From 8810e260f591a780feb251fa5dcaa195b2dcc366 Mon Sep 17 00:00:00 2001
From: Thomas Voss <mail@thomasvoss.com>
Date: Mon, 24 Jun 2024 09:05:56 +0200
Subject: Try to handle floating point numbers better

---
 src/analyzer.c | 21 ++++++++++++++++-----
 src/codegen.c  | 31 +++++++++++++++++++++++++++----
 2 files changed, 43 insertions(+), 9 deletions(-)

(limited to 'src')

diff --git a/src/analyzer.c b/src/analyzer.c
index a449653..c7cf3a5 100644
--- a/src/analyzer.c
+++ b/src/analyzer.c
@@ -1,5 +1,7 @@
 #include <assert.h>
 #include <ctype.h>
+#include <limits.h>
+#include <math.h>
 #include <stdbool.h>
 #include <stddef.h>
 #include <stdint.h>
@@ -17,6 +19,10 @@
 #include "symtab.h"
 #include "types.h"
 
+#define LOG2_10       (3.321928)
+#define MP_BITCNT_MAX ((mp_bitcnt_t)-1)
+#define MIN(x, y)     ((x) < (y) ? (x) : (y))
+
 /* In debug builds we want to actually alloc a new mpq_t so that it’s
    easier to free memory without doing a double free */
 #if DEBUG
@@ -259,11 +265,14 @@ analyzeexpr(struct azctx ctx, scope_t *scps, type_t *types, ast_t ast,
             aux_t aux, lexemes_t toks, idx_t i)
 {
 	switch (ast.kinds[i]) {
-	case ASTNUMLIT:
+	case ASTNUMLIT: {
+		strview_t sv = toks.strs[ast.lexemes[i]];
 		types[i].kind = TYPE_NUM;
 		types[i].size = 0;
 		types[i].issigned = true;
+		types[i].isfloat = memchr(sv.p, '.', sv.len) != NULL;
 		return fwdnode(ast, i);
+	}
 	case ASTIDENT: {
 		strview_t sv = toks.strs[ast.lexemes[i]];
 
@@ -403,11 +412,13 @@ constfoldexpr(struct cfctx ctx, mpq_t *folds, scope_t *scps, type_t *types,
 		buf[len] = 0;
 
 		if (isfloat) {
+			/* TODO: Is this correct?  It seems to work… but I don’t know
+			   if there is a better way to do this; sometimes the
+			   precision is a few bits more than it could be. */
 			mpf_t x;
-			/* FIXME: This uses global precision, not thread safe! */
-			/* FIXME: This doesn’t try to figure out what the correct
-			   precision is. */
-			int ret = mpf_init_set_str(x, buf, 10);
+			double prec = ceil((sv.len - 1) * LOG2_10);
+			mpf_init2(x, MIN(MP_BITCNT_MAX, (mp_bitcnt_t)prec));
+			int ret = mpf_set_str(x, buf, 10);
 			assert(ret == 0);
 			mpq_set_f(folds[i], x);
 			mpf_clear(x);
diff --git a/src/codegen.c b/src/codegen.c
index 69f0214..d036b74 100644
--- a/src/codegen.c
+++ b/src/codegen.c
@@ -155,13 +155,36 @@ codegentypedexpr(struct cgctx ctx, idx_t i, type_t type, LLVMValueRef *outv)
 		*outv = LLVMConstIntOfString(type2llvm(ctx, type), buf, 10);
 		return fwdnode(ctx.ast, i);
 	} else if (MPQ_IS_INIT(ctx.folds[i]) /* && type.isfloat */) {
-		/* FIXME: This is bad and broken */
+		char *s, *buf;
+		size_t len;
 		mpf_t x;
-		mpf_init(x);
+		mp_exp_t e;
+		mp_bitcnt_t prec;
+
+		/* TODO: Is this even correct? */
+		switch (type.size) {
+		case 2:  prec =  5; break;
+		case 4:  prec =  8; break;
+		case 8:  prec = 11; break;
+		case 16: prec = 16; break;
+		default: __builtin_unreachable();
+		}
+
+		mpf_init2(x, prec);
 		mpf_set_q(x, ctx.folds[i]);
-		char buf[256];
-		gmp_snprintf(buf, sizeof(buf), "%Ff", x);
+
+		s = mpf_get_str(NULL, &e, 10, 0, x);
+		len = strlen(s);
+		buf = tmpalloc(ctx.s, len + 2, 1);
+		for (size_t i = 0; i < (size_t)e; i++)
+			buf[i] = s[i];
+		buf[e] = '.';
+		for (size_t i = e; i < len; i++)
+			buf[i + 1] = s[i];
+		buf[len + 1] = 0;
 		*outv = LLVMConstRealOfString(type2llvm(ctx, type), buf);
+
+		free(s);
 		mpf_clear(x);
 		return fwdnode(ctx.ast, i);
 	}
-- 
cgit v1.2.3