CGT bug with SPLOOPW on C674x?

Michael Pocino

I've been porting some code from the C672x to the C674x, and I ran into something that looks like a compiler bug. Specifically, a speculative ADDSP in a SPLOOPW had a write conflict with a single-cycle operation using the same unit after the end of the loop.

Here's the original C code (it's populating some vectors for a unit test):

f = 1.0f;

g = -1.0f;

for (i = 0; i < TEST_VECT_LENGTH; i++)

{

ctest_src_1[i].re = f * g;

ctest_src_1[i].im = -2.0f * f * g;

f += 1.0f;

g *= -1.0f;

}

f = 1.0f;

g = -1.0f;

for (i = 0; i < TEST_VECT_LENGTH; i++)

{

ctest_ref[i].re = -f * g;

ctest_ref[i].im = 2.0f * f * g;

f += 1.0f;

g *= -1.0f;

}

There are a bunch of loops like this in the code, but in this case the second loop failed to run at all, and the ctest_ref array did not get its new values.

Here's the resulting code that was generated. I've highlighted what I think are the offending lines. The first highlighted line (inside the loop) is speculatively executed during the last iteration of the loop. Its result is written to A6 from the L1 unit in the same cycle as the second highlighted line executes after the loop. The second line tries to initialize A0 to 1, also using the L1 unit, which fails to happen. Because the condition register still contains 0 from the previous loop, the second loop never runs.

$C$L65: ; PIPED LOOP PROLOG

.dwpsn file "../../src/testmdspcvect.c",line 632,column 0,is_stmt

[ A0] SPLOOPW 8 ;16 ; (P)

;** --------------------------------------------------------------------------*

$C$L66: ; PIPED LOOP KERNEL

$C$DW$L$_test_single_in$146$B:

.dwpsn file "../../src/testmdspcvect.c",line 633,column 0,is_stmt

NOP 1

SPMASK L1,S1,D1

|| MV .L1 A10,A7

|| MV .S1 A10,A6 ; |629|

|| MV .D1X B12,A8

SPMASK L2,S2

|| MVKL .S2 _ctest_src_1,B7

|| ZERO .L2 B8

|| MPYSP .M1 A8,A6,A6 ; |635| (P) <0,2>

|| ADDSP .L1 A7,A6,A9 ; |635| (P) <0,2> ^

SPMASK S2

|| MVKH .S2 0xbf800000,B8

SPMASK L2,S2

|| ZERO .L2 B9 ; |637|

|| MVKH .S2 _ctest_src_1,B7

|| MPYSP .M2X B8,A6,B4 ; |632| (P) <0,4>

SET .S2 B9,31,31,B5 ; |637| (P) <0,5>

SPMASK S1,D1

|| ZERO .S1 A5 ; |637|

|| ADD .D1X 4,B7,A3

|| XOR .L2 B8,B5,B16 ; |637| (P) <0,6>

|| ADDSP .L1 A7,A9,A6 ; |636| (P) <0,6> ^

|| MPYSP .M1 A8,A9,A16 ; |635| (P) <0,6>

SPMASK S1,L2

|| MVK .S1 64,A4 ; |632|

|| ADD .L2 8,B7,B6

|| ADD .L1 2,A5,A5 ; |632| (P) <0,7>

|| MPYSP .M1X B8,A6,A17 ; |635| (P) <0,7>

|| MPYSP .M2X B16,A9,B4 ; |635| (P) <0,7>

STW .D2T2 B4,*B7++(16) ; |632| <0,8>

|| CMPLT .L1 A5,A4,A9 ; |632| <0,8>

NOP 1

SET .S2 B9,31,31,B5 ; |637| <0,10>

MV .L1 A9,A0 ; |632| <0,11>

|| XOR .L2 B16,B5,B8 ; |637| <0,11>

|| STW .D1T1 A17,*A3++(16) ; |635| <0,11>

|| MPYSP .M2X B16,A16,B5 ; |635| <0,11>

NOP 2

NOP 1

SPKERNEL 0,0

|| STDW .D2T2 B5:B4,*B6++(16) ; |635| <0,15>

.dwpsn file "../../src/testmdspcvect.c",line 638,column 0,is_stmt

NOP 1 ; SDSCM00012367 HW bug workaround

$C$DW$L$_test_single_in$146$E:

;** --------------------------------------------------------------------------*

$C$L67: ; PIPED LOOP EPILOG

;** --------------------------------------------------------------------------*

MVK .L1 0x1,A0 ; init sploopw predicate

;*----------------------------------------------------------------------------*

$C$L68: ; PIPED LOOP PROLOG

.dwpsn file "../../src/testmdspcvect.c",line 643,column 0,is_stmt

[ A0] SPLOOPW 8 ;16 ; (P)

I'm sorry if this is a false alarm, or if this bug has already been reported. If my understanding of the problem is incorrect, please let me know.

Thanks,

Michael

over 15 years ago

0 George Mock over 15 years ago

TI__Guru**** 252300 points

Please submit a test case which compiles, though it does not need to run. You might have to preprocess it as described in http://processors.wiki.ti.com/index.php/Preprocess_Complex_Source_Code_for_Bug_Submissions . Include the exact compiler version and build options you use.

Thanks and regards,

-George

0 Michael Pocino over 15 years ago in reply to George Mock

Prodigy 115 points

George,

I've attached a test case. The compiler version is 6.1.16, and the options used are -c -ax -mv6740 -mi200 -mh1024 -qq -o2 -eo .o67 -k.

The problem does not occur when I compile with -mu.

Thanks,

Michael

sploopw_bug.tar.gz

0 George Mock over 15 years ago in reply to Michael Pocino

TI__Guru**** 252300 points

Thank you for submitting a test case. I filed record id SDSCM00037551 in our SDOWP system. Your test case is included in that record as an attachment. Feel free to track it at the SDOWP link in my sig below.

Thanks and regards,

-George

Code Composer Studio™︎

Code Composer Studio forum

CGT bug with SPLOOPW on C674x?