Support 16-bit IEEE floating point; used in SSE5

SSE5 supports standard IEEE 16-bit floating point, so we should
support that too.

2 files changed, 82 insertions, 13 deletions

diff --git a/float.c b/float.c
index 099e23f2..afa84d2e 100644
--- a/float.c
+++ b/float.c
@@ -213,6 +213,7 @@ static int ieee_round(uint16_t *mant, int i)
 
 #define put(a,b) ( (*(a)=(b)), ((a)[1]=(b)>>8) )
 
+/* 64-bit format with 52-bit mantissa and 11-bit exponent */
 static int to_double(char *str, int32_t sign, uint8_t *result,
                      efunc error)
 {
@@ -275,6 +276,7 @@ static int to_double(char *str, int32_t sign, uint8_t *result,
     return 1;                   /* success */
 }
 
+/* 32-bit format with 23-bit mantissa and 8-bit exponent */
 static int to_float(char *str, int32_t sign, uint8_t *result,
                     efunc error)
 {
@@ -330,6 +332,64 @@ static int to_float(char *str, int32_t sign, uint8_t *result,
     return 1;
 }
 
+/* 16-bit format with 10-bit mantissa and 5-bit exponent.
+   Defined in IEEE 754r.  Used in SSE5.  See the AMD SSE5 manual, AMD
+   document number 43479. */
+static int to_float16(char *str, int32_t sign, uint8_t *result,
+		      efunc error)
+{
+    uint16_t mant[MANT_WORDS];
+    int32_t exponent;
+
+    sign = (sign < 0 ? 0x8000L : 0L);
+
+    ieee_flconvert(str, mant, &exponent, error);
+    if (mant[0] & 0x8000) {
+        /*
+         * Non-zero.
+         */
+        exponent--;
+        if (exponent >= -14 && exponent <= 16) {
+            /*
+             * Normalised.
+             */
+            exponent += 15;
+            ieee_shr(mant, 5);
+            ieee_round(mant, 1);
+            if (mant[0] & 0x800)        /* did we scale up by one? */
+                ieee_shr(mant, 1), exponent++;
+            mant[0] &= 0x3FF;    /* remove leading one */
+            put(result + 0, (exponent << 7) | mant[0] | sign);
+        } else if (exponent < -14 && exponent >= -24) {
+            /*
+             * Denormal.
+             */
+            int shift = -(exponent + 8);
+            int sh = shift % 16, wds = shift / 16;
+            ieee_shr(mant, sh);
+            if (ieee_round(mant, 1 - wds)
+                || (sh > 0 && (mant[0] & (0x8000 >> (sh - 1))))) {
+                ieee_shr(mant, 1);
+                if (sh == 0)
+                    mant[0] |= 0x8000;
+                exponent++;
+            }
+            put(result + 0, (wds == 0 ? mant[0] : 0) | sign);
+        } else {
+            if (exponent > 0) {
+                error(ERR_NONFATAL, "overflow in floating-point constant");
+                return 0;
+            } else
+                memset(result, 0, 2);
+        }
+    } else {
+        memset(result, 0, 2);
+    }
+    return 1;
+}
+
+/* 80-bit format with 64-bit mantissa *including an explicit integer 1*
+   and 15-bit exponent. */
 static int to_ldoub(char *str, int32_t sign, uint8_t *result,
                     efunc error)
 {
@@ -394,13 +454,16 @@ static int to_ldoub(char *str, int32_t sign, uint8_t *result,
 int float_const(char *number, int32_t sign, uint8_t *result, int bytes,
                 efunc error)
 {
-    if (bytes == 4)
+    switch (bytes) {
+    case 2:
+	return to_float16(number, sign, result, error);
+    case 4:
         return to_float(number, sign, result, error);
-    else if (bytes == 8)
+    case 8:
         return to_double(number, sign, result, error);
-    else if (bytes == 10)
+    case 10:
         return to_ldoub(number, sign, result, error);
-    else {
+    default:
         error(ERR_PANIC, "strange value %d passed to float_const", bytes);
         return 0;
     }
diff --git a/parser.c b/parser.c
index ca12a097..69ae3790 100644
--- a/parser.c
+++ b/parser.c
@@ -230,30 +230,36 @@ insn *parse_line(int pass, char *buffer, insn * result,
                 if (i == TOKEN_FLOAT) {
                     eop->type = EOT_DB_STRING;
                     result->eops_float = TRUE;
-                    if (result->opcode == I_DD)
+		    switch (result->opcode) {
+		    case I_DW:
+			eop->stringlen = 2;
+			break;
+		    case I_DD:
                         eop->stringlen = 4;
-                    else if (result->opcode == I_DQ)
+			break;
+		    case I_DQ:
                         eop->stringlen = 8;
-                    else if (result->opcode == I_DT)
+			break;
+		    case I_DT:
                         eop->stringlen = 10;
-		    else if (result->opcode == I_DO)
-			eop->stringlen = 16;
-                    else {
+			break;
+		    default:
                         error(ERR_NONFATAL, "floating-point constant"
-                              " encountered in `D%c' instruction",
-                              result->opcode == I_DW ? 'W' : 'B');
+                              " encountered in `d%c' instruction"
+			      ? (result->opcode == I_DO) ? 'o' : 'b');
                         /*
                          * fix suggested by Pedro Gimeno... original line
                          * was:
                          * eop->type = EOT_NOTHING;
                          */
                         eop->stringlen = 0;
+			break;
                     }
                     eop = nasm_realloc(eop, sizeof(extop) + eop->stringlen);
                     tail = &eop->next;
                     *fixptr = eop;
                     eop->stringval = (char *)eop + sizeof(extop);
-                    if (eop->stringlen < 4 ||
+                    if (!eop->stringlen ||
                         !float_const(tokval.t_charptr, sign,
                                      (uint8_t *)eop->stringval,
                                      eop->stringlen, error))