diff mbox series

[02/25] target/i386: rewrite flags writeback for ADCX/ADOX

Message ID 20240608084113.2770363-3-pbonzini@redhat.com
State New
Headers show
Series target/i386: more progress towards new decoder | expand

Commit Message

Paolo Bonzini June 8, 2024, 8:40 a.m. UTC
Avoid using set_cc_op() in preparation for implementing APX; treat
CC_OP_EFLAGS similar to the case where we have the "opposite" cc_op
(CC_OP_ADOX for ADCX and CC_OP_ADCX for ADOX), except the resulting
cc_op is not CC_OP_ADCOX. This is written easily as two "if"s, whose
conditions are both false for CC_OP_EFLAGS, both true for CC_OP_ADCOX,
and one each true for CC_OP_ADCX/ADOX.

The new logic also makes it easy to drop usage of tmp0.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 target/i386/cpu.h          |  9 +++---
 target/i386/tcg/emit.c.inc | 61 ++++++++++++++++++++++----------------
 2 files changed, 40 insertions(+), 30 deletions(-)

Comments

Richard Henderson June 8, 2024, 6:05 p.m. UTC | #1
On 6/8/24 01:40, Paolo Bonzini wrote:
> Avoid using set_cc_op() in preparation for implementing APX; treat
> CC_OP_EFLAGS similar to the case where we have the "opposite" cc_op
> (CC_OP_ADOX for ADCX and CC_OP_ADCX for ADOX), except the resulting
> cc_op is not CC_OP_ADCOX. This is written easily as two "if"s, whose
> conditions are both false for CC_OP_EFLAGS, both true for CC_OP_ADCOX,
> and one each true for CC_OP_ADCX/ADOX.
> 
> The new logic also makes it easy to drop usage of tmp0.
> 
> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
> ---
>   target/i386/cpu.h          |  9 +++---
>   target/i386/tcg/emit.c.inc | 61 ++++++++++++++++++++++----------------
>   2 files changed, 40 insertions(+), 30 deletions(-)
> 
> diff --git a/target/i386/cpu.h b/target/i386/cpu.h
> index 8fe28b67e0f..ee873a0ed84 100644
> --- a/target/i386/cpu.h
> +++ b/target/i386/cpu.h
> @@ -1260,6 +1260,8 @@ uint64_t x86_cpu_get_supported_feature_word(FeatureWord w,
>   /* Use a clearer name for this.  */
>   #define CPU_INTERRUPT_INIT      CPU_INTERRUPT_RESET
>   
> +#define CC_OP_HAS_EFLAGS(op) ((op) >= CC_OP_EFLAGS && (op) <= CC_OP_ADCOX)
> +
>   /* Instead of computing the condition codes after each x86 instruction,
>    * QEMU just stores one operand (called CC_SRC), the result
>    * (called CC_DST) and the type of operation (called CC_OP). When the
> @@ -1270,6 +1272,9 @@ uint64_t x86_cpu_get_supported_feature_word(FeatureWord w,
>   typedef enum {
>       CC_OP_DYNAMIC, /* must use dynamic code to get cc_op */
>       CC_OP_EFLAGS,  /* all cc are explicitly computed, CC_SRC = flags */
> +    CC_OP_ADCX, /* CC_DST = C, CC_SRC = rest.  */
> +    CC_OP_ADOX, /* CC_DST = O, CC_SRC = rest.  */
> +    CC_OP_ADCOX, /* CC_DST = C, CC_SRC2 = O, CC_SRC = rest.  */

Comment for ADOX is (and was) wrong -- SRC2 = O.
That made review a bit confusing... :-)

Otherwise,
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>


r~
diff mbox series

Patch

diff --git a/target/i386/cpu.h b/target/i386/cpu.h
index 8fe28b67e0f..ee873a0ed84 100644
--- a/target/i386/cpu.h
+++ b/target/i386/cpu.h
@@ -1260,6 +1260,8 @@  uint64_t x86_cpu_get_supported_feature_word(FeatureWord w,
 /* Use a clearer name for this.  */
 #define CPU_INTERRUPT_INIT      CPU_INTERRUPT_RESET
 
+#define CC_OP_HAS_EFLAGS(op) ((op) >= CC_OP_EFLAGS && (op) <= CC_OP_ADCOX)
+
 /* Instead of computing the condition codes after each x86 instruction,
  * QEMU just stores one operand (called CC_SRC), the result
  * (called CC_DST) and the type of operation (called CC_OP). When the
@@ -1270,6 +1272,9 @@  uint64_t x86_cpu_get_supported_feature_word(FeatureWord w,
 typedef enum {
     CC_OP_DYNAMIC, /* must use dynamic code to get cc_op */
     CC_OP_EFLAGS,  /* all cc are explicitly computed, CC_SRC = flags */
+    CC_OP_ADCX, /* CC_DST = C, CC_SRC = rest.  */
+    CC_OP_ADOX, /* CC_DST = O, CC_SRC = rest.  */
+    CC_OP_ADCOX, /* CC_DST = C, CC_SRC2 = O, CC_SRC = rest.  */
 
     CC_OP_MULB, /* modify all flags, C, O = (CC_SRC != 0) */
     CC_OP_MULW,
@@ -1326,10 +1331,6 @@  typedef enum {
     CC_OP_BMILGL,
     CC_OP_BMILGQ,
 
