On Fri, Sep 23, 2011 at 12:09:18AM +0200, Uros Bizjak wrote:
> > --- gcc/config/i386/i386.c.jj 2011-09-02 16:29:38.000000000 +0200
> > +++ gcc/config/i386/i386.c 2011-09-07 21:57:52.000000000 +0200
> > @@ -18304,6 +18304,11 @@ ix86_prepare_sse_fp_compare_args (rtx de
> > {
> > rtx tmp;
> >
> > + /* AVX supports all the needed comparisons, no need to swap arguments
> > + nor help reload. */
> > + if (TARGET_AVX)
> > + return code;
> > +
>
> Unfortunately, this part prevents generation of vmin/vmax instructions
> for TARGET_AVX. In ix86_expand_fp_movcc, we call
> ix86_prepare_sse_fp_compare_args, where we previously converted GT
> into LT. LT is recognized in ix86_expand_sse_fp_minmax as valid
> operand for MIN/MAX, whereas GT is not.
>
> I'm not sure if we can swap operands in ix86_expand_sse_fp_minmax,
> there is a scary comment in front of ix86_expand_sse_fp_minmax w.r.t.
> to IEEE safety.
swap_condition is documented to be IEEE safe:
/* Similar, but return the code when two operands of a comparison are swapped.
This IS safe for IEEE floating-point. */
So, do you prefer this?
2011-09-23 Jakub Jelinek <ja...@redhat.com>
* config/i386/i386.c (ix86_prepare_sse_fp_compare_args): For
GE/GT/UNLE/UNLT swap arguments and condition even for TARGET_AVX.
* gcc.target/i386/avxfp-1.c: New test.
* gcc.target/i386/avxfp-2.c: New test.
--- gcc/config/i386/i386.c.jj 2011-09-22 18:55:45.000000000 +0200
+++ gcc/config/i386/i386.c 2011-09-23 10:04:35.000000000 +0200
@@ -18739,15 +18739,13 @@ ix86_prepare_sse_fp_compare_args (rtx de
{
rtx tmp;
- /* AVX supports all the needed comparisons, no need to swap arguments
- nor help reload. */
- if (TARGET_AVX)
- return code;
-
switch (code)
{
case LTGT:
case UNEQ:
+ /* AVX supports all the needed comparisons. */
+ if (TARGET_AVX)
+ break;
/* We have no LTGT as an operator. We could implement it with
NE & ORDERED, but this requires an extra temporary. It's
not clear that it's worth it. */
@@ -18764,6 +18762,9 @@ ix86_prepare_sse_fp_compare_args (rtx de
case NE:
case UNORDERED:
case ORDERED:
+ /* AVX has 3 operand comparisons, no need to swap anything. */
+ if (TARGET_AVX)
+ break;
/* For commutative operators, try to canonicalize the destination
operand to be first in the comparison - this helps reload to
avoid extra moves. */
@@ -18775,8 +18776,10 @@ ix86_prepare_sse_fp_compare_args (rtx de
case GT:
case UNLE:
case UNLT:
- /* These are not supported directly. Swap the comparison operands
- to transform into something that is supported. */
+ /* These are not supported directly before AVX, and furthermore
+ ix86_expand_sse_fp_minmax only optimizes LT/UNGE. Swap the
+ comparison operands to transform into something that is
+ supported. */
tmp = *pop0;
*pop0 = *pop1;
*pop1 = tmp;
--- gcc/testsuite/gcc.target/i386/avxfp-1.c.jj 2011-09-23 10:07:48.000000000
+0200
+++ gcc/testsuite/gcc.target/i386/avxfp-1.c 2011-09-23 10:08:02.000000000
+0200
@@ -0,0 +1,15 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -mavx -mfpmath=sse" } */
+/* { dg-final { scan-assembler "vmaxsd" } } */
+/* { dg-final { scan-assembler "vminsd" } } */
+double x;
+t()
+{
+ x=x>5?x:5;
+}
+
+double x;
+q()
+{
+ x=x<5?x:5;
+}
--- gcc/testsuite/gcc.target/i386/avxfp-2.c.jj 2011-09-23 10:07:51.000000000
+0200
+++ gcc/testsuite/gcc.target/i386/avxfp-2.c 2011-09-23 10:08:17.000000000
+0200
@@ -0,0 +1,15 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -mavx -mfpmath=sse" } */
+/* { dg-final { scan-assembler "vmaxsd" } } */
+/* { dg-final { scan-assembler "vminsd" } } */
+double x;
+q()
+{
+ x=x<5?5:x;
+}
+
+double x;
+q1()
+{
+ x=x>5?5:x;
+}
Jakub
