On 11/11/2026 12:50 PM, Daniel Henrique Barboza wrote:
> gen_cmpxchg* and trans_amocas_q() load amocas registers using get_gpr().
> When the register is x0/zero get_gpr() returns ctx->zero, the shared TCG
> constant that every x0 read uses.  As it is today, the amocas helpers
> uses tcg_gen_atomic_cmpxchg_* helpers using the dest register retrieved
> by get_gpr().
> 
> So, if we have something like:
> 
> amocas.d                zero,s0,(a0)
> 
> gen_cmpxchg() uses ctx->zero as 'dest' for tcg_gen_atomic_cmpxchg_tl(),
> which will end up clobbering it with whatever is in 'rs1'.  ctx->zero
> will be reset back to normal in init_disas_context(), but if there's
> something else that ends up reading it in the same TB, e.g.:
> 
> beqz                    a2,24
> 
> This will read a clobbered ctx->zero and then all bets are off.
> 
> The fix is to use a temp variable as 'dest' for gen_atomic_cmpxchg_*
> calls instead of the result of get_gpr().  We are already executing
> gen_set_grp() in all cases using 'dest', making this change a bit
> simpler.
> 
> Fixes: b52d49e97fc4 ("target/riscv: Add support for Zacas extension")
> Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/3526
> Signed-off-by: Daniel Henrique Barboza <[email protected]>
> ---
>  .../riscv/tcg/insn_trans/trans_rvzacas.c.inc  | 15 ++++--
>  target/riscv/tcg/translate.c                  |  5 +-
>  tests/tcg/riscv64/system/meson.build          |  8 +++
>  tests/tcg/riscv64/test-amocas-x0.S            | 49 +++++++++++++++++++
>  4 files changed, 70 insertions(+), 7 deletions(-)
>  create mode 100644 tests/tcg/riscv64/test-amocas-x0.S
> 
> diff --git a/target/riscv/tcg/insn_trans/trans_rvzacas.c.inc 
> b/target/riscv/tcg/insn_trans/trans_rvzacas.c.inc
> index 79bca1e957..e908ed2377 100644
> --- a/target/riscv/tcg/insn_trans/trans_rvzacas.c.inc
> +++ b/target/riscv/tcg/insn_trans/trans_rvzacas.c.inc
> @@ -72,13 +72,14 @@ static bool gen_cmpxchg64(DisasContext *ctx, arg_atomic 
> *a, MemOp mop)
>          return false;
>      }
>  
> -    TCGv_i64 dest = get_gpr_pair(ctx, a->rd);
> +    TCGv_i64 cmp = get_gpr_pair(ctx, a->rd);
>      TCGv src1 = get_address(ctx, a->rs1, 0);
>      TCGv_i64 src2 = get_gpr_pair(ctx, a->rs2);
> +    TCGv_i64 dest = tcg_temp_new_i64();
>  
>      mop |= ctx->mo_endianness;
>      decode_save_opc(ctx, RISCV_UW2_ALWAYS_STORE_AMO);
> -    tcg_gen_atomic_cmpxchg_i64(dest, src1, dest, src2, ctx->mem_idx, mop);
> +    tcg_gen_atomic_cmpxchg_i64(dest, src1, cmp, src2, ctx->mem_idx, mop);
>  
>      gen_set_gpr_pair(ctx, a->rd, dest);
>      return true;
> @@ -117,15 +118,19 @@ static bool trans_amocas_q(DisasContext *ctx, 
> arg_amocas_q *a)
>      TCGv_i128 src2 = tcg_temp_new_i128();
>      TCGv_i64 src2l = get_gpr(ctx, a->rs2, EXT_NONE);
>      TCGv_i64 src2h = get_gpr(ctx, a->rs2 == 0 ? 0 : a->rs2 + 1, EXT_NONE);
> -    TCGv_i64 destl = get_gpr(ctx, a->rd, EXT_NONE);
> -    TCGv_i64 desth = get_gpr(ctx, a->rd == 0 ? 0 : a->rd + 1, EXT_NONE);
> +    TCGv_i64 cmpl = get_gpr(ctx, a->rd, EXT_NONE);
> +    TCGv_i64 cmph = get_gpr(ctx, a->rd == 0 ? 0 : a->rd + 1, EXT_NONE);
> +    TCGv_i64 destl = tcg_temp_new_i64();
> +    TCGv_i64 desth = tcg_temp_new_i64();
> +    TCGv_i128 cmp = tcg_temp_new_i128();
>      MemOp memop = MO_ALIGN | MO_UO;
>  
>      memop |= ctx->mo_endianness;
> +    tcg_gen_concat_i64_i128(cmp, cmpl, cmph);
>      tcg_gen_concat_i64_i128(src2, src2l, src2h);
>      tcg_gen_concat_i64_i128(dest, destl, desth);
>      decode_save_opc(ctx, RISCV_UW2_ALWAYS_STORE_AMO);
> -    tcg_gen_atomic_cmpxchg_i128(dest, src1, dest, src2, ctx->mem_idx, memop);
> +    tcg_gen_atomic_cmpxchg_i128(dest, src1, cmp, src2, ctx->mem_idx, memop);
>  
>      tcg_gen_extr_i128_i64(destl, desth, dest);
>  
> diff --git a/target/riscv/tcg/translate.c b/target/riscv/tcg/translate.c
> index 79b0f444ef..b9f47a4fc2 100644
> --- a/target/riscv/tcg/translate.c
> +++ b/target/riscv/tcg/translate.c
> @@ -1175,13 +1175,14 @@ static bool gen_amo(DisasContext *ctx, arg_atomic *a,
>  
>  static bool gen_cmpxchg(DisasContext *ctx, arg_atomic *a, MemOp mop)
>  {
> -    TCGv dest = get_gpr(ctx, a->rd, EXT_NONE);
> +    TCGv cmp = get_gpr(ctx, a->rd, EXT_NONE);
>      TCGv src1 = get_address(ctx, a->rs1, 0);
>      TCGv src2 = get_gpr(ctx, a->rs2, EXT_NONE);
> +    TCGv dest = tcg_temp_new();
>  
>      mop |= ctx->mo_endianness;
>      decode_save_opc(ctx, RISCV_UW2_ALWAYS_STORE_AMO);
> -    tcg_gen_atomic_cmpxchg_tl(dest, src1, dest, src2, ctx->mem_idx, mop);
> +    tcg_gen_atomic_cmpxchg_tl(dest, src1, cmp, src2, ctx->mem_idx, mop);
>  
>      gen_set_gpr(ctx, a->rd, dest);
>      return true;
> diff --git a/tests/tcg/riscv64/system/meson.build 
> b/tests/tcg/riscv64/system/meson.build
> index 3c8b018d9e..985c492a00 100644
> --- a/tests/tcg/riscv64/system/meson.build
> +++ b/tests/tcg/riscv64/system/meson.build
> @@ -89,6 +89,14 @@ tests += {
>    }
>  }
>  
> +tests += {
> +  'test-amocas-x0.S': {
> +    'exe_name': 'test-amocas-x0',
> +    'cflags': cflags + ['-march=rv64gc'],
> +    'qemu_args': ['-cpu', 'max', qemu_args],
> +  }
> +}

