gcc.dg/pr28796-2.c execution test fails on sh for a while. There is a wrong code like fcmp/eq dr2,dr4 bf/s .L93 fcmp/gt dr2,dr4 bra .L92 movt r2 which should be fcmp/eq dr2,dr4 bf/s .L53 movt r2 .L92: ... .L53: fcmp/gt dr2,dr4 .L93: i.e. a delay slot is filled with a wrong insn. It stated to fail between revision 152230 and 152317 but it seems there are no related changes. Perhaps some changes revealed this latent problem.
For a slightly reduced case, the insns before dbr pass are configured like as: (insn 460 456 461 (parallel [(set (reg:SI 147 t) (eq:SI (reg:DF 68 fr4) (reg:DF 66 fr2))) (use (reg/v:PSI 151 ))]) {cmpeqdf_t}) (jump_insn 461 460 575 (set (pc) (if_then_else (eq (reg:SI 147 t) (const_int 0)) (label_ref:SI 571) (pc))) {branch_false}) (insn 575 461 576 (parallel [(set (pc) (unspec [(const_int 200) (pc)] 4)) (set (reg:SI 147 t) (const_int 1 [0x1]))]) {stuff_delay_slot}) (insn 576 575 572 (set (pc) (unspec [(const_int 200)] 4))) (jump_insn 572 576 573 (set (pc) (label_ref 200))) ... (code_label 200 198 263 25 "" [1 uses]) (note 263 200 202 [bb 20] NOTE_INSN_BASIC_BLOCK) (insn 202 263 203 (set (reg:QI 2 r2) (reg:QI 147 t)) {movqi_i}) ... (code_label 571 573 501 85 "" [1 uses]) (note 501 571 502 [bb 32] NOTE_INSN_BASIC_BLOCK) (jump_insn 502 501 503 (set (pc) (label_ref 471)) {jump_compact}) ... (code_label 471 497 262 53 "" [2 uses]) (note 262 471 199 [bb 29] NOTE_INSN_BASIC_BLOCK) (insn 199 262 449 (parallel [(set (reg:SI 147 t) (gt:SI (reg:DF 68 fr4) (reg:DF 66 fr2))) (use (reg/v:PSI 151 ))]) {cmpgtdf_t}) and fill_slots_from_thread takes insn 199 as an insn filling the delay slot of jump_insn 461. When this is done, the opposite thread is started with stuff_delay_slot having (set (reg T_REG) (const_int 1)) sub-pattern which makes the live register analysis consider T_REG is defined here. Thus fill_slots_from_thread takes insn 199 which changes T_REG in spite of insn 202 in the opposite thread which expects the original T_REG value. It seems that the pattern (set (reg T_REG) ... in stuff_delay_slot insn is the cause of this problem and stuff_delay_slot is simply used as a unique marker. I'm testing the patch below. --- ORIG/trunk/gcc/config/sh/sh.md 2009-08-13 09:46:10.000000000 +0900 +++ trunk/gcc/config/sh/sh.md 2009-10-23 17:07:19.000000000 +0900 @@ -6825,7 +6825,7 @@ label: (define_insn "stuff_delay_slot" [(set (pc) (unspec [(match_operand:SI 0 "const_int_operand" "") (pc)] UNSPEC_BBR)) - (set (reg:SI T_REG) (match_operand:SI 1 "const_int_operand" ""))] + (match_operand:SI 1 "const_int_operand" "")] "TARGET_SH1" "" [(set_attr "length" "0")
Subject: Bug 41813 Author: kkojima Date: Sun Oct 25 23:10:22 2009 New Revision: 153543 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=153543 Log: PR target/41813 * config/sh/sh.md (stuff_delay_slot): Don't set T_REG in pattern. Modified: trunk/gcc/ChangeLog trunk/gcc/config/sh/sh.md
Fixed.