tcg/i386: Support 128-bit load/store

Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

2 files changed, 190 insertions, 5 deletions

diff --git a/tcg/i386/tcg-target.c.inc b/tcg/i386/tcg-target.c.inc
index bfe9d98..ae54e5f 100644
--- a/tcg/i386/tcg-target.c.inc
+++ b/tcg/i386/tcg-target.c.inc
@@ -91,6 +91,8 @@ static const int tcg_target_reg_alloc_order[] = {
 #endif
 };
 
+#define TCG_TMP_VEC  TCG_REG_XMM5
+
 static const int tcg_target_call_iarg_regs[] = {
 #if TCG_TARGET_REG_BITS == 64
 #if defined(_WIN64)
@@ -319,6 +321,8 @@ static bool tcg_target_const_match(int64_t val, TCGType type, int ct)
 #define OPC_PCMPGTW     (0x65 | P_EXT | P_DATA16)
 #define OPC_PCMPGTD     (0x66 | P_EXT | P_DATA16)
 #define OPC_PCMPGTQ     (0x37 | P_EXT38 | P_DATA16)
+#define OPC_PEXTRD      (0x16 | P_EXT3A | P_DATA16)
+#define OPC_PINSRD      (0x22 | P_EXT3A | P_DATA16)
 #define OPC_PMAXSB      (0x3c | P_EXT38 | P_DATA16)
 #define OPC_PMAXSW      (0xee | P_EXT | P_DATA16)
 #define OPC_PMAXSD      (0x3d | P_EXT38 | P_DATA16)
@@ -1753,7 +1757,21 @@ typedef struct {
 
 bool tcg_target_has_memory_bswap(MemOp memop)
 {
-    return have_movbe;
+    TCGAtomAlign aa;
+
+    if (!have_movbe) {
+        return false;
+    }
+    if ((memop & MO_SIZE) < MO_128) {
+        return true;
+    }
+
+    /*
+     * Reject 16-byte memop with 16-byte atomicity, i.e. VMOVDQA,
+     * but do allow a pair of 64-bit operations, i.e. MOVBEQ.
+     */
+    aa = atom_and_align_for_opc(tcg_ctx, memop, MO_ATOM_IFALIGN, true);
+    return aa.atom < MO_128;
 }
 
 /*
@@ -1781,6 +1799,30 @@ static const TCGLdstHelperParam ldst_helper_param = {
 static const TCGLdstHelperParam ldst_helper_param = { };
 #endif
 
+static void tcg_out_vec_to_pair(TCGContext *s, TCGType type,
+                                TCGReg l, TCGReg h, TCGReg v)
+{
+    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
+
+    /* vpmov{d,q} %v, %l */
+    tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, v, 0, l);
+    /* vpextr{d,q} $1, %v, %h */
+    tcg_out_vex_modrm(s, OPC_PEXTRD + rexw, v, 0, h);
+    tcg_out8(s, 1);
+}
+
+static void tcg_out_pair_to_vec(TCGContext *s, TCGType type,
+                                TCGReg v, TCGReg l, TCGReg h)
+{
+    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
+
+    /* vmov{d,q} %l, %v */
+    tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, v, 0, l);
+    /* vpinsr{d,q} $1, %h, %v, %v */
+    tcg_out_vex_modrm(s, OPC_PINSRD + rexw, v, v, h);
+    tcg_out8(s, 1);
+}
+
 /*
  * Generate code for the slow path for a load at the end of block
  */
@@ -1870,6 +1912,7 @@ static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
 {
     TCGLabelQemuLdst *ldst = NULL;
     MemOp opc = get_memop(oi);
+    MemOp s_bits = opc & MO_SIZE;
     unsigned a_mask;
 
 #ifdef CONFIG_SOFTMMU
@@ -1880,7 +1923,7 @@ static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
     *h = x86_guest_base;
 #endif
     h->base = addrlo;
-    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
+    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, s_bits == MO_128);
     a_mask = (1 << h->aa.align) - 1;
 
 #ifdef CONFIG_SOFTMMU
@@ -1890,7 +1933,6 @@ static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
     TCGType tlbtype = TCG_TYPE_I32;
     int trexw = 0, hrexw = 0, tlbrexw = 0;
     unsigned mem_index = get_mmuidx(oi);
-    unsigned s_bits = opc & MO_SIZE;
     unsigned s_mask = (1 << s_bits) - 1;
     int tlb_mask;
 
