Do not generate vmaddfp and vnmsubfp

This is version 3 of the patch.  This is essentially version 1 with the removal
of changes to altivec.md, and cleanup of the comments.

Version 2 generated the vmaddfp and vnmsubfp instructions if -Ofast was used,
and those changes are deleted in this patch.

The Altivec instructions vmaddfp and vnmsubfp have different rounding behaviors
than the VSX xvmaddsp and xvnmsubsp instructions due to VSCR[NJ] and other
corner cases.  In particular, generating these instructions seems to break
Eigen on big endian systems.

2023-04-09   Michael Meissner  <meissner@linux.ibm.com>

gcc/

	PR target/70243
	* config/rs6000/vsx.md (vsx_fmav4sf4): Do not generate vmaddfp.
	(vsx_nfmsv4sf4): Do not generate vnmsubfp.

gcc/testsuite/

	PR target/70243
	* gcc.target/powerpc/pr70243.c: New test.

diff --git a/gcc/config/rs6000/vsx.md b/gcc/config/rs6000/vsx.md
index 0865608..806ee43 100644 (file)
--- a/gcc/config/rs6000/vsx.md
+++ b/gcc/config/rs6000/vsx.md
@@ -2009,22 +2009,23 @@
   "x<VSv>tsqrt<sd>p %0,%x1"
   [(set_attr "type" "<VStype_simple>")])
 
-;; Fused vector multiply/add instructions. Support the classical Altivec
-;; versions of fma, which allows the target to be a separate register from the
-;; 3 inputs.  Under VSX, the target must be either the addend or the first
-;; multiply.
-
+;; Fused vector multiply/add instructions. Do not generate the Altivec versions
+;; of fma (vmaddfp and vnmsubfp).  These instructions allows the target to be a
+;; separate register from the 3 inputs, which can possibly save an extra move
+;; being generated (assuming all registers are AltiVec registers).  However,
+;; vmaddfp and vnmsubfp can have different behaviors than the VSX instructions
+;; in some corner cases due to VSCR[NJ] being set or if the addend is +0.0
+;; instead of -0.0.
 (define_insn "*vsx_fmav4sf4"
-  [(set (match_operand:V4SF 0 "vsx_register_operand" "=wa,wa,v")
+  [(set (match_operand:V4SF 0 "vsx_register_operand" "=wa,wa")
        (fma:V4SF
-         (match_operand:V4SF 1 "vsx_register_operand" "%wa,wa,v")
-         (match_operand:V4SF 2 "vsx_register_operand" "wa,0,v")
-         (match_operand:V4SF 3 "vsx_register_operand" "0,wa,v")))]
+         (match_operand:V4SF 1 "vsx_register_operand" "%wa,wa")
+         (match_operand:V4SF 2 "vsx_register_operand" "wa,0")
+         (match_operand:V4SF 3 "vsx_register_operand" "0,wa")))]
   "VECTOR_UNIT_VSX_P (V4SFmode)"
   "@
    xvmaddasp %x0,%x1,%x2
-   xvmaddmsp %x0,%x1,%x3
-   vmaddfp %0,%1,%2,%3"
+   xvmaddmsp %x0,%x1,%x3"
   [(set_attr "type" "vecfloat")])
 
 (define_insn "*vsx_fmav2df4"
@@ -2066,18 +2067,17 @@
   [(set_attr "type" "<VStype_mul>")])
 
 (define_insn "*vsx_nfmsv4sf4"
-  [(set (match_operand:V4SF 0 "vsx_register_operand" "=wa,wa,v")
+  [(set (match_operand:V4SF 0 "vsx_register_operand" "=wa,wa")
        (neg:V4SF
         (fma:V4SF
-          (match_operand:V4SF 1 "vsx_register_operand" "%wa,wa,v")
-          (match_operand:V4SF 2 "vsx_register_operand" "wa,0,v")
+          (match_operand:V4SF 1 "vsx_register_operand" "%wa,wa")
+          (match_operand:V4SF 2 "vsx_register_operand" "wa,0")
           (neg:V4SF
-            (match_operand:V4SF 3 "vsx_register_operand" "0,wa,v")))))]
+            (match_operand:V4SF 3 "vsx_register_operand" "0,wa")))))]
   "VECTOR_UNIT_VSX_P (V4SFmode)"
   "@
    xvnmsubasp %x0,%x1,%x2
-   xvnmsubmsp %x0,%x1,%x3
-   vnmsubfp %0,%1,%2,%3"
+   xvnmsubmsp %x0,%x1,%x3"
   [(set_attr "type" "vecfloat")])
 
 (define_insn "*vsx_nfmsv2df4"

diff --git a/gcc/testsuite/gcc.target/powerpc/pr70243.c b/gcc/testsuite/gcc.target/powerpc/pr70243.c
new file mode 100644 (file)
index 0000000..18a5ce7
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/pr70243.c
@@ -0,0 +1,41 @@
+/* { dg-do compile */
+/* { dg-require-effective-target powerpc_vsx_ok } */
+/* { dg-options "-O2 -mvsx" } */
+
+/* PR 70423, Make sure we don't generate vmaddfp or vnmsubfp.  These
+   instructions have different rounding modes than the VSX instructions
+   xvmaddsp and xvnmsubsp.  These tests are written where the 3 inputs and
+   target are all separate registers.  Because vmaddfp and vnmsubfp are no
+   longer generated the compiler will have to generate an xsmaddsp or xsnmsubsp
+   instruction followed by a move operation.  */
+
+#include <altivec.h>
+
+vector float
+do_add1 (vector float dummy, vector float a, vector float b, vector float c)
+{
+  return (a * b) + c;
+}
+
+vector float
+do_nsub1 (vector float dummy, vector float a, vector float b, vector float c)
+{
+  return -((a * b) - c);
+}
+
+vector float
+do_add2 (vector float dummy, vector float a, vector float b, vector float c)
+{
+  return vec_madd (a, b, c);
+}
+
+vector float
+do_nsub2 (vector float dummy, vector float a, vector float b, vector float c)
+{
+  return vec_nmsub (a, b, c);
+}
+
+/* { dg-final { scan-assembler     {\mxvmadd[am]sp\M}  } } */
+/* { dg-final { scan-assembler     {\mxvnmsub[am]sp\M} } } */
+/* { dg-final { scan-assembler-not {\mvmaddfp\M}       } } */
+/* { dg-final { scan-assembler-not {\mvnmsubfp\M}      } } */