Changeset 7962

Timestamp:

Dec 27, 2007, 4:31:01 PM (17 years ago)

Author:

R. Matthew Emerson

Message:

Add save[0-3] fields to the TCR. These are a caller-saved spill area
for node registers. A function should clear out these fields when its
done with them so that they don't make the GC hang on to stuff that's
otherwise garbage.

As the comment says, the save0 word must be aligned on an 16 byte boundary.
This is so we can use (movapd (% fpzero) (@ (% :rcontext) tcr.save0)) to
clear the entire area quickly.

One possible wrinkle: although malloc/calloc will return suitably
aligned tcr structs, on systems with TLS, a mechanism like
attribute ((aligned (16))) or something might be needed. (Or we
could just use movupd.)

Define next-method-context and a couple of storage layouts. Define
constants used when looking for (movl ($ imm32) (% fn)) at tagged
return addresses.

File:

• branches/ia32/compiler/X86/X8632/x8632-arch.lisp (modified) (4 diffs)

branches/ia32/compiler/X86/X8632/x8632-arch.lisp

@@ -170 +170 @@
 (defx86reg xfn temp1)
 
+(defx86reg next-method-context temp0)
+
 ;;; This follows the ppc32 scheme pretty closely.
 
@@ -489 +491 @@
 (define-fixedsized-object value-cell
   value)
+
+(define-storage-layout lisp-frame 0
+  backptr
+  return-address
+  xtra)
+
+(define-storage-layout xcf 0            ;"exception callback frame"
+  backptr
+  return-address                        ; always 0
+  nominal-function
+  relative-pc
+  containing-object
+  xp
+  ra0
+  )
 
 ;;; The kernel uses these (rather generically named) structures
@@ -534 +551 @@
   node-regs-mask                        ; bit set means corresponding reg contains node
   linear
+  ;; save0 *must* be aligned on a 16-byte boundary!
+  save0                                 ;spill area for node registers
+  save1                                 ; (caller saved)
+  save2                                 ; probably saved/restored in
+  save3                                 ; callout/trap handlers
   save-ebp                              ; lisp frame ptr for foreign code
   lisp-mxcsr
@@ -1148 +1170 @@
       ,@body)))
 
+;;; When found at a tagged return address, the instruction
+;;; (movl ($ imm32) (% fn))
+;;; lets the runtime easily map a return address to the containing
+;;; function.
+;;;
+;;; The notation ($ :self) is used in the assembler to mean "a 32-bit
+;;; immediate whose offset will be remembered in a table at the end of
+;;; the function object."
+;;;
+;;; Before the function is made executable (or when the GC moves the
+;;; function), these :self immediates are filled in with the actual
+;;; address of the function.
+
+(defconstant recover-fn-opcode-byte #b10111111) ;when %fn is %edi
+(defconstant recover-fn-address-offset 1)
+
 (provide "X8632-ARCH")