https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93996
--- Comment #10 from Jeffrey A. Law <law at redhat dot com> ---
So what seems to be happening here for the original test is we have something
like this just before split3:
(insn 14 13 28 2 (parallel [
(set (mem/v:SI (reg/f:DI 0 x0 [97]) [-1 S4 A32])
(unspec_volatile:SI [
(plus:SI (mem/v:SI (reg/f:DI 0 x0 [97]) [-1 S4 A32])
(const_int 1 [0x1]))
(const_int 5 [0x5])
] UNSPECV_ATOMIC_OP))
(clobber (reg:CC 66 cc))
(clobber (reg:SI 1 x1 [98]))
(clobber (reg:SI 2 x2 [99]))
]) "k.c":2:17 3913 {aarch64_atomic_addsi}
(expr_list:REG_DEAD (reg/f:DI 0 x0 [97])
(expr_list:REG_UNUSED (reg:CC 66 cc)
(expr_list:REG_UNUSED (reg:SI 2 x2 [99])
(expr_list:REG_UNUSED (reg:SI 1 x1 [98])
(nil))))))
(note 28 14 23 2 NOTE_INSN_EPILOGUE_BEG)
(insn 23 28 24 2 (use (reg:DI 30 x30)) "k.c":2:1 -1
(nil))
That would work fine in terms of what's after the NOTE_INSN_EPILOGUE_BEG.
insn 14 gets split into a loop and we end up with a block note after the
epilogue begin note:
;; basic block 3, loop depth 0, count 3221225464 (estimated locally), maybe hot
;; Invalid sum of incoming counts 1075352416 (estimated locally), should be
3221225464 (estimated locally)
;; prev block 2, next block 4, flags: (NEW, RTL, MODIFIED)
;; pred: 2 [always] count:1073741824 (estimated locally) (FALLTHRU)
;; 3 [0.0% (guessed)] count:1610592 (estimated locally)
;; bb 3 artificial_defs: { }
;; bb 3 artificial_uses: { u-1(31){ }}
;; lr in
;; lr use
;; lr def
;; live in
;; live gen
;; live kill
(code_label 29 13 34 3 3 (nil) [1 uses])
(note 34 29 30 3 [bb 3] NOTE_INSN_BASIC_BLOCK)
(insn 30 34 31 3 (set (reg:SI 1 x1 [98])
(unspec_volatile:SI [
(mem/v:SI (reg/f:DI 0 x0 [97]) [-1 S4 A32])
(const_int 5 [0x5])
] UNSPECV_LX)) "k.c":2:17 4017 {aarch64_load_exclusivesi}
(nil))
(insn 31 30 32 3 (set (reg:SI 1 x1 [98])
(plus:SI (reg:SI 1 x1 [98])
(const_int 1 [0x1]))) "k.c":2:17 120 {*addsi3_aarch64}
(nil))
(insn 32 31 33 3 (parallel [
(set (reg:SI 2 x2 [99])
(unspec_volatile:SI [
(const_int 0 [0])
] UNSPECV_SX))
(set (mem/v:SI (reg/f:DI 0 x0 [97]) [-1 S4 A32])
(unspec_volatile:SI [
(reg:SI 1 x1 [98])
(const_int 5 [0x5])
] UNSPECV_SX))
]) "k.c":2:17 4022 {aarch64_store_exclusivesi}
(nil))
(jump_insn 33 32 28 3 (set (pc)
(if_then_else (ne (reg:SI 2 x2 [99])
(const_int 0 [0]))
(label_ref:DI 29)
(pc))) "k.c":2:17 31 {*cbnesi1}
(int_list:REG_BR_PROB 536868 (nil))
-> 29)
;; succ: 3 [0.0% (guessed)] count:1610592 (estimated locally)
;; 4 [100.0% (guessed)] count:3219614872 (estimated locally)
(FALLTHRU)
;; lr out
;; live out
(note 28 33 35 NOTE_INSN_EPILOGUE_BEG)
;; basic block 4, loop depth 0, count 3219614872 (estimated locally), maybe hot
;; prev block 3, next block 1, flags: (NEW, RTL, MODIFIED)
;; pred: 3 [100.0% (guessed)] count:3219614872 (estimated locally)
(FALLTHRU)
;; bb 4 artificial_defs: { }
;; bb 4 artificial_uses: { u-1(31){ }}
;; lr in
;; lr use
;; lr def
;; live in
;; live gen
;; live kill
(note 35 28 23 4 [bb 4] NOTE_INSN_BASIC_BLOCK)
(insn 23 35 24 4 (use (reg:DI 30 x30)) "k.c":2:1 -1
(nil))
The second testcase is essentially the same. We're splitting at atomic_add
that appears immediately before the EPILOGUE_BEG note.