Default exe_name is derived from source name by default, so no need to
add it here.

> +
>  if 'qemu-system-riscv64' in emulators
>    tcg_tests += {
>      'riscv64-softmmu': {
> diff --git a/tests/tcg/riscv64/test-amocas-x0.S 
> b/tests/tcg/riscv64/test-amocas-x0.S
> new file mode 100644
> index 0000000000..d71814035c
> --- /dev/null
> +++ b/tests/tcg/riscv64/test-amocas-x0.S
> @@ -0,0 +1,49 @@
> +/*
> + * Test if amocas.d will clobber x0.  Based on Gitlab 3526.
> + *
> + * Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
> + * SPDX-License-Identifier: GPL-2.0-or-later
> + */
> +
> +     .option norvc
> +     .text
> +
> +     .global _start
> +
> +_start:
> +     li     t0, 5
> +     lla    t1, amo_target
> +     li     t2, 0
> +
> +     /* amocas.d zero, t0, (t1) */
> +     .word  0x2853302f
> +
> +        /* If x0 is clobbered this bnez will fail.  */
> +     bnez   t2, _fail
> +
> +     li     a0, 0
> +     j      _exit
> +
> +_fail:
> +     li    a0, 1
> +
> +/* Exit through the semihosting SYS_EXIT_EXTENDED call with a0 as the code. 
> */
> +_exit:
> +     lla     a1, semiargs
> +     li      t0, 0x20026     /* ADP_Stopped_ApplicationExit */
> +     sd      t0, 0(a1)
> +     sd      a0, 8(a1)
> +     li      a0, 0x20        /* TARGET_SYS_EXIT_EXTENDED */
> +     .balign 16
> +     slli    zero, zero, 0x1f
> +     ebreak
> +     srai    zero, zero, 0x7
> +     j       .
> +
> +     .data
> +     .balign 8
> +semiargs:
> +     .space  16
> +/* An amo target must be a valid, 8-byte aligned address.  */
> +amo_target:
> +     .dword  0x1234


Reply via email to