[Qemu-devel] [PATCH v15 30/33] target-tilegx: Handle atomic instructions

[Richard Henderson <rth@twiddle.net>]

Signed-off-by: Richard Henderson <rth@twiddle.net>
---
 linux-user/main.c         | 84 +++++++----------------------------------------
 target-tilegx/cpu.h       |  4 ++-
 target-tilegx/translate.c | 80 +++++++++++++++++++++++++++++++++++++++++++-
 3 files changed, 94 insertions(+), 74 deletions(-)

diff --git a/linux-user/main.c b/linux-user/main.c
index 4fe18c0..04e5bbe 100644
--- a/linux-user/main.c
+++ b/linux-user/main.c
@@ -3433,51 +3433,26 @@ static void gen_sigill_reg(CPUTLGState *env)
     queue_signal(env, info.si_signo, &info);
 }
 
-static int get_regval(CPUTLGState *env, uint8_t reg, target_ulong *val)
-{
-    if (likely(reg < TILEGX_R_COUNT)) {
-        *val = env->regs[reg];
-        return 0;
-    }
-
-    switch (reg) {
-    case TILEGX_R_SN:
-    case TILEGX_R_ZERO:
-        *val = 0;
-        return 0;
-    case TILEGX_R_IDN0:
-    case TILEGX_R_IDN1:
-    case TILEGX_R_UDN0:
-    case TILEGX_R_UDN1:
-    case TILEGX_R_UDN2:
-    case TILEGX_R_UDN3:
-        return -1;
-    default:
-        g_assert_not_reached();
-    }
-}
-
-static int set_regval(CPUTLGState *env, uint8_t reg, uint64_t val)
+static void set_regval(CPUTLGState *env, uint8_t reg, uint64_t val)
 {
     if (unlikely(reg >= TILEGX_R_COUNT)) {
         switch (reg) {
         case TILEGX_R_SN:
         case TILEGX_R_ZERO:
-            return 0;
+            return;
         case TILEGX_R_IDN0:
         case TILEGX_R_IDN1:
         case TILEGX_R_UDN0:
         case TILEGX_R_UDN1:
         case TILEGX_R_UDN2:
         case TILEGX_R_UDN3:
-            return -1;
+            gen_sigill_reg(env);
+            return;
         default:
             g_assert_not_reached();
         }
     }
-
     env->regs[reg] = val;
-    return 0;
 }
 
 /*
@@ -3508,19 +3483,12 @@ static int set_regval(CPUTLGState *env, uint8_t reg, uint64_t val)
  */
 static void do_exch(CPUTLGState *env, bool quad, bool cmp)
 {
-    uint8_t rdst, rsrc, rsrcb;
     target_ulong addr;
     target_long val, sprval;
 
     start_exclusive();
 
-    rdst = extract32(env->excparam, 16, 8);
-    rsrc = extract32(env->excparam, 8, 8);
-    rsrcb = extract32(env->excparam, 0, 8);
-
-    if (get_regval(env, rsrc, &addr)) {
-        goto sigill_reg;
-    }
+    addr = env->atomic_srca;
     if (quad ? get_user_s64(val, addr) : get_user_s32(val, addr)) {
         goto sigsegv_mapper;
     }
@@ -3534,56 +3502,35 @@ static void do_exch(CPUTLGState *env, bool quad, bool cmp)
     }
 
     if (!cmp || val == sprval) {
-        target_long valb;
-
-        if (get_regval(env, rsrcb, (target_ulong *)&valb)) {
-            goto sigill_reg;
-        }
+        target_long valb = env->atomic_srcb;
         if (quad ? put_user_u64(valb, addr) : put_user_u32(valb, addr)) {
             goto sigsegv_mapper;
         }
     }
 
-    if (set_regval(env, rdst, val)) {
-        goto sigill_reg;
-    }
+    set_regval(env, env->atomic_dstr, val);
     end_exclusive();
     return;
 
-sigill_reg:
-    end_exclusive();
-    gen_sigill_reg(env);
-    return;
-
-sigsegv_mapper:
+ sigsegv_mapper:
     end_exclusive();
     gen_sigsegv_mapper(env, addr);
 }
 
 static void do_fetch(CPUTLGState *env, int trapnr, bool quad)
 {
-    uint8_t rdst, rsrc, rsrcb;
     int8_t write = 1;
     target_ulong addr;
     target_long val, valb;
 
     start_exclusive();
 
-    rdst = extract32(env->excparam, 16, 8);
-    rsrc = extract32(env->excparam, 8, 8);
-    rsrcb = extract32(env->excparam, 0, 8);
-
-
-    if (get_regval(env, rsrc, &addr)) {
-        goto sigill_reg;
-    }
+    addr = env->atomic_srca;
+    valb = env->atomic_srcb;
     if (quad ? get_user_s64(val, addr) : get_user_s32(val, addr)) {
         goto sigsegv_mapper;
     }
 
