cplxlower: Avoid a transform when looking at a default definition

Message ID ri6pn5ihuuo.fsf@suse.cz
State New
Headers show
Series
  • cplxlower: Avoid a transform when looking at a default definition
Related show

Commit Message

Martin Jambor Oct. 16, 2020, 6:46 p.m.
Hi,

in PR 97456, IPA-SRA triggers a bug in tree-complex.c where it
converts:

 <bb 2>
   a$_M_value_21 = COMPLEX_EXPR <ISRA.18_10(D), ISRA.18_10(D)>;

(where ISRA.18 is IPA-SRA created PARM_DECL with DECL_IGNORED_P set,
which is why it only happens with IPA-SRA) into:

  <bb 2>
    a$_M_value_21 = COMPLEX_EXPR <a$_M_value$real_10(D), a$_M_value$real_10(D)>;

i.e. it replaces two uses of the parameter default-def with two
uninitialized default-defs of a new variable - all in hope to produce
code with better debug info.

This patch fixes it by avoiding the transform when the SSA_NAME to be
replaced is a default definition.  Bootstrapped and tested on
x86_64-linux.  OK for trunk and the gcc-10 branch?

Thanks,

Martin


gcc/ChangeLog:

2020-10-16  Martin Jambor  <mjambor@suse.cz>

	PR tree-optimization/97456
	* tree-complex.c (set_component_ssa_name): Do not replace ignored decl
	default definitions with new component vars.  Reorder if conditions.

gcc/testsuite/ChangeLog:

2020-10-16  Martin Jambor  <mjambor@suse.cz>

	PR tree-optimization/97456
	* gcc.dg/tree-ssa/pr97456.c: New test.
---
 gcc/testsuite/gcc.dg/tree-ssa/pr97456.c | 40 +++++++++++++++++++++++++
 gcc/tree-complex.c                      |  6 ++--
 2 files changed, 44 insertions(+), 2 deletions(-)
 create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/pr97456.c

-- 
2.28.0

Comments

Richard Biener Oct. 17, 2020, 9:46 a.m. | #1
On October 16, 2020 8:46:39 PM GMT+02:00, Martin Jambor <mjambor@suse.cz> wrote:
>Hi,

>

>in PR 97456, IPA-SRA triggers a bug in tree-complex.c where it

>converts:

>

> <bb 2>

>   a$_M_value_21 = COMPLEX_EXPR <ISRA.18_10(D), ISRA.18_10(D)>;

>

>(where ISRA.18 is IPA-SRA created PARM_DECL with DECL_IGNORED_P set,

>which is why it only happens with IPA-SRA) into:

>

>  <bb 2>

>a$_M_value_21 = COMPLEX_EXPR <a$_M_value$real_10(D),

>a$_M_value$real_10(D)>;

>

>i.e. it replaces two uses of the parameter default-def with two

>uninitialized default-defs of a new variable - all in hope to produce

>code with better debug info.

>

>This patch fixes it by avoiding the transform when the SSA_NAME to be

>replaced is a default definition.  Bootstrapped and tested on

>x86_64-linux.  OK for trunk and the gcc-10 branch?


It's never OK to replace SSA_NAME_VAR of a default Def so may I suggest to put this restriction first, independent of any other predicate?

Ok with that change. Thanks, 
Richard. 

>Thanks,

>

>Martin

>

>

>gcc/ChangeLog:

>

>2020-10-16  Martin Jambor  <mjambor@suse.cz>

>

>	PR tree-optimization/97456

>	* tree-complex.c (set_component_ssa_name): Do not replace ignored decl

>	default definitions with new component vars.  Reorder if conditions.

>

>gcc/testsuite/ChangeLog:

>

>2020-10-16  Martin Jambor  <mjambor@suse.cz>

>

>	PR tree-optimization/97456

>	* gcc.dg/tree-ssa/pr97456.c: New test.

