Changeset 13937 for branches

Timestamp:

Jul 7, 2010, 12:59:31 PM (14 years ago)

Author:

R. Matthew Emerson

Message:

New scheme for dealing with fp exceptions in foreign code.

Instead of masking fp exceptions around FF calls, leave them enabled.
handle_exception() and handle_floating_point_exception() are now
prepared to deal with them by preserving the currently masked exceptions,
masking them all, and then resuming execution (after setting a flag
to make note of this). At the end of .SPffcall, if the flag is set
the fp status is saved, and the masked exceptions are restored
to what they were before the FF call.

This may not make a huge difference in ff call performance, but
unconditionally saving and restoring the MXCSR around every call
is probably not free (according to one source I consulted, ldmxcsr
takes 42 clock cycles; stmxcsr 19).

Location:

branches/rme-fpe/lisp-kernel

Files:

• x86-exceptions.c (modified) (4 diffs)
• x86-spentry32.s (modified) (4 diffs)
• x86-spentry64.s (modified) (8 diffs)

branches/rme-fpe/lisp-kernel/x86-exceptions.c

@@ -942 +942 @@
 
 Boolean
-handle_floating_point_exception(TCR *tcr, ExceptionInformation *xp, siginfo_t *info)
+handle_floating_point_exception(TCR *tcr, ExceptionInformation *xp, siginfo_t *info, int old_valence)
 {
   int code,skip;
@@ -954 +954 @@
 #endif  
 
-  if ((fulltag_of(cmain) == fulltag_misc) &&
-      (header_subtag(header_of(cmain)) == subtag_macptr)) {
+  if (old_valence == TCR_STATE_FOREIGN) {
+    /* We got a floating point exception in foreign code.  Set a flag
+       in the TCR to indicate this, and resume with FP exceptions
+       disabled.  .SPffcall will check the flag, and if it's set,
+       will save the FP status and re-enable exceptions. */
+
+    /*
+     * We might get multiple foreign FP exceptions, so take care
+     * to preseve lisp's original MXCSR.
+     */
+    if (! (tcr->flags & (1<<TCR_FLAG_BIT_FOREIGN_FPE))) {
+      tcr->lisp_mxcsr = xpMXCSR(xp) & ~MXCSR_STATUS_MASK;
+      tcr->flags |= (1<<TCR_FLAG_BIT_FOREIGN_FPE);
+    }
+    xpMXCSR(xp) &= ~MXCSR_STATUS_MASK;
+    xpMXCSR(xp) |= MXCSR_CONTROL_MASK;
+    return true;
+  } else if ((fulltag_of(cmain) == fulltag_misc) &&
+             (header_subtag(header_of(cmain)) == subtag_macptr)) {
     xcf = create_exception_callback_frame(xp, tcr);
     skip = callback_to_lisp(tcr, cmain, xp, xcf, SIGFPE, code, 0, 0);
@@ -1078 +1095 @@
   pc program_counter = (pc)xpPC(context);
 
-  if (old_valence != TCR_STATE_LISP) {
+  if ((old_valence != TCR_STATE_LISP) && signum != SIGFPE) {
     return false;
   }
@@ -1237 +1254 @@
 #endif
 
-    return handle_floating_point_exception(tcr, context, info);
+    return handle_floating_point_exception(tcr, context, info, old_valence);
 
