[PATCH, rs6000] Disable gimple fold for float or double vec_minmax when fast-math is not set

HAO CHEN GUI guihaoc@linux.ibm.com
Tue Nov 2 02:40:21 GMT 2021


David,

    My patch file was broken. I am sorry for it.  Here is the correct one. Thanks a lot.

ChangeLog

2021-11-01 Haochen Gui <guihaoc@linux.ibm.com>

gcc/
        * config/rs6000/rs6000-call.c (rs6000_gimple_fold_builtin): Disable
        gimple fold for VSX_BUILTIN_XVMINDP, ALTIVEC_BUILTIN_VMINFP,
        VSX_BUILTIN_XVMAXDP, ALTIVEC_BUILTIN_VMAXFP when fast-math is not
        set.

gcc/testsuite/
        * gcc.target/powerpc/vec-minmax-1.c: New test.
        * gcc.target/powerpc/vec-minmax-2.c: Likewise.


patch.diff

diff --git a/gcc/config/rs6000/rs6000-call.c b/gcc/config/rs6000/rs6000-call.c
index 7d485480225..a8e193a0089 100644
--- a/gcc/config/rs6000/rs6000-call.c
+++ b/gcc/config/rs6000/rs6000-call.c
@@ -12333,6 +12333,14 @@ rs6000_gimple_fold_builtin (gimple_stmt_iterator *gsi)
       return true;
     /* flavors of vec_min.  */
     case VSX_BUILTIN_XVMINDP:
+    case ALTIVEC_BUILTIN_VMINFP:
+      {
+       lhs = gimple_call_lhs (stmt);
+       tree type = TREE_TYPE (lhs);
+       if (HONOR_NANS (type))
+         return false;
+       gcc_fallthrough ();
+      }
     case P8V_BUILTIN_VMINSD:
     case P8V_BUILTIN_VMINUD:
     case ALTIVEC_BUILTIN_VMINSB:
@@ -12341,7 +12349,6 @@ rs6000_gimple_fold_builtin (gimple_stmt_iterator *gsi)
     case ALTIVEC_BUILTIN_VMINUB:
     case ALTIVEC_BUILTIN_VMINUH:
     case ALTIVEC_BUILTIN_VMINUW:
-    case ALTIVEC_BUILTIN_VMINFP:
       arg0 = gimple_call_arg (stmt, 0);
       arg1 = gimple_call_arg (stmt, 1);
       lhs = gimple_call_lhs (stmt);
@@ -12351,6 +12358,14 @@ rs6000_gimple_fold_builtin (gimple_stmt_iterator *gsi)
       return true;
     /* flavors of vec_max.  */
     case VSX_BUILTIN_XVMAXDP:
+    case ALTIVEC_BUILTIN_VMAXFP:
+      {
+       lhs = gimple_call_lhs (stmt);
+       tree type = TREE_TYPE (lhs);
+       if (HONOR_NANS (type))
+         return false;
+       gcc_fallthrough ();
+      }
     case P8V_BUILTIN_VMAXSD:
     case P8V_BUILTIN_VMAXUD:
     case ALTIVEC_BUILTIN_VMAXSB:
@@ -12359,7 +12374,6 @@ rs6000_gimple_fold_builtin (gimple_stmt_iterator *gsi)
     case ALTIVEC_BUILTIN_VMAXUB:
     case ALTIVEC_BUILTIN_VMAXUH:
     case ALTIVEC_BUILTIN_VMAXUW:
-    case ALTIVEC_BUILTIN_VMAXFP:
       arg0 = gimple_call_arg (stmt, 0);
       arg1 = gimple_call_arg (stmt, 1);
       lhs = gimple_call_lhs (stmt);
diff --git a/gcc/testsuite/gcc.target/powerpc/vec-minmax-1.c b/gcc/testsuite/gcc.target/powerpc/vec-minmax-1.c
new file mode 100644
index 00000000000..e238659c9be
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/vec-minmax-1.c
@@ -0,0 +1,52 @@
+/* { dg-require-effective-target powerpc_p9vector_ok } */
+/* { dg-options "-O2 -mdejagnu-cpu=power9" } */
