Skip to contentSkip to conversation

target/mips: Translate MSA vector load/store opcodes

RFC PATCH In review Philippe Mathieu-Daudé 2 participants 14 messages Apr 15 – Apr 16

lore
Share & tools
lore
#0 lore raw

v5 cover target/mips: Translate MSA vector load/store opcodes

Since v4:
- No algorithm change
- Split to ease review

Hi Richard,

I tried to implement your suggestion from v3 [*]. Unfortunately the box where I had the MSA test suite got erased, so it will take me some time to restore the setup. I was surprised these patches pass our CI tests so I added a pair of abort() which didn't fire, so this code path isn't tested at all there. Anyhow, do you mind having a quick look to see this is going in the right direction?

Thanks!

Phil.

[*] https://lore.kernel.org/qemu-devel/[email protected]/

Philippe Mathieu-Daudé (6):
target/mips: Expand TRANS_DF_iv() macro for MSA Load/Store target/mips: Convert MSA LD/ST.B (Byte Vector) target/mips: Convert MSA LD/ST.H (Halfword Vector) target/mips: Convert MSA LD/ST.W (Word Vector) target/mips: Convert MSA LD/ST.D (Doubleword Vector) target/mips: Endian-swap MSA LD/ST.D (Doubleword Vector)

 target/mips/tcg/msa_helper.h.inc |   9 --
 target/mips/tcg/msa_helper.c     | 176 -------------------------------
 target/mips/tcg/msa_translate.c  |  93 +++++++++++++---
 3 files changed, 81 insertions(+), 197 deletions(-)
Signature · 1 line
2.53.0
#1 lore raw

RFC PATCH v5 1/6 target/mips: Expand TRANS_DF_iv() macro for MSA Load/Store

In order to ease the conversion of MSA Load/Store runtime helpers to translated code, expand the TRANS_DF_iv() macro.

---
 target/mips/tcg/msa_translate.c | 59 ++++++++++++++++++++++++++-------
 1 file changed, 47 insertions(+), 12 deletions(-)
target/mips/tcg/msa_translate.c +47 -12
index 82b149922fa..2e1a769e032 100644
@@ -164,7 +164,6 @@ static inline bool check_msa_enabled(DisasContext *ctx)
     return true;
 }
 
-typedef void gen_helper_piv(TCGv_ptr, TCGv_i32, TCGv);
 typedef void gen_helper_pii(TCGv_ptr, TCGv_i32, TCGv_i32);
 typedef void gen_helper_piii(TCGv_ptr, TCGv_i32, TCGv_i32, TCGv_i32);
 typedef void gen_helper_piiii(TCGv_ptr, TCGv_i32, TCGv_i32, TCGv_i32, TCGv_i32);
@@ -175,9 +174,6 @@ typedef void gen_helper_piiii(TCGv_ptr, TCGv_i32, TCGv_i32, TCGv_i32, TCGv_i32);
     }; \
     TRANS(NAME, trans_func, NAME##_tab[a->df])
 
-#define TRANS_DF_iv(NAME, trans_func, gen_func) \
-    TRANS_DF_x(iv, NAME, trans_func, gen_func)
-
 #define TRANS_DF_ii(NAME, trans_func, gen_func) \
     TRANS_DF_x(ii, NAME, trans_func, gen_func)
 
@@ -758,25 +754,64 @@ TRANS(FTINT_U,  trans_msa_2rf, gen_helper_msa_ftint_u_df);
 TRANS(FFINT_S,  trans_msa_2rf, gen_helper_msa_ffint_s_df);
 TRANS(FFINT_U,  trans_msa_2rf, gen_helper_msa_ffint_u_df);
 
-static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a,
-                           gen_helper_piv *gen_msa_ldst)
+static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
 {
-    TCGv taddr;
+    TCGv_i32 wd;
+    TCGv_va addr;
 
     if (!check_msa_enabled(ctx)) {
         return true;
     }
 
-    taddr = tcg_temp_new();
+    addr = tcgv_va_temp_new();
+    gen_base_offset_addr(ctx, addr, a->ws, a->sa << a->df);
 
-    gen_base_offset_addr(ctx, taddr, a->ws, a->sa << a->df);
-    gen_msa_ldst(tcg_env, tcg_constant_i32(a->wd), taddr);
+    wd = tcg_constant_i32(a->wd);
+
+    if (is_load) {
+        switch (a->df) {
+        case 0:
+            gen_helper_msa_ld_b(tcg_env, wd, addr);
+            return true;
+        case 1:
+            gen_helper_msa_ld_h(tcg_env, wd, addr);
+            return true;
+        case 2:
+            gen_helper_msa_ld_w(tcg_env, wd, addr);
+            return true;
+        case 3:
+            gen_helper_msa_ld_d(tcg_env, wd, addr);
+            return true;
+        }
+    } else {
+        switch (a->df) {
+        case 0:
+            gen_helper_msa_st_b(tcg_env, wd, addr);
+            return true;
+        case 1:
+            gen_helper_msa_st_h(tcg_env, wd, addr);
+            return true;
+        case 2:
+            gen_helper_msa_st_w(tcg_env, wd, addr);
+            return true;
+        case 3:
+            gen_helper_msa_st_d(tcg_env, wd, addr);
+            return true;
+        }
+    }
 
     return true;
 }
 
-TRANS_DF_iv(LD, trans_msa_ldst, gen_helper_msa_ld);
-TRANS_DF_iv(ST, trans_msa_ldst, gen_helper_msa_st);
+static bool trans_LD(DisasContext *ctx, arg_msa_i *a)
+{
+    return trans_msa_ldst(ctx, a, true);
+}
+
+static bool trans_ST(DisasContext *ctx, arg_msa_i *a)
+{
+    return trans_msa_ldst(ctx, a, false);
+}
 
 static bool trans_LSA(DisasContext *ctx, arg_r *a)
 {
Signature · 1 line
2.53.0
Reviewed-by on v5 1/6
#2 lore raw

