JIT/AOT to interpreter calls support (#116353)

* JIT/AOT to interpreter calls support

Until now, we only had support for calling interpreted void returning
method with no arguments by the JIT/AOT. This change enables support for
passing all possible kinds of arguments and returning all types.

It reuses the `CallStubGenerator` that was implemented for the interpreter
to native code usage and adds support for the other direction to it in
mostly trivial manner.
In addition to that, assembler routines for storing argument values to
the interpreter stack were added.
The `CallStubGenerator` generates a list of routines to copy the
arguments from CPU registers and stack to the interpreter stack.
The last one makes call to the `ExecuteInterpretedMethod` and then puts
the result into appropriate registers.
For functions that return result via a return buffer, the buffer is
passed to the `ExecuteInterpretedMethod` so that the IR opcode to return
valuetype stores it directly to the return buffer.

The ARM64 for Apple OSes is the most optimized version, as it is the
primary target where the performance matters the most. It eliminates
argument registers saving to stack on the fast path when we already have
the call stub.

* Fix m_pInterpThreadContext offset after rebase to main

* Fix m_pInterpThreadContext for more targets

* Move the CreateNativeToInterpreterCallStub to prestub

* Fix build break and add a comment

* Create stub only when interp. method called from JIT/AOT code

* Unify some asm macros

* Update src/coreclr/pal/inc/unixasmmacrosamd64.inc

* Fix OSX x64 build

---------

Co-authored-by: Jan Kotas <[email protected]>

Author: janvorli

src/coreclr/interpreter/interpretershared.h

@@ -19,6 +19,8 @@
 
 #define INTERP_INDIRECT_HELPER_TAG 1 // When a helper ftn's address is indirect we tag it with this tag bit
 
+struct CallStubHeader;
+
 struct InterpMethod
 {
 #if DEBUG
@@ -27,6 +29,8 @@ struct InterpMethod
     CORINFO_METHOD_HANDLE methodHnd;
     int32_t allocaSize;
     void** pDataItems;
+    // This stub is used for calling the interpreted method from JITted/AOTed code
+    CallStubHeader *pCallStub;
     bool initLocals;
 
     InterpMethod(CORINFO_METHOD_HANDLE methodHnd, int32_t allocaSize, void** pDataItems, bool initLocals)
@@ -38,6 +42,7 @@ struct InterpMethod
         this->allocaSize = allocaSize;
         this->pDataItems = pDataItems;
         this->initLocals = initLocals;
+        pCallStub = NULL;
     }
 
     bool CheckIntegrity()

src/coreclr/pal/inc/unixasmmacrosamd64.inc

@@ -222,6 +222,12 @@ C_FUNC(\Name\()_End):
 
 .endm
 
+.macro SKIP_ARGUMENT_REGISTERS
+
+        add rsp, 6 * 8
+
+.endm
+
 .macro SAVE_FLOAT_ARGUMENT_REGISTERS ofs
 
         save_xmm128_postrsp xmm0, \ofs
@@ -344,7 +350,7 @@ C_FUNC(\Name\()_End):
 .macro EPILOG_WITH_TRANSITION_BLOCK_RETURN
 
         free_stack      __PWTB_StackAlloc
-        POP_ARGUMENT_REGISTERS
+        SKIP_ARGUMENT_REGISTERS
         POP_CALLEE_SAVED_REGISTERS
         ret
 
@@ -368,19 +374,29 @@ C_FUNC(\Name\()_End):
 
 .endm
 
-// Inlined version of GetThreadEEAllocContext. Trashes volatile registers.
-.macro INLINE_GET_ALLOC_CONTEXT_BASE
-#if defined(FEATURE_EMULATED_TLS) || defined(__APPLE__)
-        call    C_FUNC(GetThreadEEAllocContext)
+.macro INLINE_GET_TLS_VAR Var
+       .att_syntax
+#if defined(__APPLE__)
+        movq    _\Var@TLVP(%rip), %rdi
+        callq   *(%rdi)
 #else
-        .att_syntax
         .byte 0x66  // data16 prefix - padding to have space for linker relaxations
-        leaq    t_runtime_thread_locals@TLSGD(%rip), %rdi
+        leaq    \Var@TLSGD(%rip), %rdi
         .byte 0x66  //
         .byte 0x66  //
         .byte 0x48  // rex.W prefix, also for padding
         callq   __tls_get_addr@PLT
-        .intel_syntax noprefix
+#endif
+       .intel_syntax noprefix
+.endm
+
+
+// Inlined version of GetThreadEEAllocContext. Trashes volatile registers.
+.macro INLINE_GET_ALLOC_CONTEXT_BASE
+#ifdef FEATURE_EMULATED_TLS
+        call    C_FUNC(GetThreadEEAllocContext)
+#else
+        INLINE_GET_TLS_VAR t_runtime_thread_locals
 
