[PATCH 17/18] target/i386/tcg: add a CCOp for SBB x,x

[Paolo Bonzini <pbonzini@redhat.com>]

This is more efficient both when generating code and when testing
flags.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 target/i386/cpu.h           | 13 ++++++++++++-
 target/i386/cpu-dump.c      |  2 ++
 target/i386/tcg/cc_helper.c |  6 ++++++
 target/i386/tcg/translate.c | 13 +++++++++++++
 target/i386/tcg/emit.c.inc  | 33 ++++++---------------------------
 5 files changed, 39 insertions(+), 28 deletions(-)

diff --git a/target/i386/cpu.h b/target/i386/cpu.h
index ecca38ed0b5..314e773a5d4 100644
--- a/target/i386/cpu.h
+++ b/target/i386/cpu.h
@@ -1515,7 +1515,18 @@ typedef enum {
     CC_OP_POPCNTL__,
     CC_OP_POPCNTQ__,
     CC_OP_POPCNT = sizeof(target_ulong) == 8 ? CC_OP_POPCNTQ__ : CC_OP_POPCNTL__,
-#define CC_OP_LAST_BWLQ CC_OP_POPCNTQ__
+
+    /*
+     * Note that only CC_OP_SBB_SELF (i.e. the one with MO_TL size)
+     * is used or implemented, because the translation produces a
+     * sign-extended CC_DST.
+     */
+    CC_OP_SBB_SELFB__, /* S/Z/C/A via CC_DST, O clear, P set.  */
+    CC_OP_SBB_SELFW__,
+    CC_OP_SBB_SELFL__,
+    CC_OP_SBB_SELFQ__,
+    CC_OP_SBB_SELF = sizeof(target_ulong) == 8 ? CC_OP_SBB_SELFQ__ : CC_OP_SBB_SELFL__,
+#define CC_OP_LAST_BWLQ CC_OP_SBB_SELFQ__
 
     CC_OP_DYNAMIC, /* must use dynamic code to get cc_op */
 } CCOp;
diff --git a/target/i386/cpu-dump.c b/target/i386/cpu-dump.c
index 67bf31e0caa..20a3002f013 100644
--- a/target/i386/cpu-dump.c
+++ b/target/i386/cpu-dump.c
@@ -91,6 +91,8 @@ static const char * const cc_op_str[] = {
     [CC_OP_BMILGQ] = "BMILGQ",
 
     [CC_OP_POPCNT] = "POPCNT",
+
+    [CC_OP_SBB_SELF] = "SBBx,x",
 };
 
 static void
diff --git a/target/i386/tcg/cc_helper.c b/target/i386/tcg/cc_helper.c
index 2c4170b5b77..91e492196af 100644
--- a/target/i386/tcg/cc_helper.c
+++ b/target/i386/tcg/cc_helper.c
@@ -100,6 +100,9 @@ target_ulong helper_cc_compute_all(target_ulong dst, target_ulong src1,
         return src1;
     case CC_OP_POPCNT:
         return dst ? 0 : CC_Z;
+    case CC_OP_SBB_SELF:
+	/* dst is either all zeros (--Z-P-) or all ones (-S-APC) */
+        return (dst & (CC_Z|CC_A|CC_C|CC_S)) ^ (CC_P | CC_Z);
 
     case CC_OP_ADCX:
         return compute_all_adcx(dst, src1, src2);
@@ -326,6 +329,9 @@ target_ulong helper_cc_compute_c(target_ulong dst, target_ulong src1,
     case CC_OP_MULQ:
         return src1 != 0;
 
+    case CC_OP_SBB_SELF:
+        return dst & 1;
+
     case CC_OP_ADCX:
     case CC_OP_ADCOX:
         return dst;
diff --git a/target/i386/tcg/translate.c b/target/i386/tcg/translate.c
index e91715af817..17ad4ccacaf 100644
--- a/target/i386/tcg/translate.c
+++ b/target/i386/tcg/translate.c
@@ -304,6 +304,7 @@ static const uint8_t cc_op_live_[] = {
     [CC_OP_ADOX] = USES_CC_SRC | USES_CC_SRC2,
     [CC_OP_ADCOX] = USES_CC_DST | USES_CC_SRC | USES_CC_SRC2,
     [CC_OP_POPCNT] = USES_CC_DST,
+    [CC_OP_SBB_SELF] = USES_CC_DST,
 };
 
 static uint8_t cc_op_live(CCOp op)
@@ -938,6 +939,9 @@ static CCPrepare gen_prepare_eflags_c(DisasContext *s, TCGv reg)
         size = cc_op_size(s->cc_op);
         return gen_prepare_val_nz(cpu_cc_src, size, false);
 