+/* { dg-final { scan-assembler-times {\mxvmaxdp\M} 1 } } */
+/* { dg-final { scan-assembler-times {\mxvmaxsp\M} 1 } } */
+/* { dg-final { scan-assembler-times {\mxvmindp\M} 1 } } */
+/* { dg-final { scan-assembler-times {\mxvminsp\M} 1 } } */
+
+/* This test verifies that float or double vec_min/max are bound to
+   xv[min|max][d|s]p instructions when fast-math is not set.  */
+
+
+#include <altivec.h>
+
+#ifdef _BIG_ENDIAN
+   const int PREF_D = 0;
+#else
+   const int PREF_D = 1;
+#endif
+
+double vmaxd (double a, double b)
+{
+  vector double va = vec_promote (a, PREF_D);
+  vector double vb = vec_promote (b, PREF_D);
+  return vec_extract (vec_max (va, vb), PREF_D);
+}
+
+double vmind (double a, double b)
+{
+  vector double va = vec_promote (a, PREF_D);
+  vector double vb = vec_promote (b, PREF_D);
+  return vec_extract (vec_min (va, vb), PREF_D);
+}
+
+#ifdef _BIG_ENDIAN
+   const int PREF_F = 0;
+#else
+   const int PREF_F = 3;
+#endif
+
+float vmaxf (float a, float b)
+{
+  vector float va = vec_promote (a, PREF_F);
+  vector float vb = vec_promote (b, PREF_F);
+  return vec_extract (vec_max (va, vb), PREF_F);
+}
+
+float vminf (float a, float b)
+{
+  vector float va = vec_promote (a, PREF_F);
+  vector float vb = vec_promote (b, PREF_F);
+  return vec_extract (vec_min (va, vb), PREF_F);
+}
diff --git a/gcc/testsuite/gcc.target/powerpc/vec-minmax-2.c b/gcc/testsuite/gcc.target/powerpc/vec-minmax-2.c
new file mode 100644
index 00000000000..149275d8709
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/vec-minmax-2.c
@@ -0,0 +1,50 @@
+/* { dg-require-effective-target powerpc_p9vector_ok } */
+/* { dg-options "-O2 -mdejagnu-cpu=power9 -ffast-math" } */
+/* { dg-final { scan-assembler-times {\mxsmaxcdp\M} 2 } } */
+/* { dg-final { scan-assembler-times {\mxsmincdp\M} 2 } } */
+
+/* This test verifies that float or double vec_min/max can be converted
+   to scalar comparison when fast-math is set.  */
+
+
+#include <altivec.h>
+
+#ifdef _BIG_ENDIAN
+   const int PREF_D = 0;
+#else
+   const int PREF_D = 1;
+#endif
+
+double vmaxd (double a, double b)
+{
+  vector double va = vec_promote (a, PREF_D);
+  vector double vb = vec_promote (b, PREF_D);
+  return vec_extract (vec_max (va, vb), PREF_D);
+}
+
+double vmind (double a, double b)
+{
+  vector double va = vec_promote (a, PREF_D);
+  vector double vb = vec_promote (b, PREF_D);
+  return vec_extract (vec_min (va, vb), PREF_D);
+}
+
+#ifdef _BIG_ENDIAN
+   const int PREF_F = 0;
+#else
+   const int PREF_F = 3;
+#endif
+
+float vmaxf (float a, float b)
+{
+  vector float va = vec_promote (a, PREF_F);
+  vector float vb = vec_promote (b, PREF_F);
+  return vec_extract (vec_max (va, vb), PREF_F);
+}
+
+float vminf (float a, float b)
+{
+  vector float va = vec_promote (a, PREF_F);
+  vector float vb = vec_promote (b, PREF_F);
+  return vec_extract (vec_min (va, vb), PREF_F);
+}


