]> git.ipfire.org Git - thirdparty/gcc.git/commitdiff
i386: Fix ICE with conditional QI/HI vector maxmin [PR118776]
authorJakub Jelinek <jakub@redhat.com>
Sat, 8 Feb 2025 07:54:31 +0000 (08:54 +0100)
committerJakub Jelinek <jakub@gcc.gnu.org>
Sat, 8 Feb 2025 07:54:31 +0000 (08:54 +0100)
The following testcase ICEs starting with GCC 12 since r12-4526
although the bug has been introduced already in r12-2751.
The problem was in the addition of cond_<code><mode> define_expand
which uses nonimmediate_operand predicates for both maxmin operands
for all VI1248_AVX512VLBW modes.  It works fine with
VI48_AVX512VL modes because the <code><mode>3_mask VI48_AVX512VL
define_expand uses ix86_fixup_binary_operands_no_copy and the
*avx512f_<code><mode>3<mask_name> VI48_AVX512VL define_insn uses
% in constraint and !(MEM_P && MEM_P) check in condition (and
<code><mode>3 define_expand with VI124_256_AVX512F_AVX512BW iterator
does that too), but eventhough the 8-bit and 16-bit element maxmin
is commutative too, the <mask_codefor><code><mode>3<mask_name>
define_insn with VI12_AVX512VL iterator didn't use % in constraint
to make it commutative.  So, e.g. cond_umaxv32qi define_expand
allowed nonimmediate_operand for both umax operands, but used
gen_umaxv32qi_mask which wasn't commutative and only allowed
nonimmediate_operand for the second operand.

The following patch fixes it by keeping the <code><mode>3
VI124_256_AVX512F_AVX512BW define_expand as is (it does
ix86_fixup_binary_operands_no_copy) but extending the
<code><mode>3_mask define_expand from VI48_AVX512VL to
VI1248_AVX512VLBW which keeps the current modes with their
ISA conditions and adds the VI12_AVX512VL modes under additional
TARGET_AVX512BW condition, and turning the actual define_insn
into an * prefixed name (which it was before just for the non-masked
case) and having the same commutative operand handling as in other
define_insns.

2025-02-08  Jakub Jelinek  <jakub@redhat.com>

PR target/118776
* config/i386/sse.md (<code><mode>3_mask): Use VI1248_AVX512VLBW
iterator rather than VI48_AVX512VL.
(<mask_codefor><code><mode>3<mask_name>): Rename to ...
(*avx512bw_<code><mode>3<mask_name>): ... this.  Use
nonimmediate_operand rather than register_operand predicate and %v
rather than v constraint for operand 1 and adjust condition to reject
MEMs in both operand 1 and 2.

* gcc.target/i386/pr118776.c: New test.

gcc/config/i386/sse.md
gcc/testsuite/gcc.target/i386/pr118776.c [new file with mode: 0644]

index 9c495ffa40c688d85389d2df854d958ed1b1d1b9..7b39103d99f02f0738978adf3dc0680c10f2b2fc 100644 (file)
 })
 
 (define_expand "<code><mode>3_mask"
-  [(set (match_operand:VI48_AVX512VL 0 "register_operand")
-       (vec_merge:VI48_AVX512VL
-         (maxmin:VI48_AVX512VL
-           (match_operand:VI48_AVX512VL 1 "nonimmediate_operand")
-           (match_operand:VI48_AVX512VL 2 "nonimmediate_operand"))
-         (match_operand:VI48_AVX512VL 3 "nonimm_or_0_operand")
+  [(set (match_operand:VI1248_AVX512VLBW 0 "register_operand")
+       (vec_merge:VI1248_AVX512VLBW
+         (maxmin:VI1248_AVX512VLBW
+           (match_operand:VI1248_AVX512VLBW 1 "nonimmediate_operand")
+           (match_operand:VI1248_AVX512VLBW 2 "nonimmediate_operand"))
+         (match_operand:VI1248_AVX512VLBW 3 "nonimm_or_0_operand")
          (match_operand:<avx512fmaskmode> 4 "register_operand")))]
   "TARGET_AVX512F"
   "ix86_fixup_binary_operands_no_copy (<CODE>, <MODE>mode, operands);")
    (set_attr "prefix" "maybe_evex")
    (set_attr "mode" "<sseinsnmode>")])
 
-(define_insn "<mask_codefor><code><mode>3<mask_name>"
+(define_insn "*avx512bw_<code><mode>3<mask_name>"
   [(set (match_operand:VI12_AVX512VL 0 "register_operand" "=v")
         (maxmin:VI12_AVX512VL
-          (match_operand:VI12_AVX512VL 1 "register_operand" "v")
+          (match_operand:VI12_AVX512VL 1 "nonimmediate_operand" "%v")
           (match_operand:VI12_AVX512VL 2 "nonimmediate_operand" "vm")))]
-  "TARGET_AVX512BW"
+  "TARGET_AVX512BW && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
   "vp<maxmin_int><ssemodesuffix>\t{%2, %1, %0<mask_operand3>|%0<mask_operand3>, %1, %2}"
   [(set_attr "type" "sseiadd")
    (set_attr "prefix" "evex")
diff --git a/gcc/testsuite/gcc.target/i386/pr118776.c b/gcc/testsuite/gcc.target/i386/pr118776.c
new file mode 100644 (file)
index 0000000..44c18ca
--- /dev/null
@@ -0,0 +1,23 @@
+/* PR target/118776 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -mavx512bw -mavx512vl" } */
+
+void bar (unsigned char *);
+
+void
+foo (unsigned char *x)
+{
+  unsigned char b[32];
+  bar (b);
+  for (int i = 0; i < 32; i++)
+    {
+      unsigned char c = 8;
+      if (i > 3)
+       {
+         unsigned char d = b[i];
+         d = 1 > d ? 1 : d;
+         c = d;
+       }
+      x[i] = c;
+    }
+}