Re: [RFC PATCH v5 1/6] target/mips: Expand TRANS_DF_iv() macro for MSA Load/Store

Quoted 7 lines · Philippe Mathieu-Daudé

In order to ease the conversion of MSA Load/Store runtime helpers to translated code, expand the TRANS_DF_iv() macro.

---
  target/mips/tcg/msa_translate.c | 59 ++++++++++++++++++++++++++-------
  1 file changed, 47 insertions(+), 12 deletions(-)

r~

Quoted 96 lines
target/mips/tcg/msa_translate.c +47 -12
index 82b149922fa..2e1a769e032 100644
@@ -164,7 +164,6 @@ static inline bool check_msa_enabled(DisasContext *ctx)
      return true;
  }
 
-typedef void gen_helper_piv(TCGv_ptr, TCGv_i32, TCGv);
  typedef void gen_helper_pii(TCGv_ptr, TCGv_i32, TCGv_i32);
  typedef void gen_helper_piii(TCGv_ptr, TCGv_i32, TCGv_i32, TCGv_i32);
  typedef void gen_helper_piiii(TCGv_ptr, TCGv_i32, TCGv_i32, TCGv_i32, TCGv_i32);
@@ -175,9 +174,6 @@ typedef void gen_helper_piiii(TCGv_ptr, TCGv_i32, TCGv_i32, TCGv_i32, TCGv_i32);
      }; \
      TRANS(NAME, trans_func, NAME##_tab[a->df])
 
-#define TRANS_DF_iv(NAME, trans_func, gen_func) \
-    TRANS_DF_x(iv, NAME, trans_func, gen_func)
-
  #define TRANS_DF_ii(NAME, trans_func, gen_func) \
      TRANS_DF_x(ii, NAME, trans_func, gen_func)
 
@@ -758,25 +754,64 @@ TRANS(FTINT_U,  trans_msa_2rf, gen_helper_msa_ftint_u_df);
  TRANS(FFINT_S,  trans_msa_2rf, gen_helper_msa_ffint_s_df);
  TRANS(FFINT_U,  trans_msa_2rf, gen_helper_msa_ffint_u_df);
 
-static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a,
-                           gen_helper_piv *gen_msa_ldst)
+static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
  {
-    TCGv taddr;
+    TCGv_i32 wd;
+    TCGv_va addr;
 
      if (!check_msa_enabled(ctx)) {
          return true;
      }
 
-    taddr = tcg_temp_new();
+    addr = tcgv_va_temp_new();
+    gen_base_offset_addr(ctx, addr, a->ws, a->sa << a->df);
 
-    gen_base_offset_addr(ctx, taddr, a->ws, a->sa << a->df);
-    gen_msa_ldst(tcg_env, tcg_constant_i32(a->wd), taddr);
+    wd = tcg_constant_i32(a->wd);
+
+    if (is_load) {
+        switch (a->df) {
+        case 0:
+            gen_helper_msa_ld_b(tcg_env, wd, addr);
+            return true;
+        case 1:
+            gen_helper_msa_ld_h(tcg_env, wd, addr);
+            return true;
+        case 2:
+            gen_helper_msa_ld_w(tcg_env, wd, addr);
+            return true;
+        case 3:
+            gen_helper_msa_ld_d(tcg_env, wd, addr);
+            return true;
+        }
+    } else {
+        switch (a->df) {
+        case 0:
+            gen_helper_msa_st_b(tcg_env, wd, addr);
+            return true;
+        case 1:
+            gen_helper_msa_st_h(tcg_env, wd, addr);
+            return true;
+        case 2:
+            gen_helper_msa_st_w(tcg_env, wd, addr);
+            return true;
+        case 3:
+            gen_helper_msa_st_d(tcg_env, wd, addr);
+            return true;
+        }
+    }
 
      return true;
  }
 
-TRANS_DF_iv(LD, trans_msa_ldst, gen_helper_msa_ld);
-TRANS_DF_iv(ST, trans_msa_ldst, gen_helper_msa_st);
+static bool trans_LD(DisasContext *ctx, arg_msa_i *a)
+{
+    return trans_msa_ldst(ctx, a, true);
+}
+
+static bool trans_ST(DisasContext *ctx, arg_msa_i *a)
+{
+    return trans_msa_ldst(ctx, a, false);
+}
 
  static bool trans_LSA(DisasContext *ctx, arg_r *a)
  {
#3 lore raw

RFC PATCH v5 2/6 target/mips: Convert MSA LD/ST.B (Byte Vector)

In reply to Philippe Mathieu-Daudé

Replace runtime helpers by translation.

Replace the legacy cpu_ld/st*_data_ra() calls by tcg_gen_qemu_ld/st() which allow to respect atomicity.

Consider the host endianness.

