[PATCH v4 17/27] tcg/s390x: Tighten constraints for or_i64 and xor_i64

[Richard Henderson]

Drop support for sequential OR and XOR, as the serial dependency is
slower than loading the constant first.  Let the register allocator
handle such immediates by matching only what one insn can achieve.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/s390x/tcg-target-con-set.h |   1 +
 tcg/s390x/tcg-target-con-str.h |   1 +
 tcg/s390x/tcg-target.c.inc     | 114 ++++++++++++++++-----------------
 3 files changed, 56 insertions(+), 60 deletions(-)

diff --git a/tcg/s390x/tcg-target-con-set.h b/tcg/s390x/tcg-target-con-set.h
index b1a89a88ba..34ae4c7743 100644
--- a/tcg/s390x/tcg-target-con-set.h
+++ b/tcg/s390x/tcg-target-con-set.h
@@ -24,6 +24,7 @@ C_O1_I2(r, 0, rI)
 C_O1_I2(r, 0, rJ)
 C_O1_I2(r, r, ri)
 C_O1_I2(r, r, rJ)
+C_O1_I2(r, r, rK)
 C_O1_I2(r, rZ, r)
 C_O1_I2(v, v, r)
 C_O1_I2(v, v, v)
diff --git a/tcg/s390x/tcg-target-con-str.h b/tcg/s390x/tcg-target-con-str.h
index 76446aecae..7b910d6d11 100644
--- a/tcg/s390x/tcg-target-con-str.h
+++ b/tcg/s390x/tcg-target-con-str.h
@@ -20,4 +20,5 @@ REGS('o', 0xaaaa) /* odd numbered general regs */
 CONST('A', TCG_CT_CONST_S33)
 CONST('I', TCG_CT_CONST_S16)
 CONST('J', TCG_CT_CONST_S32)
+CONST('K', TCG_CT_CONST_P32)
 CONST('Z', TCG_CT_CONST_ZERO)
diff --git a/tcg/s390x/tcg-target.c.inc b/tcg/s390x/tcg-target.c.inc
index fc304327fc..2a7410ba58 100644
--- a/tcg/s390x/tcg-target.c.inc
+++ b/tcg/s390x/tcg-target.c.inc
@@ -37,6 +37,7 @@
 #define TCG_CT_CONST_S32   0x200
 #define TCG_CT_CONST_S33   0x400
 #define TCG_CT_CONST_ZERO  0x800
+#define TCG_CT_CONST_P32   0x1000
 
 #define ALL_GENERAL_REGS     MAKE_64BIT_MASK(0, 16)
 #define ALL_VECTOR_REGS      MAKE_64BIT_MASK(32, 32)
@@ -507,6 +508,28 @@ static bool patch_reloc(tcg_insn_unit *src_rw, int type,
     return false;
 }
 
+static int is_const_p16(uint64_t val)
+{
+    for (int i = 0; i < 4; ++i) {
+        uint64_t mask = 0xffffull << (i * 16);
+        if ((val & ~mask) == 0) {
+            return i;
+        }
+    }
+    return -1;
+}
+
+static int is_const_p32(uint64_t val)
+{
+    if ((val & 0xffffffff00000000ull) == 0) {
+        return 0;
+    }
+    if ((val & 0x00000000ffffffffull) == 0) {
+        return 1;
+    }
+    return -1;
+}
+
 /* Test if a constant matches the constraint. */
 static bool tcg_target_const_match(int64_t val, TCGType type, int ct)
 {
@@ -529,6 +552,14 @@ static bool tcg_target_const_match(int64_t val, TCGType 
type, int ct)
         return val == 0;
     }
 
+    /*
+     * Note that is_const_p16 is a subset of is_const_p32,
+     * so we don't need both constraints.
+     */
+    if ((ct & TCG_CT_CONST_P32) && is_const_p32(val) >= 0) {
+        return true;
+    }
+
     return 0;
 }
 
@@ -1125,7 +1156,7 @@ static void tgen_andi(TCGContext *s, TCGType type, TCGReg 
dest, uint64_t val)
     }
 }
 
-static void tgen_ori(TCGContext *s, TCGType type, TCGReg dest, uint64_t val)
+static void tgen_ori(TCGContext *s, TCGReg dest, uint64_t val)
 {
     static const S390Opcode oi_insns[4] = {
         RI_OILL, RI_OILH, RI_OIHL, RI_OIHH
@@ -1136,70 +1167,32 @@ static void tgen_ori(TCGContext *s, TCGType type, 
TCGReg dest, uint64_t val)
 
     int i;
 
-    /* Look for no-op.  */
-    if (unlikely(val == 0)) {
+    i = is_const_p16(val);
+    if (i >= 0) {
+        tcg_out_insn_RI(s, oi_insns[i], dest, val >> (i * 16));
         return;
     }
 
