[kernel] Rewrite printk for speed, add __divmod fast 32/16-bit divide

elks/arch/i86/lib/Makefile

@@ -37,6 +37,7 @@ SRCS1 = \
 	peekpoke.S \
 	string.S \
 	printreg.S \
+	divmod.S \
 	# end of list
 
 ifeq ($(CONFIG_ARCH_PC98), y)

elks/arch/i86/lib/divmod.S

@@ -0,0 +1,47 @@
+// Fast 32-bit combined divide and modulo routine
+//
+// unsigned long __divmod(unsigned long val, unsigned int *baserem)
+//  Unsigned divide 32-bits by 16-bits
+//      Store denominator in *baserem before calling
+//      Returns 32-bit quotient in DX:AX and remainder in *baserem
+//
+//  Designed for a fast replacement of the following code which calls __udivsi3/__umodsi3:
+//      unsigned int rem, base;
+//      rem = val % base;
+//      val = val / base;
+//  New code:
+//      rem = base;
+//      val = __divmod(val, &rem);
+//
+//  inspired by OpenWatcom ltoa.c __uldiv routine
+//  13 Sep 2024 Greg Haerr
+
+#define NUMLO   2
+#define NUMHI   4
+#define ADDR    6
+
+        .arch   i8086, nojumps
+        .code16
+        .text
+
+        .global __divmod
+__divmod:
+        mov     %sp,%bx
+        mov     NUMLO(%bx),%ax
+        mov     NUMHI(%bx),%dx
+        mov     ADDR(%bx),%bx
+
+//  divides DX:AX / [BX]
+//  returns DX:AX with remainder in [BX]
+
+        xor     %cx,%cx             // temp CX = 0
+        cmp     (%bx),%dx           // is upper 16 bits numerator less than denominator
+        jb      1f                  // yes - only one DIV needed
+        xchg    %dx,%ax             // AX = upper numerator, DX = lower numerator
+        xchg    %dx,%cx             // DX = 0, CX = lower numerator
+        divw    (%bx)               // AX = upper numerator / base, DX = remainder
+        xchg    %cx,%ax             // AX = lower numerator, CX = high quotient
+1:      divw    (%bx)               // AX = lower numerator / base, DX = remainder
+        mov     %dx,(%bx)           // store remainder
+        mov     %cx,%dx             // DX = high quotient, AX = low quotient
+        ret

elks/include/arch/divmod.h

@@ -0,0 +1,6 @@
+#ifndef __ARCH_8086_DIVMOD_H
+#define __ARCH_8086_DIVMOD_H
+
+unsigned long __divmod(unsigned long val, unsigned int *baserem);
+
+#endif

elks/init/main.c

@@ -114,6 +114,25 @@ void start_kernel(void)
     kfork_proc(init_task);
     wake_up_process(&task[1]);
 
+#if TIMER_TEST
+    /* printk rewrite debug statements */
+    printk("#04X: '%#04X'\n", 0x2ab);
+    printk("04X: '%04X'\n", 0x2ab);
+    printk("04x: '%04x'\n", 0x2ab);
+    printk(" 4x: '%4x'\n", 0x2ab);
+    printk("04d: '%04d'\n", 0x200);
+    printk(" 4d: '%4d'\n", 0x200);
+    printk("05d: '%05d'\n", -200);
+    printk(" 5d: '%5d'\n", -200);
+    printk(" ld: '%ld'\n", -123456789L);
+    printk(" lx: '%lx'\n", 0x87654321L);
+    printk(" lo: '%lo'\n", 0xFFFFFFFFL);
+    printk("  s: '%s'\n", "thisisatest");
+    printk(" 6s: '%6s'\n", "thisisatest");
+    printk("20s: '%20s'\n", "thisisatest");
+    test_ptime_print();
+#endif
+
     /*
      * We are now the idle task. We won't run unless no other process can run.
      * The idle task always runs with _gint_count == 1 (switched from user mode syscall)

elks/kernel/printk.c

@@ -26,6 +26,7 @@
  *              Alan Cox.
  *
  *      MTK:    Sep 97 - Misc hacks to shrink generated code
+ *      GRH:    Sep 24 - Rewritten for speed
  */
 
 #include <linuxmt/config.h>
@@ -40,6 +41,7 @@
 #include <linuxmt/prectimer.h>
 #include <arch/segment.h>
 #include <arch/irq.h>
+#include <arch/divmod.h>
 #include <stdarg.h>
 