---
 target/mips/tcg/msa_helper.h.inc |  1 -
 target/mips/tcg/msa_helper.c     | 28 ----------------------------
 target/mips/tcg/msa_translate.c  | 30 ++++++++++++++++++++++++------
 3 files changed, 24 insertions(+), 35 deletions(-)
target/mips/tcg/msa_helper.h.inc +0 -1
index 4963d1553a0..50a2732e916 100644
@@ -436,7 +436,6 @@ DEF_HELPER_4(msa_ffint_u_df, void, env, i32, i32, i32)
 #define MSALDST_PROTO(type)                         \
 DEF_HELPER_3(msa_ld_ ## type, void, env, i32, tl)   \
 DEF_HELPER_3(msa_st_ ## type, void, env, i32, tl)
-MSALDST_PROTO(b)
 MSALDST_PROTO(h)
 MSALDST_PROTO(w)
 MSALDST_PROTO(d)
target/mips/tcg/msa_helper.c +0 -28
index f554b3d10ee..d9f35fca657 100644
@@ -8223,20 +8223,6 @@ static inline uint64_t bswap32x2(uint64_t x)
     return ror64(bswap64(x), 32);
 }
 
-void helper_msa_ld_b(CPUMIPSState *env, uint32_t wd,
-                     target_ulong addr)
-{
-    wr_t *pwd = &(env->active_fpu.fpr[wd].wr);
-    uintptr_t ra = GETPC();
-    uint64_t d0, d1;
-
-    /* Load 8 bytes at a time.  Vector element ordering makes this LE.  */
-    d0 = cpu_ldq_le_data_ra(env, addr + 0, ra);
-    d1 = cpu_ldq_le_data_ra(env, addr + 8, ra);
-    pwd->d[0] = d0;
-    pwd->d[1] = d1;
-}
-
 void helper_msa_ld_h(CPUMIPSState *env, uint32_t wd,
                      target_ulong addr)
 {
@@ -8310,20 +8296,6 @@ static inline void ensure_writable_pages(CPUMIPSState *env,
     }
 }
 
-void helper_msa_st_b(CPUMIPSState *env, uint32_t wd,
-                     target_ulong addr)
-{
-    wr_t *pwd = &(env->active_fpu.fpr[wd].wr);
-    int mmu_idx = mips_env_mmu_index(env);
-    uintptr_t ra = GETPC();
-
-    ensure_writable_pages(env, addr, mmu_idx, ra);
-
-    /* Store 8 bytes at a time.  Vector element ordering makes this LE.  */
-    cpu_stq_le_data_ra(env, addr + 0, pwd->d[0], ra);
-    cpu_stq_le_data_ra(env, addr + 8, pwd->d[1], ra);
-}
-
 void helper_msa_st_h(CPUMIPSState *env, uint32_t wd,
                      target_ulong addr)
 {
target/mips/tcg/msa_translate.c +24 -6
index 2e1a769e032..340a0a14aaf 100644
@@ -756,8 +756,16 @@ TRANS(FFINT_U,  trans_msa_2rf, gen_helper_msa_ffint_u_df);
 
 static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
 {
+    static const MemOp mo_atom_df[4] = {
+        MO_ATOM_NONE,
+    };
     TCGv_i32 wd;
+    TCGv_i128 t16;
     TCGv_va addr;
+    MemOp mop;
+    int idx = a->wd << 1;
+    TCGv_i64 d0 = msa_wr_d[idx +  FP_ENDIAN_IDX];
+    TCGv_i64 d1 = msa_wr_d[idx + !FP_ENDIAN_IDX];
 
     if (!check_msa_enabled(ctx)) {
         return true;
@@ -770,9 +778,6 @@ static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
 
     if (is_load) {
         switch (a->df) {
-        case 0:
-            gen_helper_msa_ld_b(tcg_env, wd, addr);
-            return true;
         case 1:
             gen_helper_msa_ld_h(tcg_env, wd, addr);
             return true;
@@ -785,9 +790,6 @@ static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
         }
     } else {
         switch (a->df) {
-        case 0:
-            gen_helper_msa_st_b(tcg_env, wd, addr);
-            return true;
         case 1:
             gen_helper_msa_st_h(tcg_env, wd, addr);
             return true;
@@ -800,6 +802,22 @@ static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
         }
     }
 
+    t16 = tcg_temp_new_i128();
+
+    mop = MO_128 | MO_LE;
+    mop |= a->df << MO_ASHIFT; /* MO_ALIGN */
+    mop |= mo_atom_df[a->df];
+
+    if (is_load) {
+        tcg_gen_qemu_ld_i128(t16, addr, ctx->mem_idx, mop);
+        tcg_gen_extr_i128_i64(d0, d1, t16);
+    }
+
+    if (!is_load) {
+        tcg_gen_concat_i64_i128(t16, d0, d1);
+        tcg_gen_qemu_st_i128(t16, addr, ctx->mem_idx, mop);
+    }
+
     return true;
 }
 
Signature · 1 line
2.53.0
Reviewed-by on v5 2/6
#4 lore raw

Re: [RFC PATCH v5 2/6] target/mips: Convert MSA LD/ST.B (Byte Vector)

Quoted 5 lines · Philippe Mathieu-Daudé

Replace runtime helpers by translation.

Replace the legacy cpu_ld/st*_data_ra() calls by tcg_gen_qemu_ld/st() which allow to respect atomicity.

Quoted 8 lines

Consider the host endianness.

