--- Comment #2 from Jakub Jelinek <jakub at gcc dot> ---
The problem is likely with the caching of ix86_compute_frame_layout results, at
least if I call this function at the beginning of ix86_expand_prologue, it
doesn't ICE anymore.

What changes in particular is ix86_nsaved_regs () on this testcase, during lra
-> lra_eliminate -> update_reg_eliminate it still returns 1, but during
pro_and_epilogue already 2.
The difference is in register 41.

What happens is:
1) fwprop1 adds a REG_EQUAL to the umulditi3_1 instruction:
(insn 13 12 111 3 (parallel [
            (set (reg:TI 99)
                (mult:TI (zero_extend:TI (subreg/s/v:DI (reg:TI 94) 0))
                    (zero_extend:TI (reg:DI 100))))
            (clobber (reg:CC 17 flags))
        ]) "pr85414.c":5 366 {*umulditi3_1}
     (expr_list:REG_EQUAL (mult:TI (zero_extend:TI (subreg/s/v:DI (reg:TI 94)
            (const_int 59 [0x3b]))
        (expr_list:REG_DEAD (reg:DI 100)
            (expr_list:REG_DEAD (reg:TI 94)
                (expr_list:REG_UNUSED (reg:CC 17 flags)
2) then probably based on the /s/v flags on the subreg that indicate
"optimizes" the REG_EQUAL note to (mult:TI (subreg:TI (subreg/s/v:DI (reg:TI
94) 0) 0) (const_int 59 [0x3b]))
That actually looks to me like invalid RTL, subreg of a subreg is not valid.
I guess it should have been just (mult:TI (reg:TI 94) (const_int 59 [0x3b])).
3) then the subreg2 pass just replaces uses of (reg:TI 94) subregs with (reg:DI
4) during LRA we have (mult:TI (subreg:TI (reg:DI 40 r11 [146]) 0) (const_int
59 [0x3b])) and that refers to the r12 too, but nothing earlier recognized it
as used (nor it is actually used in the insn).

Reply via email to