On 10/30/2026 12:47 PM, Marco Liebel wrote:
On Thu, Oct 29, 2026 at 9:41 PM Brian Cain <[email protected]> wrote:
Implement an HMX accumulator clear instruction - mxclracc.
Signed-off-by: Brian Cain <[email protected]>
---
target/hexagon/gen_tcg_hmx.h | 13 +++++++++++++
target/hexagon/helper.h | 2 ++
target/hexagon/genptr.c | 1 +
target/hexagon/op_helper.c | 10 ++++++++++
target/hexagon/translate.c | 3 ++-
target/hexagon/gen_idef_parser_funcs.py | 6 ++++++
target/hexagon/hex_common.py | 2 ++
target/hexagon/meson.build | 17 +++++++++--------
8 files changed, 45 insertions(+), 9 deletions(-)
create mode 100644 target/hexagon/gen_tcg_hmx.h
+void HELPER(mxclracc)(CPUHexagonState *env)
+{
+ HmxState *hmx = env->hmx_state;
+ int i;
+
+ for (i = 0; i < HMX_NUM_ACC_SETS; i++) {
+ memset(&hmx->acc[i].fxp_primary, 0, sizeof(hmx->acc[i].fxp_primary));
+ }
+}
The helper clears live state before packet commit. A store fault in the same
packet can leave the accumulators cleared without a commit.
It also dereferences hmx_state without checking that an HMX context exists.
env->hmx_state can be NULL here, because gen_start_packet() only checks
XE2.
Sent a v3 with fixes for these.
