[PATCH v4 36/36] target/i386: Inline cmpxchg16b

[Richard Henderson <richard.henderson@linaro.org>]

Use tcg_gen_atomic_cmpxchg_i128 for the atomic case,
and tcg_gen_qemu_ld/st_i128 otherwise.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 target/i386/helper.h         |  4 ---
 target/i386/tcg/mem_helper.c | 69 ------------------------------------
 target/i386/tcg/translate.c  | 44 ++++++++++++++++++++---
 3 files changed, 39 insertions(+), 78 deletions(-)

diff --git a/target/i386/helper.h b/target/i386/helper.h
index 2df8049f91..e627a93107 100644
--- a/target/i386/helper.h
+++ b/target/i386/helper.h
@@ -66,10 +66,6 @@ DEF_HELPER_1(rsm, void, env)
 #endif /* !CONFIG_USER_ONLY */
 
 DEF_HELPER_2(into, void, env, int)
-#ifdef TARGET_X86_64
-DEF_HELPER_2(cmpxchg16b_unlocked, void, env, tl)
-DEF_HELPER_2(cmpxchg16b, void, env, tl)
-#endif
 DEF_HELPER_FLAGS_1(single_step, TCG_CALL_NO_WG, noreturn, env)
 DEF_HELPER_1(rechecking_single_step, void, env)
 DEF_HELPER_1(cpuid, void, env)
diff --git a/target/i386/tcg/mem_helper.c b/target/i386/tcg/mem_helper.c
index 814786bb87..3ef84e90d9 100644
--- a/target/i386/tcg/mem_helper.c
+++ b/target/i386/tcg/mem_helper.c
@@ -27,75 +27,6 @@
 #include "tcg/tcg.h"
 #include "helper-tcg.h"
 
-#ifdef TARGET_X86_64
-void helper_cmpxchg16b_unlocked(CPUX86State *env, target_ulong a0)
-{
-    uintptr_t ra = GETPC();
-    Int128 oldv, cmpv, newv;
-    uint64_t o0, o1;
-    int eflags;
-    bool success;
-
-    if ((a0 & 0xf) != 0) {
-        raise_exception_ra(env, EXCP0D_GPF, GETPC());
-    }
-    eflags = cpu_cc_compute_all(env, CC_OP);
-
-    cmpv = int128_make128(env->regs[R_EAX], env->regs[R_EDX]);
-    newv = int128_make128(env->regs[R_EBX], env->regs[R_ECX]);
-
-    o0 = cpu_ldq_data_ra(env, a0 + 0, ra);
-    o1 = cpu_ldq_data_ra(env, a0 + 8, ra);
-
-    oldv = int128_make128(o0, o1);
-    success = int128_eq(oldv, cmpv);
-    if (!success) {
-        newv = oldv;
-    }
-
-    cpu_stq_data_ra(env, a0 + 0, int128_getlo(newv), ra);
-    cpu_stq_data_ra(env, a0 + 8, int128_gethi(newv), ra);
-
-    if (success) {
-        eflags |= CC_Z;
-    } else {
-        env->regs[R_EAX] = int128_getlo(oldv);
-        env->regs[R_EDX] = int128_gethi(oldv);
-        eflags &= ~CC_Z;
-    }
-    CC_SRC = eflags;
-}
-
-void helper_cmpxchg16b(CPUX86State *env, target_ulong a0)
-{
-    uintptr_t ra = GETPC();
-
-    if ((a0 & 0xf) != 0) {
-        raise_exception_ra(env, EXCP0D_GPF, ra);
-    } else if (HAVE_CMPXCHG128) {
-        int eflags = cpu_cc_compute_all(env, CC_OP);
-
-        Int128 cmpv = int128_make128(env->regs[R_EAX], env->regs[R_EDX]);
-        Int128 newv = int128_make128(env->regs[R_EBX], env->regs[R_ECX]);
-
-        int mem_idx = cpu_mmu_index(env, false);
-        MemOpIdx oi = make_memop_idx(MO_TE | MO_128 | MO_ALIGN, mem_idx);
-        Int128 oldv = cpu_atomic_cmpxchgo_le_mmu(env, a0, cmpv, newv, oi, ra);
-
-        if (int128_eq(oldv, cmpv)) {
-            eflags |= CC_Z;
-        } else {
-            env->regs[R_EAX] = int128_getlo(oldv);
-            env->regs[R_EDX] = int128_gethi(oldv);
-            eflags &= ~CC_Z;
-        }
-        CC_SRC = eflags;
-    } else {
-        cpu_loop_exit_atomic(env_cpu(env), ra);
-    }
-}
-#endif
-
 void helper_boundw(CPUX86State *env, target_ulong a0, int v)
 {
     int low, high;
diff --git a/target/i386/tcg/translate.c b/target/i386/tcg/translate.c
index b542b084a6..9d9392b009 100644
--- a/target/i386/tcg/translate.c
+++ b/target/i386/tcg/translate.c
@@ -3053,15 +3053,49 @@ static void gen_cmpxchg8b(DisasContext *s, CPUX86State *env, int modrm)
 #ifdef TARGET_X86_64
 static void gen_cmpxchg16b(DisasContext *s, CPUX86State *env, int modrm)
 {
+    MemOp mop = MO_TE | MO_128 | MO_ALIGN;
+    TCGv_i64 t0, t1;
+    TCGv_i128 cmp, val;
+
     gen_lea_modrm(env, s, modrm);
 
