Make ByRefWriteBarrier to use the same call convention as in standard GC for R2R compatibility. (#56)

src/coreclr/jit/targetamd64.h

@@ -204,12 +204,11 @@
   // Registers no longer containing GC pointers after CORINFO_HELP_ASSIGN_REF and CORINFO_HELP_CHECKED_ASSIGN_REF.
   #define RBM_CALLEE_GCTRASH_WRITEBARRIER       RBM_CALLEE_TRASH_NOGC
 
-  // TODO: Satori make more precise?
-  // Registers killed by CORINFO_HELP_ASSIGN_BYREF.
-  #define RBM_CALLEE_TRASH_WRITEBARRIER_BYREF   (RBM_RSI | RBM_RDI | RBM_CALLEE_TRASH_NOGC)
-
   // Registers no longer containing GC pointers after CORINFO_HELP_ASSIGN_BYREF.
-  #define RBM_CALLEE_GCTRASH_WRITEBARRIER_BYREF (RBM_CALLEE_TRASH_NOGC & ~(RBM_RDI | RBM_RSI))
+  #define RBM_CALLEE_GCTRASH_WRITEBARRIER_BYREF (RBM_RAX | RBM_RCX)
+
+  // Registers killed by CORINFO_HELP_ASSIGN_BYREF.
+  #define RBM_CALLEE_TRASH_WRITEBARRIER_BYREF   (RBM_RSI | RBM_RDI | RBM_CALLEE_GCTRASH_WRITEBARRIER_BYREF)
 
   // We have two register classifications
   // * callee trash: aka     volatile or caller saved

src/coreclr/nativeaot/Runtime/amd64/WriteBarriers.S

@@ -400,7 +400,7 @@ LEAF_ENTRY RhpAssignRef, _TEXT
 
     JustAssign:
 ALTERNATE_ENTRY RhpAssignRefAVLocationNotHeap
-       mov     [rdi], rsi              // no card marking, src is not a heap object
+        mov     [rdi], rsi              // no card marking, src is not a heap object
         // set rdi, rsi per contract with JIT_ByRefWriteBarrier
         add     rdi, 8
         mov     rsi, r10
@@ -515,9 +515,17 @@ ALTERNATE_ENTRY RhpAssignRefAVLocation
         push rdx
         push r10
 
+        // also save xmm0, in case it is used for stack clearing, as JIT_ByRefWriteBarrier should not trash xmm0
+        // Hopefully EscapeFn cannot corrupt other xmm regs, since there is no float math or vectorizable code in there.
+        sub     rsp, 16
+        movdqu  [rsp], xmm0
+
         // void SatoriRegion::EscapeFn(SatoriObject** dst, SatoriObject* src, SatoriRegion* region)
         call    qword ptr [rdx + 8]
 
+        movdqu  xmm0, [rsp]
+        add     rsp, 16
+
         pop     r10
         pop     rdx
         pop     rsi
@@ -529,31 +537,48 @@ LEAF_END RhpAssignRef, _TEXT
 //
 // RhpByRefAssignRef simulates movs instruction for object references.
 //
-// On entry:
+// Entry:
-//      rdi: address of ref-field (assigned to)
+//   RDI - address of ref-field (assigned to)
-//      rsi: address of the data (source)
+//   RSI - address of the data  (source)
-//
+// Exit:
-// On exit:
+//   RCX is trashed
-//      rdi, rsi are incremented by 8,
+//   RAX is trashed
-//      rdi, rdx, r9, r10, r11: trashed
+//   RDI, RSI are incremented by SIZEOF(LPVOID)
-//
 LEAF_ENTRY RhpByRefAssignRef, _TEXT
-        lea     r10, [rsi + 8]
+     // See if dst is in GCHeap
+        mov     rax, [C_VAR(g_card_bundle_table)] // fetch the page byte map
+        mov     rcx,  rdi
+        shr     rcx,  30                    // dst page index
+        cmp     byte ptr [rax + rcx], 0
+        jne     InHeap
+
 ALTERNATE_ENTRY RhpByRefAssignRefAVLocation1
+        mov     rcx, [rsi]
+ALTERNATE_ENTRY RhpByRefAssignRefAVLocation2
+        mov     [rdi], rcx
+        add     rdi, 8
+        add     rsi, 8
+        ret
+
+    InHeap:
+        // JIT_WriteBarrier may trash these registers 
+        push    rdx
+        push    r8
+        push    r9
+        push    r10
+        push    r11
+
+        // save preadjusted rsi
+        lea     r10, [rsi + 8]
         mov     rsi, [rsi]
 
-        // See if dst is in GCHeap
+        call    CheckedEntry
-        mov     rax, [C_VAR(g_card_bundle_table)] // fetch the page byte map
-        mov     r8,  rdi
-        shr     r8,  30                           // dst page index
-        cmp     byte ptr [rax + r8], 0
-        jne     C_FUNC(CheckedEntry)
 