>---

> gcc/testsuite/gcc.dg/tree-ssa/pr97456.c | 40 +++++++++++++++++++++++++

> gcc/tree-complex.c                      |  6 ++--

> 2 files changed, 44 insertions(+), 2 deletions(-)

> create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/pr97456.c

>

>diff --git a/gcc/testsuite/gcc.dg/tree-ssa/pr97456.c

>b/gcc/testsuite/gcc.dg/tree-ssa/pr97456.c

>new file mode 100644

>index 00000000000..5171c9b4577

>--- /dev/null

>+++ b/gcc/testsuite/gcc.dg/tree-ssa/pr97456.c

>@@ -0,0 +1,40 @@

>+/* { dg-do run } */

>+/* { dg-options "-O2 -fwhole-program" } */

>+

>+

>+float val2 = 1.710780f;

>+float val3;

>+volatile float vf;

>+

>+int __attribute__((noipa))

>+get_bool (void)

>+{

>+  return 1;

>+}

>+

>+int __attribute__((noinline))

>+wrong (float *pos)

>+{

>+  _Complex float a;

>+

>+  __real__ a = *pos;

>+  __imag__ a = *pos;

>+

>+  _Complex float b = 0 + 0i;

>+

>+  b = b + a;

>+

>+  if (b == 0.0f)

>+    return 1;

>+

>+  vf = __imag__ b;

>+  return 0;

>+}

>+

>+int main(int argc, char **argv) {

>+  float val = get_bool () == 1 ? val2 : val3;

>+

>+  if ((wrong(&val), wrong(&val)))

>+    __builtin_abort ();

>+  return 0;

>+}

>diff --git a/gcc/tree-complex.c b/gcc/tree-complex.c

>index 2e54bbb917c..71ad7c18523 100644

>--- a/gcc/tree-complex.c

>+++ b/gcc/tree-complex.c

>@@ -570,8 +570,10 @@ set_component_ssa_name (tree ssa_name, bool

>imag_p, tree value)

>  /* Replace an anonymous base value with the variable from cvc_lookup.

> 	 This should result in better debug info.  */

>       if (SSA_NAME_VAR (ssa_name)

>-	  && (!SSA_NAME_VAR (value) || DECL_IGNORED_P (SSA_NAME_VAR (value)))

>-	  && !DECL_IGNORED_P (SSA_NAME_VAR (ssa_name)))

>+	  && !DECL_IGNORED_P (SSA_NAME_VAR (ssa_name))

>+	  && (!SSA_NAME_VAR (value)

>+	      || (DECL_IGNORED_P (SSA_NAME_VAR (value))

>+		  && !SSA_NAME_IS_DEFAULT_DEF (value))))

> 	{

> 	  comp = get_component_var (SSA_NAME_VAR (ssa_name), imag_p);

> 	  replace_ssa_name_symbol (value, comp);
Martin Jambor Oct. 19, 2020, 2:43 p.m. | #2
Hi,

On Sat, Oct 17 2020, Richard Biener wrote:
> On October 16, 2020 8:46:39 PM GMT+02:00, Martin Jambor <mjambor@suse.cz> wrote:

>>Hi,

>>

>>in PR 97456, IPA-SRA triggers a bug in tree-complex.c where it

>>converts:

>>

>> <bb 2>

>>   a$_M_value_21 = COMPLEX_EXPR <ISRA.18_10(D), ISRA.18_10(D)>;

>>

>>(where ISRA.18 is IPA-SRA created PARM_DECL with DECL_IGNORED_P set,

>>which is why it only happens with IPA-SRA) into:

>>

>>  <bb 2>

>>a$_M_value_21 = COMPLEX_EXPR <a$_M_value$real_10(D),

>>a$_M_value$real_10(D)>;

>>

>>i.e. it replaces two uses of the parameter default-def with two

>>uninitialized default-defs of a new variable - all in hope to produce

>>code with better debug info.

>>

>>This patch fixes it by avoiding the transform when the SSA_NAME to be

>>replaced is a default definition.  Bootstrapped and tested on

>>x86_64-linux.  OK for trunk and the gcc-10 branch?

>

> It's never OK to replace SSA_NAME_VAR of a default Def so may I suggest to put this restriction first, independent of any other predicate?

>


OK, I've rebased the following on the current master and will commit it
after the final round of testing.

