Hello,
the SPU __multi3 and __udivmodti4 routines were using code that violates
the aliasing rules for type-punning between TImode and vector types.
This now causes errors when building those files. The patch below
replaces this code by type-punning via unions.
Tested on spu-elf, committed to mainline.
Bye,
Ulrich
ChangeLog:
* config/spu/divmovti4.c (union qword_UTItype): New data type.
(si_from_UTItype, si_to_UTItype): New functions.
(__udivmodti4): Use them to implement type-punning.
* config/spu/multi3.c (union qword_TItype): New data type.
(si_from_TItype, si_to_TItype): New functions.
(__multi3): Use them to implement type-punning.
Index: gcc/config/spu/divmodti4.c
===================================================================
*** gcc/config/spu/divmodti4.c (revision 172953)
--- gcc/config/spu/divmodti4.c (working copy)
*************** UTItype __udivti3 (UTItype u, UTItype v)
*** 29,34 ****
--- 29,56 ----
UTItype __umodti3 (UTItype u, UTItype v);
UTItype __udivmodti4 (UTItype u, UTItype v, UTItype *w);
+ union qword_UTItype
+ {
+ qword q;
+ UTItype t;
+ };
+
+ inline static qword
+ si_from_UTItype (UTItype t)
+ {
+ union qword_UTItype u;
+ u.t = t;
+ return u.q;
+ }
+
+ inline static UTItype
+ si_to_UTItype (qword q)
+ {
+ union qword_UTItype u;
+ u.q = q;
+ return u.t;
+ }
+
inline static unsigned int
count_leading_zeros (UTItype x)
{
*************** __udivmodti4 (UTItype num, UTItype den,
*** 67,74 ****
{
qword shift =
si_from_uint (count_leading_zeros (den) - count_leading_zeros (num));
! qword n0 = *(qword *) & num;
! qword d0 = *(qword *) & den;
qword bit = si_andi (si_fsmbi (1), 1);
qword r0 = si_il (0);
qword m1 = si_fsmbi (0x000f);
--- 89,96 ----
{
qword shift =
si_from_uint (count_leading_zeros (den) - count_leading_zeros (num));
! qword n0 = si_from_UTItype (num);
! qword d0 = si_from_UTItype (den);
qword bit = si_andi (si_fsmbi (1), 1);
qword r0 = si_il (0);
qword m1 = si_fsmbi (0x000f);
*************** __udivmodti4 (UTItype num, UTItype den,
*** 101,108 ****
}
while (si_to_uint (si_orx (bit)));
if (rp)
! *rp = *(UTItype *) & n0;
! return *(UTItype *) & r0;
}
UTItype
--- 123,130 ----
}
while (si_to_uint (si_orx (bit)));
if (rp)
! *rp = si_to_UTItype (n0);
! return si_to_UTItype (r0);
}
UTItype
Index: gcc/config/spu/multi3.c
===================================================================
*** gcc/config/spu/multi3.c (revision 172953)
--- gcc/config/spu/multi3.c (working copy)
***************
*** 23,28 ****
--- 23,50 ----
typedef int TItype __attribute__ ((mode (TI)));
+ union qword_TItype
+ {
+ qword q;
+ TItype t;
+ };
+
+ inline static qword
+ si_from_TItype (TItype t)
+ {
+ union qword_TItype u;
+ u.t = t;
+ return u.q;
+ }
+
+ inline static TItype
+ si_to_TItype (qword q)
+ {
+ union qword_TItype u;
+ u.q = q;
+ return u.t;
+ }
+
/* A straight forward vectorization and unrolling of
* short l[8], r[8];
* TItype total = 0;
*************** typedef int TItype __attribute__ ((mode
*** 33,40 ****
TItype
__multi3 (TItype l, TItype r)
{
! qword u = *(qword *) & l;
! qword v = *(qword *) & r;
qword splat0 = si_shufb (v, v, si_ilh (0x0001));
qword splat1 = si_shufb (v, v, si_ilh (0x0203));
qword splat2 = si_shufb (v, v, si_ilh (0x0405));
--- 55,62 ----
TItype
__multi3 (TItype l, TItype r)
{
! qword u = si_from_TItype (l);
! qword v = si_from_TItype (r);
qword splat0 = si_shufb (v, v, si_ilh (0x0001));
qword splat1 = si_shufb (v, v, si_ilh (0x0203));
qword splat2 = si_shufb (v, v, si_ilh (0x0405));
*************** __multi3 (TItype l, TItype r)
*** 93,97 ****
total = si_cgx (total10, carry, total);
total = si_shlqbyi (total, 4);
total = si_addx (total10, carry, total);
! return *(TItype *) & total;
}
--- 115,119 ----
total = si_cgx (total10, carry, total);
total = si_shlqbyi (total, 4);
total = si_addx (total10, carry, total);
! return si_to_TItype (total);
}
--
Dr. Ulrich Weigand
GNU Toolchain for Linux on System z and Cell BE
ulrich.weig...@de.ibm.com
