Compiler/TMS320F28379D: memcpy for >256 words does generate RPT || PREAD but does lcr memcpy => how to avoid

Please see this forum thread.

Thanks and regards,

-George

Hello George,

I know this thread and read it again but I could not find an answer to my question.

Here is a small example code of my issue (basically a smaller and a bigger struct are copied):

typedef struct {
   float a;
} small_type ;

typedef struct {
   float a[150];
} big_type ;

big_type big1_struct;
big_type big2_struct;
small_type small1_struct;
small_type small2_struct;

int main(void)
{
    big1_struct = big2_struct;
    small1_struct = small2_struct;

	return 0;
}

The resulting assembly is:

_main:
	.dwcfi	cfa_offset, -2
	.dwcfi	save_reg_to_mem, 26, 0
	.dwpsn	file "../main.c",line 22,column 5,is_stmt,isa 0
;----------------------------------------------------------------------
;  22 | big1_struct = big2_struct;                                             
;----------------------------------------------------------------------
        MOVL      XAR4,#300             ; [CPU_ARAU] |22| 
        MOVL      ACC,XAR4              ; [CPU_ALU] |22| 
        MOVL      XAR5,#_big2_struct    ; [CPU_ARAU] |22| 
        MOVL      XAR4,#_big1_struct    ; [CPU_ARAU] |22| 
$C$DW$6	.dwtag  DW_TAG_TI_branch
	.dwattr $C$DW$6, DW_AT_low_pc(0x00)
	.dwattr $C$DW$6, DW_AT_name("_memcpy")
	.dwattr $C$DW$6, DW_AT_TI_call

        LCR       #_memcpy              ; [CPU_ALU] |22| 
        ; call occurs [#_memcpy] ; [] |22| 
	.dwpsn	file "../main.c",line 23,column 5,is_stmt,isa 0
;----------------------------------------------------------------------
;  23 | small1_struct = small2_struct;                                         
;----------------------------------------------------------------------
        MOVL      XAR4,#_small1_struct  ; [CPU_ARAU] |23| 
        MOVL      XAR7,#_small2_struct  ; [CPU_ARAU] |23| 
	.dwpsn	file "../main.c",line 25,column 2,is_stmt,isa 0
;----------------------------------------------------------------------
;  25 | return 0;                                                              
;----------------------------------------------------------------------
        MOVB      AL,#0                 ; [CPU_ALU] |25| 
	.dwpsn	file "../main.c",line 23,column 5,is_stmt,isa 0
        RPT       #1
||     PREAD     *XAR4++,*XAR7         ; [CPU_ALU] |23| 
$C$DW$7	.dwtag  DW_TAG_TI_branch
	.dwattr $C$DW$7, DW_AT_low_pc(0x00)
	.dwattr $C$DW$7, DW_AT_TI_return

        LRETR     ; [CPU_ALU] 
        ; return occurs ; [] 

As you can see the smaller struct is copied by an inlined memcpy using RPT and PREAD, while the bigger one is copied by a function call to memcpy. But performacewise it would be better to use multiple RPTs.

According to the compiler manual the code may be gereated using multiple RPTs, but I could not find how:

--rpt_threshold=k
Generates RPT loops that iterate k times or less (k is a constant
between 0 and 256). Multiple RPT’s may be generated for the same
loop, if iteration count is more than k and if code size does not
increase too much. Using this option when optimizing for code size
disables RPT loop generation for loops whose iteration count can be
greater than k.

This is the compiler command that generated the cited code:

"C:/ti/ccsv8/tools/compiler/ti-cgt-c2000_18.1.4.LTS/bin/cl2000" -v28 -ml -mt --cla_support=cla1 --float_support=fpu32 --tmu_support=tmu0 --vcu_support=vcu2 -O2 --opt_for_speed=5 --include_path="D:/repos/merlin-gansolarwr/CCS_workspace/test_struc_copy" --include_path="C:/ti/ccsv8/tools/compiler/ti-cgt-c2000_18.1.4.LTS/include" -g --c99 --diag_warning=225 --diag_wrap=off --display_error_number -k --asm_listing --c_src_interlist --asm_cross_reference_listing --preproc_with_compile --preproc_dependency="main.d_raw"  "../main.c"

Thanks,
Thomas.