[Qemu-devel] [PULL 05/16] target-i386: Implement XSAVEOPT

[Richard Henderson <rth@twiddle.net>]

Signed-off-by: Richard Henderson <rth@twiddle.net>
---
 target-i386/cpu.c        |  6 ++++--
 target-i386/fpu_helper.c | 29 +++++++++++++++++++++++------
 target-i386/helper.h     |  1 +
 target-i386/translate.c  | 18 +++++++++++++++---
 4 files changed, 43 insertions(+), 11 deletions(-)

diff --git a/target-i386/cpu.c b/target-i386/cpu.c
index fb8a646..dcc6e9c 100644
--- a/target-i386/cpu.c
+++ b/target-i386/cpu.c
@@ -366,7 +366,9 @@ static const char *cpuid_6_feature_name[] = {
 #define TCG_7_0_ECX_FEATURES 0
 #define TCG_APM_FEATURES 0
 #define TCG_6_EAX_FEATURES CPUID_6_EAX_ARAT
-
+#define TCG_XSAVE_FEATURES (CPUID_XSAVE_XSAVEOPT | CPUID_XSAVE_XGETBV1)
+          /* missing:
+          CPUID_XSAVE_XSAVEC, CPUID_XSAVE_XSAVES */
 
 typedef struct FeatureWordInfo {
     const char **feat_names;
@@ -440,7 +442,7 @@ static FeatureWordInfo feature_word_info[FEATURE_WORDS] = {
         .cpuid_eax = 0xd,
         .cpuid_needs_ecx = true, .cpuid_ecx = 1,
         .cpuid_reg = R_EAX,
-        .tcg_features = 0,
+        .tcg_features = TCG_XSAVE_FEATURES,
     },
     [FEAT_6_EAX] = {
         .feat_names = cpuid_6_feature_name,
diff --git a/target-i386/fpu_helper.c b/target-i386/fpu_helper.c
index 159e1df..9750cd2 100644
--- a/target-i386/fpu_helper.c
+++ b/target-i386/fpu_helper.c
@@ -1197,9 +1197,9 @@ static uint64_t get_xinuse(CPUX86State *env)
     return -1;
 }
 
-void helper_xsave(CPUX86State *env, target_ulong ptr, uint64_t rfbm)
+static void do_xsave(CPUX86State *env, target_ulong ptr, uint64_t rfbm,
+                     uint64_t inuse, uint64_t opt, uintptr_t ra)
 {
-    uintptr_t ra = GETPC();
     uint64_t old_bv, new_bv;
 
     /* The OS must have enabled XSAVE.  */
@@ -1214,21 +1214,36 @@ void helper_xsave(CPUX86State *env, target_ulong ptr, uint64_t rfbm)
 
     /* Never save anything not enabled by XCR0.  */
     rfbm &= env->xcr0;
+    opt &= rfbm;
 
-    if (rfbm & XSTATE_FP) {
+    if (opt & XSTATE_FP) {
         do_xsave_fpu(env, ptr, ra);
     }
     if (rfbm & XSTATE_SSE) {
+        /* Note that saving MXCSR is not suppressed by XSAVEOPT.  */
         do_xsave_mxcsr(env, ptr, ra);
+    }
+    if (opt & XSTATE_SSE) {
         do_xsave_sse(env, ptr, ra);
     }
 
     /* Update the XSTATE_BV field.  */
     old_bv = cpu_ldq_data_ra(env, ptr + 512, ra);
-    new_bv = (old_bv & ~rfbm) | (get_xinuse(env) & rfbm);
+    new_bv = (old_bv & ~rfbm) | (inuse & rfbm);
     cpu_stq_data_ra(env, ptr + 512, new_bv, ra);
 }
 
