[gcc r12-5319] rs6000: MMA test case emits wrong code when building a vector pair [PR102976]

[Peter Bergner <bergner@gcc.gnu.org>]

https://gcc.gnu.org/g:4cdf7db9a39d18bd536d816a5751d4d3cf23808b

commit r12-5319-g4cdf7db9a39d18bd536d816a5751d4d3cf23808b
Author: Peter Bergner <bergner@linux.ibm.com>
Date:   Tue Nov 16 12:14:22 2021 -0600

    rs6000: MMA test case emits wrong code when building a vector pair [PR102976]
    
    PR102976 shows a test case where we generate wrong code when building
    a vector pair from 2 vector registers.  The bug here is that with unlucky
    register assignments, we can clobber one of the input operands before
    we write both registers of the output operand.  The solution is to use
    early-clobbers in the assemble pair and accumulator patterns.
    
    2021-11-16  Peter Bergner  <bergner@linux.ibm.com>
    
    gcc/
            PR target/102976
            * config/rs6000/mma.md (*vsx_assemble_pair): Add early-clobber for
            output operand.
            (*mma_assemble_acc): Likewise.
    
    gcc/testsuite/
            PR target/102976
            * gcc.target/powerpc/pr102976.c: New test.

Diff:
---
 gcc/config/rs6000/mma.md                    | 10 ++++++++--
 gcc/testsuite/gcc.target/powerpc/pr102976.c | 14 ++++++++++++++
 2 files changed, 22 insertions(+), 2 deletions(-)

diff --git a/gcc/config/rs6000/mma.md b/gcc/config/rs6000/mma.md
index 1990a2183f6..fa081608c4c 100644
--- a/gcc/config/rs6000/mma.md
+++ b/gcc/config/rs6000/mma.md
@@ -338,8 +338,11 @@
   DONE;
 })
 
+;; We cannot update the two output registers atomically, so mark the output
+;; as an early clobber so we don't accidentally clobber the input operands.  */
+
 (define_insn_and_split "*vsx_assemble_pair"
-  [(set (match_operand:OO 0 "vsx_register_operand" "=wa")
+  [(set (match_operand:OO 0 "vsx_register_operand" "=&wa")
 	(unspec:OO [(match_operand:V16QI 1 "mma_assemble_input_operand" "mwa")
 		    (match_operand:V16QI 2 "mma_assemble_input_operand" "mwa")]
 		    UNSPEC_MMA_ASSEMBLE))]
@@ -404,8 +407,11 @@
   DONE;
 })
 
+;; We cannot update the four output registers atomically, so mark the output
+;; as an early clobber so we don't accidentally clobber the input operands.  */
+
 (define_insn_and_split "*mma_assemble_acc"
-  [(set (match_operand:XO 0 "fpr_reg_operand" "=d")
+  [(set (match_operand:XO 0 "fpr_reg_operand" "=&d")
 	(unspec:XO [(match_operand:V16QI 1 "mma_assemble_input_operand" "mwa")
 		    (match_operand:V16QI 2 "mma_assemble_input_operand" "mwa")
 		    (match_operand:V16QI 3 "mma_assemble_input_operand" "mwa")
diff --git a/gcc/testsuite/gcc.target/powerpc/pr102976.c b/gcc/testsuite/gcc.target/powerpc/pr102976.c
new file mode 100644
index 00000000000..5a4320f8e0a
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/pr102976.c
@@ -0,0 +1,14 @@
+/* { dg-require-effective-target power10_ok } */
+/* { dg-options "-O2 -mdejagnu-cpu=power10" } */
+
+#include <altivec.h>
+void
+bug (__vector_pair *dst)
+{
+  register vector unsigned char vec0 asm ("vs44");
+  register vector unsigned char vec1 asm ("vs32");
+  __builtin_vsx_build_pair (dst, vec0, vec1);
+}
+
+/* { dg-final { scan-assembler-times {(?p)\mxxlor \d+,44,44\M} 1 } } */
+/* { dg-final { scan-assembler-times {(?p)\mxxlor \d+,32,32\M} 1 } } */