[clr-interp] Runtime async on armel

src/coreclr/vm/arm/asmhelpers.S

@@ -986,11 +986,22 @@ LEAF_ENTRY Load_R0_R1_4B
     EPILOG_BRANCH_REG r5
 LEAF_END Load_R0_R1_4B
 
+LEAF_ENTRY Load_R0_R1_R2_4B
+    ldr r0, [r7], #4
+ALTERNATE_ENTRY Load_R1_R2_4B
+    ldr r1, [r7], #4
+    ldr r2, [r7], #4
+    ldr r5, [r6], #4
+    EPILOG_BRANCH_REG r5
+LEAF_END Load_R0_R1_R2_4B
+
 LEAF_ENTRY Load_R0_R1_R2_R3_4B
     ldr r0, [r7], #4
+ALTERNATE_ENTRY Load_R1_R2_R3_4B
     ldr r1, [r7], #4
 ALTERNATE_ENTRY Load_R2_R3_4B
     ldr r2, [r7], #4
+ALTERNATE_ENTRY Load_R3_4B
     ldr r3, [r7], #4
     ldr r5, [r6], #4
     EPILOG_BRANCH_REG r5
@@ -1039,11 +1050,22 @@ LEAF_ENTRY Store_R0_R1_4B
     EPILOG_BRANCH_REG r5
 LEAF_END Store_R0_R1_4B
 
+LEAF_ENTRY Store_R0_R1_R2_4B
+    str r0, [r7], #4
+ALTERNATE_ENTRY Store_R1_R2_4B
+    str r1, [r7], #4
+    str r2, [r7], #4
+    ldr r5, [r6], #4
+    EPILOG_BRANCH_REG r5
+LEAF_END Store_R0_R1_R2_4B
+
 LEAF_ENTRY Store_R0_R1_R2_R3_4B
     str r0, [r7], #4
+ALTERNATE_ENTRY Store_R1_R2_R3_4B
     str r1, [r7], #4
 ALTERNATE_ENTRY Store_R2_R3_4B
     str r2, [r7], #4
+ALTERNATE_ENTRY Store_R3_4B
     str r3, [r7], #4
     ldr r5, [r6], #4
     EPILOG_BRANCH_REG r5
@@ -1081,34 +1103,6 @@ LOCAL_LABEL(CopyLoop_4B):
     EPILOG_BRANCH_REG r5
 LEAF_END Load_Stack_4B
 
-LEAF_ENTRY Store_Stack
-    ldr r9, [r6], #4 // SP offset
-    ldr r8, [r6], #4 // number of stack slots
-    add r9, sp, r9
-    add r9, r9, #__PWTB_TransitionBlock + SIZEOF__TransitionBlock
-LOCAL_LABEL(StoreCopyLoop):
-    ldr r5, [r9], #4
-    str r5, [r7], #8
-    subs r8, r8, #4
-    bne LOCAL_LABEL(StoreCopyLoop)
-    ldr r5, [r6], #4
-    EPILOG_BRANCH_REG r5
-LEAF_END Store_Stack
-
-LEAF_ENTRY Store_Stack_4B
-    ldr r9, [r6], #4 // SP offset
-    ldr r8, [r6], #4 // number of stack slots
-    add r9, sp, r9
-    add r9, r9, #__PWTB_TransitionBlock + SIZEOF__TransitionBlock
-LOCAL_LABEL(StoreCopyLoop_4B):
-    ldr r5, [r9], #4
-    str r5, [r7], #4
-    subs r8, r8, #4
-    bne LOCAL_LABEL(StoreCopyLoop_4B)
-    ldr r5, [r6], #4
-    EPILOG_BRANCH_REG r5
-LEAF_END Store_Stack_4B
-
 NESTED_ENTRY CallJittedMethodRetVoid, _TEXT, NoHandler
     PROLOG_PUSH "{r4-r10,lr}"
     mov r4, sp
@@ -1119,6 +1113,8 @@ NESTED_ENTRY CallJittedMethodRetVoid, _TEXT, NoHandler
     CHECK_STACK_ALIGNMENT
     blx r5
     mov sp, r4
+    ldr r4, [r4, 32]
+    str r2, [r4]
     EPILOG_POP "{r4-r10,pc}"
 NESTED_END CallJittedMethodRetVoid, _TEXT
 