 

On 2/11/2021 上午 8:21, David Edelsohn wrote:
> Hi, Hao
>
> Neither the inlined patch nor the attached patch seem to contain the
> change to rs6000-call.c.  I only see the new testcases.
>
> Please resend the complete patch.
>
> Thanks David
>
> On Mon, Nov 1, 2021 at 2:48 AM HAO CHEN GUI <guihaoc@linux.ibm.com> wrote:
>> Hi,
>>
>>   This patch disables gimple folding for VSX_BUILTIN_XVMINDP, VSX_BUILTIN_XVMAXDP, ALTIVEC_BUILTIN_VMINFP and  ALTIVEC_BUILTIN_VMAXFP when fast-math is not set.  With the gimple folding is enabled, the four built-ins will be implemented by c-type instructions - xs[min|max]cdp on P9 and P10 if they can be converted to scalar comparisons.  While they are implemented by xv[min|max][s|d]p on P8 and P7 as P8 and P7 don't have corresponding scalar comparison instructions.  The patch binds these four built-ins to xv[min|max][s|d]p when fast-math is not set. The two new test cases illustrate it.
>>
>>   ALTIVEC_BUILTIN_VMINFP and  ALTIVEC_BUILTIN_VMAXFP are not implemented by vminfp or vmaxfp.
>>
>> rs6000-builtin.def:BU_ALTIVEC_2 (VMAXFP,              "vmaxfp",         CONST, smaxv4sf3)
>>
>> rs6000-builtin.def:BU_ALTIVEC_2 (VMINFP,              "vminfp",         CONST, sminv4sf3)
>>
>> Bootstrapped and tested on powerpc64le-linux with no regressions. Is this okay for trunk? Any recommendations? Thanks a lot.
>>
>>
>> ChangeLog
>>
>> 2021-11-01 Haochen Gui <guihaoc@linux.ibm.com>
>>
>> gcc/
>>         * config/rs6000/rs6000-call.c (rs6000_gimple_fold_builtin): Disable
>>         gimple fold for VSX_BUILTIN_XVMINDP, ALTIVEC_BUILTIN_VMINFP,
>>         VSX_BUILTIN_XVMAXDP, ALTIVEC_BUILTIN_VMAXFP when fast-math is not
>>         set.
>>
>> gcc/testsuite/
>>         * gcc.target/powerpc/vec-minmax-1.c: New test.
>>         * gcc.target/powerpc/vec-minmax-2.c: Likewise.
>>
>>
>> patch.diff
>>
>> diff --git a/gcc/testsuite/gcc.target/powerpc/vec-minmax-1.c b/gcc/testsuite/gcc.target/powerpc/vec-minmax-1.c
>> new file mode 100644
>> index 00000000000..e238659c9be
>> --- /dev/null
>> +++ b/gcc/testsuite/gcc.target/powerpc/vec-minmax-1.c
>> @@ -0,0 +1,52 @@
>> +/* { dg-require-effective-target powerpc_p9vector_ok } */
>> +/* { dg-options "-O2 -mdejagnu-cpu=power9" } */
>> +/* { dg-final { scan-assembler-times {\mxvmaxdp\M} 1 } } */
>> +/* { dg-final { scan-assembler-times {\mxvmaxsp\M} 1 } } */
>> +/* { dg-final { scan-assembler-times {\mxvmindp\M} 1 } } */
>> +/* { dg-final { scan-assembler-times {\mxvminsp\M} 1 } } */
>> +
>> +/* This test verifies that float or double vec_min/max are bound to
>> +   xv[min|max][d|s]p instructions when fast-math is not set.  */
>> +
>> +
>> +#include <altivec.h>
>> +
>> +#ifdef _BIG_ENDIAN
>> +   const int PREF_D = 0;
>> +#else
>> +   const int PREF_D = 1;
>> +#endif
>> +
>> +double vmaxd (double a, double b)
>> +{
>> +  vector double va = vec_promote (a, PREF_D);
>> +  vector double vb = vec_promote (b, PREF_D);
>> +  return vec_extract (vec_max (va, vb), PREF_D);
>> +}
>> +
>> +double vmind (double a, double b)
>> +{
>> +  vector double va = vec_promote (a, PREF_D);
>> +  vector double vb = vec_promote (b, PREF_D);
>> +  return vec_extract (vec_min (va, vb), PREF_D);
>> +}
>> +
>> +#ifdef _BIG_ENDIAN
>> +   const int PREF_F = 0;
>> +#else
>> +   const int PREF_F = 3;
>> +#endif
>> +
>> +float vmaxf (float a, float b)
>> +{
>> +  vector float va = vec_promote (a, PREF_F);
>> +  vector float vb = vec_promote (b, PREF_F);
>> +  return vec_extract (vec_max (va, vb), PREF_F);
>> +}
>> +
>> +float vminf (float a, float b)
>> +{
>> +  vector float va = vec_promote (a, PREF_F);
>> +  vector float vb = vec_promote (b, PREF_F);
>> +  return vec_extract (vec_min (va, vb), PREF_F);
>> +}
>> diff --git a/gcc/testsuite/gcc.target/powerpc/vec-minmax-2.c b/gcc/testsuite/gcc.target/powerpc/vec-minmax-2.c
>> new file mode 100644
>> index 00000000000..149275d8709
>> --- /dev/null
>> +++ b/gcc/testsuite/gcc.target/powerpc/vec-minmax-2.c
>> @@ -0,0 +1,50 @@
>> +/* { dg-require-effective-target powerpc_p9vector_ok } */
>> +/* { dg-options "-O2 -mdejagnu-cpu=power9 -ffast-math" } */
>> +/* { dg-final { scan-assembler-times {\mxsmaxcdp\M} 2 } } */
>> +/* { dg-final { scan-assembler-times {\mxsmincdp\M} 2 } } */
>> +
>> +/* This test verifies that float or double vec_min/max can be converted
>> +   to scalar comparison when fast-math is set.  */
>> +
>> +
>> +#include <altivec.h>
>> +
>> +#ifdef _BIG_ENDIAN
>> +   const int PREF_D = 0;
>> +#else
>> +   const int PREF_D = 1;
>> +#endif
>> +
>> +double vmaxd (double a, double b)
>> +{
>> +  vector double va = vec_promote (a, PREF_D);
>> +  vector double vb = vec_promote (b, PREF_D);
>> +  return vec_extract (vec_max (va, vb), PREF_D);
>> +}
>> +
>> +double vmind (double a, double b)
>> +{
>> +  vector double va = vec_promote (a, PREF_D);
>> +  vector double vb = vec_promote (b, PREF_D);
>> +  return vec_extract (vec_min (va, vb), PREF_D);
>> +}
>> +
>> +#ifdef _BIG_ENDIAN
>> +   const int PREF_F = 0;
>> +#else
>> +   const int PREF_F = 3;
>> +#endif
>> +
>> +float vmaxf (float a, float b)
>> +{
>> +  vector float va = vec_promote (a, PREF_F);
>> +  vector float vb = vec_promote (b, PREF_F);
>> +  return vec_extract (vec_max (va, vb), PREF_F);
>> +}
>> +
>> +float vminf (float a, float b)
>> +{
>> +  vector float va = vec_promote (a, PREF_F);
>> +  vector float vb = vec_promote (b, PREF_F);
>> +  return vec_extract (vec_min (va, vb), PREF_F);
>> +}
-------------- next part --------------
2021-11-01 Haochen Gui <guihaoc@linux.ibm.com>

