https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83920
--- Comment #9 from cesar at gcc dot gnu.org ---
I figured out why my patch does work. Here's the assembly code for your C test
case at -O0:
{
.reg.u32 %x;
mov.u32 %x, %tid.x;
setp.ne.u32 %r81, %x, 0;
}
@%r81 bra $L11;
mov.u64 %r58, %ar0;
st.u64 [%frame+32], %r58;
mov.u64 %r59, %ar1;
st.u64 [%frame+40], %r59;
ld.u64 %r60, [%frame+40];
cvt.u32.u64 %r26, %r60;
st.u32 [%frame], %r26;
ld.u64 %r61, [%frame+32];
st.u64 [%frame+8], %r61;
mov.u32 %r22, 0;
mov.u32 %r28, 1;
mov.u32 %r29, 1;
$L11:
$L9:
mov.pred %r82, %r62;
setp.eq.u32 %r62, 1, 0;
@%r81 bra $L12;
mov.u32 %r55, %nctaid.x;
mov.u32 %r56, %ctaid.x;
mov.u32 %r48, 9;
add.u32 %r49, %r55, %r48;
div.s32 %r50, %r49, %r55;
mul.lo.u32 %r23, %r56, %r50;
mov.u32 %r57, %nctaid.x;
mov.u32 %r51, 9;
add.u32 %r52, %r57, %r51;
div.s32 %r53, %r52, %r57;
add.u32 %r54, %r23, %r53;
min.s32 %r30, %r54, 10;
setp.ge.s32 %r62, %r23, %r30;
$L12:
mov.pred %r62, %r82;
selp.u32 %r83, 1, 0, %r62;
shfl.idx.b32 %r83, %r83, 0, 31;
setp.ne.u32 %r62, %r83, 0;
@%r62 bra.uni $L2;
$L4:
mov.pred %r84, %r64;
setp.eq.u32 %r64, 1, 0;
@%r81 bra $L13;
The predicate register in question here is %r62. Notice how the JIT workaround
clobbers %r62 much earlier than it's defined. My patch just copied the register
predicate register before it was clobbered. That's fine, but when it restores
the value of %r62 in L12, r62 gets an uninitialized value.