@@ -1134,6 +1130,8 @@ NESTED_ENTRY CallJittedMethodRetI4, _TEXT, NoHandler
     blx r5
     mov sp, r4
     str r0, [r10]
+    ldr r4, [r4, 32]
+    str r2, [r4]
     EPILOG_POP "{r4-r10,pc}"
 NESTED_END CallJittedMethodRetI4, _TEXT
 
@@ -1150,6 +1148,8 @@ NESTED_ENTRY CallJittedMethodRetI8, _TEXT, NoHandler
     mov sp, r4
     str r0, [r10]
     str r1, [r10, 4]
+    ldr r4, [r4, 32]
+    str r2, [r4]
     EPILOG_POP "{r4-r10,pc}"
 NESTED_END CallJittedMethodRetI8, _TEXT
 
@@ -1172,6 +1172,8 @@ NESTED_ENTRY CallJittedMethodRetBuffR0, _TEXT, NoHandler
     CHECK_STACK_ALIGNMENT
     blx r5
     mov sp, r4
+    ldr r4, [r4, 32]
+    str r2, [r4]
     EPILOG_POP "{r4-r10,pc}"
 NESTED_END CallJittedMethodRetBuffR0, _TEXT
 
@@ -1186,6 +1188,8 @@ NESTED_ENTRY CallJittedMethodRetBuffR1, _TEXT, NoHandler
     CHECK_STACK_ALIGNMENT
     blx r5
     mov sp, r4
+    ldr r4, [r4, 32]
+    str r2, [r4]
     EPILOG_POP "{r4-r10,pc}"
 NESTED_END CallJittedMethodRetBuffR1, _TEXT
 
@@ -1202,6 +1206,8 @@ NESTED_ENTRY CallJittedMethodRetI1, _TEXT, NoHandler
     mov sp, r4
     sxtb r0, r0
     str r0, [r10]
+    ldr r4, [r4, 32]
+    str r2, [r4]
     EPILOG_POP "{r4-r10,pc}"
 NESTED_END CallJittedMethodRetI1, _TEXT
 
@@ -1218,6 +1224,8 @@ NESTED_ENTRY CallJittedMethodRetI2, _TEXT, NoHandler
     mov sp, r4
     sxth r0, r0
     str r0, [r10]
+    ldr r4, [r4, 32]
+    str r2, [r4]
     EPILOG_POP "{r4-r10,pc}"
 NESTED_END CallJittedMethodRetI2, _TEXT
 
@@ -1234,6 +1242,8 @@ NESTED_ENTRY CallJittedMethodRetU1, _TEXT, NoHandler
     mov sp, r4
     uxtb r0, r0
     str r0, [r10]
+    ldr r4, [r4, 32]
+    str r2, [r4]
     EPILOG_POP "{r4-r10,pc}"
 NESTED_END CallJittedMethodRetU1, _TEXT
 
@@ -1250,6 +1260,8 @@ NESTED_ENTRY CallJittedMethodRetU2, _TEXT, NoHandler
     mov sp, r4
     uxth r0, r0
     str r0, [r10]
+    ldr r4, [r4, 32]
+    str r2, [r4]
     EPILOG_POP "{r4-r10,pc}"
 NESTED_END CallJittedMethodRetU2, _TEXT
 