---
  target/mips/tcg/msa_helper.h.inc |  1 -
  target/mips/tcg/msa_helper.c     | 28 ----------------------------
  target/mips/tcg/msa_translate.c  | 30 ++++++++++++++++++++++++------
  3 files changed, 24 insertions(+), 35 deletions(-)

r~

#5 lore raw

RFC PATCH v5 3/6 target/mips: Convert MSA LD/ST.H (Halfword Vector)

In reply to Philippe Mathieu-Daudé

Replace runtime helpers by translation.

Replace the legacy cpu_ld/st*_data_ra() calls by tcg_gen_qemu_ld/st() which allow to respect atomicity.

---
 target/mips/tcg/msa_helper.h.inc |  1 -
 target/mips/tcg/msa_helper.c     | 48 --------------------------------
 target/mips/tcg/msa_translate.c  | 32 +++++++++++++++++----
 3 files changed, 26 insertions(+), 55 deletions(-)
target/mips/tcg/msa_helper.h.inc +0 -1
index 50a2732e916..86a79bf804c 100644
@@ -436,7 +436,6 @@ DEF_HELPER_4(msa_ffint_u_df, void, env, i32, i32, i32)
 #define MSALDST_PROTO(type)                         \
 DEF_HELPER_3(msa_ld_ ## type, void, env, i32, tl)   \
 DEF_HELPER_3(msa_st_ ## type, void, env, i32, tl)
-MSALDST_PROTO(h)
 MSALDST_PROTO(w)
 MSALDST_PROTO(d)
 #undef MSALDST_PROTO
target/mips/tcg/msa_helper.c +0 -48

65 lines not on this page · raw message

target/mips/tcg/msa_translate.c +26 -6
index 340a0a14aaf..111d9eedd11 100644
@@ -754,10 +754,27 @@ TRANS(FTINT_U,  trans_msa_2rf, gen_helper_msa_ftint_u_df);
 TRANS(FFINT_S,  trans_msa_2rf, gen_helper_msa_ffint_s_df);
 TRANS(FFINT_U,  trans_msa_2rf, gen_helper_msa_ffint_u_df);
 
+/*
+ * Byte pattern: ab.cd.ef.gh -> ba.dc.fe.bg
+ */
+static void gen_bswap16x4_i64(TCGv_i64 dst, TCGv_i64 src)
+{
+    uint64_t m = 0x00ff00ff00ff00ffull;
+    TCGv_i64 t0 = tcg_temp_new_i64();
+    TCGv_i64 t1 = tcg_temp_new_i64();
+                                        /* src = abcdefgh */
+    tcg_gen_andi_i64(t0, src, m);       /*  t0 = .b.d.f.h */
+    tcg_gen_shli_i64(t0, t0, 8);        /*  t0 = b.d.f.h. */
+    tcg_gen_shri_i64(t1, src, 8);       /*  t1 = habcdefg */
+    tcg_gen_andi_i64(t1, t1, m);        /*  t1 = .a.c.e.g */
+    tcg_gen_or_i64(dst, t0, t1);        /* dst = badcfebg */
+}
+
 static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
 {
     static const MemOp mo_atom_df[4] = {
         MO_ATOM_NONE,
+        MO_ATOM_SUBALIGN, /* Slightly stronger than required */
     };
     TCGv_i32 wd;
     TCGv_i128 t16;
@@ -778,9 +795,6 @@ static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
 
     if (is_load) {
         switch (a->df) {
-        case 1:
-            gen_helper_msa_ld_h(tcg_env, wd, addr);
-            return true;
         case 2:
             gen_helper_msa_ld_w(tcg_env, wd, addr);
             return true;
@@ -790,9 +804,6 @@ static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
         }
     } else {
         switch (a->df) {
-        case 1:
-            gen_helper_msa_st_h(tcg_env, wd, addr);
-            return true;
         case 2:
             gen_helper_msa_st_w(tcg_env, wd, addr);
             return true;
@@ -813,6 +824,15 @@ static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
         tcg_gen_extr_i128_i64(d0, d1, t16);
     }
 
+    if (mo_endian(ctx) != MO_LE) {
+        switch (a->df) {
+        case 1:
+            gen_bswap16x4_i64(d0, d0);
+            gen_bswap16x4_i64(d1, d1);
+            break;
+        }
+    }
+
     if (!is_load) {
         tcg_gen_concat_i64_i128(t16, d0, d1);
         tcg_gen_qemu_st_i128(t16, addr, ctx->mem_idx, mop);
Signature · 1 line
2.53.0
#6 lore raw

Re: [RFC PATCH v5 3/6] target/mips: Convert MSA LD/ST.H (Halfword Vector)

Quoted 6 lines · Philippe Mathieu-Daudé
target/mips/tcg/msa_translate.c +2 -0
@@ -813,6 +824,15 @@ static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
          tcg_gen_extr_i128_i64(d0, d1, t16);
      }
 
+    if (mo_endian(ctx) != MO_LE) {
+        switch (a->df) {

target/mips/tcg/msa_translate.c static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load) · line 838

Quoted 10 lines
target/mips/tcg/msa_translate.c +7 -0
+        case 1:
+            gen_bswap16x4_i64(d0, d0);
+            gen_bswap16x4_i64(d1, d1);
+            break;
+        }
+    }
+
      if (!is_load) {
          tcg_gen_concat_i64_i128(t16, d0, d1);
          tcg_gen_qemu_st_i128(t16, addr, ctx->mem_idx, mop);

You've just bswaped the architectural registers.
You need new temporaries here.

