Workaround a 88100 errata where a FPU imprecise exception may be raised

for floating-point-to-int conversions by mistake (these instructions are
not supposed to raise this exception).

diff --git a/sys/arch/m88k/m88k/m88100_fp.c b/sys/arch/m88k/m88k/m88100_fp.c
index 199e019..250cfe3 100644 (file)
--- a/sys/arch/m88k/m88k/m88100_fp.c
+++ b/sys/arch/m88k/m88k/m88100_fp.c
@@ -1,4 +1,4 @@
-/*     $OpenBSD: m88100_fp.c,v 1.5 2020/08/19 10:10:58 mpi Exp $       */
+/*     $OpenBSD: m88100_fp.c,v 1.6 2024/03/03 11:03:13 miod Exp $      */
 
 /*
  * Copyright (c) 2007, 2014, Miodrag Vallat.
@@ -355,6 +355,24 @@ m88100_fpu_imprecise_exception(struct trapframe *frame)
        /* Reset the exception cause register */
        __asm__ volatile ("fstcr %r0, %fcr0");
 
+       /*
+        * The 88100 errata for mask C82N (rev 0x0a) documents that an
+        * imprecise exception may be raised for integer instructions
+        * returning an inexact result.
+        * However, there is nothing to do in this case, since the result
+        * is not a floating-point value, and has been correctly put in
+        * the destination register; we simply need to to ignore that
+        * exception.
+        */
+       switch ((frame->tf_fpit >> 11) & 0x1f) {
+       case 0x09:      /* int */
+       case 0x0a:      /* nint */
+       case 0x0b:      /* trnc */
+               return;
+       default:
+               break;
+       }
+
        /*
         * Pick the inexact result, build a float32 or a float64 out of it, and
         * normalize it to the destination width.