-    if ((s->prefix & PREFIX_LOCK) &&
-        (tb_cflags(s->base.tb) & CF_PARALLEL)) {
-        gen_helper_cmpxchg16b(cpu_env, s->A0);
+    cmp = tcg_temp_new_i128();
+    val = tcg_temp_new_i128();
+    tcg_gen_concat_i64_i128(cmp, cpu_regs[R_EAX], cpu_regs[R_EDX]);
+    tcg_gen_concat_i64_i128(val, cpu_regs[R_EBX], cpu_regs[R_ECX]);
+
+    /* Only require atomic with LOCK; non-parallel handled in generator. */
+    if (s->prefix & PREFIX_LOCK) {
+        tcg_gen_atomic_cmpxchg_i128(val, s->A0, cmp, val, s->mem_index, mop);
     } else {
-        gen_helper_cmpxchg16b_unlocked(cpu_env, s->A0);
+        tcg_gen_nonatomic_cmpxchg_i128(val, s->A0, cmp, val, s->mem_index, mop);
     }
-    set_cc_op(s, CC_OP_EFLAGS);
+
+    tcg_gen_extr_i128_i64(s->T0, s->T1, val);
+    tcg_temp_free_i128(cmp);
+    tcg_temp_free_i128(val);
+
+    /* Determine success after the fact. */
+    t0 = tcg_temp_new_i64();
+    t1 = tcg_temp_new_i64();
+    tcg_gen_xor_i64(t0, s->T0, cpu_regs[R_EAX]);
+    tcg_gen_xor_i64(t1, s->T1, cpu_regs[R_EDX]);
+    tcg_gen_or_i64(t0, t0, t1);
+    tcg_temp_free_i64(t1);
+
+    /* Update Z. */
+    gen_compute_eflags(s);
+    tcg_gen_setcondi_i64(TCG_COND_EQ, t0, t0, 0);
+    tcg_gen_deposit_tl(cpu_cc_src, cpu_cc_src, t0, ctz32(CC_Z), 1);
+    tcg_temp_free_i64(t0);
+
+    /*
+     * Extract the result values for the register pair.  We may do this
+     * unconditionally, because on success (Z=1), the old value matches
+     * the previous value in RDX:RAX.
+     */
+    tcg_gen_mov_i64(cpu_regs[R_EAX], s->T0);
+    tcg_gen_mov_i64(cpu_regs[R_EDX], s->T1);
 }
 #endif
 
-- 
2.34.1