Merge pull request #19344 from adityamandaleeka/aa64_vm_callconv_changes

CallDescr/TransitionBlock changes for ARM64 calling convention

7 files changed, 51 insertions, 35 deletions

diff --git a/src/vm/arm64/CallDescrWorkerARM64.asm b/src/vm/arm64/CallDescrWorkerARM64.asm
index b9f8a6090d..65c7db6f3f 100644
--- a/src/vm/arm64/CallDescrWorkerARM64.asm
+++ b/src/vm/arm64/CallDescrWorkerARM64.asm
@@ -62,15 +62,18 @@ Ldonestack
         ldp     d6, d7, [x9, #48]
 LNoFloatingPoint
 
-        ;; Copy [pArgumentRegisters, ..., pArgumentRegisters + 64]
-        ;; into x0, ..., x7, x8
+        ;; Copy [pArgumentRegisters, ..., pArgumentRegisters + 56]
+        ;; into x0, ..., x7
 
         ldr     x9, [x19,#CallDescrData__pArgumentRegisters]
         ldp     x0, x1, [x9]
         ldp     x2, x3, [x9, #16]
         ldp     x4, x5, [x9, #32]
         ldp     x6, x7, [x9, #48]
-        ldr     x8, [x9, #64]
+
+        ;; Copy pRetBuffArg into x8
+        ldr     x9, [x19,#CallDescrData__pRetBuffArg]
+        ldr     x8, [x9]
 
         ;; call pTarget
         ldr     x9, [x19,#CallDescrData__pTarget]
diff --git a/src/vm/arm64/asmconstants.h b/src/vm/arm64/asmconstants.h
index dca845d000..262e481b53 100644
--- a/src/vm/arm64/asmconstants.h
+++ b/src/vm/arm64/asmconstants.h
@@ -55,19 +55,20 @@ ASMCONSTANTS_C_ASSERT(AppDomain__m_dwId == offsetof(AppDomain, m_dwId));
 
 #define METHODDESC_REGISTER            x12
 
-#define SIZEOF__ArgumentRegisters 0x48
+#define SIZEOF__ArgumentRegisters 0x40
 ASMCONSTANTS_C_ASSERT(SIZEOF__ArgumentRegisters == sizeof(ArgumentRegisters))
 
 #define SIZEOF__FloatArgumentRegisters 0x40
 ASMCONSTANTS_C_ASSERT(SIZEOF__FloatArgumentRegisters == sizeof(FloatArgumentRegisters))
 
-#define CallDescrData__pSrc                0x00
-#define CallDescrData__numStackSlots       0x08
-#define CallDescrData__pArgumentRegisters  0x10
-#define CallDescrData__pFloatArgumentRegisters 0x18
-#define CallDescrData__fpReturnSize        0x20
-#define CallDescrData__pTarget             0x28
-#define CallDescrData__returnValue         0x30
+#define CallDescrData__pSrc                     0x00
+#define CallDescrData__numStackSlots            0x08
+#define CallDescrData__pArgumentRegisters       0x10
+#define CallDescrData__pFloatArgumentRegisters  0x18
+#define CallDescrData__fpReturnSize             0x20
+#define CallDescrData__pTarget                  0x28
+#define CallDescrData__pRetBuffArg              0x30
+#define CallDescrData__returnValue              0x38
 
 ASMCONSTANTS_C_ASSERT(CallDescrData__pSrc                 == offsetof(CallDescrData, pSrc))
 ASMCONSTANTS_C_ASSERT(CallDescrData__numStackSlots        == offsetof(CallDescrData, numStackSlots))
@@ -75,6 +76,7 @@ ASMCONSTANTS_C_ASSERT(CallDescrData__pArgumentRegisters   == offsetof(CallDescrD
 ASMCONSTANTS_C_ASSERT(CallDescrData__pFloatArgumentRegisters == offsetof(CallDescrData, pFloatArgumentRegisters))
 ASMCONSTANTS_C_ASSERT(CallDescrData__fpReturnSize         == offsetof(CallDescrData, fpReturnSize))
 ASMCONSTANTS_C_ASSERT(CallDescrData__pTarget              == offsetof(CallDescrData, pTarget))
+ASMCONSTANTS_C_ASSERT(CallDescrData__pRetBuffArg          == offsetof(CallDescrData, pRetBuffArg))
 ASMCONSTANTS_C_ASSERT(CallDescrData__returnValue          == offsetof(CallDescrData, returnValue))
 
 #define                  CORINFO_NullReferenceException_ASM 0
diff --git a/src/vm/arm64/asmhelpers.S b/src/vm/arm64/asmhelpers.S
index 23f591fccc..e138feb402 100644
--- a/src/vm/arm64/asmhelpers.S
+++ b/src/vm/arm64/asmhelpers.S
@@ -562,7 +562,7 @@ LEAF_END SinglecastDelegateInvokeStub, _TEXT
 #ifdef FEATURE_COMINTEROP
 
