| Message ID | 69382818-fafa-bab7-31be-01bc476d60b4@linux.ibm.com |
|---|---|
| State | New |
| Headers | show |
| Series | rs6000: MMA type causes an ICE in ranger pass due to incompatible types | expand |
Hi! On Wed, Oct 21, 2020 at 01:00:20PM -0500, Peter Bergner wrote: > PR97360 shows a problem in how we create our PXI and POI modes that cause > an ICE in the ranger pass. The problem seems to be that the extra call > to build_distinct_type_copy() also creates new TYPE_{MIN,MAX}_VALUEs that > are not compatible/the same as the base type itself. The simple "fix" is > to actually remove the unneeded build_distinct_type_copy(), since according > to richi, the types returned from make_unsigned_type() are already distinct. > > The following patch from Andrew and richi fixes the ICE on Martin's test > case and passes bootstrap and regtesting on powerpc64le-linux. > Ok for trunk? Yes, okay for trunk. Thanks! > Since the ranger code that triggered this doesn't seem to be in GCC 10, > I assume we do not want to backport this this change? No, please do, in a week or so, it is a pretty serious problem that we could just asa well run into some other way, as far as I can see? Segher > gcc/ > PR target/97360 > * config/rs6000/rs6000-call.c (rs6000_init_builtins): Remove call to > build_distinct_type_copy(). > > gcc/testsuite/ > PR target/97360 > * gcc.target/powerpc/pr97360.c: New test.
On 10/21/20 1:34 PM, Segher Boessenkool wrote: >> The following patch from Andrew and richi fixes the ICE on Martin's test >> case and passes bootstrap and regtesting on powerpc64le-linux. >> Ok for trunk? > > Yes, okay for trunk. Thanks! Ok, pushed to trunk. Thanks! >> Since the ranger code that triggered this doesn't seem to be in GCC 10, >> I assume we do not want to backport this this change? > > No, please do, in a week or so, it is a pretty serious problem that we > could just asa well run into some other way, as far as I can see? Ok, I'll wait a week and then do the backport and testing. Peter
On 10/21/20 2:31 PM, Peter Bergner wrote:
> Ok, I'll wait a week and then do the backport and testing.
I just pushed this to the GCC 10 release branch now.
Peter
diff --git a/gcc/config/rs6000/rs6000-call.c b/gcc/config/rs6000/rs6000-call.c index 9fdf97bc803..7639aab171d 100644 --- a/gcc/config/rs6000/rs6000-call.c +++ b/gcc/config/rs6000/rs6000-call.c @@ -12914,15 +12914,13 @@ rs6000_init_builtins (void) /* Vector pair and vector quad support. */ if (TARGET_EXTRA_BUILTINS) { - tree oi_uns_type = make_unsigned_type (256); - vector_pair_type_node = build_distinct_type_copy (oi_uns_type); + vector_pair_type_node = make_unsigned_type (256); SET_TYPE_MODE (vector_pair_type_node, POImode); layout_type (vector_pair_type_node); lang_hooks.types.register_builtin_type (vector_pair_type_node, "__vector_pair"); - tree xi_uns_type = make_unsigned_type (512); - vector_quad_type_node = build_distinct_type_copy (xi_uns_type); + vector_quad_type_node = make_unsigned_type (512); SET_TYPE_MODE (vector_quad_type_node, PXImode); layout_type (vector_quad_type_node); lang_hooks.types.register_builtin_type (vector_quad_type_node, diff --git a/gcc/testsuite/gcc.target/powerpc/pr97360.c b/gcc/testsuite/gcc.target/powerpc/pr97360.c new file mode 100644 index 00000000000..2328d28a283 --- /dev/null +++ b/gcc/testsuite/gcc.target/powerpc/pr97360.c @@ -0,0 +1,18 @@ +/* PR target/97360 */ +/* { dg-do compile } */ +/* { dg-require-effective-target power10_ok } */ +/* { dg-options "-O2 -mdejagnu-cpu=power10" } */ + +/* Verify we do not ICE on the test below. */ + +typedef unsigned char vec_t __attribute__((vector_size(16))); + +void +foo (__vector_quad *dst, __vector_pair *vpair, vec_t *vec) +{ + __vector_quad acc = *dst; + for (;;) + { + __builtin_mma_xvf64gerpp(&acc, *vpair, vec[7]); + } +}
PR97360 shows a problem in how we create our PXI and POI modes that cause an ICE in the ranger pass. The problem seems to be that the extra call to build_distinct_type_copy() also creates new TYPE_{MIN,MAX}_VALUEs that are not compatible/the same as the base type itself. The simple "fix" is to actually remove the unneeded build_distinct_type_copy(), since according to richi, the types returned from make_unsigned_type() are already distinct. The following patch from Andrew and richi fixes the ICE on Martin's test case and passes bootstrap and regtesting on powerpc64le-linux. Ok for trunk? Since the ranger code that triggered this doesn't seem to be in GCC 10, I assume we do not want to backport this this change? Peter gcc/ PR target/97360 * config/rs6000/rs6000-call.c (rs6000_init_builtins): Remove call to build_distinct_type_copy(). gcc/testsuite/ PR target/97360 * gcc.target/powerpc/pr97360.c: New test.