Look Ma' what I found on my harddisk ...

 o New faster syscalls for 2.1.x, too
 o Upgrade to 2.1.89.

Don't try to run this.  It's flaky as hell.  But feel free to debug ...

3 files changed, 43 insertions, 122 deletions

diff --git a/arch/alpha/math-emu/Makefile b/arch/alpha/math-emu/Makefile
index 308533a02..6ea85465f 100644
--- a/arch/alpha/math-emu/Makefile
+++ b/arch/alpha/math-emu/Makefile
@@ -2,11 +2,11 @@
 # Makefile for math-emulator files...
 #
 
-OBJS  = fp-emul.o ieee-math.o
+O_TARGET := math-emu.o
+O_OBJS   := fp-emul.o ieee-math.o
 
-math-emu.a: $(OBJS)
-	$(AR) rcs $@ $(OBJS)
-
-dep:
+ifeq ($(CONFIG_MATHEMU),m)
+M_OBJS   := $(O_TARGET)
+endif
 
 include $(TOPDIR)/Rules.make
diff --git a/arch/alpha/math-emu/fp-emul.c b/arch/alpha/math-emu/fp-emul.c
index f1ff5a2be..543044019 100644
--- a/arch/alpha/math-emu/fp-emul.c
+++ b/arch/alpha/math-emu/fp-emul.c
@@ -1,5 +1,5 @@
+#include <linux/module.h>
 #include <linux/types.h>
-
 #include <linux/kernel.h>
 #include <linux/sched.h>
 
@@ -52,111 +52,33 @@
 #define MISC_TRAPB	0x0000
 #define MISC_EXCB	0x0400
 
+extern unsigned long alpha_read_fp_reg (unsigned long reg);
+extern void alpha_write_fp_reg (unsigned long reg, unsigned long val);
+
 
-extern unsigned long	rdfpcr (void);
-extern void		wrfpcr (unsigned long);
+#ifdef MODULE
 
+MODULE_DESCRIPTION("FP Software completion module");
 
-unsigned long
-alpha_read_fp_reg (unsigned long reg)
+extern long (*alpha_fp_emul_imprecise)(struct pt_regs *, unsigned long);
+static long (*save_emul)(struct pt_regs *, unsigned long);
+long do_alpha_fp_emul_imprecise(struct pt_regs *, unsigned long);
+
+int init_module(void)
 {
-	unsigned long r;
-
-	switch (reg) {
-	      case  0: asm ("stt  $f0,%0" : "m="(r)); break;
-	      case  1: asm ("stt  $f1,%0" : "m="(r)); break;
-	      case  2: asm ("stt  $f2,%0" : "m="(r)); break;
-	      case  3: asm ("stt  $f3,%0" : "m="(r)); break;
-	      case  4: asm ("stt  $f4,%0" : "m="(r)); break;
-	      case  5: asm ("stt  $f5,%0" : "m="(r)); break;
-	      case  6: asm ("stt  $f6,%0" : "m="(r)); break;
-	      case  7: asm ("stt  $f7,%0" : "m="(r)); break;
-	      case  8: asm ("stt  $f8,%0" : "m="(r)); break;
-	      case  9: asm ("stt  $f9,%0" : "m="(r)); break;
-	      case 10: asm ("stt $f10,%0" : "m="(r)); break;
-	      case 11: asm ("stt $f11,%0" : "m="(r)); break;
-	      case 12: asm ("stt $f12,%0" : "m="(r)); break;
-	      case 13: asm ("stt $f13,%0" : "m="(r)); break;
-	      case 14: asm ("stt $f14,%0" : "m="(r)); break;
-	      case 15: asm ("stt $f15,%0" : "m="(r)); break;
-	      case 16: asm ("stt $f16,%0" : "m="(r)); break;
-	      case 17: asm ("stt $f17,%0" : "m="(r)); break;
-	      case 18: asm ("stt $f18,%0" : "m="(r)); break;
-	      case 19: asm ("stt $f19,%0" : "m="(r)); break;
-	      case 20: asm ("stt $f20,%0" : "m="(r)); break;
-	      case 21: asm ("stt $f21,%0" : "m="(r)); break;
-	      case 22: asm ("stt $f22,%0" : "m="(r)); break;
-	      case 23: asm ("stt $f23,%0" : "m="(r)); break;
-	      case 24: asm ("stt $f24,%0" : "m="(r)); break;
-	      case 25: asm ("stt $f25,%0" : "m="(r)); break;
-	      case 26: asm ("stt $f26,%0" : "m="(r)); break;
-	      case 27: asm ("stt $f27,%0" : "m="(r)); break;
-	      case 28: asm ("stt $f28,%0" : "m="(r)); break;
-	      case 29: asm ("stt $f29,%0" : "m="(r)); break;
-	      case 30: asm ("stt $f30,%0" : "m="(r)); break;
-	      case 31: asm ("stt $f31,%0" : "m="(r)); break;
-	      default:
-		break;
-	}
-	return r;
+	save_emul = alpha_fp_emul_imprecise;
+	alpha_fp_emul_imprecise = do_alpha_fp_emul_imprecise;
+	return 0;
 }
 
