoline size significantly (up to 50%, 150-170 bytes),
- reduces the compressed size of GPay by 2.5kb on arm64,
- regresses performance on arm64 (up to 2.5%).
For more arguments, percentage-wise size gains and speed regressions
are smaller.

Only applied on arm and arm64, we care about code size for these.

Note: On Raspberry Pie (arm), the performance regression on single-
argument calls regresses up to 30%.

TEST=tests/ffi/*

Design doc: https://go/dart-ffi-outline-state-transitions
Closes: dart-lang/sdk#50094
Change-Id: I8b8d7da45f69be6ac1432b11b695de71e56acfd1
Cq-Include-Trybots: luci.dart.try:vm-precomp-ffi-qemu-linux-release-arm-try,vm-ffi-android-debug-arm64c-try,vm-ffi-android-debug-arm-try,vm-kernel-nnbd-mac-debug-arm64-try
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/262343
Reviewed-by: Martin Kustermann <[email protected]>
Commit-Queue: Daco Harkes <[email protected]>

Author: dcharkes

runtime/vm/compiler/backend/il_arm.cc

@@ -1501,32 +1501,18 @@ void FfiCallInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
                                    env());
 
     // Update information in the thread object and enter a safepoint.
-    if (CanExecuteGeneratedCodeInSafepoint()) {
-      __ LoadImmediate(temp1, compiler::target::Thread::exit_through_ffi());
-      __ TransitionGeneratedToNative(branch, FPREG, temp1, saved_fp_or_sp,
-                                     /*enter_safepoint=*/true);
-
-      __ blx(branch);
-
-      // Update information in the thread object and leave the safepoint.
-      __ TransitionNativeToGenerated(saved_fp_or_sp, temp1,
-                                     /*leave_safepoint=*/true);
-    } else {
-      // We cannot trust that this code will be executable within a safepoint.
-      // Therefore we delegate the responsibility of entering/exiting the
-      // safepoint to a stub which in the VM isolate's heap, which will never
-      // lose execute permission.
-      __ ldr(temp1,
-             compiler::Address(
-                 THR, compiler::target::Thread::
-                          call_native_through_safepoint_entry_point_offset()));
-
-      // Calls R8 in a safepoint and clobbers R4 and NOTFP.
-      ASSERT(branch == R8);
-      static_assert((kReservedCpuRegisters & (1 << NOTFP)) != 0,
-                    "NOTFP should be a reserved register");
-      __ blx(temp1);
-    }
+    // Outline state transition. In AOT, for code size. In JIT, because we
+    // cannot trust that code will be executable.
+    __ ldr(temp1,
+           compiler::Address(
+               THR, compiler::target::Thread::
+                        call_native_through_safepoint_entry_point_offset()));
+
+    // Calls R8 in a safepoint and clobbers R4 and NOTFP.
+    ASSERT(branch == R8);
+    static_assert((kReservedCpuRegisters & (1 << NOTFP)) != 0,
+                  "NOTFP should be a reserved register");
+    __ blx(temp1);
 
     if (marshaller_.IsHandle(compiler::ffi::kResultIndex)) {
       __ Comment("Check Dart_Handle for Error.");

runtime/vm/compiler/backend/il_arm64.cc

@@ -1371,39 +1371,17 @@ void FfiCallInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
 
     __ StoreToOffset(temp1, FPREG, kSavedCallerPcSlotFromFp * kWordSize);
 
-    if (CanExecuteGeneratedCodeInSafepoint()) {
-      // Update information in the thread object and enter a safepoint.
-      __ LoadImmediate(temp1, compiler::target::Thread::exit_through_ffi());
-      __ TransitionGeneratedToNative(branch, FPREG, temp1,
-                                     /*enter_safepoint=*/true);
-
-      // We are entering runtime code, so the C stack pointer must be restored
-      // from the stack limit to the top of the stack.
-      __ mov(temp_csp, CSP);
-      __ mov(CSP, SP);
-
-      __ blr(branch);
-
-      // Restore the Dart stack pointer.
-      __ mov(SP, CSP);
-      __ mov(CSP, temp_csp);
-
-      // Update information in the thread object and leave the safepoint.
-      __ TransitionNativeToGenerated(temp1, /*leave_safepoint=*/true);
-    } else {
-      // We cannot trust that this code will be executable within a safepoint.
-      // Therefore we delegate the responsibility of entering/exiting the
-      // safepoint to a stub which in the VM isolate's heap, which will never
-      // lose execute permission.
-      __ ldr(temp1,
-             compiler::Address(
-                 THR, compiler::target::Thread::
-                          call_native_through_safepoint_entry_point_offset()));
+    // Update information in the thread object and enter a safepoint.
+    // Outline state transition. In AOT, for code size. In JIT, because we
+    // cannot trust that code will be executable.
+    __ ldr(temp1,
+           compiler::Address(
+               THR, compiler::target::Thread::
+                        call_native_through_safepoint_entry_point_offset()));
 
-      // Calls R9 and clobbers R19 (along with volatile registers).
-      ASSERT(branch == R9);
-      __ blr(temp1);
-    }
+    // Calls R9 and clobbers R19 (along with volatile registers).
+    ASSERT(branch == R9);
+    __ blr(temp1);
 
     if (marshaller_.IsHandle(compiler::ffi::kResultIndex)) {
       __ Comment("Check Dart_Handle for Error.");