Thanks,

Martin


gcc/ChangeLog:

2020-10-19  Martin Jambor  <mjambor@suse.cz>

	PR tree-optimization/97456
	* tree-complex.c (set_component_ssa_name): Do not replace ignored decl
	default definitions with new component vars.

gcc/testsuite/ChangeLog:

2020-10-19  Martin Jambor  <mjambor@suse.cz>

	PR tree-optimization/97456
	* gcc.dg/tree-ssa/pr97456.c: New test.
---
 gcc/testsuite/gcc.dg/tree-ssa/pr97456.c | 40 +++++++++++++++++++++++++
 gcc/tree-complex.c                      |  3 +-
 2 files changed, 42 insertions(+), 1 deletion(-)
 create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/pr97456.c

diff --git a/gcc/testsuite/gcc.dg/tree-ssa/pr97456.c b/gcc/testsuite/gcc.dg/tree-ssa/pr97456.c
new file mode 100644
index 00000000000..5171c9b4577
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/tree-ssa/pr97456.c
@@ -0,0 +1,40 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -fwhole-program" } */
+
+
+float val2 = 1.710780f;
+float val3;
+volatile float vf;
+
+int __attribute__((noipa))
+get_bool (void)
+{
+  return 1;
+}
+
+int __attribute__((noinline))
+wrong (float *pos)
+{
+  _Complex float a;
+
+  __real__ a = *pos;
+  __imag__ a = *pos;
+
+  _Complex float b = 0 + 0i;
+
+  b = b + a;
+
+  if (b == 0.0f)
+    return 1;
+
+  vf = __imag__ b;
+  return 0;
+}
+
+int main(int argc, char **argv) {
+  float val = get_bool () == 1 ? val2 : val3;
+
+  if ((wrong(&val), wrong(&val)))
+    __builtin_abort ();
+  return 0;
+}
diff --git a/gcc/tree-complex.c b/gcc/tree-complex.c
index 2e54bbb917c..f132e0f41c7 100644
--- a/gcc/tree-complex.c
+++ b/gcc/tree-complex.c
@@ -569,7 +569,8 @@ set_component_ssa_name (tree ssa_name, bool imag_p, tree value)
     {
       /* Replace an anonymous base value with the variable from cvc_lookup.
 	 This should result in better debug info.  */
-      if (SSA_NAME_VAR (ssa_name)
+      if (!SSA_NAME_IS_DEFAULT_DEF (value)
+	  && SSA_NAME_VAR (ssa_name)
 	  && (!SSA_NAME_VAR (value) || DECL_IGNORED_P (SSA_NAME_VAR (value)))
 	  && !DECL_IGNORED_P (SSA_NAME_VAR (ssa_name)))
 	{
-- 
2.28.0

Patch

diff --git a/gcc/testsuite/gcc.dg/tree-ssa/pr97456.c b/gcc/testsuite/gcc.dg/tree-ssa/pr97456.c
new file mode 100644
index 00000000000..5171c9b4577
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/tree-ssa/pr97456.c
@@ -0,0 +1,40 @@ 
+/* { dg-do run } */
+/* { dg-options "-O2 -fwhole-program" } */
+
+
+float val2 = 1.710780f;
+float val3;
+volatile float vf;
+
+int __attribute__((noipa))
+get_bool (void)
+{
+  return 1;
+}
+
+int __attribute__((noinline))
+wrong (float *pos)
+{
+  _Complex float a;
+
+  __real__ a = *pos;
+  __imag__ a = *pos;
+
+  _Complex float b = 0 + 0i;
+
+  b = b + a;
+
+  if (b == 0.0f)
+    return 1;
+
+  vf = __imag__ b;
+  return 0;
+}
+
+int main(int argc, char **argv) {
+  float val = get_bool () == 1 ? val2 : val3;
+
+  if ((wrong(&val), wrong(&val)))
+    __builtin_abort ();
+  return 0;
+}
diff --git a/gcc/tree-complex.c b/gcc/tree-complex.c
index 2e54bbb917c..71ad7c18523 100644
--- a/gcc/tree-complex.c
+++ b/gcc/tree-complex.c
@@ -570,8 +570,10 @@  set_component_ssa_name (tree ssa_name, bool imag_p, tree value)
       /* Replace an anonymous base value with the variable from cvc_lookup.
 	 This should result in better debug info.  */
       if (SSA_NAME_VAR (ssa_name)
-	  && (!SSA_NAME_VAR (value) || DECL_IGNORED_P (SSA_NAME_VAR (value)))
-	  && !DECL_IGNORED_P (SSA_NAME_VAR (ssa_name)))
+	  && !DECL_IGNORED_P (SSA_NAME_VAR (ssa_name))
+	  && (!SSA_NAME_VAR (value)
+	      || (DECL_IGNORED_P (SSA_NAME_VAR (value))
+		  && !SSA_NAME_IS_DEFAULT_DEF (value))))
 	{
 	  comp = get_component_var (SSA_NAME_VAR (ssa_name), imag_p);
 	  replace_ssa_name_symbol (value, comp);