-    if (get_regval(env, rsrcb, (target_ulong *)&valb)) {
-        goto sigill_reg;
-    }
     switch (trapnr) {
     case TILEGX_EXCP_OPCODE_FETCHADD:
     case TILEGX_EXCP_OPCODE_FETCHADD4:
@@ -3619,18 +3566,11 @@ static void do_fetch(CPUTLGState *env, int trapnr, bool quad)
         }
     }
 
-    if (set_regval(env, rdst, val)) {
-        goto sigill_reg;
-    }
-    end_exclusive();
-    return;
-
-sigill_reg:
+    set_regval(env, env->atomic_dstr, val);
     end_exclusive();
-    gen_sigill_reg(env);
     return;
 
-sigsegv_mapper:
+ sigsegv_mapper:
     end_exclusive();
     gen_sigsegv_mapper(env, addr);
 }
diff --git a/target-tilegx/cpu.h b/target-tilegx/cpu.h
index 3a62d20..b9f5082 100644
--- a/target-tilegx/cpu.h
+++ b/target-tilegx/cpu.h
@@ -87,7 +87,9 @@ typedef struct CPUTLGState {
     uint64_t pc;                       /* Current pc */
 
 #if defined(CONFIG_USER_ONLY)
-    uint32_t excparam;                 /* exception parameter */
+    uint64_t atomic_srca;              /* Arguments to atomic "exceptions" */
+    uint64_t atomic_srcb;
+    uint32_t atomic_dstr;
     uint64_t excaddr;                  /* exception address */
 #endif
 
diff --git a/target-tilegx/translate.c b/target-tilegx/translate.c
index 9418799..702289e 100644
--- a/target-tilegx/translate.c
+++ b/target-tilegx/translate.c
@@ -61,6 +61,7 @@ typedef struct {
     int num_wb;
     int mmuidx;
     bool exit_tb;
+    TileExcp atomic_excp;
 
     struct {
         TCGCond cond;    /* branch condition */
@@ -180,6 +181,32 @@ static void gen_saturate_op(TCGv tdest, TCGv tsrca, TCGv tsrcb,
     tcg_temp_free(t0);
 }
 
