[PATCH v2 18/54] tcg: Introduce tcg_out_movext2

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

This is common code in most qemu_{ld,st} slow paths, moving two
registers when there may be overlap between sources and destinations.
At present, this is only used by 32-bit hosts for 64-bit data,
but will shortly be used for more than that.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/tcg.c                 | 69 ++++++++++++++++++++++++++++++++++++---
 tcg/arm/tcg-target.c.inc  | 42 ++++++++++--------------
 tcg/i386/tcg-target.c.inc | 19 +++++------
 3 files changed, 89 insertions(+), 41 deletions(-)

diff --git a/tcg/tcg.c b/tcg/tcg.c
index fde5ccc57c..cfd3262a4a 100644
--- a/tcg/tcg.c
+++ b/tcg/tcg.c
@@ -115,8 +115,7 @@ static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg);
 static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg);
 static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg ret, TCGReg arg);
 static void tcg_out_addi_ptr(TCGContext *s, TCGReg, TCGReg, tcg_target_long);
-static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
-    __attribute__((unused));
+static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2);
 static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg);
 static void tcg_out_goto_tb(TCGContext *s, int which);
 static void tcg_out_op(TCGContext *s, TCGOpcode opc,
@@ -354,6 +353,14 @@ void tcg_raise_tb_overflow(TCGContext *s)
     siglongjmp(s->jmp_trans, -2);
 }
 
+typedef struct TCGMovExtend {
+    TCGReg dst;
+    TCGReg src;
+    TCGType dst_type;
+    TCGType src_type;
+    MemOp src_ext;
+} TCGMovExtend;
+
 /**
  * tcg_out_movext -- move and extend
  * @s: tcg context
@@ -365,9 +372,8 @@ void tcg_raise_tb_overflow(TCGContext *s)
  *
  * Move or extend @src into @dst, depending on @src_ext and the types.
  */
-static void __attribute__((unused))
-tcg_out_movext(TCGContext *s, TCGType dst_type, TCGReg dst,
-               TCGType src_type, MemOp src_ext, TCGReg src)
+static void tcg_out_movext(TCGContext *s, TCGType dst_type, TCGReg dst,
+                           TCGType src_type, MemOp src_ext, TCGReg src)
 {
     switch (src_ext) {
     case MO_UB:
@@ -417,6 +423,59 @@ tcg_out_movext(TCGContext *s, TCGType dst_type, TCGReg dst,
     }
 }
 
+/* Minor variations on a theme, using a structure. */
+static void tcg_out_movext1_new_src(TCGContext *s, const TCGMovExtend *i,
+                                    TCGReg src)
+{
+    tcg_out_movext(s, i->dst_type, i->dst, i->src_type, i->src_ext, src);
+}
+
+static void tcg_out_movext1(TCGContext *s, const TCGMovExtend *i)
+{
+    tcg_out_movext1_new_src(s, i, i->src);
+}
+
+/**
+ * tcg_out_movext2 -- move and extend two pair
+ * @s: tcg context
+ * @i1: first move description
+ * @i2: second move description
+ * @scratch: temporary register, or -1 for none
+ *
+ * As tcg_out_movext, for both @i1 and @i2, caring for overlap
+ * between the sources and destinations.
+ */
+
+static void __attribute__((unused))
+tcg_out_movext2(TCGContext *s, const TCGMovExtend *i1,
+                const TCGMovExtend *i2, int scratch)
+{
+    TCGReg src1 = i1->src;
+    TCGReg src2 = i2->src;
+
+    if (i1->dst != src2) {
+        tcg_out_movext1(s, i1);
+        tcg_out_movext1(s, i2);
+        return;
+    }
+    if (i2->dst == src1) {
+        TCGType src1_type = i1->src_type;
+        TCGType src2_type = i2->src_type;
+
+        if (tcg_out_xchg(s, MAX(src1_type, src2_type), src1, src2)) {
+            /* The data is now in the correct registers, now extend. */
+            src1 = i2->src;
+            src2 = i1->src;
+        } else {
+            tcg_debug_assert(scratch >= 0);
+            tcg_out_mov(s, src1_type, scratch, src1);
+            src1 = scratch;
+        }
+    }
+    tcg_out_movext1_new_src(s, i2, src2);
+    tcg_out_movext1_new_src(s, i1, src1);
+}
+
 #define C_PFX1(P, A)                    P##A
 #define C_PFX2(P, A, B)                 P##A##_##B
 #define C_PFX3(P, A, B, C)              P##A##_##B##_##C
diff --git a/tcg/arm/tcg-target.c.inc b/tcg/arm/tcg-target.c.inc
index 4a5d57a41c..83c818a58b 100644
--- a/tcg/arm/tcg-target.c.inc
+++ b/tcg/arm/tcg-target.c.inc
@@ -1545,7 +1545,7 @@ static void add_qemu_ldst_label(TCGContext *s, bool is_ld, MemOpIdx oi,
 
 static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
 {
-    TCGReg argreg, datalo, datahi;
+    TCGReg argreg;
     MemOpIdx oi = lb->oi;
     MemOp opc = get_memop(oi);
 
@@ -1565,20 +1565,14 @@ static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
     /* Use the canonical unsigned helpers and minimize icache usage. */
     tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE]);
 