-
-#if 0
-/*
- * This is IMHO the better way of implementing LDT().  But it
- * has the disadvantage that gcc 2.7.0 refuses to compile it
- * (invalid operand constraints), so instead, we use the uglier
- * macro below.
- */
-# define LDT(reg,val)	\
-  asm volatile ("ldt $f"#reg",%0" :: "m"(val));
-#else
-# define LDT(reg,val)	\
-  asm volatile ("ldt $f"#reg",0(%0)" :: "r"(&val));
-#endif
-
-void
-alpha_write_fp_reg (unsigned long reg, unsigned long val)
+void cleanup_module(void)
 {
-	switch (reg) {
-	      case  0: LDT( 0, val); break;
-	      case  1: LDT( 1, val); break;
-	      case  2: LDT( 2, val); break;
-	      case  3: LDT( 3, val); break;
-	      case  4: LDT( 4, val); break;
-	      case  5: LDT( 5, val); break;
-	      case  6: LDT( 6, val); break;
-	      case  7: LDT( 7, val); break;
-	      case  8: LDT( 8, val); break;
-	      case  9: LDT( 9, val); break;
-	      case 10: LDT(10, val); break;
-	      case 11: LDT(11, val); break;
-	      case 12: LDT(12, val); break;
-	      case 13: LDT(13, val); break;
-	      case 14: LDT(14, val); break;
-	      case 15: LDT(15, val); break;
-	      case 16: LDT(16, val); break;
-	      case 17: LDT(17, val); break;
-	      case 18: LDT(18, val); break;
-	      case 19: LDT(19, val); break;
-	      case 20: LDT(20, val); break;
-	      case 21: LDT(21, val); break;
-	      case 22: LDT(22, val); break;
-	      case 23: LDT(23, val); break;
-	      case 24: LDT(24, val); break;
-	      case 25: LDT(25, val); break;
-	      case 26: LDT(26, val); break;
-	      case 27: LDT(27, val); break;
-	      case 28: LDT(28, val); break;
-	      case 29: LDT(29, val); break;
-	      case 30: LDT(30, val); break;
-	      case 31: LDT(31, val); break;
-	      default:
-		break;
-	}
+	alpha_fp_emul_imprecise = save_emul;
 }
 
+#define alpha_fp_emul_imprecise  do_alpha_fp_emul_imprecise
+
+#endif /* MODULE */
 
 /*
  * Emulate the floating point instruction at address PC.  Returns 0 if
@@ -334,6 +256,8 @@ alpha_fp_emul_imprecise (struct pt_regs *regs, unsigned long write_mask)
 	unsigned long trigger_pc = regs->pc - 4;
 	unsigned long insn, opcode, rc;
 
+	MOD_INC_USE_COUNT;
+
 	/*
 	 * Turn off the bits corresponding to registers that are the
 	 * target of instructions that set bits in the exception
@@ -354,12 +278,14 @@ alpha_fp_emul_imprecise (struct pt_regs *regs, unsigned long write_mask)
 		      case OPC_PAL:
 		      case OPC_JSR:
 		      case 0x30 ... 0x3f:	/* branches */
+			MOD_DEC_USE_COUNT;
 			return 0;
 
 		      case OPC_MISC:
 			switch (insn & 0xffff) {
 			      case MISC_TRAPB:
 			      case MISC_EXCB:
+				MOD_DEC_USE_COUNT;
 				return 0;
 
