The region-length bounds check compared (EA + i0) <= (EA + LEN), which reduces to i0 <= LEN and never actually depends on the region base. This let vgather/vscatter keep elements whose offset was beyond the declared region instead of dropping them. Compare the offset to the region length directly instead.
Only load bytes for gather lanes that are kept. A dropped or predicate-false lane must not access memory and raise an exception. Signed-off-by: Brian Cain <[email protected]> --- target/hexagon/mmvec/macros.h | 31 +++++++++++++++++-------------- 1 file changed, 17 insertions(+), 14 deletions(-) diff --git a/target/hexagon/mmvec/macros.h b/target/hexagon/mmvec/macros.h index e85beed3387..858afd995d9 100644 --- a/target/hexagon/mmvec/macros.h +++ b/target/hexagon/mmvec/macros.h @@ -130,9 +130,9 @@ do { \ int log_byte = 0; \ target_ulong va = EA; \ - target_ulong va_high = EA + LEN; \ + int in_region = (OFFSET) <= (LEN); \ for (int i0 = 0; i0 < 4; i0++) { \ - log_byte = (va + i0) <= va_high; \ + log_byte = in_region; \ LOG_VTCM_BYTE(va + i0, log_byte, \ hexagon_mmvec_get_byte(&(INC), 4 * IDX + i0), \ 4 * IDX + i0); \ @@ -142,9 +142,9 @@ do { \ int log_byte = 0; \ target_ulong va = EA; \ - target_ulong va_high = EA + LEN; \ + int in_region = (OFFSET) <= (LEN); \ for (int i0 = 0; i0 < 2; i0++) { \ - log_byte = (va + i0) <= va_high; \ + log_byte = in_region; \ LOG_VTCM_BYTE(va + i0, log_byte, \ hexagon_mmvec_get_byte(&(INC), 2 * IDX + i0), \ 2 * IDX + i0); \ @@ -156,9 +156,9 @@ do { \ int log_byte = 0; \ target_ulong va = EA; \ - target_ulong va_high = EA + LEN; \ + int in_region = (OFFSET) <= (LEN); \ for (int i0 = 0; i0 < 2; i0++) { \ - log_byte = (va + i0) <= va_high; \ + log_byte = in_region; \ LOG_VTCM_BYTE(va + i0, log_byte, \ hexagon_mmvec_get_byte(&(INC), 2 * IDX + i0), \ 2 * IDX + i0); \ @@ -170,15 +170,18 @@ do { \ int i0; \ target_ulong va = EA; \ - target_ulong va_high = EA + LEN; \ uintptr_t ra = GETPC(); \ int log_byte = 0; \ + int in_region = (OFFSET) <= (LEN); \ for (i0 = 0; i0 < ELEMENT_SIZE; i0++) { \ - log_byte = ((va + i0) <= va_high) && QVAL; \ - uint8_t B; \ - B = cpu_ldub_data_ra(env, EA + i0, ra); \ - hexagon_mmvec_set_byte(&env->tmp_VRegs[0], \ - ELEMENT_SIZE * IDX + i0, B); \ + uint8_t B = 0; \ + \ + log_byte = in_region && QVAL; \ + if (log_byte) { \ + B = cpu_ldub_data_ra(env, va + i0, ra); \ + hexagon_mmvec_set_byte(&env->tmp_VRegs[0], \ + ELEMENT_SIZE * IDX + i0, B); \ + } \ LOG_VTCM_BYTE(va + i0, log_byte, B, ELEMENT_SIZE * IDX + i0); \ } \ } while (0) @@ -250,10 +253,10 @@ do { \ int i0; \ target_ulong va = EA; \ - target_ulong va_high = EA + LEN; \ int log_byte = 0; \ + int in_region = (OFFSET) <= (LEN); \ for (i0 = 0; i0 < ELEM_SIZE; i0++) { \ - log_byte = ((va + i0) <= va_high) && QVAL; \ + log_byte = in_region && QVAL; \ LOG_VTCM_BYTE(va + i0, log_byte, \ hexagon_mmvec_get_byte(&(IN), \ ELEM_SIZE * IDX + i0), \ -- 2.34.1