gcc/
	* config/rs6000/rs6000-call.c (rs6000_gimple_fold_builtin): Disable
	gimple fold for VSX_BUILTIN_XVMINDP, ALTIVEC_BUILTIN_VMINFP,
	VSX_BUILTIN_XVMAXDP, ALTIVEC_BUILTIN_VMAXFP when fast-math is not
	set.

gcc/testsuite/
	* gcc.target/powerpc/vec-minmax-1.c: New test.
	* gcc.target/powerpc/vec-minmax-2.c: Likewise.
-------------- next part --------------
diff --git a/gcc/config/rs6000/rs6000-call.c b/gcc/config/rs6000/rs6000-call.c
index 7d485480225..a8e193a0089 100644
--- a/gcc/config/rs6000/rs6000-call.c
+++ b/gcc/config/rs6000/rs6000-call.c
@@ -12333,6 +12333,14 @@ rs6000_gimple_fold_builtin (gimple_stmt_iterator *gsi)
       return true;
     /* flavors of vec_min.  */
     case VSX_BUILTIN_XVMINDP:
+    case ALTIVEC_BUILTIN_VMINFP:
+      {
+	lhs = gimple_call_lhs (stmt);
+	tree type = TREE_TYPE (lhs);
+	if (HONOR_NANS (type))
+	  return false;
+	gcc_fallthrough ();
+      }
     case P8V_BUILTIN_VMINSD:
     case P8V_BUILTIN_VMINUD:
     case ALTIVEC_BUILTIN_VMINSB:
@@ -12341,7 +12349,6 @@ rs6000_gimple_fold_builtin (gimple_stmt_iterator *gsi)
     case ALTIVEC_BUILTIN_VMINUB:
     case ALTIVEC_BUILTIN_VMINUH:
     case ALTIVEC_BUILTIN_VMINUW:
-    case ALTIVEC_BUILTIN_VMINFP:
       arg0 = gimple_call_arg (stmt, 0);
       arg1 = gimple_call_arg (stmt, 1);
       lhs = gimple_call_lhs (stmt);
@@ -12351,6 +12358,14 @@ rs6000_gimple_fold_builtin (gimple_stmt_iterator *gsi)
       return true;
     /* flavors of vec_max.  */
     case VSX_BUILTIN_XVMAXDP:
+    case ALTIVEC_BUILTIN_VMAXFP:
+      {
+	lhs = gimple_call_lhs (stmt);
+	tree type = TREE_TYPE (lhs);
+	if (HONOR_NANS (type))
+	  return false;
+	gcc_fallthrough ();
+      }
     case P8V_BUILTIN_VMAXSD:
     case P8V_BUILTIN_VMAXUD:
     case ALTIVEC_BUILTIN_VMAXSB:
@@ -12359,7 +12374,6 @@ rs6000_gimple_fold_builtin (gimple_stmt_iterator *gsi)
     case ALTIVEC_BUILTIN_VMAXUB:
     case ALTIVEC_BUILTIN_VMAXUH:
     case ALTIVEC_BUILTIN_VMAXUW:
-    case ALTIVEC_BUILTIN_VMAXFP:
       arg0 = gimple_call_arg (stmt, 0);
       arg1 = gimple_call_arg (stmt, 1);
       lhs = gimple_call_lhs (stmt);
diff --git a/gcc/testsuite/gcc.target/powerpc/vec-minmax-1.c b/gcc/testsuite/gcc.target/powerpc/vec-minmax-1.c
new file mode 100644
index 00000000000..e238659c9be
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/vec-minmax-1.c
@@ -0,0 +1,52 @@
+/* { dg-require-effective-target powerpc_p9vector_ok } */
+/* { dg-options "-O2 -mdejagnu-cpu=power9" } */
+/* { dg-final { scan-assembler-times {\mxvmaxdp\M} 1 } } */
+/* { dg-final { scan-assembler-times {\mxvmaxsp\M} 1 } } */
+/* { dg-final { scan-assembler-times {\mxvmindp\M} 1 } } */
+/* { dg-final { scan-assembler-times {\mxvminsp\M} 1 } } */
