This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[patch, avr] Fix ICE PR64452 pushing eliminated rtxes
- From: Georg-Johann Lay <avr at gjlay dot de>
- To: GCC Patches <gcc-patches at gcc dot gnu dot org>
- Cc: Denis Chertykov <chertykov at gmail dot com>
- Date: Tue, 17 Feb 2015 12:12:59 +0100
- Subject: [patch, avr] Fix ICE PR64452 pushing eliminated rtxes
- Authentication-results: sourceware.org; auth=none
Byte-wise pushing virtual regs like arg pointer migth result in patterns like
(set (mem:QI (post_dec:HI (reg:HI 32 SP)))
(subreg:QI (plus:HI (reg:HI 28)
(const_int 17)) 0))
after elimination.
Attached patch uses new pushhi1_insn to push virtuals in HImode so that
expressions like in subreg_reg from above can be reloaded.
Ok to commit ?
Johann
PR target/64452
* config/avr/avr.md (pushhi_insn): New insn.
(push<mode>1): Push virtual regs in one chunk using pushhi1_insn.
Index: config/avr/avr.md
===================================================================
--- config/avr/avr.md (revision 220738)
+++ config/avr/avr.md (working copy)
@@ -371,6 +371,13 @@ (define_insn "push<mode>1"
push __zero_reg__"
[(set_attr "length" "1,1")])
+(define_insn "pushhi1_insn"
+ [(set (mem:HI (post_dec:HI (reg:HI REG_SP)))
+ (match_operand:HI 0 "register_operand" "r"))]
+ ""
+ "push %B0\;push %A0"
+ [(set_attr "length" "2")])
+
;; All modes for a multi-byte push. We must include complex modes here too,
;; lest emit_single_push_insn "helpfully" create the auto-inc itself.
(define_mode_iterator MPUSH
@@ -386,17 +393,42 @@ (define_expand "push<mode>1"
[(match_operand:MPUSH 0 "" "")]
""
{
- int i;
-
- // Avoid (subreg (mem)) for non-generic address spaces below. Because
- // of the poor addressing capabilities of these spaces it's better to
- // load them in one chunk. And it avoids PR61443.
-
if (MEM_P (operands[0])
&& !ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (operands[0])))
- operands[0] = copy_to_mode_reg (<MODE>mode, operands[0]);
+ {
+ // Avoid (subreg (mem)) for non-generic address spaces. Because
+ // of the poor addressing capabilities of these spaces it's better to
+ // load them in one chunk. And it avoids PR61443.
+
+ operands[0] = copy_to_mode_reg (<MODE>mode, operands[0]);
+ }
+ else if (REG_P (operands[0])
+ && IN_RANGE (REGNO (operands[0]), FIRST_VIRTUAL_REGISTER,
+ LAST_VIRTUAL_REGISTER))
+ {
+ // Byte-wise pushing of virtual regs might result in something like
+ //
+ // (set (mem:QI (post_dec:HI (reg:HI 32 SP)))
+ // (subreg:QI (plus:HI (reg:HI 28)
+ // (const_int 17)) 0))
+ //
+ // after elimination. This cannot be handled by reload, cf. PR64452.
+ // Reload virtuals in one chunk. That way it's possible to reload
+ // above situation and finally
+ //
+ // (set (reg:HI **)
+ // (const_int 17))
+ // (set (reg:HI **)
+ // (plus:HI (reg:HI **)
+ // (reg:HI 28)))
+ // (set (mem:HI (post_dec:HI (reg:HI 32 SP))
+ // (reg:HI **)))
+
+ emit_insn (gen_pushhi1_insn (operands[0]));
+ DONE;
+ }
- for (i = GET_MODE_SIZE (<MODE>mode) - 1; i >= 0; --i)
+ for (int i = GET_MODE_SIZE (<MODE>mode) - 1; i >= 0; --i)
{
rtx part = simplify_gen_subreg (QImode, operands[0], <MODE>mode, i);
if (part != const0_rtx)