+void helper_xsave(CPUX86State *env, target_ulong ptr, uint64_t rfbm)
+{
+    do_xsave(env, ptr, rfbm, get_xinuse(env), -1, GETPC());
+}
+
+void helper_xsaveopt(CPUX86State *env, target_ulong ptr, uint64_t rfbm)
+{
+    uint64_t inuse = get_xinuse(env);
+    do_xsave(env, ptr, rfbm, inuse, inuse, GETPC());
+}
+
 static void do_xrstor_fpu(CPUX86State *env, target_ulong ptr, uintptr_t ra)
 {
     int i, fpus, fptag;
@@ -1369,8 +1384,10 @@ uint64_t helper_xgetbv(CPUX86State *env, uint32_t ecx)
     case 0:
         return env->xcr0;
     case 1:
-        /* FIXME: #GP if !CPUID.(EAX=0DH,ECX=1):EAX.XG1[bit 2].  */
-        return env->xcr0 & get_xinuse(env);
+        if (env->features[FEAT_XSAVE] & CPUID_XSAVE_XGETBV1) {
+            return env->xcr0 & get_xinuse(env);
+        }
+        break;
     }
     raise_exception_ra(env, EXCP0D_GPF, GETPC());
 }
diff --git a/target-i386/helper.h b/target-i386/helper.h
index 9dfc735..9a83955 100644
--- a/target-i386/helper.h
+++ b/target-i386/helper.h
@@ -188,6 +188,7 @@ DEF_HELPER_3(frstor, void, env, tl, int)
 DEF_HELPER_FLAGS_2(fxsave, TCG_CALL_NO_WG, void, env, tl)
 DEF_HELPER_FLAGS_2(fxrstor, TCG_CALL_NO_WG, void, env, tl)
 DEF_HELPER_FLAGS_3(xsave, TCG_CALL_NO_WG, void, env, tl, i64)
+DEF_HELPER_FLAGS_3(xsaveopt, TCG_CALL_NO_WG, void, env, tl, i64)
 DEF_HELPER_FLAGS_3(xrstor, TCG_CALL_NO_WG, void, env, tl, i64)
 DEF_HELPER_FLAGS_2(xgetbv, TCG_CALL_NO_WG, i64, env, i32)
 DEF_HELPER_FLAGS_3(xsetbv, TCG_CALL_NO_WG, void, env, i32, i64)
diff --git a/target-i386/translate.c b/target-i386/translate.c
index 7571e85..ce2ffc1 100644
--- a/target-i386/translate.c
+++ b/target-i386/translate.c
@@ -127,6 +127,7 @@ typedef struct DisasContext {
     int cpuid_ext2_features;
     int cpuid_ext3_features;
     int cpuid_7_0_ebx_features;
+    int cpuid_xsave_features;
 } DisasContext;
 
 static void gen_eob(DisasContext *s);
@@ -7634,7 +7635,7 @@ static target_ulong disas_insn(CPUX86State *env, DisasContext *s,
             gen_helper_xrstor(cpu_env, cpu_A0, cpu_tmp1_i64);
             break;
 
-        CASE_MEM_OP(6): /* clwb */
+        CASE_MEM_OP(6): /* xsaveopt / clwb */
             if (prefixes & PREFIX_LOCK) {
                 goto illegal_op;
             }
@@ -7644,9 +7645,19 @@ static target_ulong disas_insn(CPUX86State *env, DisasContext *s,
                     goto illegal_op;
                 }
                 gen_nop_modrm(env, s, modrm);
-                break;
+            } else {
+                /* xsaveopt */
+                if ((s->cpuid_ext_features & CPUID_EXT_XSAVE) == 0
+                    || (s->cpuid_xsave_features & CPUID_XSAVE_XSAVEOPT) == 0
+                    || (prefixes & (PREFIX_REPZ | PREFIX_REPNZ))) {
+                    goto illegal_op;
+                }
+                gen_lea_modrm(env, s, modrm);
+                tcg_gen_concat_tl_i64(cpu_tmp1_i64, cpu_regs[R_EAX],
+                                      cpu_regs[R_EDX]);
+                gen_helper_xsaveopt(cpu_env, cpu_A0, cpu_tmp1_i64);
             }
-            goto illegal_op;
+            break;
 
         CASE_MEM_OP(7): /* clflush / clflushopt */
             if (prefixes & PREFIX_LOCK) {
@@ -7868,6 +7879,7 @@ void gen_intermediate_code(CPUX86State *env, TranslationBlock *tb)
     dc->cpuid_ext2_features = env->features[FEAT_8000_0001_EDX];
     dc->cpuid_ext3_features = env->features[FEAT_8000_0001_ECX];
     dc->cpuid_7_0_ebx_features = env->features[FEAT_7_0_EBX];
+    dc->cpuid_xsave_features = env->features[FEAT_XSAVE];
 #ifdef TARGET_X86_64
     dc->lma = (flags >> HF_LMA_SHIFT) & 1;
     dc->code64 = (flags >> HF_CS64_SHIFT) & 1;
-- 
2.5.0