r~

#7 lore raw

Re: [RFC PATCH v5 3/6] target/mips: Convert MSA LD/ST.H (Halfword Vector)

Quoted 12 lines · Richard Henderson
Quoted 11 lines · Philippe Mathieu-Daudé
target/mips/tcg/msa_translate.c +7 -0
@@ -813,6 +824,15 @@ static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool 
 is_load)
           tcg_gen_extr_i128_i64(d0, d1, t16);
       }
+    if (mo_endian(ctx) != MO_LE) {
+        switch (a->df) {
+        case 1:
+            gen_bswap16x4_i64(d0, d0);
+            gen_bswap16x4_i64(d1, d1);
+            break;
+        }
Quoted 8 lines

target/mips/tcg/msa_translate.c static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load) · line 838

Quoted 5 lines
target/mips/tcg/msa_translate.c +2 -0
+    }
+
       if (!is_load) {
           tcg_gen_concat_i64_i128(t16, d0, d1);
           tcg_gen_qemu_st_i128(t16, addr, ctx->mem_idx, mop);

You've just bswaped the architectural registers.
You need new temporaries here...

... for stores.

Hit send too soon. :-)

r~

#8 lore raw

RFC PATCH v5 4/6 target/mips: Convert MSA LD/ST.W (Word Vector)

In reply to Philippe Mathieu-Daudé

Replace runtime helpers by translation.

Replace the legacy cpu_ld/st*_data_ra() calls by tcg_gen_qemu_ld/st() which allow to respect atomicity.

---
 target/mips/tcg/msa_helper.h.inc |  1 -
 target/mips/tcg/msa_helper.c     | 47 --------------------------------
 target/mips/tcg/msa_translate.c  | 20 ++++++++++----
 3 files changed, 14 insertions(+), 54 deletions(-)
target/mips/tcg/msa_helper.h.inc +0 -1
index 86a79bf804c..f0d31ceef29 100644
@@ -436,6 +436,5 @@ DEF_HELPER_4(msa_ffint_u_df, void, env, i32, i32, i32)
 #define MSALDST_PROTO(type)                         \
 DEF_HELPER_3(msa_ld_ ## type, void, env, i32, tl)   \
 DEF_HELPER_3(msa_st_ ## type, void, env, i32, tl)
-MSALDST_PROTO(w)
 MSALDST_PROTO(d)
 #undef MSALDST_PROTO
target/mips/tcg/msa_helper.c +0 -47

59 lines not on this page · raw message

target/mips/tcg/msa_translate.c +14 -6
index 111d9eedd11..8486508f5ba 100644
@@ -770,11 +770,21 @@ static void gen_bswap16x4_i64(TCGv_i64 dst, TCGv_i64 src)
     tcg_gen_or_i64(dst, t0, t1);        /* dst = badcfebg */
 }
 
+/*
+ * Byte pattern: abcd.efgh -> dcba.hgfe
+ */
+static void gen_bswap32x2_i64(TCGv_i64 dst, TCGv_i64 src)
+{
+    tcg_gen_bswap64_i64(dst, src);
+    tcg_gen_rotri_i64(dst, dst, 32);
+}
+
 static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
 {
     static const MemOp mo_atom_df[4] = {
         MO_ATOM_NONE,
         MO_ATOM_SUBALIGN, /* Slightly stronger than required */
+        MO_ATOM_SUBALIGN, /* Slightly stronger than required */
     };
     TCGv_i32 wd;
     TCGv_i128 t16;
@@ -795,18 +805,12 @@ static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
 
     if (is_load) {
         switch (a->df) {
-        case 2:
-            gen_helper_msa_ld_w(tcg_env, wd, addr);
-            return true;
         case 3:
             gen_helper_msa_ld_d(tcg_env, wd, addr);
             return true;
         }
     } else {
         switch (a->df) {
-        case 2:
-            gen_helper_msa_st_w(tcg_env, wd, addr);
-            return true;
         case 3:
             gen_helper_msa_st_d(tcg_env, wd, addr);
             return true;
@@ -830,6 +834,10 @@ static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
             gen_bswap16x4_i64(d0, d0);
             gen_bswap16x4_i64(d1, d1);
             break;
+        case 2:
+            gen_bswap32x2_i64(d0, d0);
+            gen_bswap32x2_i64(d1, d1);
+            break;
         }
     }
 
Signature · 1 line
2.53.0
Reviewed-by on v5 4/6
#9 lore raw

Re: [RFC PATCH v5 4/6] target/mips: Convert MSA LD/ST.W (Word Vector)

Quoted 11 lines · Philippe Mathieu-Daudé

Replace runtime helpers by translation.

Replace the legacy cpu_ld/st*_data_ra() calls by tcg_gen_qemu_ld/st() which allow to respect atomicity.