 #if SIGBUS != SIGNUM_FOR_INTN_TRAP

branches/rme-fpe/lisp-kernel/x86-spentry32.s

@@ -4184 +4184 @@
         __(popl rcontext(tcr.save_eflags))
         __(cld)        
-        __(stmxcsr rcontext(tcr.lisp_mxcsr))
         __(emms)
-        __(ldmxcsr rcontext(tcr.foreign_mxcsr))
         __(movl (%esp),%ebp)
 LocalLabelPrefix`'ffcall_setup:
@@ -4215 +4213 @@
         __(clr %fn)
         __(pxor %fpzero,%fpzero)
+        /* If we got a floating-point exception during the ff-call,
+           our handler will have set a flag, preserved lisp's MXCSR,
+           and resumed execution with fp exceptions masked. */
+        __(btrq $TCR_FLAG_BIT_FOREIGN_FPE,rcontext(tcr.flags))
+        __(jnc 1f)
         __(cmpb $0,C(bogus_fp_exceptions))
         __(je 0f)
@@ -4222 +4225 @@
         __ifdef(`SSE2_MATH_LIB')
         __(stmxcsr rcontext(tcr.ffi_exception))
+        __(ldmxcsr rcontext(tcr.lisp_mxcsr)) /* preserved by the handler */
         __else
         __(fnstsw rcontext(tcr.ffi_exception))
@@ -4235 +4239 @@
         __(pop %arg_y)           
         __(pop %temp1) 
-        __(ldmxcsr rcontext(tcr.lisp_mxcsr))
         __(check_pending_interrupt(%temp0))
         __(pop %temp0)

branches/rme-fpe/lisp-kernel/x86-spentry64.s

@@ -3991 +3991 @@
         __(movq $TCR_STATE_FOREIGN,rcontext(tcr.valence))
         __(movq rcontext(tcr.foreign_sp),%rsp)
-        __(stmxcsr rcontext(tcr.lisp_mxcsr))
         __(emms)
-        __(ldmxcsr rcontext(tcr.foreign_mxcsr))
         __(movq (%rsp),%rbp)
         __ifdef(`DARWIN_GS_HACK')
@@ -4070 +4068 @@
         __(clr %fn)
         __(pxor %fpzero,%fpzero)
+
+        /* If we got a floating-point exception during the ff-call,
+           our handler will have set a flag, preserved lisp's MXCSR,
+           and resumed execution with fp exceptions masked. */
+        __(btrq $TCR_FLAG_BIT_FOREIGN_FPE,rcontext(tcr.flags))
+        __(jnc 1f)
         __(cmpb $0,C(bogus_fp_exceptions)(%rip))
         __(je 0f)
@@ -4075 +4079 @@
         __(jmp 1f)
 0:      __(stmxcsr rcontext(tcr.ffi_exception))
+        __(ldmxcsr rcontext(tcr.lisp_mxcsr)) /* preserved by the handler */
 1:      __(movq rcontext(tcr.save_vsp),%rsp)
         __(movq rcontext(tcr.save_rbp),%rbp)
@@ -4090 +4095 @@
         __(pop %temp2)
         __(pop %temp1)
-        __(ldmxcsr rcontext(tcr.lisp_mxcsr))
         __(check_pending_interrupt(%temp0))
         __(pop %temp0)
@@ -4211 +4215 @@
         __(movq $TCR_STATE_FOREIGN,rcontext(tcr.valence))
         __(movq rcontext(tcr.foreign_sp),%rsp)
-        __(stmxcsr rcontext(tcr.lisp_mxcsr))
         __(emms)
-        __(ldmxcsr rcontext(tcr.foreign_mxcsr))
         __(movq (%rsp),%rbp)
         __ifdef(`DARWIN_GS_HACK')
@@ -4290 +4292 @@
         __(clr %fn)
         __(pxor %fpzero,%fpzero)
+        /* Check for fp exceptions as in .SPffcall, above. */
+        __(btrq $TCR_FLAG_BIT_FOREIGN_FPE,rcontext(tcr.flags))
+        __(jnc 1f)
         __(cmpb $0,C(bogus_fp_exceptions)(%rip))
         __(je 0f)
@@ -4295 +4300 @@
         __(jmp 1f)
 0:      __(stmxcsr rcontext(tcr.ffi_exception))
+        __(ldmxcsr rcontext(tcr.lisp_mxcsr))
 1:      __(movq rcontext(tcr.save_vsp),%rsp)
         __(movq rcontext(tcr.save_rbp),%rbp)
@@ -4310 +4316 @@
         __(pop %temp2)
         __(pop %temp1)
-        __(ldmxcsr rcontext(tcr.lisp_mxcsr))
         __(check_pending_interrupt(%temp0))
         __(pop %temp0)