Hi!
ix86_expand_sse_fp_minmax bypasses the expanders and so might end up with
both if_true and if_false being a MEM, which violates the condition of the
insn it wants to match.
The following patch makes sure at most one of the operands is a MEM.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2019-04-24 Jakub Jelinek <ja...@redhat.com>
PR target/90187
* config/i386/i386.c (ix86_expand_sse_fp_minmax): Force if_true into
a register if both if_true and if_false are MEMs.
* g++.target/i386/pr90187.C: New test.
--- gcc/config/i386/i386.c.jj 2019-04-16 10:40:15.077091789 +0200
+++ gcc/config/i386/i386.c 2019-04-23 11:42:32.088334896 +0200
@@ -23712,6 +23712,8 @@ ix86_expand_sse_fp_minmax (rtx dest, enu
else
{
code = is_min ? SMIN : SMAX;
+ if (MEM_P (if_true) && MEM_P (if_false))
+ if_true = force_reg (mode, if_true);
tmp = gen_rtx_fmt_ee (code, mode, if_true, if_false);
}
--- gcc/testsuite/g++.target/i386/pr90187.C.jj 2019-04-23 11:42:51.175025002
+0200
+++ gcc/testsuite/g++.target/i386/pr90187.C 2019-04-23 11:41:56.405914260
+0200
@@ -0,0 +1,15 @@
+// PR target/90187
+// { dg-do compile }
+// { dg-options "-Ofast -ffloat-store" }
+
+double a[64];
+double *foo (void);
+
+void
+bar (int x, const double *y)
+{
+ int i;
+ for (i = 0; i < x; i++)
+ if (y[i] < a[i])
+ a[i] = y[i];
+}
Jakub