 #define ComCallPreStub_FrameSize (SIZEOF__GSCookie + SIZEOF__ComMethodFrame)
-#define ComCallPreStub_FirstStackAdjust  (SIZEOF__ArgumentRegisters + 2 * 8) // reg args , fp & lr already pushed
+#define ComCallPreStub_FirstStackAdjust  (8 + SIZEOF__ArgumentRegisters + 2 * 8) // x8, reg args , fp & lr already pushed
 #define ComCallPreStub_StackAlloc0        (ComCallPreStub_FrameSize - ComCallPreStub_FirstStackAdjust)
 #define ComCallPreStub_StackAlloc1        (ComCallPreStub_StackAlloc0 + SIZEOF__FloatArgumentRegisters + 8)// 8 for ErrorReturn
 #define ComCallPreStub_StackAlloc    (ComCallPreStub_StackAlloc1 + (ComCallPreStub_StackAlloc1 & 8))
@@ -692,7 +692,7 @@ COMToCLRDispatchHelper_RegSetup
     ldp x2, x3, [x1, #(SIZEOF__ComMethodFrame - SIZEOF__ArgumentRegisters + 16)]
     ldp x4, x5, [x1, #(SIZEOF__ComMethodFrame - SIZEOF__ArgumentRegisters + 32)]
     ldp x6, x7, [x1, #(SIZEOF__ComMethodFrame - SIZEOF__ArgumentRegisters + 48)]
-    ldr x8, [x1, #(SIZEOF__ComMethodFrame - SIZEOF__ArgumentRegisters + 64)]
+    ldr x8, [x1, #(SIZEOF__ComMethodFrame - SIZEOF__ArgumentRegisters - 8)]
 
     ldr x1, [x1, #(SIZEOF__ComMethodFrame - SIZEOF__ArgumentRegisters + 8)]
 
@@ -1253,7 +1253,7 @@ DelayLoad_Helper\suffix:
     mov x4, \frameFlags
     bl DynamicHelperWorker
     cbnz x0, LOCAL_LABEL(FakeProlog\suffix\()_0)
-    ldr x0, [sp, #__PWTB_ArgumentRegisters]
+    ldr x0, [sp, #__PWTB_ArgumentRegister_FirstArg]
     EPILOG_WITH_TRANSITION_BLOCK_RETURN
 LOCAL_LABEL(FakeProlog\suffix\()_0):
     mov x12, x0
diff --git a/src/vm/arm64/asmhelpers.asm b/src/vm/arm64/asmhelpers.asm
index 376bbbcec7..7346c468a0 100644
--- a/src/vm/arm64/asmhelpers.asm
+++ b/src/vm/arm64/asmhelpers.asm
@@ -533,7 +533,7 @@ LNullThis
     GBLA ComCallPreStub_FirstStackAdjust
 
 ComCallPreStub_FrameSize         SETA (SIZEOF__GSCookie + SIZEOF__ComMethodFrame)
-ComCallPreStub_FirstStackAdjust  SETA (SIZEOF__ArgumentRegisters + 2 * 8) ; reg args , fp & lr already pushed
+ComCallPreStub_FirstStackAdjust  SETA (8 + SIZEOF__ArgumentRegisters + 2 * 8) ; x8, reg args , fp & lr already pushed
 ComCallPreStub_StackAlloc        SETA ComCallPreStub_FrameSize - ComCallPreStub_FirstStackAdjust 
 ComCallPreStub_StackAlloc        SETA ComCallPreStub_StackAlloc + SIZEOF__FloatArgumentRegisters + 8; 8 for ErrorReturn
     IF ComCallPreStub_StackAlloc:MOD:16 != 0
@@ -603,7 +603,7 @@ ComCallPreStub_ErrorExit
     GBLA GenericComCallStub_FirstStackAdjust
 
 GenericComCallStub_FrameSize         SETA (SIZEOF__GSCookie + SIZEOF__ComMethodFrame)
-GenericComCallStub_FirstStackAdjust  SETA (SIZEOF__ArgumentRegisters + 2 * 8)
+GenericComCallStub_FirstStackAdjust  SETA (8 + SIZEOF__ArgumentRegisters + 2 * 8)
 GenericComCallStub_StackAlloc        SETA GenericComCallStub_FrameSize - GenericComCallStub_FirstStackAdjust
 GenericComCallStub_StackAlloc        SETA GenericComCallStub_StackAlloc + SIZEOF__FloatArgumentRegisters
 