         .ifnc OFFSETOF__RuntimeThreadLocals__ee_alloc_context, 0
                 lea     rax, [rax + OFFSETOF__RuntimeThreadLocals__ee_alloc_context]
@@ -405,3 +421,9 @@ C_FUNC(\Name\()_End):
         free_stack      56
         POP_CALLEE_SAVED_REGISTERS
 .endm
+
+.macro INLINE_GETTHREAD
+        // Inlined version of call C_FUNC(RhpGetThread)
+        INLINE_GET_TLS_VAR t_CurrentThreadInfo
+        mov rax, [rax + OFFSETOF__ThreadLocalInfo__m_pThread]
+.endm

src/coreclr/pal/inc/unixasmmacrosarm64.inc

@@ -184,6 +184,7 @@ C_FUNC(\Name\()_End):
 // ArgumentRegisters::x2
 // ArgumentRegisters::x1
 // ArgumentRegisters::x0
+// ArgumentRegisters::x8
 // FloatRegisters::q7
 // FloatRegisters::q6
 // FloatRegisters::q5
@@ -192,7 +193,7 @@ C_FUNC(\Name\()_End):
 // FloatRegisters::q2
 // FloatRegisters::q1
 // FloatRegisters::q0
-.macro PROLOG_WITH_TRANSITION_BLOCK extraLocals = 0, SaveFPArgs = 1
+.macro PROLOG_WITH_TRANSITION_BLOCK extraLocals = 0, SaveFPArgs = 1, SaveGPArgs = 1
 
         __PWTB_FloatArgumentRegisters = \extraLocals
         __PWTB_SaveFPArgs = \SaveFPArgs
@@ -222,8 +223,10 @@ C_FUNC(\Name\()_End):
         // Allocate space for the rest of the frame
         PROLOG_STACK_ALLOC  __PWTB_StackAlloc
 
-        // Spill argument registers.
-        SAVE_ARGUMENT_REGISTERS        sp, __PWTB_ArgumentRegisters
+        .if (\SaveGPArgs == 1)
+                // Spill argument registers.
+                SAVE_ARGUMENT_REGISTERS        sp, __PWTB_ArgumentRegisters
+        .endif
 
         .if (__PWTB_SaveFPArgs == 1)
                 SAVE_FLOAT_ARGUMENT_REGISTERS sp, \extraLocals
@@ -301,7 +304,6 @@ C_FUNC(\Name\()_End):
 
 .endm
 
-
 //-----------------------------------------------------------------------------
 // Provides a matching epilog to PROLOG_WITH_TRANSITION_BLOCK and ends by preparing for tail-calling.
 // Since this is a tail call argument registers are restored.
@@ -325,6 +327,41 @@ C_FUNC(\Name\()_End):
 
 .endm
 
+// Loads the address of a thread-local variable into the target register,
+// which cannot be x0.
+// Preserves registers except for xip0 and xip1 on Apple
+.macro INLINE_GET_TLS_VAR target, var
+    .ifc \target, x0
+        .error "target cannot be x0"
+    .endif
+
+    // This sequence of instructions is recognized and potentially patched
+    // by the linker (GD->IE/LE relaxation).
+#if defined(__APPLE__)
+
+    adrp    x0, \var@TLVPPAGE
+    ldr     x0, [x0, \var@TLVPPAGEOFF]
+    ldr     \target, [x0]
+
+    blr     \target
+    // End of the sequence
+
+    mov     \target, x0
+#else
+    adrp    x0, :tlsdesc:\var
+    ldr     \target, [x0, #:tlsdesc_lo12:\var]
+    add     x0, x0, :tlsdesc_lo12:\var
+    .tlsdesccall \var
+    blr     \target
+    // End of the sequence
+
+    mrs     \target, tpidr_el0
+    add     \target, \target, x0
+#endif
+
+.endm
+
+
 // Inlined version of GetThreadEEAllocContext. Target cannot be x0 or x1.
 .macro INLINE_GET_ALLOC_CONTEXT_BASE target
         .ifc  \target, x0
@@ -345,17 +382,7 @@ C_FUNC(\Name\()_End):
         EPILOG_RESTORE_REG_PAIR_INDEXED   fp, lr, 0x20
 #else
         PROLOG_SAVE_REG_PAIR_INDEXED      x0, lr, -0x10
-
-        // This sequence of instructions is recognized and potentially patched
-        // by the linker (GD->IE/LE relaxation).
-        adrp    x0, :tlsdesc:t_runtime_thread_locals
-        ldr     \target, [x0, :tlsdesc_lo12:t_runtime_thread_locals]
-        add     x0, x0, :tlsdesc_lo12:t_runtime_thread_locals
-        blr     \target
-        // End of the sequence
-
-        mrs     \target, TPIDR_EL0
-        add     \target, \target, x0
+        INLINE_GET_TLS_VAR \target, t_runtime_thread_locals
 
         .ifnc OFFSETOF__RuntimeThreadLocals__ee_alloc_context, 0
                 add     \target, x0, OFFSETOF__RuntimeThreadLocals__ee_alloc_context
@@ -470,3 +497,13 @@ $__RedirectionStubEndFuncName
 0:
 #endif
 .endm
+
+#define xip0 x16
+#define xip1 x17
+#define xpr x18
+
+// Inlined version of RhpGetThread. Target cannot be x0.
+.macro INLINE_GETTHREAD target
+    INLINE_GET_TLS_VAR \target, C_FUNC(t_CurrentThreadInfo)
+    ldr     \target, [\target, #OFFSETOF__ThreadLocalInfo__m_pThread]
+.endm