---
  target/mips/tcg/msa_helper.h.inc |  1 -
  target/mips/tcg/msa_helper.c     | 47 --------------------------------
  target/mips/tcg/msa_translate.c  | 20 ++++++++++----
  3 files changed, 14 insertions(+), 54 deletions(-)

Modulo new temps for store,

r~

Quoted 133 lines
target/mips/tcg/msa_helper.h.inc +0 -1
index 86a79bf804c..f0d31ceef29 100644
@@ -436,6 +436,5 @@ DEF_HELPER_4(msa_ffint_u_df, void, env, i32, i32, i32)
  #define MSALDST_PROTO(type)                         \
  DEF_HELPER_3(msa_ld_ ## type, void, env, i32, tl)   \
  DEF_HELPER_3(msa_st_ ## type, void, env, i32, tl)
-MSALDST_PROTO(w)
  MSALDST_PROTO(d)
  #undef MSALDST_PROTO
target/mips/tcg/msa_helper.c +0 -47
index 5489d8830dd..c93c0cddebc 100644
@@ -8212,32 +8212,6 @@ void helper_msa_ffint_u_df(CPUMIPSState *env, uint32_t df, uint32_t wd,
  /* Element-by-element access macros */
  #define DF_ELEMENTS(df) (MSA_WRLEN / DF_BITS(df))
 
-static inline uint64_t bswap32x2(uint64_t x)
-{
-    return ror64(bswap64(x), 32);
-}
-
-void helper_msa_ld_w(CPUMIPSState *env, uint32_t wd,
-                     target_ulong addr)
-{
-    wr_t *pwd = &(env->active_fpu.fpr[wd].wr);
-    uintptr_t ra = GETPC();
-    uint64_t d0, d1;
-
-    /*
-     * Load 8 bytes at a time.  Use little-endian load, then for
-     * big-endian target, we must then bswap the two words.
-     */
-    d0 = cpu_ldq_le_data_ra(env, addr + 0, ra);
-    d1 = cpu_ldq_le_data_ra(env, addr + 8, ra);
-    if (mips_env_is_bigendian(env)) {
-        d0 = bswap32x2(d0);
-        d1 = bswap32x2(d1);
-    }
-    pwd->d[0] = d0;
-    pwd->d[1] = d1;
-}
-
  void helper_msa_ld_d(CPUMIPSState *env, uint32_t wd,
                       target_ulong addr)
  {
@@ -8269,27 +8243,6 @@ static inline void ensure_writable_pages(CPUMIPSState *env,
      }
  }
 
-void helper_msa_st_w(CPUMIPSState *env, uint32_t wd,
-                     target_ulong addr)
-{
-    wr_t *pwd = &(env->active_fpu.fpr[wd].wr);
-    int mmu_idx = mips_env_mmu_index(env);
-    uintptr_t ra = GETPC();
-    uint64_t d0, d1;
-
-    ensure_writable_pages(env, addr, mmu_idx, ra);
-
-    /* Store 8 bytes at a time.  See helper_msa_ld_w. */
-    d0 = pwd->d[0];
-    d1 = pwd->d[1];
-    if (mips_env_is_bigendian(env)) {
-        d0 = bswap32x2(d0);
-        d1 = bswap32x2(d1);
-    }
-    cpu_stq_le_data_ra(env, addr + 0, d0, ra);
-    cpu_stq_le_data_ra(env, addr + 8, d1, ra);
-}
-
  void helper_msa_st_d(CPUMIPSState *env, uint32_t wd,
                       target_ulong addr)
  {
target/mips/tcg/msa_translate.c +14 -6
index 111d9eedd11..8486508f5ba 100644
@@ -770,11 +770,21 @@ static void gen_bswap16x4_i64(TCGv_i64 dst, TCGv_i64 src)
      tcg_gen_or_i64(dst, t0, t1);        /* dst = badcfebg */
  }
 
+/*
+ * Byte pattern: abcd.efgh -> dcba.hgfe
+ */
+static void gen_bswap32x2_i64(TCGv_i64 dst, TCGv_i64 src)
+{
+    tcg_gen_bswap64_i64(dst, src);
+    tcg_gen_rotri_i64(dst, dst, 32);
+}
+
  static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
  {
      static const MemOp mo_atom_df[4] = {
          MO_ATOM_NONE,
          MO_ATOM_SUBALIGN, /* Slightly stronger than required */
+        MO_ATOM_SUBALIGN, /* Slightly stronger than required */
      };
      TCGv_i32 wd;
      TCGv_i128 t16;
@@ -795,18 +805,12 @@ static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
 
      if (is_load) {
          switch (a->df) {
-        case 2:
-            gen_helper_msa_ld_w(tcg_env, wd, addr);
-            return true;
          case 3:
              gen_helper_msa_ld_d(tcg_env, wd, addr);
              return true;
          }
      } else {
          switch (a->df) {
-        case 2:
-            gen_helper_msa_st_w(tcg_env, wd, addr);
-            return true;
          case 3:
              gen_helper_msa_st_d(tcg_env, wd, addr);
              return true;
@@ -830,6 +834,10 @@ static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
              gen_bswap16x4_i64(d0, d0);
              gen_bswap16x4_i64(d1, d1);
              break;
+        case 2:
+            gen_bswap32x2_i64(d0, d0);
+            gen_bswap32x2_i64(d1, d1);
+            break;
          }
      }
 
#10 lore raw

RFC PATCH v5 5/6 target/mips: Convert MSA LD/ST.D (Doubleword Vector)

In reply to Philippe Mathieu-Daudé

Replace runtime helpers by translation.

Replace the legacy cpu_ld/st*_data_ra() calls by tcg_gen_qemu_ld/st() which allow to respect atomicity.

Remove the ensure_writable_pages() hack and MSALDST_PROTO macro.