+static void gen_atomic_excp(DisasContext *dc, unsigned dest, TCGv tdest,
+                            TCGv tsrca, TCGv tsrcb, TileExcp excp)
+{
+#ifdef CONFIG_USER_ONLY
+    TCGv_i32 t;
+
+    tcg_gen_st_tl(tsrca, cpu_env, offsetof(CPUTLGState, atomic_srca));
+    tcg_gen_st_tl(tsrcb, cpu_env, offsetof(CPUTLGState, atomic_srcb));
+    t = tcg_const_i32(dest);
+    tcg_gen_st_i32(t, cpu_env, offsetof(CPUTLGState, atomic_dstr));
+    tcg_temp_free_i32(t);
+
+    /* We're going to write the real result in the exception.  But in
+       the meantime we've already created a writeback register, and
+       we don't want that to remain uninitialized.  */
+    tcg_gen_movi_tl(tdest, 0);
+
+    /* Note that we need to delay issuing the exception that implements
+       the atomic operation until after writing back the results of the
+       instruction occupying the X0 pipe.  */
+    dc->atomic_excp = excp;
+#else
+    gen_exception(dc, TILEGX_EXCP_OPCODE_UNIMPLEMENTED);
+#endif
+}
+
 /* Shift the 128-bit value TSRCA:TSRCD riht by the number of bytes
    specified by the bottom 3 bits of TSRCB, and set TDEST to the
    low 64 bits of the resulting value.  */
@@ -591,8 +618,15 @@ static TileExcp gen_rrr_opcode(DisasContext *dc, unsigned opext,
         mnemonic = "cmpeq";
         break;
     case OE_RRR(CMPEXCH4, 0, X1):
+        gen_atomic_excp(dc, dest, tdest, tsrca, tsrcb,
+                        TILEGX_EXCP_OPCODE_CMPEXCH4);
+        mnemonic = "cmpexch4";
+        break;
     case OE_RRR(CMPEXCH, 0, X1):
-        return TILEGX_EXCP_OPCODE_UNIMPLEMENTED;
+        gen_atomic_excp(dc, dest, tdest, tsrca, tsrcb,
+                        TILEGX_EXCP_OPCODE_CMPEXCH);
+        mnemonic = "cmpexch";
+        break;
     case OE_RRR(CMPLES, 0, X0):
     case OE_RRR(CMPLES, 0, X1):
     case OE_RRR(CMPLES, 2, Y0):
@@ -658,7 +692,15 @@ static TileExcp gen_rrr_opcode(DisasContext *dc, unsigned opext,
         mnemonic = "dblalign";
         break;
     case OE_RRR(EXCH4, 0, X1):
+        gen_atomic_excp(dc, dest, tdest, tsrca, tsrcb,
+                        TILEGX_EXCP_OPCODE_EXCH4);
+        mnemonic = "exch4";
+        break;
     case OE_RRR(EXCH, 0, X1):
+        gen_atomic_excp(dc, dest, tdest, tsrca, tsrcb,
+                        TILEGX_EXCP_OPCODE_EXCH);
+        mnemonic = "exch";
+        break;
     case OE_RRR(FDOUBLE_ADDSUB, 0, X0):
     case OE_RRR(FDOUBLE_ADD_FLAGS, 0, X0):
     case OE_RRR(FDOUBLE_MUL_FLAGS, 0, X0):
@@ -667,14 +709,47 @@ static TileExcp gen_rrr_opcode(DisasContext *dc, unsigned opext,
     case OE_RRR(FDOUBLE_SUB_FLAGS, 0, X0):
     case OE_RRR(FDOUBLE_UNPACK_MAX, 0, X0):
     case OE_RRR(FDOUBLE_UNPACK_MIN, 0, X0):
+        return TILEGX_EXCP_OPCODE_UNIMPLEMENTED;
     case OE_RRR(FETCHADD4, 0, X1):
+        gen_atomic_excp(dc, dest, tdest, tsrca, tsrcb,
+                        TILEGX_EXCP_OPCODE_FETCHADD4);
+        mnemonic = "fetchadd4";
+        break;
     case OE_RRR(FETCHADDGEZ4, 0, X1):
+        gen_atomic_excp(dc, dest, tdest, tsrca, tsrcb,
+                        TILEGX_EXCP_OPCODE_FETCHADDGEZ4);
+        mnemonic = "fetchaddgez4";
+        break;
     case OE_RRR(FETCHADDGEZ, 0, X1):
+        gen_atomic_excp(dc, dest, tdest, tsrca, tsrcb,
+                        TILEGX_EXCP_OPCODE_FETCHADDGEZ);
+        mnemonic = "fetchaddgez";
+        break;
     case OE_RRR(FETCHADD, 0, X1):
+        gen_atomic_excp(dc, dest, tdest, tsrca, tsrcb,
+                        TILEGX_EXCP_OPCODE_FETCHADD);
+        mnemonic = "fetchadd";
+        break;
     case OE_RRR(FETCHAND4, 0, X1):
+        gen_atomic_excp(dc, dest, tdest, tsrca, tsrcb,
+                        TILEGX_EXCP_OPCODE_FETCHAND4);
+        mnemonic = "fetchand4";
+        break;
     case OE_RRR(FETCHAND, 0, X1):
+        gen_atomic_excp(dc, dest, tdest, tsrca, tsrcb,
+                        TILEGX_EXCP_OPCODE_FETCHAND);
+        mnemonic = "fetchand";
+        break;
     case OE_RRR(FETCHOR4, 0, X1):
+        gen_atomic_excp(dc, dest, tdest, tsrca, tsrcb,
+                        TILEGX_EXCP_OPCODE_FETCHOR4);
+        mnemonic = "fetchor4";
+        break;
     case OE_RRR(FETCHOR, 0, X1):
+        gen_atomic_excp(dc, dest, tdest, tsrca, tsrcb,
+                        TILEGX_EXCP_OPCODE_FETCHOR);
+        mnemonic = "fetchor";
+        break;
     case OE_RRR(FSINGLE_ADD1, 0, X0):
     case OE_RRR(FSINGLE_ADDSUB2, 0, X0):
     case OE_RRR(FSINGLE_MUL1, 0, X0):
@@ -1936,6 +2011,8 @@ static void translate_one_bundle(DisasContext *dc, uint64_t bundle)
         tcg_temp_free_i64(dc->jmp.dest);
         tcg_gen_exit_tb(0);
         dc->exit_tb = true;
+    } else if (dc->atomic_excp != TILEGX_EXCP_NONE) {
+        gen_exception(dc, dc->atomic_excp);
     }
 }
 
@@ -1956,6 +2033,7 @@ static inline void gen_intermediate_code_internal(TileGXCPU *cpu,
     dc->pc = pc_start;
     dc->mmuidx = 0;
     dc->exit_tb = false;
+    dc->atomic_excp = TILEGX_EXCP_NONE;
     dc->jmp.cond = TCG_COND_NEVER;
     TCGV_UNUSED_I64(dc->jmp.dest);
     TCGV_UNUSED_I64(dc->jmp.val1);
-- 
2.4.3