-    NotInHeap_RhpByRefAssignRef:
+        pop     r11
-ALTERNATE_ENTRY RhpByRefAssignRefAVLocation2
+        pop     r10
-        mov     [rdi], rsi
+        pop     r9
-        add     rdi, 8
+        pop     r8
-        mov     rsi, r10
+        pop     rdx
         ret
 LEAF_END RhpByRefAssignRef, _TEXT
 

src/coreclr/nativeaot/Runtime/amd64/WriteBarriers.asm

@@ -508,17 +508,28 @@ ALTERNATE_ENTRY RhpAssignRefAVLocation
         mov  r9, rsp
         and  rsp, -16
 
-        ; save rsp, rcx, rdx, r8 and have enough stack for the callee
+        ; save rsp, rcx, rdx, r8
         push r9
         push rcx
         push rdx
         push r8
+
+        ; also save xmm0, in case it is used for stack clearing, as JIT_ByRefWriteBarrier should not trash xmm0
+        ; Hopefully EscapeFn cannot corrupt other xmm regs, since there is no float math or vectorizable code in there.
+        sub     rsp, 16
+        movdqu  [rsp], xmm0
+
+        ; shadow space
         sub  rsp, 20h
 
         ; void SatoriRegion::EscapeFn(SatoriObject** dst, SatoriObject* src, SatoriRegion* region)
         call    qword ptr [r8 + 8]
 
         add     rsp, 20h
+
+        movdqu  xmm0, [rsp]
+        add     rsp, 16
+
         pop     r8
         pop     rdx
         pop     rcx
@@ -528,33 +539,51 @@ LEAF_END RhpAssignRef, _TEXT
 
 ;;
 ;; RhpByRefAssignRef simulates movs instruction for object references.
-;;
+;; Entry:
-;; On entry:
+;;   RDI - address of ref-field (assigned to)
-;;      rdi: address of ref-field (assigned to)
+;;   RSI - address of the data  (source)
-;;      rsi: address of the data (source)
+;; Exit:
-;;
+;;   RCX is trashed
-;; On exit:
+;;   RAX is trashed
-;;      rdi, rsi are incremented by 8,
+;;   RDI, RSI are incremented by SIZEOF(LPVOID)
-;;      rcx, r8, r9, r11: trashed
-;;
 LEAF_ENTRY RhpByRefAssignRef, _TEXT
-    mov     rcx, rdi
-ALTERNATE_ENTRY RhpByRefAssignRefAVLocation1
-    mov     rdx, [rsi]
-    add     rdi, 8h
-    add     rsi, 8h
-
     ; See if dst is in GCHeap
-    mov     rax, [g_card_bundle_table] ; fetch the page byte map
+        mov     rax, [g_card_bundle_table]  ; fetch the page byte map
-    mov     r8,  rcx
+        mov     rcx,  rdi
-    shr     r8,  30                    ; dst page index
+        shr     rcx,  30                    ; dst page index
-    cmp     byte ptr [rax + r8], 0
+        cmp     byte ptr [rax + rcx], 0
-    jne     CheckedEntry
+        jne     InHeap
 
-    NotInHeap:
+ALTERNATE_ENTRY RhpByRefAssignRefAVLocation1
+        mov     rcx, [rsi]
 ALTERNATE_ENTRY RhpByRefAssignRefAVLocation2
-    mov     [rcx], rdx
+        mov     [rdi], rcx
-    ret
+        add     rdi, 8h
+        add     rsi, 8h
+        ret
+
+    InHeap:
+
+        ; JIT_WriteBarrier may trash these registers 
+        push    rdx
+        push    r8
+        push    r9
+        push    r10
+        push    r11
+
+        mov     rcx, rdi
+        mov     rdx, [rsi]
+        add     rdi, 8h
+        add     rsi, 8h
+
+        call    CheckedEntry
+
+        pop     r11
+        pop     r10
+        pop     r9
+        pop     r8
+        pop     rdx
+        ret
 LEAF_END RhpByRefAssignRef, _TEXT
 
 LEAF_ENTRY RhpCheckedLockCmpXchg, _TEXT

src/coreclr/vm/amd64/patchedcode.S

@@ -443,9 +443,17 @@ LEAF_ENTRY JIT_WriteBarrier, _TEXT
         push rdx
         push r10
 
+        // also save xmm0, in case it is used for stack clearing, as JIT_ByRefWriteBarrier should not trash xmm0
+        // Hopefully EscapeFn cannot corrupt other xmm regs, since there is no float math or vectorizable code in there.
+        sub     rsp, 16
+        movdqu  [rsp], xmm0
+
         // void SatoriRegion::EscapeFn(SatoriObject** dst, SatoriObject* src, SatoriRegion* region)
         call    qword ptr [rdx + 8]
 
