Hi
I ran into a curious problem where the compiler generates a less-than-ideal software pipelined loop after inlining a function call. Consider the following code:
#include <stdint.h>
//#define EXPAND
typedef struct X
{
uint32_t *mem;
} X;
void f1(const X* x, uint16_t a, uint16_t b)
{
volatile uint32_t * ptr = (volatile uint32_t*) x->mem;
*ptr = a + b;
}
void f2(const X* restrict x, const uint16_t * restrict data1, const uint16_t * restrict data2, const uint32_t len)
{
if(len < 16)
{
uint32_t i;
for(i=0; i<len; ++i)
{
#ifdef EXPAND
volatile uint32_t * ptr = (volatile uint32_t*) x->mem;
*ptr = data1[i] + data2[i];
#else
f1(x, data1[i], data2[i]);
#endif
}
}
}
The compiler behaves very differently depending on whether or not EXPAND is defined (yes, I am aware that the compiler internally takes completely very different execution paths).
When EXPAND is defined, the compiler generates the following for f2:
;******************************************************************************
;* TMS320C6x C/C++ Codegen PC v7.4.7 *
;* Date/Time created: Mon Nov 17 10:08:23 2014 *
;******************************************************************************
.compiler_opts --abi=eabi --c64p_l1d_workaround=default --endian=little --hll_source=on --long_precision_bits=32 --mem_model:code=near --mem_model:const=data --mem_model:data=far_aggregates --object_format=elf --silicon_version=6500 --symdebug:none
;******************************************************************************
;* GLOBAL FILE PARAMETERS *
;* *
;* Architecture : TMS320C64x+ *
;* Optimization : Enabled at level 3 *
;* Optimizing for : Speed *
;* Based on options: -o3, no -ms *
;* Endian : Little *
;* Interrupt Thrshld : Disabled *
;* Data Access Model : Far Aggregate Data *
;* Pipelining : Enabled *
;* Speculate Loads : Enabled with threshold = 0 *
;* Memory Aliases : Presume are aliases (pessimistic) *
;* Debug Info : No Debug Info *
;* *
;******************************************************************************
;******************************************************************************
;* FUNCTION NAME: f2 *
;* *
;* Regs Modified : A0,A3,A4,A5,A6,B4,B5 *
;* Regs Used : A0,A3,A4,A5,A6,B3,B4,B5,B6 *
;* Local Frame Size : 0 Args + 0 Auto + 0 Save = 0 byte *
;******************************************************************************
f2:
;** --------------------------------------------------------------------------*
MV .L1X B4,A5 ; |16|
|| CMPEQ .L2 B6,0,B4 ; |17|
|| MV .S2X A6,B5 ; |16|
XOR .L2 1,B4,B4 ; |17|
CMPLTU .L1X B6,16,A3 ; |17|
AND .L1X B4,A3,A0 ; |17|
[!A0] BNOP .S1 $C$L4,5 ; |17|
|| [ A0] MVC .S2 B6,ILC
; BRANCHCC OCCURS {$C$L4} ; |17|
;*----------------------------------------------------------------------------*
;* SOFTWARE PIPELINE INFORMATION
;*
;* Loop found in file : C:/test.c
;* Loop source line : 20
;* Loop opening brace source line : 21
;* Loop closing brace source line : 29
;* Known Minimum Trip Count : 1
;* Known Maximum Trip Count : 15
;* Known Max Trip Count Factor : 1
;* Loop Carried Dependency Bound(^) : 0
;* Unpartitioned Resource Bound : 2
;* Partitioned Resource Bound(*) : 2
;* Resource Partition:
;* A-side B-side
;* .L units 0 0
;* .S units 0 0
;* .D units 2* 1
;* .M units 0 0
;* .X cross paths 1 0
;* .T address paths 2* 1
;* Long read paths 0 0
;* Long write paths 0 0
;* Logical ops (.LS) 0 0 (.L or .S unit)
;* Addition ops (.LSD) 1 0 (.L or .S or .D unit)
;* Bound(.L .S .LS) 0 0
;* Bound(.L .S .D .LS .LSD) 1 1
;*
;* Searching for software pipeline schedule at ...
