[PR103097] tolerate reg-stack cross-block malformed asms

  The testcase shows malformed asms in one block confuse reg-stack logic
in another block.  Moving the resetting of any_malformed_asm to the
end of the pass enables it to take effect throughout the affected
function.

Regstrapped on x86_64-linux-gnu.  Ok to install?


for  gcc/ChangeLog

	PR target/103097
	* reg-stack.c (convert_regs_1): Move any_malformed_asm
	resetting...
	(reg_to_stack): ... here.

for  gcc/testsuite/ChangeLog

	PR target/103097
	* gcc.target/i386/pr103097.c: New.
---
 gcc/reg-stack.c                          |    3 +--
 gcc/testsuite/gcc.target/i386/pr103097.c |   30 ++++++++++++++++++++++++++++++
 2 files changed, 31 insertions(+), 2 deletions(-)
 create mode 100644 gcc/testsuite/gcc.target/i386/pr103097.c
  

On 12/7/2021 7:00 PM, Alexandre Oliva via Gcc-patches wrote:
> The testcase shows malformed asms in one block confuse reg-stack logic
> in another block.  Moving the resetting of any_malformed_asm to the
> end of the pass enables it to take effect throughout the affected
> function.
>
> Regstrapped on x86_64-linux-gnu.  Ok to install?
>
>
> for  gcc/ChangeLog
>
> 	PR target/103097
> 	* reg-stack.c (convert_regs_1): Move any_malformed_asm
> 	resetting...
> 	(reg_to_stack): ... here.
>
> for  gcc/testsuite/ChangeLog
>
> 	PR target/103097
> 	* gcc.target/i386/pr103097.c: New.
So it's "stickier" after your change.  ie, instead of indicating if 
there was a malformed insn in a block, it's did we find a malformed insn 
anywhere.  Which implies the comment before the declaration of 
any_malformed_asm needs a trivial update since it stated "malformed 
insns in a block".

OK with the trivial comment update.

jeff
  

On Dec  8, 2021, Jeff Law <jeffreyalaw@gmail.com> wrote:

> On 12/7/2021 7:00 PM, Alexandre Oliva via Gcc-patches wrote:
>> PR target/103097
>> * reg-stack.c (convert_regs_1): Move any_malformed_asm
>> resetting...
>> (reg_to_stack): ... here.

> So it's "stickier" after your change.  ie, instead of indicating if
> there was a malformed insn in a block, it's did we find a malformed
> insn anywhere.

Yeah, anywhere in the same function.

> Which implies the comment before the declaration of any_malformed_asm
> needs a trivial update since it stated "malformed insns in a block".

(-: But, but...  we still find it in a block!  ;-D

> OK with the trivial comment update.

Thanks, I'm making it s/block/function/.

Here's what I'm installing momentarily.


[PR103097] tolerate reg-stack cross-block malformed asms

The testcase shows malformed asms in one block confuse reg-stack logic
in another block.  Moving the resetting of any_malformed_asm to the
end of the pass enables it to take effect throughout the affected
function.


for  gcc/ChangeLog

	PR target/103097
	* reg-stack.c (convert_regs_1): Move any_malformed_asm
	resetting...
	(reg_to_stack): ... here.

for  gcc/testsuite/ChangeLog

	PR target/103097
	* gcc.target/i386/pr103097.c: New.
---
 gcc/reg-stack.c                          |    5 ++---
 gcc/testsuite/gcc.target/i386/pr103097.c |   30 ++++++++++++++++++++++++++++++
 2 files changed, 32 insertions(+), 3 deletions(-)
 create mode 100644 gcc/testsuite/gcc.target/i386/pr103097.c

diff --git a/gcc/reg-stack.c b/gcc/reg-stack.c
index 1d9ea035cf44f..cc369f0b635a0 100644
--- a/gcc/reg-stack.c
+++ b/gcc/reg-stack.c
@@ -453,7 +453,7 @@ get_true_reg (rtx *pat)
       }
 }
 
-/* Set if we find any malformed asms in a block.  */
+/* Set if we find any malformed asms in a function.  */
 static bool any_malformed_asm;
 
 /* There are many rules that an asm statement for stack-like regs must
@@ -3014,8 +3014,6 @@ convert_regs_1 (basic_block block)
   bool cfg_altered = false;
   int debug_insns_with_starting_stack = 0;
 
-  any_malformed_asm = false;
-
   /* Choose an initial stack layout, if one hasn't already been chosen.  */
   if (bi->stack_in.top == -2)
     {
@@ -3385,6 +3383,7 @@ reg_to_stack (void)
 	  0, sizeof (char) * (max_uid + 1));
 
   convert_regs ();
+  any_malformed_asm = false;
 
   free_aux_for_blocks ();
   return true;
diff --git a/gcc/testsuite/gcc.target/i386/pr103097.c b/gcc/testsuite/gcc.target/i386/pr103097.c
new file mode 100644
index 0000000000000..2b7a307deec9a
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr103097.c
@@ -0,0 +1,30 @@
+/* { dg-do compile } */
+/* { dg-options "-O1 -fharden-conditional-branches" } */
+
+/* This is a slightly simplified version of
+   gcc.target/s390/vector/long-double-asm-earlyclobber.c.  On x86, the f
+   constraints in asm statements imposes some requirements that the testcase
+   doesn't meet.  What's unusual is that -fharden-conditional-branches extends
+   the effects of the malformed asm onto a different basic blocks, which
+   reg-stack did not expect.  */
+
+#include <assert.h>
+#include <stdint.h>
+
+void
+f (void)
+{
+  long double res, x = 0;
+  asm("" : "=f"(res) /* { dg-error "must specify a single register" } */
+      : "0"(x));
+  assert (res == x);
+}  
+
+void
+g (void)
+{
+  long double res, x = 0;
+  asm("" : "=g"(res) /* this is ok.  */
+      : "0"(x));
+  assert (res == x);
+}