 #define CONFIG_PREC_TIMER   1   /* =1 to include %k precision timer printk format */
@@ -56,7 +58,7 @@ static void (*kputc)(dev_t, int) = 0;
 
 void set_console(dev_t dev)
 {
-    register struct tty *ttyp;
+    struct tty *ttyp;
 
     if (dev == 0)
             dev = DEVCONSOLE;
@@ -86,29 +88,18 @@ static void kputs(const char *buf)
  *      Output a number
  */
 
-static char hex_string[] = "0123456789ABCDEF 0123456789abcdef ";
-
 static void numout(unsigned long v, int width, unsigned int base, int type,
     int Zero, int alt)
 {
-    unsigned long dvr;
-    int c, vch, i;
-    int useSign, Lower, Decimal;
-
-    i = 10;
-    dvr = 1000000000L;
-    if (base > 10) {
-        i = 8;
-        dvr = 0x10000000L;
-    }
-    if (base < 10) {
-        i = 11;
-        dvr = 0x40000000L;
-    }
-
-    Decimal = 0;
+    int n, i;
+    unsigned int c;
+    char *p;
+    int Sign, Suffix, Decimal;
+    char buf[12];                       /* small stack: good up to max long octal v */
+
+    Decimal = -1;
+    Sign = Suffix = 0;
 #if CONFIG_PREC_TIMER
-    int Msecs = 0;
     /* display 1/1193182s get_time*() pticks in range 0.838usec through 42.85sec */
     if (type == 'k') {                  /* format works w/limited ranges only */
         Decimal = 3;
@@ -121,59 +112,57 @@ static void numout(unsigned long v, int width, unsigned int base, int type,
             v = v * 838UL;              /* convert to nanosecs w/o 32-bit overflow */
         if (v > 1000000000UL) {         /* display x.xxx secs */
             v /= 1000000UL;             /* divide using _udivsi3 for speed */
+            Suffix = 's';
         } else if (v > 1000000UL) {
             v /= 1000UL;                /* display xx.xxx msecs */
-            Msecs = 1;
+            Suffix = ('s' << 8) | 'm';
         } else {
-            Msecs = 2;                  /* display xx.xxx usecs */
+            Suffix = ('s' << 8) | 'u';  /* display xx.xxx usecs */
         }
     }
 #endif
 
-    useSign = (type == 'd');
-    if (useSign && ((long)v < 0L))
-        v = (-(long)v);
-    else
-        useSign = 0;
+    if (type == 'd'  && (long)v < 0L) {
+        v = -(long)v;
+        Sign = 1;
+    }
 
-    Lower = (type != 'X');
-    if (Lower)
-        Lower = 17;
-    vch = 0;
     if (alt && base == 16) {
         kputchar('0');
         kputchar('x');
     }
+
+    p = buf + sizeof(buf) - 1;
+    *p = '\0';
     do {
-        c = (int)(v / dvr);
-        v %= dvr;
-        dvr /= base;
-        if (c || (i <= width) || (i < 2)) {
-            if (i > width)
-                width = i;
-            if (!Zero && !c && (i > 1))
-                c = 16;
-            else {
-                Zero = 1;
-                if (useSign) {
-                    useSign = 0;
-                    vch = '-';
-                }
-            }
-            if (vch)
-                kputchar(vch);
-            vch = *(hex_string + Lower + c);
-            if (type == 'k' && i == Decimal) kputchar('.');
-        }
-    } while (--i);
-    kputchar(vch);
-#if CONFIG_PREC_TIMER
-    if (type == 'k') {
-        if (Msecs)
-            kputchar(Msecs == 1? 'm': 'u');
-        kputchar('s');
+        c = base;
+        v = __divmod(v, &c);            /* remainder returned in c */
+        if (c > 9)
+            *--p = ((type == 'X')? 'A': 'a') - 10 + c;
+        else
+            *--p = '0' + c;
+    } while (v != 0);
+    i = n = buf + sizeof(buf) - 1 - p + Sign;   /* string length */
+    if (n > width)                      /* expand field width */
+        width = n;
+    if (Zero && Sign) {
+        kputchar('-');
+        Sign = 0;
+    }
+    while (n++ < width)
+        kputchar(Zero? '0': ' ');
+    if (Sign)
+        kputchar('-');
+    while (*p) {
+        if (i == Decimal)
+            kputchar('.');
+        i--;
+        kputchar(*p++);
+    }
+    while (Suffix) {
+        kputchar(Suffix & 255);
+        Suffix >>= 8;
     }
-#endif
 }
 
 static void vprintk(const char *fmt, va_list p)