dojump: Improve float != comparisons on x86 [PR98212]

The x86 backend doesn't have EQ or NE floating point comparisons,
so splits x != y into x unord y || x <> y.  The problem with that is
that unord comparison doesn't trap on qNaN operands but LTGT does.
The end effect is that it doesn't trap on qNaN operands, because x unord y
will be true for those and so LTGT will not be performed, but as the backend
is currently unable to merge signalling and non-signalling comparisons (and
after all, with this exact exception it shouldn't unless the first one is
signalling and the second one is non-signalling) it means we end up with:
        ucomiss %xmm1, %xmm0
        jp      .L4
        comiss  %xmm1, %xmm0
        jne     .L4
        ret
        .p2align 4,,10
        .p2align 3
.L4:
        xorl    %eax, %eax
        jmp     foo
where the comiss is the signalling comparison, but we already know that
the right flags bits are already computed by the ucomiss insn.

The following patch, if target supports UNEQ comparisons, splits NE
as x unord y || !(x uneq y) instead, which in the end means we end up with
just:
        ucomiss %xmm1, %xmm0
        jp      .L4
        jne     .L4
        ret
        .p2align 4,,10
        .p2align 3
.L4:
        jmp     foo
because UNEQ is like UNORDERED non-signalling.

2020-12-10  Jakub Jelinek  <jakub@redhat.com>

	PR rtl-optimization/98212
	* dojump.c (do_compare_rtx_and_jump): When splitting NE and backend
	can do UNEQ, prefer splitting x != y into x unord y || !(x uneq y)
	instead of into x unord y || x ltgt y.

	* gcc.target/i386/pr98212.c: New test.

diff --git a/gcc/dojump.c b/gcc/dojump.c
index 4c7fafe296540ec97feb5c69f8896ee70ca90da4..b12bcea7468af1ef5e127dd1e9af6c2bd7baf072 100644 (file)
--- a/gcc/dojump.c
+++ b/gcc/dojump.c
@@ -1168,6 +1168,23 @@ do_compare_rtx_and_jump (rtx op0, rtx op1, enum rtx_code code, int unsignedp,
                  profile_probability first_prob = prob.split (cprob);
                  do_compare_rtx_and_jump (op0, op1, first_code, unsignedp, mode,
                                           size, NULL, if_true_label, first_prob);
+                 if (orig_code == NE && can_compare_p (UNEQ, mode, ccp_jump))
+                   {
+                     /* x != y can be split into x unord y || x ltgt y
+                        or x unord y || !(x uneq y).  The latter has the
+                        advantage that both comparisons are non-signalling and
+                        so there is a higher chance that the RTL optimizations
+                        merge the two comparisons into just one.  */
+                     code = UNEQ;
+                     prob = prob.invert ();
+                     if (! if_false_label)
+                       {
+                         if (! dummy_label)
+                           dummy_label = gen_label_rtx ();
+                         if_false_label = dummy_label;
+                       }
+                     std::swap (if_false_label, if_true_label);
+                   }
                }
            }
        }

diff --git a/gcc/testsuite/gcc.target/i386/pr98212.c b/gcc/testsuite/gcc.target/i386/pr98212.c
new file mode 100644 (file)
index 0000000..b8ed023
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr98212.c
@@ -0,0 +1,21 @@
+/* PR rtl-optimization/98212 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -msse2 -mfpmath=sse -mno-avx" } */
+/* { dg-final { scan-assembler-times "\tucomiss\t" 2 } } */
+/* { dg-final { scan-assembler-not "\tcomiss\t" } } */
+
+void foo (void);
+
+void
+bar (float a, float b)
+{
+  if (a != b)
+    foo ();
+}
+
+void
+baz (float a, float b)
+{
+  if (a == b)
+    foo ();
+}