---
 target/mips/tcg/msa_helper.h.inc |  6 ----
 target/mips/tcg/msa_helper.c     | 53 --------------------------------
 target/mips/tcg/msa_translate.c  | 18 +----------
 3 files changed, 1 insertion(+), 76 deletions(-)
target/mips/tcg/msa_helper.h.inc +0 -6
index f0d31ceef29..e994353056f 100644
@@ -432,9 +432,3 @@ DEF_HELPER_4(msa_ftint_s_df, void, env, i32, i32, i32)
 DEF_HELPER_4(msa_ftint_u_df, void, env, i32, i32, i32)
 DEF_HELPER_4(msa_ffint_s_df, void, env, i32, i32, i32)
 DEF_HELPER_4(msa_ffint_u_df, void, env, i32, i32, i32)
-
-#define MSALDST_PROTO(type)                         \
-DEF_HELPER_3(msa_ld_ ## type, void, env, i32, tl)   \
-DEF_HELPER_3(msa_st_ ## type, void, env, i32, tl)
-MSALDST_PROTO(d)
-#undef MSALDST_PROTO
target/mips/tcg/msa_helper.c +0 -53

64 lines not on this page · raw message

target/mips/tcg/msa_translate.c +1 -17
index 8486508f5ba..b1d764bd179 100644
@@ -785,8 +785,8 @@ static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
         MO_ATOM_NONE,
         MO_ATOM_SUBALIGN, /* Slightly stronger than required */
         MO_ATOM_SUBALIGN, /* Slightly stronger than required */
+        MO_ATOM_IFALIGN_PAIR
     };
-    TCGv_i32 wd;
     TCGv_i128 t16;
     TCGv_va addr;
     MemOp mop;
@@ -801,22 +801,6 @@ static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
     addr = tcgv_va_temp_new();
     gen_base_offset_addr(ctx, addr, a->ws, a->sa << a->df);
 
-    wd = tcg_constant_i32(a->wd);
-
-    if (is_load) {
-        switch (a->df) {
-        case 3:
-            gen_helper_msa_ld_d(tcg_env, wd, addr);
-            return true;
-        }
-    } else {
-        switch (a->df) {
-        case 3:
-            gen_helper_msa_st_d(tcg_env, wd, addr);
-            return true;
-        }
-    }
-
     t16 = tcg_temp_new_i128();
 
     mop = MO_128 | MO_LE;
Signature · 1 line
2.53.0
#11 lore raw

Re: [RFC PATCH v5 5/6] target/mips: Convert MSA LD/ST.D (Doubleword Vector)

Quoted 88 lines · Philippe Mathieu-Daudé

Replace runtime helpers by translation.

Replace the legacy cpu_ld/st*_data_ra() calls by tcg_gen_qemu_ld/st() which allow to respect atomicity.

Remove the ensure_writable_pages() hack and MSALDST_PROTO macro.

---
  target/mips/tcg/msa_helper.h.inc |  6 ----
  target/mips/tcg/msa_helper.c     | 53 --------------------------------
  target/mips/tcg/msa_translate.c  | 18 +----------
  3 files changed, 1 insertion(+), 76 deletions(-)
target/mips/tcg/msa_helper.h.inc +0 -6
index f0d31ceef29..e994353056f 100644
@@ -432,9 +432,3 @@ DEF_HELPER_4(msa_ftint_s_df, void, env, i32, i32, i32)
  DEF_HELPER_4(msa_ftint_u_df, void, env, i32, i32, i32)
  DEF_HELPER_4(msa_ffint_s_df, void, env, i32, i32, i32)
  DEF_HELPER_4(msa_ffint_u_df, void, env, i32, i32, i32)
-
-#define MSALDST_PROTO(type)                         \
-DEF_HELPER_3(msa_ld_ ## type, void, env, i32, tl)   \
-DEF_HELPER_3(msa_st_ ## type, void, env, i32, tl)
-MSALDST_PROTO(d)
-#undef MSALDST_PROTO
target/mips/tcg/msa_helper.c +0 -43
index c93c0cddebc..67ec19ea95c 100644
@@ -21,11 +21,8 @@
  #include "cpu.h"
  #include "internal.h"
  #include "tcg/tcg.h"
-#include "accel/tcg/cpu-ldst.h"
-#include "accel/tcg/probe.h"
  #include "exec/helper-proto.h"
  #include "exec/memop.h"
-#include "exec/target_page.h"
  #include "fpu/softfloat.h"
  #include "fpu_helper.h"
 
@@ -8205,53 +8202,3 @@ void helper_msa_ffint_u_df(CPUMIPSState *env, uint32_t df, uint32_t wd,
 
      msa_move_v(pwd, pwx);
  }
-
-/* Data format min and max values */
-#define DF_BITS(df) (1 << ((df) + 3))
-
-/* Element-by-element access macros */
-#define DF_ELEMENTS(df) (MSA_WRLEN / DF_BITS(df))
-
-void helper_msa_ld_d(CPUMIPSState *env, uint32_t wd,
-                     target_ulong addr)
-{
-    wr_t *pwd = &(env->active_fpu.fpr[wd].wr);
-    uintptr_t ra = GETPC();
-    uint64_t d0, d1;
-
-    d0 = cpu_ldq_data_ra(env, addr + 0, ra);
-    d1 = cpu_ldq_data_ra(env, addr + 8, ra);
-    pwd->d[0] = d0;
-    pwd->d[1] = d1;
-}
-
-#define MSA_PAGESPAN(x) \
-        ((((x) & ~TARGET_PAGE_MASK) + MSA_WRLEN / 8 - 1) >= TARGET_PAGE_SIZE)
-
-static inline void ensure_writable_pages(CPUMIPSState *env,
-                                         target_ulong addr,
-                                         int mmu_idx,
-                                         uintptr_t retaddr)
-{
-    /* FIXME: Probe the actual accesses (pass and use a size) */
-    if (unlikely(MSA_PAGESPAN(addr))) {
-        /* first page */
-        probe_write(env, addr, 0, mmu_idx, retaddr);
-        /* second page */
-        addr = (addr & TARGET_PAGE_MASK) + TARGET_PAGE_SIZE;
-        probe_write(env, addr, 0, mmu_idx, retaddr);
-    }
-}
-
-void helper_msa_st_d(CPUMIPSState *env, uint32_t wd,
-                     target_ulong addr)

target/mips/tcg/msa_helper.c void helper_msa_ffint_u_df(CPUMIPSState *env, uint32_t df, uint32_t wd, · line 8257

