On Wed, Feb 20, 2019 at 11:13:54AM +0000, Andrew Stubbs wrote:
> This patch causes a number of test regressions for (at least) arm and
> amdgcn. I've run a bisect and confirmed this is the first commit where it
> doesn't work.
The following, so far untested (except for x86_64->armv7hl-linux-gnueabi
cross on the builtin-math-7.c testcase and x86_64-linux native on the
pr88074* testcases) patch should cure this, will bootstrap/regtest it
tonight.
2019-02-20 Jakub Jelinek <ja...@redhat.com>
PR middle-end/88074
* toplev.c (do_compile): Double the emin/emax exponents to workaround
buggy mpc_norm.
* gcc.dg/pr88074-2.c: New test.
--- gcc/toplev.c.jj 2019-02-20 10:00:49.290492694 +0100
+++ gcc/toplev.c 2019-02-20 14:04:27.037769362 +0100
@@ -2173,8 +2173,12 @@ do_compile ()
max_exp = fmt->emax;
}
}
- if (mpfr_set_emin (min_exp)
- || mpfr_set_emax (max_exp))
+ /* mpc_norm assumes it can square a number without bothering with
+ with range scaling, so until that is fixed, double the minimum
+ and maximum exponents, plus add some buffer for arithmetics
+ on the squared numbers. */
+ if (mpfr_set_emin (2 * (min_exp - 1))
+ || mpfr_set_emax (2 * (max_exp + 1)))
sorry ("mpfr not configured to handle all float modes");
/* Set up the back-end if requested. */
--- gcc/testsuite/gcc.dg/pr88074-2.c.jj 2019-02-20 14:10:07.109140297 +0100
+++ gcc/testsuite/gcc.dg/pr88074-2.c 2019-02-20 14:09:46.696478179 +0100
@@ -0,0 +1,17 @@
+/* PR middle-end/88074 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -fdump-tree-optimized" } */
+/* { dg-add-options float128 } */
+/* { dg-require-effective-target float128 } */
+/* { dg-final { scan-tree-dump-not "link_error " "optimized" } } */
+
+extern void link_error (void);
+int
+main ()
+{
+ if (((__FLT128_MAX__ * 0.5 + __FLT128_MAX__ * 0.5i)
+ / (__FLT128_MAX__ * 0.25 + __FLT128_MAX__ * 0.25i))
+ != (_Complex _Float128) 2)
+ link_error ();
+ return 0;
+}
Jakub
