[PATCH] lra: Updates of biggest mode for hard regs [PR112278]

[PATCH] lra: Updates of biggest mode for hard regs [PR112278]

[Richard Sandiford]

[Gah.  In my head I'd sent this a few weeks ago, but it turns out
 that I hadn't even got to the stage of writing the changlog...]

LRA keeps track of the biggest mode for both hard registers and
pseudos.  The updates assume that the modes are ordered, i.e. that
we can tell whether one is no bigger than the other at compile time.

That is (or at least seemed to be) a reasonable restriction for pseudos.
But it isn't necessarily so for hard registers, since the uses of hard
registers can be logically distinct.  The testcase is an example of this.

The biggest mode of hard registers is also special for other reasons.
As the existing comment says:

  /* A reg can have a biggest_mode of VOIDmode if it was only ever seen as
     part of a multi-word register.  In that case, just use the reg_rtx
     mode.  Do the same also if the biggest mode was larger than a register
     or we can not compare the modes.  Otherwise, limit the size to that of
     the biggest access in the function or to the natural mode at least.  */

This patch applies the same approach to the updates.

Tested on aarch64-linus-gnu (with and without SVE) and on x86_64-linux-gnu.
OK to install?

Richard


gcc/
	PR rtl-optimization/112278
	* lra-int.h (lra_update_biggest_mode): New function.
	* lra-coalesce.cc (merge_pseudos): Use it.
	* lra-lives.cc (process_bb_lives): Likewise.
	* lra.cc (new_insn_reg): Likewise.

gcc/testsuite/
	PR rtl-optimization/112278
	* gcc.target/aarch64/sve/pr112278.c: New test.
---
 gcc/lra-coalesce.cc                             |  4 +---
 gcc/lra-int.h                                   | 15 +++++++++++++++
 gcc/lra-lives.cc                                |  4 +---
 gcc/lra.cc                                      |  5 ++---
 gcc/testsuite/gcc.target/aarch64/sve/pr112278.c | 15 +++++++++++++++
 5 files changed, 34 insertions(+), 9 deletions(-)
 create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/pr112278.c

diff --git a/gcc/lra-coalesce.cc b/gcc/lra-coalesce.cc
index 04a5bbd714b..d8ca096c35f 100644
--- a/gcc/lra-coalesce.cc
+++ b/gcc/lra-coalesce.cc
@@ -112,9 +112,7 @@ merge_pseudos (int regno1, int regno2)
     = (lra_merge_live_ranges
        (lra_reg_info[first].live_ranges,
 	lra_copy_live_range_list (lra_reg_info[first2].live_ranges)));
-  if (partial_subreg_p (lra_reg_info[first].biggest_mode,
-			lra_reg_info[first2].biggest_mode))
-    lra_reg_info[first].biggest_mode = lra_reg_info[first2].biggest_mode;
+  lra_update_biggest_mode (first, lra_reg_info[first2].biggest_mode);
 }
 
 /* Change pseudos in *LOC on their coalescing group
diff --git a/gcc/lra-int.h b/gcc/lra-int.h
index d7ec7c7dc7f..5cdf92be7fc 100644
--- a/gcc/lra-int.h
+++ b/gcc/lra-int.h
@@ -535,4 +535,19 @@ lra_assign_reg_val (int from, int to)
   lra_reg_info[to].offset = lra_reg_info[from].offset;
 }
 