@@ -676,7 +676,7 @@ COMToCLRDispatchHelper_RegSetup
     ldp x2, x3, [x1, #(SIZEOF__ComMethodFrame - SIZEOF__ArgumentRegisters + 16)]
     ldp x4, x5, [x1, #(SIZEOF__ComMethodFrame - SIZEOF__ArgumentRegisters + 32)]
     ldp x6, x7, [x1, #(SIZEOF__ComMethodFrame - SIZEOF__ArgumentRegisters + 48)]
-    ldr x8, [x1, #(SIZEOF__ComMethodFrame - SIZEOF__ArgumentRegisters + 64)]
+    ldr x8, [x1, #(SIZEOF__ComMethodFrame - SIZEOF__ArgumentRegisters - 8)]
 
     ldr x1, [x1, #(SIZEOF__ComMethodFrame - SIZEOF__ArgumentRegisters + 8)]
 
@@ -1262,9 +1262,9 @@ Fail
         mov x4, $frameFlags
         bl DynamicHelperWorker
         cbnz x0, %FT0
-        ldr x0, [sp, #__PWTB_ArgumentRegisters]
+        ldr x0, [sp, #__PWTB_ArgumentRegister_FirstArg]
         EPILOG_WITH_TRANSITION_BLOCK_RETURN
-0		
+0
         mov x12, x0
         EPILOG_WITH_TRANSITION_BLOCK_TAILCALL
         EPILOG_BRANCH_REG  x12
diff --git a/src/vm/arm64/asmmacros.h b/src/vm/arm64/asmmacros.h
index 74613daa8f..291fcf8e70 100644
--- a/src/vm/arm64/asmmacros.h
+++ b/src/vm/arm64/asmmacros.h
@@ -44,7 +44,9 @@ $name   SETS    "|$symbol|"
         PROLOG_WITH_TRANSITION_BLOCK $extraLocals, $SaveFPArgs
 
         GBLA __PWTB_FloatArgumentRegisters
-        GBLA __PWTB_ArgumentRegisters 
+        GBLA __PWTB_ArgumentRegisters
+        GBLA __PWTB_ArgumentRegister_FirstArg ; We save the x8 register ahead of the first argument, so this
+                                              ; is different from the start of the argument register save area.
         GBLA __PWTB_StackAlloc
         GBLA __PWTB_TransitionBlock
         GBLL __PWTB_SaveFPArgs
@@ -72,7 +74,8 @@ __PWTB_TransitionBlock SETA __PWTB_FloatArgumentRegisters
         ENDIF
 
 __PWTB_StackAlloc SETA __PWTB_TransitionBlock
-__PWTB_ArgumentRegisters SETA __PWTB_StackAlloc + 96 
+__PWTB_ArgumentRegisters SETA __PWTB_StackAlloc + 104
+__PWTB_ArgumentRegister_FirstArg SETA __PWTB_ArgumentRegisters + 8
 
         PROLOG_SAVE_REG_PAIR   fp, lr, #-176!
         ; Spill callee saved registers 
@@ -172,11 +175,12 @@ __PWTB_SAVE_ARGUMENT_REGISTERS_OFFSET SETA $offset
 __PWTB_SAVE_ARGUMENT_REGISTERS_OFFSET SETA 0
        ENDIF
 
-        stp                    x0, x1, [$reg, #(__PWTB_SAVE_ARGUMENT_REGISTERS_OFFSET)]
-        stp                    x2, x3, [$reg, #(__PWTB_SAVE_ARGUMENT_REGISTERS_OFFSET + 16)]
-        stp                    x4, x5, [$reg, #(__PWTB_SAVE_ARGUMENT_REGISTERS_OFFSET + 32)]
-        stp                    x6, x7, [$reg, #(__PWTB_SAVE_ARGUMENT_REGISTERS_OFFSET + 48)]
-        str                    x8, [$reg, #(__PWTB_SAVE_ARGUMENT_REGISTERS_OFFSET + 64)]
+        str                    x8, [$reg, #(__PWTB_SAVE_ARGUMENT_REGISTERS_OFFSET)]
+        stp                    x0, x1, [$reg, #(__PWTB_SAVE_ARGUMENT_REGISTERS_OFFSET + 8)]
+        stp                    x2, x3, [$reg, #(__PWTB_SAVE_ARGUMENT_REGISTERS_OFFSET + 24)]
+        stp                    x4, x5, [$reg, #(__PWTB_SAVE_ARGUMENT_REGISTERS_OFFSET + 40)]
+        stp                    x6, x7, [$reg, #(__PWTB_SAVE_ARGUMENT_REGISTERS_OFFSET + 56)]
+
     MEND
 
 ; Reserve 64 bytes of memory before calling  SAVE_FLOAT_ARGUMENT_REGISTERS
@@ -208,11 +212,12 @@ __PWTB_RESTORE_ARGUMENT_REGISTERS_OFFSET SETA $offset
 __PWTB_RESTORE_ARGUMENT_REGISTERS_OFFSET SETA 0
        ENDIF
 
