Merge branch 'master' into feature/api-mpfr

1 files changed, 19 insertions, 18 deletions

diff --git a/awk.h b/awk.h
index e2a400a6..2b3a0dea 100644
--- a/awk.h
+++ b/awk.h
@@ -365,21 +365,6 @@ typedef struct exp_node {
 		} nodep;
 
 		struct {
-			// April 2017:
-			// The NODE union will be the death of me yet. :-(
-			//
-			// On 64 bit Intel systems, at least, if compiling
-			// without MPFR, the pointers need to come before the 
-			// the double. Otherwise things break pretty
-			// badly.  This doesn't happen on 32 bit compiles.
-			// I saw this on GCC 5.4.0, GCC 6.3.0, PCC, TCC
-			// and clang 3.8.0. It's way too much work to
-			// try to figure out why. So I just put pointers,
-			// then the double, then the size_t's, and then
-			// the rest. Sigh.
-			char *sp;
-			wchar_t *wsp;
-			struct exp_node *typre;
 #ifdef HAVE_MPFR
 			union {
 				AWKNUM fltnum;
@@ -389,14 +374,31 @@ typedef struct exp_node {
 #else
 			AWKNUM fltnum;
 #endif
+			char *sp;
 			size_t slen;
-			size_t wslen;
-			long sref;
 			int idx;
+			wchar_t *wsp;
+			size_t wslen;
+			struct exp_node *typre;
 		} val;
 	} sub;
 	NODETYPE type;
 	unsigned int flags;
+	// April 2017:
+	// The NODE union will be the death of me yet. :-(
+	//
+	// On 64 bit Intel systems, at least, if compiling without MPFR,
+	// the valref (formerly) sref field needs to be at the end. In its
+	// original position it overlapped with stuff in the nodep part of
+	// the union causing things to break pretty badly.  This doesn't
+	// happen on 32 bit compiles.
+	//
+	// I saw this on GCC 4.9.0, GCC 5.4.0, GCC 6.3.0, PCC, TCC and
+	// clang 3.8.0.
+	//
+	// We simply move valref out of the unions entirely to avoid future
+	// problems.
+	long valref;
 
 /* type = Node_val */
 	/*
@@ -506,7 +508,6 @@ typedef struct exp_node {
  */
 #define stptr	sub.val.sp
 #define stlen	sub.val.slen
-#define valref	sub.val.sref
 #define stfmt	sub.val.idx
 #define wstptr	sub.val.wsp
 #define wstlen	sub.val.wslen