Re: [PATCH] Fix PR80492

[Richard Biener <rguenther at suse dot de>]

On Tue, 25 Apr 2017, Richard Biener wrote:

> 
> The following fixes redundant hard-register "stores" to be not eliminated
> by FRE/PRE and the alias machinery to properly handle different 
> local VAR_DECLs with the same asm specification.
> 
> Comments are welcome.  I tested the testcase on x86_64, ppc64le and
> aarch64 and all seem to be happy with *4 as register specification.
> 
> Bootstrap / regtest running on x86_64-unknown-linux-gnu.

So in bugzilla the conclusion is that the DSE is valid (for the
testcase it's not because it failed to compute aliasing properly).
And I can't see any reason to not do the transform (the corresponding
CSE is not done to avoid increasing register pressure by increased
lifetime).  That is, removing a redundant setting of a hardreg
reduces the lifetime of the value stored.

Bootstrapped on x86_64-unknown-linux-gnu, testing still in progress.

Richard.

2017-04-25  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/80492
	* alias.c (compare_base_decls): Handle registers with asm
	specification conservatively.
	* tree-ssa-alias.c (decl_refs_may_alias_p): Handle
	compare_base_decls returning dont-know properly.

	* gcc.dg/pr80492.c: New testcase.

Index: gcc/alias.c
===================================================================
--- gcc/alias.c	(revision 247147)
+++ gcc/alias.c	(working copy)
@@ -2046,6 +2046,18 @@ compare_base_decls (tree base1, tree bas
   if (base1 == base2)
     return 1;
 
+  /* If we have two register decls with register specification we
+     cannot decide unless their assembler name is the same.  */
+  if (DECL_REGISTER (base1)
+      && DECL_REGISTER (base2)
+      && DECL_ASSEMBLER_NAME_SET_P (base1)
+      && DECL_ASSEMBLER_NAME_SET_P (base2))
+    {
+      if (DECL_ASSEMBLER_NAME (base1) == DECL_ASSEMBLER_NAME (base2))
+	return 1;
+      return -1;
+    }
+
   /* Declarations of non-automatic variables may have aliases.  All other
      decls are unique.  */
   if (!decl_in_symtab_p (base1)
Index: gcc/passes.c
===================================================================
--- gcc/passes.c	(revision 247147)
+++ gcc/passes.c	(working copy)
@@ -1532,7 +1532,7 @@ pass_manager::pass_manager (context *ctx
 : all_passes (NULL), all_small_ipa_passes (NULL), all_lowering_passes (NULL),
   all_regular_ipa_passes (NULL),
   all_late_ipa_passes (NULL), passes_by_id (NULL), passes_by_id_size (0),
-  m_ctxt (ctxt)
+  m_ctxt (ctxt), m_name_to_pass_map (NULL)
 {
   opt_pass **p;
 
Index: gcc/testsuite/gcc.dg/pr80492.c
===================================================================
--- gcc/testsuite/gcc.dg/pr80492.c	(nonexistent)
+++ gcc/testsuite/gcc.dg/pr80492.c	(working copy)
@@ -0,0 +1,20 @@
+/* { dg-do compile } */
+/* { dg-options "-w -O2 -fdump-tree-optimized" } */
+
+static __inline__ __attribute__((__always_inline__))
+void syscall_7 (int val)
+{
+  register int reg __asm ("4") = val;
+  __asm __volatile__ ("/* Some Code %0 */" :: "r" (reg));
+}
+
+void do_syscalls (void)
+{
+  for (int s = 0; s < 2; s++)
+    {
+      syscall_7 (0);
+      syscall_7 (1);
+    }
+}
+
+/* { dg-final { scan-tree-dump-times "reg = " 4 "optimized" } } */
Index: gcc/tree-ssa-alias.c
===================================================================
--- gcc/tree-ssa-alias.c	(revision 247147)
+++ gcc/tree-ssa-alias.c	(working copy)
@@ -1096,13 +1096,16 @@ decl_refs_may_alias_p (tree ref1, tree b
 {
   gcc_checking_assert (DECL_P (base1) && DECL_P (base2));
 
+  int cmp = compare_base_decls (base1, base2);
+
   /* If both references are based on different variables, they cannot alias.  */
-  if (compare_base_decls (base1, base2) == 0)
+  if (cmp == 0)
     return false;
 
   /* If both references are based on the same variable, they cannot alias if
      the accesses do not overlap.  */
-  if (!ranges_overlap_p (offset1, max_size1, offset2, max_size2))
+  if (cmp == 1
+      && !ranges_overlap_p (offset1, max_size1, offset2, max_size2))
     return false;
 
   /* For components with variable position, the above test isn't sufficient,