+        movdqu  xmm0, [rsp]
+        add     rsp, 16
+
         pop     r10
         pop     rdx
         pop     rsi
@@ -460,29 +468,43 @@ LEAF_END_MARKED JIT_WriteBarrier, _TEXT
 // Entry:
 //   RDI - address of ref-field (assigned to)
 //   RSI - address of the data  (source)
-//   Note: RyuJIT assumes that all volatile registers can be trashed by 
-//   the CORINFO_HELP_ASSIGN_BYREF helper (i.e. JIT_ByRefWriteBarrier)
-//   except RDI and RSI. This helper uses and defines RDI and RSI, so
-//   they remain as live GC refs or byrefs, and are not killed.
 // Exit:
+//   RCX is trashed
+//   RAX is trashed
 //   RDI, RSI are incremented by SIZEOF(LPVOID)
 LEAF_ENTRY JIT_ByRefWriteBarrier, _TEXT
+     // See if dst is in GCHeap
+        PREPARE_EXTERNAL_VAR g_card_bundle_table, rax   // fetch the page byte map
+        mov     rcx,  rdi
+        shr     rcx,  30                    // dst page index
+        cmp     byte ptr [rax + rcx], 0
+        jne     InHeap
+
+        mov     rcx, [rsi]
+        mov     [rdi], rcx
+        add     rdi, 8
+        add     rsi, 8
+        ret
+
+    InHeap:
+        // JIT_WriteBarrier may trash these registers 
+        push    rdx
+        push    r8
+        push    r9
+        push    r10
+        push    r11
+
+        // save preadjusted rsi
         lea     r10, [rsi + 8]
         mov     rsi, [rsi]
 
-     // See if dst is in GCHeap
+        call    CheckedEntry
-        PREPARE_EXTERNAL_VAR g_card_bundle_table, rax   // fetch the page byte map
-        mov     rax, [rax]
 
-        mov     r8, rdi
+        pop     r11
-        shr     r8, 30                                  // dst page index
+        pop     r10
-        cmp     byte ptr [rax + r8], 0
+        pop     r9
-        jne     C_FUNC(CheckedEntry)
+        pop     r8
-
+        pop     rdx
-    NotInHeap_ByRefWriteBarrier:
-        mov     [rdi], rsi
-        add     rdi, 8
-        mov     rsi, r10
         ret
 LEAF_END_MARKED JIT_ByRefWriteBarrier, _TEXT
 

src/coreclr/vm/amd64/patchedcode.asm

@@ -376,17 +376,28 @@ endif
         mov  r9, rsp
         and  rsp, -16
 
-        ; save rsp, rcx, rdx, r8 and have enough stack for the callee
+        ; save rsp, rcx, rdx, r8
         push r9
         push rcx
         push rdx
         push r8
+
+        ; also save xmm0, in case it is used for stack clearing, as JIT_ByRefWriteBarrier should not trash xmm0
+        ; Hopefully EscapeFn cannot corrupt other xmm regs, since there is no float math or vectorizable code in there.
+        sub     rsp, 16
+        movdqu  [rsp], xmm0
+
+        ; shadow space
         sub  rsp, 20h
 
         ; void SatoriRegion::EscapeFn(SatoriObject** dst, SatoriObject* src, SatoriRegion* region)
         call    qword ptr [r8 + 8]
 
         add     rsp, 20h
+
+        movdqu  xmm0, [rsp]
+        add     rsp, 16
+
         pop     r8
         pop     rdx
         pop     rcx
@@ -399,27 +410,45 @@ LEAF_END_MARKED JIT_WriteBarrier, _TEXT
 ; Entry:
 ;   RDI - address of ref-field (assigned to)
 ;   RSI - address of the data  (source)
-;   Note: RyuJIT assumes that all volatile registers can be trashed by 
-;   the CORINFO_HELP_ASSIGN_BYREF helper (i.e. JIT_ByRefWriteBarrier)
-;   except RDI and RSI. This helper uses and defines RDI and RSI, so
-;   they remain as live GC refs or byrefs, and are not killed.
 ; Exit:
+;   RCX is trashed
+;   RAX is trashed
 ;   RDI, RSI are incremented by SIZEOF(LPVOID)
 LEAF_ENTRY JIT_ByRefWriteBarrier, _TEXT
+    ; See if dst is in GCHeap
+        mov     rax, [g_card_bundle_table]  ; fetch the page byte map
+        mov     rcx,  rdi
+        shr     rcx,  30                    ; dst page index
+        cmp     byte ptr [rax + rcx], 0
+        jne     InHeap
+
+        mov     rcx, [rsi]
+        mov     [rdi], rcx
+        add     rdi, 8h
+        add     rsi, 8h
+        ret
+
+    InHeap:
+
+        ; JIT_WriteBarrier may trash these registers 
+        push    rdx
+        push    r8
+        push    r9
+        push    r10
+        push    r11
+
         mov     rcx, rdi
         mov     rdx, [rsi]
         add     rdi, 8h
         add     rsi, 8h
 
-    ; See if dst is in GCHeap
+        call    CheckedEntry
-        mov     rax, [g_card_bundle_table] ; fetch the page byte map
-        mov     r8,  rcx
-        shr     r8,  30                    ; dst page index
-        cmp     byte ptr [rax + r8], 0
-        jne     CheckedEntry
 
-    NotInHeap:
+        pop     r11
-        mov     [rcx], rdx
+        pop     r10
+        pop     r9
+        pop     r8
+        pop     rdx
         ret
 LEAF_END_MARKED JIT_ByRefWriteBarrier, _TEXT