-    CC_OP_ADCX, /* CC_DST = C, CC_SRC = rest.  */
-    CC_OP_ADOX, /* CC_DST = O, CC_SRC = rest.  */
-    CC_OP_ADCOX, /* CC_DST = C, CC_SRC2 = O, CC_SRC = rest.  */
-
     CC_OP_CLR, /* Z set, all other flags clear.  */
     CC_OP_POPCNT, /* Z via CC_SRC, all other flags clear.  */
 
diff --git a/target/i386/tcg/emit.c.inc b/target/i386/tcg/emit.c.inc
index df7597c7e2f..2041ea9d04a 100644
--- a/target/i386/tcg/emit.c.inc
+++ b/target/i386/tcg/emit.c.inc
@@ -1122,24 +1122,41 @@  static void gen_ADC(DisasContext *s, X86DecodedInsn *decode)
     prepare_update3_cc(decode, s, CC_OP_ADCB + ot, c_in);
 }
 
-/* ADCX/ADOX do not have memory operands and can use set_cc_op.  */
-static void gen_ADCOX(DisasContext *s, MemOp ot, int cc_op)
+static void gen_ADCOX(DisasContext *s, X86DecodedInsn *decode, int cc_op)
 {
-    int opposite_cc_op;
+    MemOp ot = decode->op[0].ot;
     TCGv carry_in = NULL;
-    TCGv carry_out = (cc_op == CC_OP_ADCX ? cpu_cc_dst : cpu_cc_src2);
+    TCGv *carry_out = (cc_op == CC_OP_ADCX ? &decode->cc_dst : &decode->cc_src2);
     TCGv zero;
 
-    if (cc_op == s->cc_op || s->cc_op == CC_OP_ADCOX) {
-        /* Re-use the carry-out from a previous round.  */
-        carry_in = carry_out;
-    } else {
-        /* We don't have a carry-in, get it out of EFLAGS.  */
-        if (s->cc_op != CC_OP_ADCX && s->cc_op != CC_OP_ADOX) {
-            gen_compute_eflags(s);
+    decode->cc_op = cc_op;
+    *carry_out = tcg_temp_new();
+    if (CC_OP_HAS_EFLAGS(s->cc_op)) {
+        decode->cc_src = cpu_cc_src;
+
+        /* Re-use the carry-out from a previous round?  */
+        if (s->cc_op == cc_op || s->cc_op == CC_OP_ADCOX) {
+            carry_in = (cc_op == CC_OP_ADCX ? cpu_cc_dst : cpu_cc_src2);
         }
-        carry_in = s->tmp0;
-        tcg_gen_extract_tl(carry_in, cpu_cc_src,
+
+        /* Preserve the opposite carry from previous rounds?  */
+        if (s->cc_op != cc_op && s->cc_op != CC_OP_EFLAGS) {
+            decode->cc_op = CC_OP_ADCOX;
+            if (carry_out == &decode->cc_dst) {
+                decode->cc_src2 = cpu_cc_src2;
+            } else {
+                decode->cc_dst = cpu_cc_dst;
+            }
+        }
+    } else {
+        decode->cc_src = tcg_temp_new();
+        gen_mov_eflags(s, decode->cc_src);
+    }
+
+    if (!carry_in) {
+        /* Get carry_in out of EFLAGS.  */
+        carry_in = tcg_temp_new();
+        tcg_gen_extract_tl(carry_in, decode->cc_src,
             ctz32(cc_op == CC_OP_ADCX ? CC_C : CC_O), 1);
     }
 
@@ -1151,28 +1168,20 @@  static void gen_ADCOX(DisasContext *s, MemOp ot, int cc_op)
         tcg_gen_ext32u_tl(s->T1, s->T1);
         tcg_gen_add_i64(s->T0, s->T0, s->T1);
         tcg_gen_add_i64(s->T0, s->T0, carry_in);
-        tcg_gen_shri_i64(carry_out, s->T0, 32);
+        tcg_gen_shri_i64(*carry_out, s->T0, 32);
         break;
 #endif
     default:
         zero = tcg_constant_tl(0);
-        tcg_gen_add2_tl(s->T0, carry_out, s->T0, zero, carry_in, zero);
-        tcg_gen_add2_tl(s->T0, carry_out, s->T0, carry_out, s->T1, zero);
+        tcg_gen_add2_tl(s->T0, *carry_out, s->T0, zero, carry_in, zero);
+        tcg_gen_add2_tl(s->T0, *carry_out, s->T0, *carry_out, s->T1, zero);
         break;
     }
-
-    opposite_cc_op = cc_op == CC_OP_ADCX ? CC_OP_ADOX : CC_OP_ADCX;
-    if (s->cc_op == CC_OP_ADCOX || s->cc_op == opposite_cc_op) {
-        /* Merge with the carry-out from the opposite instruction.  */
-        set_cc_op(s, CC_OP_ADCOX);
-    } else {
-        set_cc_op(s, cc_op);
-    }
 }
 
 static void gen_ADCX(DisasContext *s, X86DecodedInsn *decode)
 {
-    gen_ADCOX(s, decode->op[0].ot, CC_OP_ADCX);
+    gen_ADCOX(s, decode, CC_OP_ADCX);
 }
 
 static void gen_ADD(DisasContext *s, X86DecodedInsn *decode)
@@ -1190,7 +1199,7 @@  static void gen_ADD(DisasContext *s, X86DecodedInsn *decode)
 
 static void gen_ADOX(DisasContext *s, X86DecodedInsn *decode)
 {
-    gen_ADCOX(s, decode->op[0].ot, CC_OP_ADOX);
+    gen_ADCOX(s, decode, CC_OP_ADOX);
 }
 
 static void gen_AND(DisasContext *s, X86DecodedInsn *decode)