-    /* Try all 32-bit insns that can perform it in one go.  */
-    for (i = 0; i < 4; i++) {
-        tcg_target_ulong mask = (0xffffull << i * 16);
-        if ((val & mask) != 0 && (val & ~mask) == 0) {
-            tcg_out_insn_RI(s, oi_insns[i], dest, val >> i * 16);
-            return;
-        }
+    i = is_const_p32(val);
+    if (i >= 0) {
+        tcg_out_insn_RIL(s, oif_insns[i], dest, val >> (i * 32));
+        return;
     }
 
-    /* Try all 48-bit insns that can perform it in one go.  */
-    for (i = 0; i < 2; i++) {
-        tcg_target_ulong mask = (0xffffffffull << i * 32);
-        if ((val & mask) != 0 && (val & ~mask) == 0) {
-            tcg_out_insn_RIL(s, oif_insns[i], dest, val >> i * 32);
-            return;
-        }
-    }
-
-    if (maybe_out_small_movi(s, type, TCG_TMP0, val)) {
-        if (type == TCG_TYPE_I32) {
-            tcg_out_insn(s, RR, OR, dest, TCG_TMP0);
-        } else {
-            tcg_out_insn(s, RRE, OGR, dest, TCG_TMP0);
-        }
-    } else {
-        /* Perform the OR via sequential modifications to the high and
-           low parts.  Do this via recursion to handle 16-bit vs 32-bit
-           masks in each half.  */
-        tgen_ori(s, type, dest, val & 0x00000000ffffffffull);
-        tgen_ori(s, type, dest, val & 0xffffffff00000000ull);
-    }
+    g_assert_not_reached();
 }
 
-static void tgen_xori(TCGContext *s, TCGType type, TCGReg dest, uint64_t val)
+static void tgen_xori(TCGContext *s, TCGReg dest, uint64_t val)
 {
-    /* Try all 48-bit insns that can perform it in one go.  */
-    if ((val & 0xffffffff00000000ull) == 0) {
+    switch (is_const_p32(val)) {
+    case 0:
         tcg_out_insn(s, RIL, XILF, dest, val);
-        return;
-    }
-    if ((val & 0x00000000ffffffffull) == 0) {
+        break;
+    case 1:
         tcg_out_insn(s, RIL, XIHF, dest, val >> 32);
-        return;
-    }
-
-    if (maybe_out_small_movi(s, type, TCG_TMP0, val)) {
-        if (type == TCG_TYPE_I32) {
-            tcg_out_insn(s, RR, XR, dest, TCG_TMP0);
-        } else {
-            tcg_out_insn(s, RRE, XGR, dest, TCG_TMP0);
-        }
-    } else {
-        /* Perform the xor by parts.  */
-        if (val & 0xffffffff) {
-            tcg_out_insn(s, RIL, XILF, dest, val);
-        }
-        if (val > 0xffffffff) {
-            tcg_out_insn(s, RIL, XIHF, dest, val >> 32);
-        }
+        break;
+    default:
+        g_assert_not_reached();
     }
 }
 
@@ -1994,7 +1987,7 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode 
opc,
         a0 = args[0], a1 = args[1], a2 = (uint32_t)args[2];
         if (const_args[2]) {
             tcg_out_mov(s, TCG_TYPE_I32, a0, a1);
-            tgen_ori(s, TCG_TYPE_I32, a0, a2);
+            tgen_ori(s, a0, a2);
         } else if (a0 == a1) {
             tcg_out_insn(s, RR, OR, a0, a2);
         } else {
@@ -2256,7 +2249,7 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode 
opc,
         a0 = args[0], a1 = args[1], a2 = args[2];
         if (const_args[2]) {
             tcg_out_mov(s, TCG_TYPE_I64, a0, a1);
-            tgen_ori(s, TCG_TYPE_I64, a0, a2);
+            tgen_ori(s, a0, a2);
         } else {
             tcg_out_insn(s, RRFa, OGRK, a0, a1, a2);
         }
@@ -2265,7 +2258,7 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode 
opc,
         a0 = args[0], a1 = args[1], a2 = args[2];
         if (const_args[2]) {
             tcg_out_mov(s, TCG_TYPE_I64, a0, a1);
-            tgen_xori(s, TCG_TYPE_I64, a0, a2);
+            tgen_xori(s, a0, a2);
         } else {
             tcg_out_insn(s, RRFa, XGRK, a0, a1, a2);
         }
@@ -2944,10 +2937,11 @@ static TCGConstraintSetIndex 
tcg_target_op_def(TCGOpcode op)
     case INDEX_op_and_i32:
     case INDEX_op_and_i64:
     case INDEX_op_or_i32:
-    case INDEX_op_or_i64:
     case INDEX_op_xor_i32:
-    case INDEX_op_xor_i64:
         return C_O1_I2(r, r, ri);
+    case INDEX_op_or_i64:
+    case INDEX_op_xor_i64:
+        return C_O1_I2(r, r, rK);
 
     case INDEX_op_mul_i32:
         return (HAVE_FACILITY(MISC_INSN_EXT2)
-- 
2.34.1