This is alternative work-around for the x86 float-point problem that uses

"volatile" rather than compile-specific #pragmas.

FossilOrigin-Name: 1d972a690fdc70ab40862bd38427d68b48e8802ddf8e5c301f2d58ce2178b6ec

diff --git a/manifest b/manifest
index 079596dfbecd07a169fc01b648d8c450839453d6..100c0c79201fc8bbb750746c86c33e57c42b1156 100644 (file)
--- a/manifest
+++ b/manifest
@@ -1,5 +1,5 @@
-C Improved\scomments\son\sthe\swork-around\sto\sthe\sGCC\sx86\sfloating\spoint\swonkiness.
-D 2023-07-05T19:56:14.847
+C This\sis\salternative\swork-around\sfor\sthe\sx86\sfloat-point\sproblem\sthat\suses\n"volatile"\srather\sthan\scompile-specific\s#pragmas.
+D 2023-07-05T22:05:18.339
 F .fossil-settings/empty-dirs dbb81e8fc0401ac46a1491ab34a7f2c7c0452f2f06b54ebb845d024ca8283ef1
 F .fossil-settings/ignore-glob 35175cdfcf539b2318cb04a9901442804be81cd677d8b889fcc9149c21f239ea
 F LICENSE.md df5091916dbb40e6e9686186587125e1b2ff51f022cc334e886c19a0e9982724
@@ -705,7 +705,7 @@ F src/trigger.c ad6ab9452715fa9a8075442e15196022275b414b9141b566af8cdb7a1605f2b0
 F src/update.c 0aa36561167a7c40d01163238c297297962f31a15a8d742216b3c37cdf25f731
 F src/upsert.c 5303dc6c518fa7d4b280ec65170f465c7a70b7ac2b22491598f6d0b4875b3145
 F src/utf.c ee39565f0843775cc2c81135751ddd93eceb91a673ea2c57f61c76f288b041a0
-F src/util.c 49a7335c225a7019d4b81606d65d105856a181ed122f25338d2323174a6e4a93
+F src/util.c eac357cf07fd9be14e467e9f38e9888acf2bec8d6221903c5e57136725170880
 F src/vacuum.c 604fcdaebe76f3497c855afcbf91b8fa5046b32de3045bab89cc008d68e40104
 F src/vdbe.c 74282a947234513872a83b0bab1b8c644ece64b3e27b053ef17677c8ff9c81e0
 F src/vdbe.h 41485521f68e9437fdb7ec4a90f9d86ab294e9bb8281e33b235915e29122cfc0
@@ -2043,8 +2043,11 @@ F vsixtest/vsixtest.tcl 6a9a6ab600c25a91a7acc6293828957a386a8a93
 F vsixtest/vsixtest.vcxproj.data 2ed517e100c66dc455b492e1a33350c1b20fbcdc
 F vsixtest/vsixtest.vcxproj.filters 37e51ffedcdb064aad6ff33b6148725226cd608e
 F vsixtest/vsixtest_TemporaryKey.pfx e5b1b036facdb453873e7084e1cae9102ccc67a0
-P 5d9e9364808793d65925d4efbfde0f4246df258758f15e8ce1105070d1018fe6
-R 721b48a97d77f38539bb9af05ab5869d
+P 7b4c16731e7bf6f03f5adf4fcb2008c0b19be473fb1b90b405c217c08916586a
+R edd6abc545b6b11f65503dd16b4a34c2
+T *branch * fix-dekker-with-volatile
+T *sym-fix-dekker-with-volatile *
+T -sym-trunk *
 U drh
-Z c80d89ca100028001e859be8d9af6fed
+Z 5a39b72d20727cda14b03b5876342c08
 # Remove this line to create a well-formed Fossil manifest.

diff --git a/manifest.uuid b/manifest.uuid
index e14a61843666ce2d3269773388f7e5acba625f37..6d12be287b49bad2630b94c0ada471992669d68c 100644 (file)
--- a/manifest.uuid
+++ b/manifest.uuid
@@ -1 +1 @@
-7b4c16731e7bf6f03f5adf4fcb2008c0b19be473fb1b90b405c217c08916586a
\ No newline at end of file
+1d972a690fdc70ab40862bd38427d68b48e8802ddf8e5c301f2d58ce2178b6ec
\ No newline at end of file

diff --git a/src/util.c b/src/util.c
index 6b757e4fb71c9d4e0c12188b4f992a0d55dedaaa..4462f6e970ed03dee32a6b443b97057e9213d44b 100644 (file)
--- a/src/util.c
+++ b/src/util.c
@@ -386,37 +386,25 @@ u8 sqlite3StrIHash(const char *z){
   return h;
 }
 
-/*
-** Work around an issue in GCC where it generates code that stores
-** intermediate results to a higher precision than binary64.  This
-** messes up the Dekker algorithm.  See the discussion at
-** https://sqlite.org/forum/info/ee7278611394034c
-**
-** By adding the -ffloat-store option, it forces GCC to truncate
-** intermediate results to the correct precision.  The GCC devs
-** recommended -fexcess-precision=standard or -std=c99.  Those options
-** work too, from the command-line, but I could not get them to work
-** as a #pragma.  We want the "sqlite3.c" to "just work" without
-** requiring any special compiler-options, so we continue to use
-** the -ffloat-store method of working around the issue.
-*/
-
-#ifdef i386
-#pragma GCC push_options
-#pragma GCC optimize("float-store")
-#endif
-
-
 /* Double-Double multiplication.  (x[0],x[1]) *= (y,yy)
 **
 ** Reference:
 **   T. J. Dekker, "A Floating-Point Technique for Extending the
 **   Available Precision".  1971-07-26.
 */
-static void dekkerMul2(double *x, double y, double yy){
-  double hx, tx, hy, ty, p, q, c, cc;
+static void dekkerMul2(volatile double *x, double y, double yy){
+  /*
+  ** The "volatile" keywords on parameter x[] and on local variables
+  ** below are needed force intermediate results to be truncated to
+  ** binary64 rather than be carried around in an extended-precision
+  ** format.  The truncation is necessary for the Dekker algorithm to
+  ** work.  Intel x86 floating point might omit the truncation without
+  ** the use of volatile. 
+  */
+  volatile double tx, ty, p, q, c, cc;
+  double hx, hy;
   u64 m;
-  memcpy(&m, &x[0], 8);
+  memcpy(&m, (void*)&x[0], 8);
   m &= 0xfffffffffc000000L;
   memcpy(&hx, &m, 8);
   tx = x[0] - hx;
@@ -434,11 +422,6 @@ static void dekkerMul2(double *x, double y, double yy){
   x[1] += cc;
 }
 
-/* End of the GCC x86 floating-point work-around */
-#ifdef i386
-#pragma GCC pop_options
-#endif
-
 /*
 ** The string z[] is an text representation of a real number.
 ** Convert this string to a double and write it into *pResult.