+    case CC_OP_SBB_SELF:
+        return (CCPrepare) { .cond = TCG_COND_NE, .reg = cpu_cc_dst };
+
     case CC_OP_ADCX:
     case CC_OP_ADCOX:
         return (CCPrepare) { .cond = TCG_COND_NE, .reg = cpu_cc_dst,
@@ -999,6 +1003,7 @@ static CCPrepare gen_prepare_eflags_o(DisasContext *s, TCGv reg)
     case CC_OP_ADCOX:
         return (CCPrepare) { .cond = TCG_COND_NE, .reg = cpu_cc_src2,
                              .no_setcond = true };
+    case CC_OP_SBB_SELF:
     case CC_OP_LOGICB ... CC_OP_LOGICQ:
     case CC_OP_POPCNT:
         return (CCPrepare) { .cond = TCG_COND_NEVER };
@@ -1078,6 +1083,14 @@ static CCPrepare gen_prepare_cc(DisasContext *s, int b, TCGv reg)
         }
         break;
 
+    case CC_OP_SBB_SELF:
+        /* checking for nonzero is usually the most efficient */
+        if (jcc_op == JCC_L || jcc_op == JCC_B || jcc_op == JCC_S) {
+            jcc_op = JCC_Z;
+            inv = !inv;
+        }
+        goto slow_jcc;
+
     case CC_OP_LOGICB ... CC_OP_LOGICQ:
         /* Mostly used for test+jump */
         size = s->cc_op - CC_OP_LOGICB;
diff --git a/target/i386/tcg/emit.c.inc b/target/i386/tcg/emit.c.inc
index 8dac4d09da1..0fde3d669d9 100644
--- a/target/i386/tcg/emit.c.inc
+++ b/target/i386/tcg/emit.c.inc
@@ -3876,37 +3876,16 @@ static void gen_SBB(DisasContext *s, X86DecodedInsn *decode)
         return;
     }
 
-    c_in = tcg_temp_new();
-    gen_compute_eflags_c(s, c_in);
-
-    /*
-     * Here the change is as follows:
-     * CC_SBB: src1 = T0, src2 = T0, src3 = c_in
-     * CC_SUB: src1 = 0, src2 = c_in (no src3)
-     *
-     * The difference also does not matter:
-     * - AF is bit 4 of dst^src1^src2, but bit 4 of src1^src2 is zero in both cases
-     *   therefore AF comes straight from dst (in fact it is c_in)
-     * - for OF, src1 and src2 have the same sign in both cases, meaning there
-     *   can be no overflow
-     */
+    /* SBB x,x has its own CCOp so that's even easier.  */
     if (decode->e.op2 != X86_TYPE_I && !decode->op[0].has_ea && decode->op[0].n == decode->op[2].n) {
-        if (s->cc_op == CC_OP_DYNAMIC) {
-            tcg_gen_neg_tl(s->T0, c_in);
-        } else {
-            /*
-             * Do not negate c_in because it will often be dead and only the
-             * instruction generated by negsetcond will survive.
-             */
-            gen_neg_setcc(s, JCC_B << 1, s->T0);
-        }
-        tcg_gen_movi_tl(s->cc_srcT, 0);
-        decode->cc_src = c_in;
-        decode->cc_dst = s->T0;
-        decode->cc_op = CC_OP_SUBB + ot;
+        gen_neg_setcc(s, JCC_B << 1, s->T0);
+        prepare_update1_cc(decode, s, CC_OP_SBB_SELF);
         return;
     }
 
+    c_in = tcg_temp_new();
+    gen_compute_eflags_c(s, c_in);
+
     if (s->prefix & PREFIX_LOCK) {
         tcg_gen_add_tl(s->T0, s->T1, c_in);
         tcg_gen_neg_tl(s->T0, s->T0);
-- 
2.52.0