-        ldp                    x0, x1, [$reg, #(__PWTB_RESTORE_ARGUMENT_REGISTERS_OFFSET)]
-        ldp                    x2, x3, [$reg, #(__PWTB_RESTORE_ARGUMENT_REGISTERS_OFFSET + 16)]
-        ldp                    x4, x5, [$reg, #(__PWTB_RESTORE_ARGUMENT_REGISTERS_OFFSET + 32)]
-        ldp                    x6, x7, [$reg, #(__PWTB_RESTORE_ARGUMENT_REGISTERS_OFFSET + 48)]
-        ldr                    x8, [$reg, #(__PWTB_RESTORE_ARGUMENT_REGISTERS_OFFSET + 64)]
+        ldr                    x8, [$reg, #(__PWTB_RESTORE_ARGUMENT_REGISTERS_OFFSET)]
+        ldp                    x0, x1, [$reg, #(__PWTB_RESTORE_ARGUMENT_REGISTERS_OFFSET + 8)]
+        ldp                    x2, x3, [$reg, #(__PWTB_RESTORE_ARGUMENT_REGISTERS_OFFSET + 24)]
+        ldp                    x4, x5, [$reg, #(__PWTB_RESTORE_ARGUMENT_REGISTERS_OFFSET + 40)]
+        ldp                    x6, x7, [$reg, #(__PWTB_RESTORE_ARGUMENT_REGISTERS_OFFSET + 56)]
+
     MEND
 
     MACRO
diff --git a/src/vm/arm64/calldescrworkerarm64.S b/src/vm/arm64/calldescrworkerarm64.S
index 5bcad2f9c0..c3ce06aa72 100644
--- a/src/vm/arm64/calldescrworkerarm64.S
+++ b/src/vm/arm64/calldescrworkerarm64.S
@@ -62,7 +62,10 @@ LOCAL_LABEL(NoFloatingPoint):
     ldp     x2, x3, [x9, #16]
     ldp     x4, x5, [x9, #32]
     ldp     x6, x7, [x9, #48]
-    ldr     x8, [x9, #64]
+
+    // Copy pRetBuffArg into x8
+    ldr     x9, [x19,#CallDescrData__pRetBuffArg]
+    ldr     x8, [x9]
 
     // call pTarget
     ldr     x9, [x19,#CallDescrData__pTarget]
diff --git a/src/vm/arm64/cgencpu.h b/src/vm/arm64/cgencpu.h
index 6edf704b8f..a297a84e17 100644
--- a/src/vm/arm64/cgencpu.h
+++ b/src/vm/arm64/cgencpu.h
@@ -57,6 +57,7 @@ extern PCODE GetPreStubEntryPoint();
 
 #define CALLDESCR_ARGREGS                       1   // CallDescrWorker has ArgumentRegister parameter
 #define CALLDESCR_FPARGREGS                     1   // CallDescrWorker has FloatArgumentRegisters parameter
+#define CALLDESCR_RETBUFFARGREG                 1   // CallDescrWorker has RetBuffArg parameter that's separate from arg regs
 
 // Given a return address retrieved during stackwalk,
 // this is the offset by which it should be decremented to arrive at the callsite.
@@ -77,7 +78,9 @@ extern PCODE GetPreStubEntryPoint();
 typedef INT64 StackElemType;
 #define STACK_ELEM_SIZE sizeof(StackElemType)
 
-// !! This expression assumes STACK_ELEM_SIZE is a power of 2.
+// The expression below assumes STACK_ELEM_SIZE is a power of 2, so check that.
+static_assert(((STACK_ELEM_SIZE & (STACK_ELEM_SIZE-1)) == 0), "STACK_ELEM_SIZE must be a power of 2");
+
 #define StackElemSize(parmSize) (((parmSize) + STACK_ELEM_SIZE - 1) & ~((ULONG)(STACK_ELEM_SIZE - 1)))
 
 //
@@ -115,11 +118,11 @@ struct CalleeSavedRegisters {
 // will probably have to communicate this back to the PromoteCallerStack
 // routine to avoid a double promotion.
 //--------------------------------------------------------------------
+#define NUM_ARGUMENT_REGISTERS 8
 typedef DPTR(struct ArgumentRegisters) PTR_ArgumentRegisters;
 struct ArgumentRegisters {
-    INT64 x[9]; // x0 ....x7 & x8 can contain return buffer address
+    INT64 x[NUM_ARGUMENT_REGISTERS]; // x0 ....x7. Note that x8 (return buffer address) is not included.
 };
-#define NUM_ARGUMENT_REGISTERS 9
 
 #define ARGUMENTREGISTERS_SIZE sizeof(ArgumentRegisters)