Quoted 10 lines
target/mips/tcg/msa_helper.c +0 -10
-{
-    wr_t *pwd = &(env->active_fpu.fpr[wd].wr);
-    int mmu_idx = mips_env_mmu_index(env);
-    uintptr_t ra = GETPC();
-
-    ensure_writable_pages(env, addr, mmu_idx, GETPC());
-
-    cpu_stq_data_ra(env, addr + 0, pwd->d[0], ra);
-    cpu_stq_data_ra(env, addr + 8, pwd->d[1], ra);
-}

This isn't using a LE operation...

Quoted 27 lines
target/mips/tcg/msa_translate.c +1 -10
index 8486508f5ba..b1d764bd179 100644
@@ -785,8 +785,8 @@ static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
          MO_ATOM_NONE,
          MO_ATOM_SUBALIGN, /* Slightly stronger than required */
          MO_ATOM_SUBALIGN, /* Slightly stronger than required */
+        MO_ATOM_IFALIGN_PAIR
      };
-    TCGv_i32 wd;
      TCGv_i128 t16;
      TCGv_va addr;
      MemOp mop;
@@ -801,22 +801,6 @@ static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
      addr = tcgv_va_temp_new();
      gen_base_offset_addr(ctx, addr, a->ws, a->sa << a->df);
 
-    wd = tcg_constant_i32(a->wd);
-
-    if (is_load) {
-        switch (a->df) {
-        case 3:
-            gen_helper_msa_ld_d(tcg_env, wd, addr);
-            return true;
-        }
-    } else {

target/mips/tcg/msa_translate.c static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load) · line 806

Quoted 10 lines
target/mips/tcg/msa_translate.c +0 -7
-        switch (a->df) {
-        case 3:
-            gen_helper_msa_st_d(tcg_env, wd, addr);
-            return true;
-        }
-    }
-
      t16 = tcg_temp_new_i128();
 
      mop = MO_128 | MO_LE;

... which you're now getting from here.

You could bswap d0 and d1, or you could have the 64-bit path reorder d0 and d1 in t16 and mop ^= MO_BSWAP.

r~

#12 lore raw

RFC PATCH v5 6/6 target/mips: Endian-swap MSA LD/ST.D (Doubleword Vector)

In reply to Philippe Mathieu-Daudé

Doubleword Vector entries need to be swapped w.r.t. endianness.

Fixes: 83be6b54123 ("Fix MSA instructions LD.<B|H|W|D> on big endian host")
Fixes: 6decc572dce ("Fix MSA instructions ST.<B|H|W|D> on big endian host")

---
Probably optimizable since tcg_gen_qemu_ld/st_i128 handles MO_BSWAP.

---
 target/mips/tcg/msa_translate.c | 4 ++++
 1 file changed, 4 insertions(+)
target/mips/tcg/msa_translate.c +4 -0
index b1d764bd179..63d797a39b7 100644
@@ -822,6 +822,10 @@ static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
             gen_bswap32x2_i64(d0, d0);
             gen_bswap32x2_i64(d1, d1);
             break;
+        case 3:
+            tcg_gen_bswap64_i64(d0, d0);
+            tcg_gen_bswap64_i64(d1, d1);
+            break;
         }
     }
 
Signature · 1 line
2.53.0
#13 lore raw

Re: [RFC PATCH v5 6/6] target/mips: Endian-swap MSA LD/ST.D (Doubleword Vector)

Quoted 16 lines · Philippe Mathieu-Daudé

Doubleword Vector entries need to be swapped w.r.t. endianness.

Fixes: 83be6b54123 ("Fix MSA instructions LD.<B|H|W|D> on big endian host")
Fixes: 6decc572dce ("Fix MSA instructions ST.<B|H|W|D> on big endian host")

---
Probably optimizable since tcg_gen_qemu_ld/st_i128 handles MO_BSWAP.

---
  target/mips/tcg/msa_translate.c | 4 ++++
  1 file changed, 4 insertions(+)
target/mips/tcg/msa_translate.c +0 -0
index b1d764bd179..63d797a39b7 100644

target/mips/tcg/msa_translate.c static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load) · line 831

Quoted 10 lines
target/mips/tcg/msa_translate.c +4 -0
@@ -822,6 +822,10 @@ static bool trans_msa_ldst(DisasContext *ctx, arg_msa_i *a, bool is_load)
              gen_bswap32x2_i64(d0, d0);
              gen_bswap32x2_i64(d1, d1);
              break;
+        case 3:
+            tcg_gen_bswap64_i64(d0, d0);
+            tcg_gen_bswap64_i64(d1, d1);
+            break;
          }
      }
 

No, this is just a fix to patch 5.

r~