+
+/* This test verifies that float or double vec_min/max are bound to
+   xv[min|max][d|s]p instructions when fast-math is not set.  */
+
+
+#include <altivec.h>
+
+#ifdef _BIG_ENDIAN
+   const int PREF_D = 0;
+#else
+   const int PREF_D = 1;
+#endif
+
+double vmaxd (double a, double b)
+{
+  vector double va = vec_promote (a, PREF_D);
+  vector double vb = vec_promote (b, PREF_D);
+  return vec_extract (vec_max (va, vb), PREF_D);
+}
+
+double vmind (double a, double b)
+{
+  vector double va = vec_promote (a, PREF_D);
+  vector double vb = vec_promote (b, PREF_D);
+  return vec_extract (vec_min (va, vb), PREF_D);
+}
+
+#ifdef _BIG_ENDIAN
+   const int PREF_F = 0;
+#else
+   const int PREF_F = 3;
+#endif
+
+float vmaxf (float a, float b)
+{
+  vector float va = vec_promote (a, PREF_F);
+  vector float vb = vec_promote (b, PREF_F);
+  return vec_extract (vec_max (va, vb), PREF_F);
+}
+
+float vminf (float a, float b)
+{
+  vector float va = vec_promote (a, PREF_F);
+  vector float vb = vec_promote (b, PREF_F);
+  return vec_extract (vec_min (va, vb), PREF_F);
+}
diff --git a/gcc/testsuite/gcc.target/powerpc/vec-minmax-2.c b/gcc/testsuite/gcc.target/powerpc/vec-minmax-2.c
new file mode 100644
index 00000000000..149275d8709
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/vec-minmax-2.c
@@ -0,0 +1,50 @@
+/* { dg-require-effective-target powerpc_p9vector_ok } */
+/* { dg-options "-O2 -mdejagnu-cpu=power9 -ffast-math" } */
+/* { dg-final { scan-assembler-times {\mxsmaxcdp\M} 2 } } */
+/* { dg-final { scan-assembler-times {\mxsmincdp\M} 2 } } */
+
+/* This test verifies that float or double vec_min/max can be converted
+   to scalar comparison when fast-math is set.  */
+
+
+#include <altivec.h>
+
+#ifdef _BIG_ENDIAN
+   const int PREF_D = 0;
+#else
+   const int PREF_D = 1;
+#endif
+
+double vmaxd (double a, double b)
+{
+  vector double va = vec_promote (a, PREF_D);
+  vector double vb = vec_promote (b, PREF_D);
+  return vec_extract (vec_max (va, vb), PREF_D);
+}
+
+double vmind (double a, double b)
+{
+  vector double va = vec_promote (a, PREF_D);
+  vector double vb = vec_promote (b, PREF_D);
+  return vec_extract (vec_min (va, vb), PREF_D);
+}
+
+#ifdef _BIG_ENDIAN
+   const int PREF_F = 0;
+#else
+   const int PREF_F = 3;
+#endif
+
+float vmaxf (float a, float b)
+{
+  vector float va = vec_promote (a, PREF_F);
+  vector float vb = vec_promote (b, PREF_F);
+  return vec_extract (vec_max (va, vb), PREF_F);
+}
+
+float vminf (float a, float b)
+{
+  vector float va = vec_promote (a, PREF_F);
+  vector float vb = vec_promote (b, PREF_F);
+  return vec_extract (vec_min (va, vb), PREF_F);
+}


More information about the Gcc-patches mailing list