When simplifying e.g. (float_truncate:SF (float_truncate:DF (reg:XF))
or (float_truncate:SF (float_extend:XF (reg:DF)) etc. into
(float_truncate:SF (reg:XF)) or (float_truncate:SF (reg:DF)) we call
simplify_gen_unary with incorrect op_mode argument, it should be
the argument's mode, but we call it with the outer mode instead.
As these are all floating point operations, the argument always
has non-VOIDmode and so we can just use that mode (as done in similar
simplifications a few lines later), but neither FLOAT_TRUNCATE nor
FLOAT_EXTEND are operations that should have the same modes of operand
and result. This bug hasn't been a problem for years because normally
op_mode is used only if the mode of op is VOIDmode, otherwise it is
redundant, but r10-2139 added an assertion in some spots that op_mode
is right even in such cases.
2024-01-31 Jakub Jelinek <jakub@redhat.com>
PR rtl-optimization/113656
* simplify-rtx.cc (simplify_context::simplify_unary_operation_1)
<case FLOAT_TRUNCATE>: Fix up last argument to simplify_gen_unary.
* gcc.target/i386/pr113656.c: New test.
(cherry picked from commit
b59775b642bb2b1ecd2e6d52c988b9c432117bc8)
> GET_MODE_UNIT_SIZE (mode)
? FLOAT_TRUNCATE : FLOAT_EXTEND,
mode,
- XEXP (op, 0), mode);
+ XEXP (op, 0), GET_MODE (XEXP (op, 0)));
/* (float_truncate (float x)) is (float x) */
if ((GET_CODE (op) == FLOAT || GET_CODE (op) == UNSIGNED_FLOAT)
--- /dev/null
+/* PR rtl-optimization/113656 */
+/* { dg-do compile } */
+/* { dg-options "-O3 -frounding-math -funsafe-math-optimizations -mavx512fp16 -mavx512vl" } */
+
+_Float16 a[8];
+
+void
+foo ()
+{
+ for (int i = 0; i < 8; i++)
+ a[i] = i - 8.4;
+}