 			      default:
@@ -386,11 +312,13 @@ alpha_fp_emul_imprecise (struct pt_regs *regs, unsigned long write_mask)
 			{
 			    /* re-execute insns in trap-shadow: */
 			    regs->pc = trigger_pc + 4;
+			    MOD_DEC_USE_COUNT;
 			    return 1;
 			}
 			break;
 		}
 		trigger_pc -= 4;
 	}
+	MOD_DEC_USE_COUNT;
 	return 0;
 }
diff --git a/arch/alpha/math-emu/ieee-math.c b/arch/alpha/math-emu/ieee-math.c
index d408eece5..59d7dfa6d 100644
--- a/arch/alpha/math-emu/ieee-math.c
+++ b/arch/alpha/math-emu/ieee-math.c
@@ -137,9 +137,8 @@ sub128 (const unsigned long a[2], const unsigned long b[2], unsigned long c[2])
 static inline void
 mul64 (const unsigned long a, const unsigned long b, unsigned long c[2])
 {
-	asm ("mulq  %2,%3,%0\n\t"
-	     "umulh %2,%3,%1"
-	     : "r="(c[0]), "r="(c[1]) : "r"(a), "r"(b));
+	c[0] = a * b;
+	asm ("umulh %1,%2,%0" : "=r"(c[1]) : "r"(a), "r"(b));
 }
 
 
@@ -276,7 +275,7 @@ make_s_ieee (long f, EXTENDED *a, unsigned long *b)
 {
 	unsigned long res, sticky;
 
-	if (!a->f[0] && !a->f[1]) {
+	if (!a->e && !a->f[0] && !a->f[1]) {
 		*b = (unsigned long) a->s << 63;	/* return +/-0 */
 		return 0;
 	}
@@ -356,7 +355,7 @@ make_t_ieee (long f, EXTENDED *a, unsigned long *b)
 {
 	unsigned long res, sticky;
 
-	if (!a->f[0] && !a->f[1]) {
+	if (!a->e && !a->f[0] && !a->f[1]) {
 		*b = (unsigned long) a->s << 63;	/* return +/-0 */
 		return 0;
 	}
@@ -384,7 +383,7 @@ make_t_ieee (long f, EXTENDED *a, unsigned long *b)
 			a->e = -0x3ff;
 		}
 	}
-	if (a->e > 0x3ff) {
+	if (a->e >= 0x3ff) {
 		res = FPCR_OVF | FPCR_INE;
 		if (f & IEEE_TRAP_ENABLE_OVF) {
 			a->e -= 0x600;	/* scale down result by 2^alpha */
@@ -777,25 +776,25 @@ ieee_CVTTQ (int f, unsigned long a, unsigned long *b)
 			midway = (temp.f[0] & 0x003fffffffffffff) == 0;
 			if ((midway && (temp.f[0] & 0x0080000000000000)) ||
 			    !midway)
-				++b;
+				++*b;
 		}
 		break;
 
 	      case ROUND_PINF:
-		if ((temp.f[0] & 0x003fffffffffffff) != 0)
-			++b;
+		if ((temp.f[0] & 0x007fffffffffffff) != 0)
+			++*b;
 		break;
 
 	      case ROUND_NINF:
-		if ((temp.f[0] & 0x003fffffffffffff) != 0)
-			--b;
+		if ((temp.f[0] & 0x007fffffffffffff) != 0)
+			--*b;
 		break;
 
 	      case ROUND_CHOP:
 		/* no action needed */
 		break;
 	}
-	if ((temp.f[0] & 0x003fffffffffffff) != 0)
+	if ((temp.f[0] & 0x007fffffffffffff) != 0)
 		res |= FPCR_INE;
 
 	if (temp.s) {
@@ -1143,12 +1142,9 @@ ieee_MULS (int f, unsigned long a, unsigned long b, unsigned long *c)
 		return 0;
 	}
 	op_c.s = op_a.s ^ op_b.s;
-	op_c.e = op_a.e + op_b.e;
+	op_c.e = op_a.e + op_b.e - 55;
 	mul64(op_a.f[0], op_b.f[0], op_c.f);
 
-	normalize(&op_c);
-	op_c.e -= 55;		/* drop the 55 original bits. */
-
 	return round_s_ieee(f, &op_c, c);
 }
 
@@ -1200,12 +1196,9 @@ ieee_MULT (int f, unsigned long a, unsigned long b, unsigned long *c)
 		return 0;
 	}
 	op_c.s = op_a.s ^ op_b.s;
-	op_c.e = op_a.e + op_b.e;
+	op_c.e = op_a.e + op_b.e - 55;
 	mul64(op_a.f[0], op_b.f[0], op_c.f);
 
-	normalize(&op_c);
-	op_c.e -= 55;	/* drop the 55 original bits. */
-
 	return round_t_ieee(f, &op_c, c);
 }