Hi,
This patch permits CSE to add REG_EQUAL notes when single setting a constant
to a register even if REG_EQUAL constant rtx is the same as the set source rtx.
This enables optimizations in later passes looking for REG_EQUAL notes, like
jump2 pass.
For example, in arm testcase pr43920-2.c, CSE previously decided not to put
an "obvious" note on insn 9, as set value was the same as note value.
At the same time, other insns set up as -1 were set up through a register
and did get a note:
(insn 9 53 34 8 (set (reg:SI 110 [ D.4934 ])
(const_int -1 [0xffffffffffffffff]))
/work/src/gcc/gcc/testsuite/gcc.target/arm/pr43920-2.c:21 613
{*thumb2_movsi_vfp}
(nil))
(insn 8 45 50 6 (set (reg:SI 110 [ D.4934 ])
(reg/v:SI 111 [ startD.4917 ]))
/work/src/gcc/gcc/testsuite/gcc.target/arm/pr43920-2.c:21 613
{*thumb2_movsi_vfp}
(expr_list:REG_EQUAL (const_int -1 [0xffffffffffffffff])
(nil)))
(insn 6 49 54 7 (set (reg:SI 110 [ D.4934 ])
(reg/v:SI 112 [ endD.4918 ]))
/work/src/gcc/gcc/testsuite/gcc.target/arm/pr43920-2.c:21 613
{*thumb2_movsi_vfp}
(expr_list:REG_EQUAL (const_int -1 [0xffffffffffffffff])
(nil)))
Jump2 pass, optimizing common code, was looking at notes to reason about
register values and failing to recognize those insns to be equal.
Making CSE to set up REG_EQUAL notes even in "obvious" cases
fixes pr43920-2.c for arm-none-eabi and other targets.
I prefer adding notes in CSE instead of adding additional checks in Jump2
and, if any, other passes, as I think it is more uniform solution and allows
single point fix. Downside is having more notes.
Done full regression run on arm-none-eabi and bootstrapped on x86.
Is patch ok?
gcc/
2015-03-04 Alex Velenko <alex.vele...@arm.com>
* cse.c (cse_insn): Check to set REG_EQUAL note relaxed.
---
gcc/cse.c | 10 ++++------
1 file changed, 4 insertions(+), 6 deletions(-)
diff --git a/gcc/cse.c b/gcc/cse.c
index 2a33827..abaf867 100644
--- a/gcc/cse.c
+++ b/gcc/cse.c
@@ -5383,10 +5383,9 @@ cse_insn (rtx_insn *insn)
}
/* If this is a single SET, we are setting a register, and we have an
- equivalent constant, we want to add a REG_EQUAL note if the constant
- is different from the source. We don't want to do it for a constant
- pseudo since verifying that this pseudo hasn't been eliminated is a
- pain; moreover such a note won't help anything.
+ equivalent constant, we want to add a REG_EQUAL note. We don't want
+ to do it for a constant pseudo since verifying that this pseudo hasn't
+ been eliminated is a pain; moreover such a note won't help anything.
Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
which can be created for a reference to a compile time computable
@@ -5400,8 +5399,7 @@ cse_insn (rtx_insn *insn)
&& !(GET_CODE (src_const) == CONST
&& GET_CODE (XEXP (src_const, 0)) == MINUS
&& GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
- && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF)
- && !rtx_equal_p (src, src_const))
+ && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
{
/* Make sure that the rtx is not shared. */
src_const = copy_rtx (src_const);
--
1.8.1.2