;* ii = 2 Schedule found with 4 iterations in parallel
;* Done
;*
;* Loop will be splooped
;* Collapsed epilog stages : 0
;* Collapsed prolog stages : 0
;* Minimum required memory pad : 0 bytes
;*
;* Minimum safe trip count : 1
;*----------------------------------------------------------------------------*
$C$L1: ; PIPED LOOP PROLOG
SPLOOP 2 ;8 ; (P)
|| LDW .D1T1 *A4,A6
;** --------------------------------------------------------------------------*
$C$L2: ; PIPED LOOP KERNEL
LDHU .D1T1 *A5++,A4 ; |24| (P) <0,0>
LDHU .D2T2 *B5++,B4 ; |24| (P) <0,1>
NOP 4
ADD .L1X B4,A4,A3 ; |24| <0,6>
SPKERNEL 3,0
|| STW .D1T1 A3,*A6 ; |24| <0,7>
;** --------------------------------------------------------------------------*
$C$L3: ; PIPED LOOP EPILOG
;** --------------------------------------------------------------------------*
$C$L4:
RETNOP .S2 B3,5 ; |31|
; BRANCH OCCURS {B3} ; |31|
In contrast, when EXPAND is not defined, the compiler generates
;******************************************************************************
;* TMS320C6x C/C++ Codegen PC v7.4.7 *
;* Date/Time created: Mon Nov 17 10:10:07 2014 *
;******************************************************************************
.compiler_opts --abi=eabi --c64p_l1d_workaround=default --endian=little --hll_source=on --long_precision_bits=32 --mem_model:code=near --mem_model:const=data --mem_model:data=far_aggregates --object_format=elf --silicon_version=6500 --symdebug:none
;******************************************************************************
;* GLOBAL FILE PARAMETERS *
;* *
;* Architecture : TMS320C64x+ *
;* Optimization : Enabled at level 3 *
;* Optimizing for : Speed *
;* Based on options: -o3, no -ms *
;* Endian : Little *
;* Interrupt Thrshld : Disabled *
;* Data Access Model : Far Aggregate Data *
;* Pipelining : Enabled *
;* Speculate Loads : Enabled with threshold = 0 *
;* Memory Aliases : Presume are aliases (pessimistic) *
;* Debug Info : No Debug Info *
;* *
;******************************************************************************
;******************************************************************************
;* FUNCTION NAME: f2 *
;* *
;* Regs Modified : A0,A3,A4,A5,B4,B5 *
;* Regs Used : A0,A3,A4,A5,A6,B3,B4,B5,B6 *
;* Local Frame Size : 0 Args + 0 Auto + 0 Save = 0 byte *
;******************************************************************************
f2:
;** --------------------------------------------------------------------------*
CMPEQ .L2 B6,0,B5 ; |17|
XOR .L2 1,B5,B5 ; |17|
CMPLTU .L1X B6,16,A3 ; |17|
AND .L1X B5,A3,A0 ; |17|
[!A0] BNOP .S1 $C$L4,5 ; |17|
|| [ A0] SUB .L2 B6,1,B5
; BRANCHCC OCCURS {$C$L4} ; |17|
;*----------------------------------------------------------------------------*
;* SOFTWARE PIPELINE INFORMATION
;*
;* Loop found in file : C:/test.c
;* Loop source line : 20
;* Loop opening brace source line : 21
;* Loop closing brace source line : 29
;* Known Minimum Trip Count : 1
;* Known Maximum Trip Count : 15
;* Known Max Trip Count Factor : 1
;* Loop Carried Dependency Bound(^) : 7
;* Unpartitioned Resource Bound : 2
;* Partitioned Resource Bound(*) : 2
;* Resource Partition:
;* A-side B-side
;* .L units 0 0
;* .S units 0 0
;* .D units 2* 2*
;* .M units 0 0
;* .X cross paths 1 0
;* .T address paths 2* 2*
;* Long read paths 0 0
;* Long write paths 0 0
;* Logical ops (.LS) 0 0 (.L or .S unit)
;* Addition ops (.LSD) 1 0 (.L or .S or .D unit)
;* Bound(.L .S .LS) 0 0
;* Bound(.L .S .D .LS .LSD) 1 1
;*
;* Searching for software pipeline schedule at ...
;* ii = 7 Schedule found with 2 iterations in parallel
;* Done
;*
;* Loop will be splooped
;* Collapsed epilog stages : 0
;* Collapsed prolog stages : 0
;* Minimum required memory pad : 0 bytes
;*
;* Minimum safe trip count : 1
;*----------------------------------------------------------------------------*
$C$L1: ; PIPED LOOP PROLOG
SPLOOPD 7 ;14 ; (P)
|| MV .L1 A4,A5
|| MV .L2X A6,B5
|| MV .S1X B4,A4
|| MVC .S2 B5,ILC
;** --------------------------------------------------------------------------*
$C$L2: ; PIPED LOOP KERNEL
LDHU .D1T1 *A4++,A3 ; |11| (P) <0,0> ^
|| LDHU .D2T2 *B5++,B4 ; |11| (P) <0,0> ^
LDW .D1T2 *A5,B4 ; |11| (P) <0,1>
NOP 3
ADD .L1X B4,A3,A3 ; |11| (P) <0,5> ^
STW .D2T1 A3,*B4 ; |11| (P) <0,6> ^
SPKERNEL 0,0
;** --------------------------------------------------------------------------*
$C$L3: ; PIPED LOOP EPILOG
NOP 1
;** --------------------------------------------------------------------------*
$C$L4:
RETNOP .S2 B3,5 ; |31|
; BRANCH OCCURS {B3} ; |31|
So after inlining, the compiler doesn't see that a and b cannot alias x->mem, reloads x->mem in every loop iteration, and thus generates an ii=7 (2 loops in parallel) schedule as opposed to the ii=2 (4 loops in parallel) schedule above.
Now I'm not overly worried about that behavior, as it's unlikely to affect the overall performance too much, and I can always optimize where relevant, but I'd like to maintain a coding style that does not induce unnecessary performance penalties, so I wonder if there is anything I am missing and what modifications (if any) would let the compiler do a better job (enabling "optimistic" aliasing assumptions seems to help, but these assumptions are a lot stricter than what's required by ISO C and too strict for my purposes).
Regards and thanks for any help
Markus
