Re: [PATCH 2/3] target/riscv: Honor Zama16b for AMOCAS accesses

[Daniel Henrique Barboza]

On 10/24/2026 9:02 AM, GavinQin wrote:

AMOCAS accesses retain their natural alignment requirement when
Zama16b is enabled, causing misaligned accesses within a naturally
aligned 16-byte granule to trap.

Clear the alignment requirement and request MO_ATOM_WITHIN16 in
gen_cmpxchg() and in gen_cmpxchg64(), which handles AMOCAS.D using
register pairs on RV32.

Preserve the existing alignment requirement when Zama16b is disabled.

Signed-off-by: GavinQin <qinzhengxi...@foxmail.com>
---

Reviewed-by: Daniel Henrique Barboza <daniel.barb...@oss.qualcomm.com>

  target/riscv/tcg/insn_trans/trans_rvzacas.c.inc | 4 ++++
  target/riscv/tcg/translate.c                    | 4 ++++
  2 files changed, 8 insertions(+)

diff --git a/target/riscv/tcg/insn_trans/trans_rvzacas.c.inc 
b/target/riscv/tcg/insn_trans/trans_rvzacas.c.inc
index 79bca1e957..7b87f3e640 100644
--- a/target/riscv/tcg/insn_trans/trans_rvzacas.c.inc
+++ b/target/riscv/tcg/insn_trans/trans_rvzacas.c.inc
@@ -77,6 +77,10 @@ static bool gen_cmpxchg64(DisasContext *ctx, arg_atomic *a, 
MemOp mop)
      TCGv_i64 src2 = get_gpr_pair(ctx, a->rs2);

mop |= ctx->mo_endianness;

+    if (ctx->cfg_ptr->ext_zama16b) {
+        mop &= ~MO_AMASK;
+        mop |= MO_ATOM_WITHIN16;
+    }
      decode_save_opc(ctx, RISCV_UW2_ALWAYS_STORE_AMO);
      tcg_gen_atomic_cmpxchg_i64(dest, src1, dest, src2, ctx->mem_idx, mop);

diff --git a/target/riscv/tcg/translate.c b/target/riscv/tcg/translate.c

index 70c1088338..3d1bd7371d 100644
--- a/target/riscv/tcg/translate.c
+++ b/target/riscv/tcg/translate.c
@@ -1180,6 +1180,10 @@ static bool gen_cmpxchg(DisasContext *ctx, arg_atomic 
*a, MemOp mop)
      TCGv src2 = get_gpr(ctx, a->rs2, EXT_NONE);

mop |= ctx->mo_endianness;

+    if (ctx->cfg_ptr->ext_zama16b && (mop & MO_SIZE) >= MO_16) {
+        mop &= ~MO_ALIGN;
+        mop |= MO_ATOM_WITHIN16;
+    }
      decode_save_opc(ctx, RISCV_UW2_ALWAYS_STORE_AMO);
      tcg_gen_atomic_cmpxchg_tl(dest, src1, dest, src2, ctx->mem_idx, mop);