@@ -1283,13 +1295,16 @@ LOCAL_LABEL(HaveInterpThreadContext):
     cmp r7, #0
     beq LOCAL_LABEL(NoManagedThreadOrCallStub)
     add r6, r7, #OFFSETOF__CallStubHeader__Routines
-    ldr r7, [r5, #OFFSETOF__InterpThreadContext__pStackPointer] // HERE
+    ldr r7, [r5, #OFFSETOF__InterpThreadContext__pStackPointer]
     ldr r5, [r6], 4 // InterpThreadContext
 
     EPILOG_POP "{r0-r3}"
     CHECK_STACK_ALIGNMENT
     blx r5
 
+    // Fill in the ContinuationContext register
+    ldr r2, [sp, #__PWTB_ArgumentRegisters + 8]
+
     EPILOG_WITH_TRANSITION_BLOCK_RETURN
 
 NESTED_END InterpreterStub, _TEXT
@@ -1368,6 +1383,35 @@ NESTED_ENTRY InterpreterStubRetBuffR1, _TEXT, NoHandler
     EPILOG_POP {pc}
 NESTED_END InterpreterStubRetBuffR1, _TEXT
 
+LEAF_ENTRY Store_Stack
+    ldr r9, [r6], #4 // SP offset
+    ldr r8, [r6], #4 // number of stack slots
+    add r9, sp, r9
+    add r9, r9, #__PWTB_TransitionBlock + SIZEOF__TransitionBlock
+LOCAL_LABEL(StoreCopyLoop):
+    ldr r5, [r9], #4
+    str r5, [r7], #8
+    subs r8, r8, #4
+    bne LOCAL_LABEL(StoreCopyLoop)
+    ldr r5, [r6], #4
+    EPILOG_BRANCH_REG r5
+LEAF_END Store_Stack
+
+LEAF_ENTRY Store_Stack_4B
+    ldr r9, [r6], #4 // SP offset
+    ldr r8, [r6], #4 // number of stack slots
+    add r9, sp, r9
+    add r9, r9, #__PWTB_TransitionBlock + SIZEOF__TransitionBlock
+LOCAL_LABEL(StoreCopyLoop_4B):
+    ldr r5, [r9], #4
+    str r5, [r7], #4
+    subs r8, r8, #4
+    bne LOCAL_LABEL(StoreCopyLoop_4B)
+    ldr r5, [r6], #4
+    EPILOG_BRANCH_REG r5
+LEAF_END Store_Stack_4B
+
+
 // ------------------------------------------------------------------
 // Create a real TransitionBlock and call CallInterpreterFuncletWorker
 // to execute an interpreter funclet (catch/finally/filter handler).
@@ -1396,9 +1440,9 @@ NESTED_ENTRY CallInterpreterFunclet, _TEXT, NoHandler
     str         r12, [sp, #4]
 
     // Load isFilter from original stack location and store as 5th param (1st stack arg)
-    // After PROLOG_WITH_TRANSITION_BLOCK, original stack args are at __PWTB_TransitionBlock offset
+    // After PROLOG_WITH_TRANSITION_BLOCK, original stack args are past the ArgumentRegisters
     // The 5th param (isFilter) was pushed before our stack allocation
-    ldr         r12, [sp, #8 + __PWTB_TransitionBlock + SIZEOF__ArgumentRegisters]
+    ldr         r12, [sp, #8 + __PWTB_ArgumentRegisters + SIZEOF__ArgumentRegisters]
     str         r12, [sp, #0]
 
     // r0-r3 remain unchanged

src/coreclr/vm/arm/cgencpu.h

@@ -63,6 +63,7 @@ EXTERN_C void checkStack(void);
 #define COMMETHOD_PREPAD                        12   // # extra bytes to allocate in addition to sizeof(ComCallMethodDesc)
 
 #define STACK_ALIGN_SIZE                        4
+#define CALL_STACK_ALIGN_SIZE                   8
 
 #define JUMP_ALLOCATE_SIZE                      8   // # bytes to allocate for a jump instruction
 #define BACK_TO_BACK_JUMP_ALLOCATE_SIZE         8   // # bytes to allocate for a back to back jump instruction