[committed] Fix combiner's make_extraction (PR rtl-optimization/89354)

Jakub Jelinek jakub@redhat.com
Thu Feb 14 23:17:00 GMT 2019


Hi!

The following testcase is miscompiled on i686-linux, because
make_extraction is asked to make an extraction of 33 bits from DImode MEM
at position 0 and happily returns ZERO_EXTRACT with SImode (even when SImode
can hold only 32 bits), the caller (make_field_assignment) then on this
testcase because of that throws away the |= 0x100000000ULL.

Fixed thusly, bootstrapped/regtested on {x86_64,i686,powerpc64{,le}}-linux,
preapproved by Segher on IRC, committed to trunk and 8.x branch.

I've also gathered statistics and the only time during those bootstraps and
(except for still pending (second) powerpc64-linux regtest) regtests the
only time this patch made any difference was on this newly added testcase on
i686-linux.

2019-02-14  Jakub Jelinek  <jakub@redhat.com>

	PR rtl-optimization/89354
	* combine.c (make_extraction): Punt if extraction_mode is narrower
	than len bits.

	* gcc.dg/pr89354.c: New test.

--- gcc/combine.c.jj	2019-02-05 16:38:28.000000000 +0100
+++ gcc/combine.c	2019-02-14 16:45:41.445096523 +0100
@@ -7830,6 +7830,10 @@ make_extraction (machine_mode mode, rtx
       && partial_subreg_p (extraction_mode, mode))
     extraction_mode = mode;
 
+  /* Punt if len is too large for extraction_mode.  */
+  if (maybe_gt (len, GET_MODE_PRECISION (extraction_mode)))
+    return NULL_RTX;
+
   if (!MEM_P (inner))
     wanted_inner_mode = wanted_inner_reg_mode;
   else
--- gcc/testsuite/gcc.dg/pr89354.c.jj	2019-02-14 17:02:26.013552853 +0100
+++ gcc/testsuite/gcc.dg/pr89354.c	2019-02-14 17:01:44.431237813 +0100
@@ -0,0 +1,22 @@
+/* PR rtl-optimization/89354 */
+/* { dg-do run } */
+/* { dg-options "-O2" } */
+/* { dg-additional-options "-msse2" { target sse2_runtime } } */
+
+static unsigned long long q = 0;
+
+__attribute__((noipa)) static void
+foo (void)
+{
+  q = (q & ~0x1ffffffffULL) | 0x100000000ULL;
+}
+
+int
+main ()
+{
+  __asm volatile ("" : "+m" (q));
+  foo ();
+  if (q != 0x100000000ULL)
+    __builtin_abort ();
+  return 0;
+}

	Jakub



More information about the Gcc-patches mailing list