This avoids forwprop from matching DFP <-> FP vector conversions using VEC_[UN]PACK{_TRUNC,_LO,_HI}. Maybe DFP vectors shouldn't be a thing, but they appearantly are. Re-using CONVERT/NOP_EXPR for DFP <-> FP conversions was probably a mistake.
Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed. 2022-02-14 Richard Biener <rguent...@suse.de> PR tree-optimization/104511 * tree-ssa-forwprop.c (simplify_vector_constructor): Avoid touching DFP <-> FP conversions. * gcc.dg/pr104511.c: New testcase. --- gcc/testsuite/gcc.dg/pr104511.c | 16 ++++++++++++++++ gcc/tree-ssa-forwprop.cc | 9 +++++++++ 2 files changed, 25 insertions(+) create mode 100644 gcc/testsuite/gcc.dg/pr104511.c diff --git a/gcc/testsuite/gcc.dg/pr104511.c b/gcc/testsuite/gcc.dg/pr104511.c new file mode 100644 index 00000000000..ad5430c67c8 --- /dev/null +++ b/gcc/testsuite/gcc.dg/pr104511.c @@ -0,0 +1,16 @@ +/* { dg-do compile { target dfp } } */ +/* { dg-options "-O -Wno-psabi" } */ + +typedef _Float64 __attribute__((__vector_size__ (32))) F; +typedef _Decimal32 __attribute__((__vector_size__ (16))) D; + +extern void bar (void); + +D g; +void +foo (F f) +{ + D d = __builtin_convertvector (f, D); + bar (); + g = d; +} diff --git a/gcc/tree-ssa-forwprop.cc b/gcc/tree-ssa-forwprop.cc index 709bde6defa..484491fa1c5 100644 --- a/gcc/tree-ssa-forwprop.cc +++ b/gcc/tree-ssa-forwprop.cc @@ -2824,6 +2824,15 @@ simplify_vector_constructor (gimple_stmt_iterator *gsi) ? VEC_UNPACK_FLOAT_HI_EXPR : VEC_UNPACK_HI_EXPR); + /* Conversions between DFP and FP have no special tree code + but we cannot handle those since all relevant vector conversion + optabs only have a single mode. */ + if (CONVERT_EXPR_CODE_P (conv_code) + && FLOAT_TYPE_P (TREE_TYPE (type)) + && (DECIMAL_FLOAT_TYPE_P (TREE_TYPE (type)) + != DECIMAL_FLOAT_TYPE_P (TREE_TYPE (conv_src_type)))) + return false; + if (CONVERT_EXPR_CODE_P (conv_code) && (2 * TYPE_PRECISION (TREE_TYPE (TREE_TYPE (orig[0]))) == TYPE_PRECISION (TREE_TYPE (type))) -- 2.34.1