-    datalo = lb->datalo_reg;
-    datahi = lb->datahi_reg;
     if ((opc & MO_SIZE) == MO_64) {
-        if (datalo != TCG_REG_R1) {
-            tcg_out_mov_reg(s, COND_AL, datalo, TCG_REG_R0);
-            tcg_out_mov_reg(s, COND_AL, datahi, TCG_REG_R1);
-        } else if (datahi != TCG_REG_R0) {
-            tcg_out_mov_reg(s, COND_AL, datahi, TCG_REG_R1);
-            tcg_out_mov_reg(s, COND_AL, datalo, TCG_REG_R0);
-        } else {
-            tcg_out_mov_reg(s, COND_AL, TCG_REG_TMP, TCG_REG_R0);
-            tcg_out_mov_reg(s, COND_AL, datahi, TCG_REG_R1);
-            tcg_out_mov_reg(s, COND_AL, datalo, TCG_REG_TMP);
-        }
+        TCGMovExtend ext[2] = {
+            { .dst = lb->datalo_reg, .dst_type = TCG_TYPE_I32,
+              .src = TCG_REG_R0, .src_type = TCG_TYPE_I32, .src_ext = MO_UL },
+            { .dst = lb->datahi_reg, .dst_type = TCG_TYPE_I32,
+              .src = TCG_REG_R1, .src_type = TCG_TYPE_I32, .src_ext = MO_UL },
+        };
+        tcg_out_movext2(s, &ext[0], &ext[1], TCG_REG_TMP);
     } else {
         tcg_out_movext(s, TCG_TYPE_I32, lb->datalo_reg,
                        TCG_TYPE_I32, opc & MO_SSIZE, TCG_REG_R0);
@@ -1663,17 +1657,15 @@ static bool tcg_out_fail_alignment(TCGContext *s, TCGLabelQemuLdst *l)
 
     if (TARGET_LONG_BITS == 64) {
         /* 64-bit target address is aligned into R2:R3. */
-        if (l->addrhi_reg != TCG_REG_R2) {
-            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R2, l->addrlo_reg);
-            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R3, l->addrhi_reg);
-        } else if (l->addrlo_reg != TCG_REG_R3) {
-            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R3, l->addrhi_reg);
-            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R2, l->addrlo_reg);
-        } else {
-            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R1, TCG_REG_R2);
-            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R2, TCG_REG_R3);
-            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R3, TCG_REG_R1);
-        }
+        TCGMovExtend ext[2] = {
+            { .dst = TCG_REG_R2, .dst_type = TCG_TYPE_I32,
+              .src = l->addrlo_reg,
+              .src_type = TCG_TYPE_I32, .src_ext = MO_UL },
+            { .dst = TCG_REG_R3, .dst_type = TCG_TYPE_I32,
+              .src = l->addrhi_reg,
+              .src_type = TCG_TYPE_I32, .src_ext = MO_UL },
+        };
+        tcg_out_movext2(s, &ext[0], &ext[1], TCG_REG_TMP);
     } else {
         tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R1, l->addrlo_reg);
     }
diff --git a/tcg/i386/tcg-target.c.inc b/tcg/i386/tcg-target.c.inc
index ce87f8fbc9..238a75b17e 100644
--- a/tcg/i386/tcg-target.c.inc
+++ b/tcg/i386/tcg-target.c.inc
@@ -1916,7 +1916,6 @@ static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
 {
     MemOpIdx oi = l->oi;
     MemOp opc = get_memop(oi);
-    TCGReg data_reg;
     tcg_insn_unit **label_ptr = &l->label_ptr[0];
 
     /* resolve label address */
@@ -1953,18 +1952,16 @@ static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
 
     tcg_out_branch(s, 1, qemu_ld_helpers[opc & (MO_BSWAP | MO_SIZE)]);
 
-    data_reg = l->datalo_reg;
     if (TCG_TARGET_REG_BITS == 32 && (opc & MO_SIZE) == MO_64) {
-        if (data_reg == TCG_REG_EDX) {
-            /* xchg %edx, %eax */
-            tcg_out_opc(s, OPC_XCHG_ax_r32 + TCG_REG_EDX, 0, 0, 0);
-            tcg_out_mov(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_EAX);
-        } else {
-            tcg_out_mov(s, TCG_TYPE_I32, data_reg, TCG_REG_EAX);
-            tcg_out_mov(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_EDX);
-        }
+        TCGMovExtend ext[2] = {
+            { .dst = l->datalo_reg, .dst_type = TCG_TYPE_I32,
+              .src = TCG_REG_EAX, .src_type = TCG_TYPE_I32, .src_ext = MO_UL },
+            { .dst = l->datahi_reg, .dst_type = TCG_TYPE_I32,
+              .src = TCG_REG_EDX, .src_type = TCG_TYPE_I32, .src_ext = MO_UL },
+        };
+        tcg_out_movext2(s, &ext[0], &ext[1], -1);
     } else {
-        tcg_out_movext(s, l->type, data_reg,
+        tcg_out_movext(s, l->type, l->datalo_reg,
                        TCG_TYPE_REG, opc & MO_SSIZE, TCG_REG_EAX);
     }
 
-- 
2.34.1