[Pkg-opencl-devel] Bug#948563: [PATCH 3/3] Fix rotate and subgroup crashes with LLVM 8+

[Rebecca N. Palmer]

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_palmer at 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;