Block rotate to fshl optimization, as we don't implement fshl. Set reg for physical registers to avoid out-of-range index crash.
Signed-off-by: Rebecca N. Palmer <rebecca_pal...@zoho.com> --- (where patches 1 and 2 are the LLVM 8/9 ones from https://svnweb.freebsd.org/ports/head/lang/beignet/files/ ; see also https://salsa.debian.org/opencl-team/beignet ) The tests succeed with this - I'm still not sure if I like it (further testing welcome), but given that upstream is abandoned, it's probably all we're going to get. For rotate, I tried to add fshl/fshr to llvm_gen_backend.cpp (mapped to rsl/rsr, though I never got far enough to find out whether that's actually correct), but there are enough other places that don't handle rsl/rsr that the obvious ways didn't work, so I decided to block the optimization instead. For subgroup, testing with LLVM 7 found that the underlying "physical register passed to a place it maybe shouldn't be" issue is not a new problem, so I kept the previously posted fix for the crash and didn't investigate further. --- a/backend/src/libocl/tmpl/ocl_integer.tmpl.cl +++ b/backend/src/libocl/tmpl/ocl_integer.tmpl.cl @@ -216,13 +216,14 @@ OVERLOADABLE ulong mad_sat(ulong a, ulon return __gen_ocl_mad_sat(a, b, c); } -OVERLOADABLE uchar __rotate_left(uchar x, uchar y) { return (x << y) | (x >> (8 - y)); } +// the 'volatile' is to make the LLVM optimizer leave these alone, as it would convert them to intrinsics (fshl/fshr) that we don't implement +OVERLOADABLE uchar __rotate_left(uchar x, uchar y) { volatile uchar z; z = (x << y); return z | (x >> (8 - y)); } OVERLOADABLE char __rotate_left(char x, char y) { return __rotate_left((uchar)x, (uchar)y); } -OVERLOADABLE ushort __rotate_left(ushort x, ushort y) { return (x << y) | (x >> (16 - y)); } +OVERLOADABLE ushort __rotate_left(ushort x, ushort y) { volatile ushort z; z = (x << y); return z | (x >> (16 - y)); } OVERLOADABLE short __rotate_left(short x, short y) { return __rotate_left((ushort)x, (ushort)y); } -OVERLOADABLE uint __rotate_left(uint x, uint y) { return (x << y) | (x >> (32 - y)); } +OVERLOADABLE uint __rotate_left(uint x, uint y) { volatile uint z; z = (x << y); return z | (x >> (32 - y)); } OVERLOADABLE int __rotate_left(int x, int y) { return __rotate_left((uint)x, (uint)y); } -OVERLOADABLE ulong __rotate_left(ulong x, ulong y) { return (x << y) | (x >> (64 - y)); } +OVERLOADABLE ulong __rotate_left(ulong x, ulong y) { volatile ulong z; z = (x << y); return z | (x >> (64 - y)); } OVERLOADABLE long __rotate_left(long x, long y) { return __rotate_left((ulong)x, (ulong)y); } #define DEF(type, m) OVERLOADABLE type rotate(type x, type y) { return __rotate_left(x, (type)(y & m)); } DEF(char, 7) --- a/backend/src/backend/gen_register.hpp +++ b/backend/src/backend/gen_register.hpp @@ -225,6 +225,7 @@ namespace gbe uint32_t width, uint32_t hstride) { + this->value.reg = 0;//avoid subgroup crash this->type = type; this->file = file; this->nr = nr; _______________________________________________ Beignet mailing list Beignet@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/beignet
