Re: [RFC PATCH] target/m68k: Use i128 for 128-bit load/store in m68k_copy_line()

[Andreas Schwab]

On Okt 17 2023, Richard Henderson wrote:

> The manual says "burst reads and writes" without defining those terms.

Burst transfers are explained in the M68040UM (7.4.2 Line Read Transfer
and 7.4.4 Line Write Transfers).

-- 
Andreas Schwab, SUSE Labs, sch...@suse.de
GPG Key fingerprint = 0196 BAD8 1CE9 1970 F4BE  1748 E4D4 88E3 0EEA B9D7
"And now for something completely different."

Re: [RFC PATCH] target/m68k: Use i128 for 128-bit load/store in m68k_copy_line()

[Richard Henderson]

On 10/17/23 05:27, Philippe Mathieu-Daudé wrote:

Signed-off-by: Philippe Mathieu-Daudé 
---
Based-on: <20231013175109.124308-1-richard.hender...@linaro.org>
   tcg: Add tcg_gen_{ld,st}_i128

RFC because unsure and untested...
---
  target/m68k/translate.c | 14 --
  1 file changed, 4 insertions(+), 10 deletions(-)

diff --git a/target/m68k/translate.c b/target/m68k/translate.c
index 4d0110de95..1e3d155bd9 100644
--- a/target/m68k/translate.c
+++ b/target/m68k/translate.c
@@ -4293,23 +4293,17 @@ DISAS_INSN(chk2)
  
  static void m68k_copy_line(TCGv dst, TCGv src, int index)

  {
+MemOp mop = MO_128 | MO_TE;
+TCGv_i128 t = tcg_temp_new_i128();
  TCGv addr;
-TCGv_i64 t0, t1;
  
  addr = tcg_temp_new();
  
-t0 = tcg_temp_new_i64();

-t1 = tcg_temp_new_i64();
-
  tcg_gen_andi_i32(addr, src, ~15);
-tcg_gen_qemu_ld_i64(t0, addr, index, MO_TEUQ);
-tcg_gen_addi_i32(addr, addr, 8);
-tcg_gen_qemu_ld_i64(t1, addr, index, MO_TEUQ);
+tcg_gen_qemu_ld_i128(t, addr, index, mop);
  
  tcg_gen_andi_i32(addr, dst, ~15);

-tcg_gen_qemu_st_i64(t0, addr, index, MO_TEUQ);
-tcg_gen_addi_i32(addr, addr, 8);
-tcg_gen_qemu_st_i64(t1, addr, index, MO_TEUQ);
+tcg_gen_st_i128(t, addr, index);


Aside from the typo, the other thing you need to consider when introducing 16-byte 
operations is the atomicity.  Do you want or need this to be atomic?


For m68k, I strongly suspect that we don't need the entire read or write to be 
atomic.

The manual says "burst reads and writes" without defining those terms.  I suspect that 
MO_ATOM_NONE is sufficient, but MO_ATOM_IFALIGN_PAIR would preserve the atomicity of the 
current code.



r~

Re: [RFC PATCH] target/m68k: Use i128 for 128-bit load/store in m68k_copy_line()

[Richard Henderson]

On 10/17/23 05:27, Philippe Mathieu-Daudé wrote:

-tcg_gen_qemu_st_i64(t1, addr, index, MO_TEUQ);
+tcg_gen_st_i128(t, addr, index);


Lost the "qemu".


r~