+/* Update REGNO's biggest recorded mode so that it includes a reference
+   in mode MODE.  */
+inline void
+lra_update_biggest_mode (int regno, machine_mode mode)
+{
+  if (!ordered_p (GET_MODE_SIZE (lra_reg_info[regno].biggest_mode),
+		  GET_MODE_SIZE (mode)))
+    {
+      gcc_checking_assert (HARD_REGISTER_NUM_P (regno));
+      lra_reg_info[regno].biggest_mode = reg_raw_mode[regno];
+    }
+  else if (partial_subreg_p (lra_reg_info[regno].biggest_mode, mode))
+    lra_reg_info[regno].biggest_mode = mode;
+}
+
 #endif /* GCC_LRA_INT_H */
diff --git a/gcc/lra-lives.cc b/gcc/lra-lives.cc
index f60e564da82..0b204232849 100644
--- a/gcc/lra-lives.cc
+++ b/gcc/lra-lives.cc
@@ -770,9 +770,7 @@ process_bb_lives (basic_block bb, int &curr_point, bool dead_insn_p)
 	{
 	  int regno = reg->regno;
 
-	  if (partial_subreg_p (lra_reg_info[regno].biggest_mode,
-				reg->biggest_mode))
-	    lra_reg_info[regno].biggest_mode = reg->biggest_mode;
+	  lra_update_biggest_mode (regno, reg->biggest_mode);
 	  if (HARD_REGISTER_NUM_P (regno))
 	    lra_hard_reg_usage[regno] += freq;
 	}
diff --git a/gcc/lra.cc b/gcc/lra.cc
index c917a1adee2..29e2a3506e1 100644
--- a/gcc/lra.cc
+++ b/gcc/lra.cc
@@ -581,9 +581,8 @@ new_insn_reg (rtx_insn *insn, int regno, enum op_type type,
   lra_insn_reg *ir = lra_insn_reg_pool.allocate ();
   ir->type = type;
   ir->biggest_mode = mode;
-  if (NONDEBUG_INSN_P (insn)
-      && partial_subreg_p (lra_reg_info[regno].biggest_mode, mode))
-    lra_reg_info[regno].biggest_mode = mode;
+  if (NONDEBUG_INSN_P (insn))
+    lra_update_biggest_mode (regno, mode);
   ir->subreg_p = subreg_p;
   ir->early_clobber_alts = early_clobber_alts;
   ir->regno = regno;
diff --git a/gcc/testsuite/gcc.target/aarch64/sve/pr112278.c b/gcc/testsuite/gcc.target/aarch64/sve/pr112278.c
new file mode 100644
index 00000000000..4f56add2b0a
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/sve/pr112278.c
@@ -0,0 +1,15 @@
+#include <arm_neon.h>
+#include <arm_sve.h>
+
+void
+f (void)
+{
+  {
+    register svint8_t v0 asm ("z0");
+    asm volatile ("" : "=w" (v0));
+  }
+  {
+    register int8x8x4_t v0 asm ("v0");
+    asm volatile ("" : "=w" (v0));
+  }
+}
-- 
2.25.1

Re: [PATCH] lra: Updates of biggest mode for hard regs [PR112278]

[Vladimir Makarov]

On 12/3/23 05:13, Richard Sandiford wrote:
> [Gah.  In my head I'd sent this a few weeks ago, but it turns out
>   that I hadn't even got to the stage of writing the changlog...]
>
> LRA keeps track of the biggest mode for both hard registers and
> pseudos.  The updates assume that the modes are ordered, i.e. that
> we can tell whether one is no bigger than the other at compile time.
>
> That is (or at least seemed to be) a reasonable restriction for pseudos.
> But it isn't necessarily so for hard registers, since the uses of hard
> registers can be logically distinct.  The testcase is an example of this.
>
> The biggest mode of hard registers is also special for other reasons.
> As the existing comment says:
>
>    /* A reg can have a biggest_mode of VOIDmode if it was only ever seen as
>       part of a multi-word register.  In that case, just use the reg_rtx
>       mode.  Do the same also if the biggest mode was larger than a register
>       or we can not compare the modes.  Otherwise, limit the size to that of
>       the biggest access in the function or to the natural mode at least.  */
>
> This patch applies the same approach to the updates.
>
> Tested on aarch64-linus-gnu (with and without SVE) and on x86_64-linux-gnu.
> OK to install?
>
Sure.  Thank you for fixing this, Richard.