fpu: Handle m68k extended precision denormals properly

Motorola treats denormals with explicit integer bit set as
having unbiased exponent 0, unlike Intel which treats it as
having unbiased exponent 1 (more like all other IEEE formats
that have no explicit integer bit).

Add a flag on FloatFmt to differentiate the behaviour.

Reported-by: Keith Packard <[email protected]>
Reviewed-by: Philippe Mathieu-Daudé <[email protected]>
Signed-off-by: Richard Henderson <[email protected]>

fpu/softfloat-parts.c.inc

@@ -118,7 +118,8 @@ static void partsN(canonicalize)(FloatPartsN *p, float_status *status,
         } else {
             int shift = frac_normalize(p);
             p->cls = float_class_normal;
-            p->exp = fmt->frac_shift - fmt->exp_bias - shift + 1;
+            p->exp = fmt->frac_shift - fmt->exp_bias
+                   - shift + !fmt->m68k_denormal;
         }
     } else if (likely(p->exp < fmt->exp_max) || fmt->arm_althp) {
         p->cls = float_class_normal;
@@ -256,7 +257,7 @@ static void partsN(uncanon_normal)(FloatPartsN *p, float_status *s,
             is_tiny = !frac_addi(&discard, p, inc);
         }
 
-        frac_shrjam(p, 1 - exp);
+        frac_shrjam(p, !fmt->m68k_denormal - exp);
 
         if (p->frac_lo & round_mask) {
             /* Need to recompute round-to-even/round-to-odd. */
@@ -287,7 +288,7 @@ static void partsN(uncanon_normal)(FloatPartsN *p, float_status *s,
             p->frac_lo &= ~round_mask;
         }
 
-        exp = (p->frac_hi & DECOMPOSED_IMPLICIT_BIT) != 0;
+        exp = (p->frac_hi & DECOMPOSED_IMPLICIT_BIT) && !fmt->m68k_denormal;
         frac_shr(p, frac_shift);
 
         if (is_tiny && (flags & float_flag_inexact)) {

fpu/softfloat.c

@@ -517,6 +517,7 @@ typedef struct {
  *   round_mask: bits below lsb which must be rounded
  * The following optional modifiers are available:
  *   arm_althp: handle ARM Alternative Half Precision
+ *   m68k_denormal: explicit integer bit for extended precision may be 1
  */
 typedef struct {
     int exp_size;
@@ -526,6 +527,7 @@ typedef struct {
     int frac_size;
     int frac_shift;
     bool arm_althp;
+    bool m68k_denormal;
     uint64_t round_mask;
 } FloatFmt;
 
@@ -576,7 +578,12 @@ static const FloatFmt float128_params = {
 static const FloatFmt floatx80_params[3] = {
     [floatx80_precision_s] = { FLOATX80_PARAMS(23) },
     [floatx80_precision_d] = { FLOATX80_PARAMS(52) },
-    [floatx80_precision_x] = { FLOATX80_PARAMS(64) },
+    [floatx80_precision_x] = {
+        FLOATX80_PARAMS(64),
+#ifdef TARGET_M68K
+        .m68k_denormal = true,
+#endif
+    },
 };
 
 /* Unpack a float to parts, but do not canonicalize.  */

tests/tcg/m68k/Makefile.target

@@ -4,7 +4,7 @@
 #
 
 VPATH += $(SRC_PATH)/tests/tcg/m68k
-TESTS += trap
+TESTS += trap denormal
 
 # On m68k Linux supports 4k and 8k pages (but 8k is currently broken)
 EXTRA_RUNS+=run-test-mmap-4096 # run-test-mmap-8192

tests/tcg/m68k/denormal.c

@@ -0,0 +1,53 @@
+/*
+ * Test m68k extended double denormals.
+ */
+
+#include <stdio.h>
+#include <stdint.h>
+
+#define TEST(X, Y)  { X, Y, X * Y }
+
+static volatile long double test[][3] = {
+    TEST(0x1p+16383l, 0x1p-16446l),
+    TEST(0x1.1p-8223l, 0x1.1p-8224l),
+    TEST(1.0l, 0x1p-16383l),
+};
+
+#undef TEST
+
+static void dump_ld(const char *label, long double ld)
+{
+    union {
+        long double  d;
+        struct {
+            uint32_t exp:16;
+            uint32_t space:16;
+            uint32_t h;
+            uint32_t l;
+        };
+    } u;
+
+    u.d = ld;
+    printf("%12s: % -27La 0x%04x 0x%08x 0x%08x\n", label, u.d, u.exp, u.h, u.l);
+}
+
+int main(void)
+{
+    int i, n = sizeof(test) / sizeof(test[0]), err = 0;
+
+    for (i = 0; i < n; ++i) {
+        long double x = test[i][0];
+        long double y = test[i][1];
+        long double build_mul = test[i][2];
+        long double runtime_mul = x * y;
+
+        if (runtime_mul != build_mul) {
+            dump_ld("x", x);
+            dump_ld("y", y);
+            dump_ld("build_mul", build_mul);
+            dump_ld("runtime_mul", runtime_mul);
+            err = 1;
+        }
+    }
+    return err;
+}