[PATCH][AArch64] PR target/84748: Mark *compare_cstore<mode>_insn as clobbering CC reg

[PATCH][AArch64] PR target/84748: Mark *compare_cstore<mode>_insn as clobbering CC reg

[Kyrill Tkachov]

[-- Attachment #1: Type: text/plain, Size: 1116 bytes --]

Hi all,

In this wrong-code PR the combine pass ends up moving a CC-using instruction past a *compare_cstore<mode>_insn
insn_and_split. After reload the *compare_cstore<mode>_insn splitter ends up generating a SUBS instruction that
clobbers the condition flags, and things go bad.

The solution is simple, the *compare_cstore<mode>_insn pattern should specify that it clobbers the CC register
so that combine (or any other pass) does not assume that it can move CC-using patterns across it.

This patch does that and fixes the testcase.

The testcase FAILs on GCC 8 only, but the buggy pattern is in GCC 6 onwards, so we should backport this as
a latent bug fix after it's had some time to bake in trunk.

Bootstrapped and tested on aarch64-none-linux-gnu.

Ok for trunk and later backports?

Thanks,
Kyrill

2018-03-07  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

     PR target/84748
     * config/aarch64/aarch64.md (*compare_cstore<mode>_insn): Mark pattern
     as clobbering CC_REGNUM.

2018-03-07  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

     PR target/84748
     * gcc.c-torture/execute/pr84748.c: New test.

[-- Attachment #2: to-test.patch --]
[-- Type: text/x-patch, Size: 1565 bytes --]

commit 9c008e450e1cd95f2905071e6c8a0ff92b856358
Author: Kyrylo Tkachov <kyrylo.tkachov@arm.com>
Date:   Wed Mar 7 15:05:25 2018 +0000

    [AArch64] PR target/84748: Mark *compare_cstore<mode>_insn as clobbering CC reg

diff --git a/gcc/config/aarch64/aarch64.md b/gcc/config/aarch64/aarch64.md
index 69ff5ca..4463a13 100644
--- a/gcc/config/aarch64/aarch64.md
+++ b/gcc/config/aarch64/aarch64.md
@@ -3242,7 +3242,8 @@ (define_insn "aarch64_cstore<mode>"
 (define_insn_and_split "*compare_cstore<mode>_insn"
   [(set (match_operand:GPI 0 "register_operand" "=r")
 	 (EQL:GPI (match_operand:GPI 1 "register_operand" "r")
-		  (match_operand:GPI 2 "aarch64_imm24" "n")))]
+		  (match_operand:GPI 2 "aarch64_imm24" "n")))
+   (clobber (reg:CC CC_REGNUM))]
   "!aarch64_move_imm (INTVAL (operands[2]), <MODE>mode)
    && !aarch64_plus_operand (operands[2], <MODE>mode)
    && !reload_completed"
diff --git a/gcc/testsuite/gcc.c-torture/execute/pr84748.c b/gcc/testsuite/gcc.c-torture/execute/pr84748.c
new file mode 100644
index 0000000..9572ab2
--- /dev/null
+++ b/gcc/testsuite/gcc.c-torture/execute/pr84748.c
@@ -0,0 +1,34 @@
+/* { dg-require-effective-target int128 } */
+
+typedef unsigned __int128 u128;
+
+int a, c, d;
+u128 b;
+
+unsigned long long g0, g1;
+
+void
+store (unsigned long long a0, unsigned long long a1)
+{
+  g0 = a0;
+  g1 = a1;
+}
+
+void
+foo (void)
+{
+  b += a;
+  c = d != 84347;
+  b /= c;
+  u128 x = b;
+  store (x >> 0, x >> 64);
+}
+
+int
+main (void)
+{
+  foo ();
+  if (g0 != 0 || g1 != 0)
+    __builtin_abort ();
+  return 0;
+}

Re: [PATCH][AArch64] PR target/84748: Mark *compare_cstore<mode>_insn as clobbering CC reg

[James Greenhalgh]

On Wed, Mar 07, 2018 at 05:39:41PM +0000, Kyrill Tkachov wrote:
> Hi all,
> 
> In this wrong-code PR the combine pass ends up moving a CC-using instruction past a *compare_cstore<mode>_insn
> insn_and_split. After reload the *compare_cstore<mode>_insn splitter ends up generating a SUBS instruction that
> clobbers the condition flags, and things go bad.
> 
> The solution is simple, the *compare_cstore<mode>_insn pattern should specify that it clobbers the CC register
> so that combine (or any other pass) does not assume that it can move CC-using patterns across it.
> 
> This patch does that and fixes the testcase.
> 
> The testcase FAILs on GCC 8 only, but the buggy pattern is in GCC 6 onwards, so we should backport this as
> a latent bug fix after it's had some time to bake in trunk.
> 
> Bootstrapped and tested on aarch64-none-linux-gnu.
> 
> Ok for trunk and later backports?

OK.

Thanks,
James

> 2018-03-07  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
> 
>      PR target/84748
>      * config/aarch64/aarch64.md (*compare_cstore<mode>_insn): Mark pattern
>      as clobbering CC_REGNUM.
> 
> 2018-03-07  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
> 
>      PR target/84748
>      * gcc.c-torture/execute/pr84748.c: New test.

Re: [PATCH][AArch64] PR target/84748: Mark *compare_cstore<mode>_insn as clobbering CC reg

[Kyrill Tkachov]

On 08/03/18 13:32, James Greenhalgh wrote:
> On Wed, Mar 07, 2018 at 05:39:41PM +0000, Kyrill Tkachov wrote:
>> Hi all,
>>
>> In this wrong-code PR the combine pass ends up moving a CC-using instruction past a *compare_cstore<mode>_insn
>> insn_and_split. After reload the *compare_cstore<mode>_insn splitter ends up generating a SUBS instruction that
>> clobbers the condition flags, and things go bad.
>>
>> The solution is simple, the *compare_cstore<mode>_insn pattern should specify that it clobbers the CC register
>> so that combine (or any other pass) does not assume that it can move CC-using patterns across it.
>>
>> This patch does that and fixes the testcase.
>>
>> The testcase FAILs on GCC 8 only, but the buggy pattern is in GCC 6 onwards, so we should backport this as
>> a latent bug fix after it's had some time to bake in trunk.
>>
>> Bootstrapped and tested on aarch64-none-linux-gnu.
>>
>> Ok for trunk and later backports?
> OK.

I've backported the patch to the GCC 6 and 7 branches after bootstrapping and testing there.
Thanks,
Kyrill

> Thanks,
> James
>
>> 2018-03-07  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
>>
>>       PR target/84748
>>       * config/aarch64/aarch64.md (*compare_cstore<mode>_insn): Mark pattern
>>       as clobbering CC_REGNUM.
>>
>> 2018-03-07  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
>>
>>       PR target/84748
>>       * gcc.c-torture/execute/pr84748.c: New test.