@@ -2070,6 +2112,72 @@ static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
                                      h.base, h.index, 0, h.ofs + 4);
         }
         break;
+
+    case MO_128:
+        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
+
+        /*
+         * Without 16-byte atomicity, use integer regs.
+         * That is where we want the data, and it allows bswaps.
+         */
+        if (h.aa.atom < MO_128) {
+            if (use_movbe) {
+                TCGReg t = datalo;
+                datalo = datahi;
+                datahi = t;
+            }
+            if (h.base == datalo || h.index == datalo) {
+                tcg_out_modrm_sib_offset(s, OPC_LEA + P_REXW, datahi,
+                                         h.base, h.index, 0, h.ofs);
+                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
+                                     datalo, datahi, 0);
+                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
+                                     datahi, datahi, 8);
+            } else {
+                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
+                                         h.base, h.index, 0, h.ofs);
+                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
+                                         h.base, h.index, 0, h.ofs + 8);
+            }
+            break;
+        }
+
+        /*
+         * With 16-byte atomicity, a vector load is required.
+         * If we already have 16-byte alignment, then VMOVDQA always works.
+         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
+         * Else use we require a runtime test for alignment for VMOVDQA;
+         * use VMOVDQU on the unaligned nonatomic path for simplicity.
+         */
+        if (h.aa.align >= MO_128) {
+            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
+                                         TCG_TMP_VEC, 0,
+                                         h.base, h.index, 0, h.ofs);
+        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
+            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
+                                         TCG_TMP_VEC, 0,
+                                         h.base, h.index, 0, h.ofs);
+        } else {
+            TCGLabel *l1 = gen_new_label();
+            TCGLabel *l2 = gen_new_label();
+
+            tcg_out_testi(s, h.base, 15);
+            tcg_out_jxx(s, JCC_JNE, l1, true);
+
+            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
+                                         TCG_TMP_VEC, 0,
+                                         h.base, h.index, 0, h.ofs);
+            tcg_out_jxx(s, JCC_JMP, l2, true);
+
+            tcg_out_label(s, l1);
+            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
+                                         TCG_TMP_VEC, 0,
+                                         h.base, h.index, 0, h.ofs);
+            tcg_out_label(s, l2);
+        }
+        tcg_out_vec_to_pair(s, TCG_TYPE_I64, datalo, datahi, TCG_TMP_VEC);
+        break;
+
     default:
         g_assert_not_reached();
     }
@@ -2140,6 +2248,63 @@ static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
                                      h.base, h.index, 0, h.ofs + 4);
         }
         break;
+
+    case MO_128:
+        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
+
+        /*
+         * Without 16-byte atomicity, use integer regs.
+         * That is where we have the data, and it allows bswaps.
+         */
+        if (h.aa.atom < MO_128) {
+            if (use_movbe) {
+                TCGReg t = datalo;
+                datalo = datahi;
+                datahi = t;
+            }
+            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
+                                     h.base, h.index, 0, h.ofs);
+            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
+                                     h.base, h.index, 0, h.ofs + 8);
+            break;
+        }
+
+        /*
+         * With 16-byte atomicity, a vector store is required.
+         * If we already have 16-byte alignment, then VMOVDQA always works.
+         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
+         * Else use we require a runtime test for alignment for VMOVDQA;
+         * use VMOVDQU on the unaligned nonatomic path for simplicity.
+         */
+        tcg_out_pair_to_vec(s, TCG_TYPE_I64, TCG_TMP_VEC, datalo, datahi);
+        if (h.aa.align >= MO_128) {
+            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
+                                         TCG_TMP_VEC, 0,
+                                         h.base, h.index, 0, h.ofs);
+        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
+            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
+                                         TCG_TMP_VEC, 0,
+                                         h.base, h.index, 0, h.ofs);
+        } else {
+            TCGLabel *l1 = gen_new_label();
+            TCGLabel *l2 = gen_new_label();
+
+            tcg_out_testi(s, h.base, 15);
+            tcg_out_jxx(s, JCC_JNE, l1, true);
+
+            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
+                                         TCG_TMP_VEC, 0,
+                                         h.base, h.index, 0, h.ofs);
+            tcg_out_jxx(s, JCC_JMP, l2, true);
+
+            tcg_out_label(s, l1);
+            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
+                                         TCG_TMP_VEC, 0,
+                                         h.base, h.index, 0, h.ofs);
+            tcg_out_label(s, l2);
+        }
+        break;
+
     default:
         g_assert_not_reached();
     }
@@ -2470,6 +2635,11 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
             tcg_out_qemu_ld(s, a0, a1, a2, args[3], args[4], TCG_TYPE_I64);
         }
         break;
+    case INDEX_op_qemu_ld_a32_i128:
+    case INDEX_op_qemu_ld_a64_i128:
+        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
+        tcg_out_qemu_ld(s, a0, a1, a2, -1, args[3], TCG_TYPE_I128);
+        break;
 
     case INDEX_op_qemu_st_a64_i32:
     case INDEX_op_qemu_st8_a64_i32:
@@ -2496,6 +2666,11 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
             tcg_out_qemu_st(s, a0, a1, a2, args[3], args[4], TCG_TYPE_I64);
         }
         break;
+    case INDEX_op_qemu_st_a32_i128:
+    case INDEX_op_qemu_st_a64_i128:
+        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
+        tcg_out_qemu_st(s, a0, a1, a2, -1, args[3], TCG_TYPE_I128);
+        break;
 
     OP_32_64(mulu2):
         tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_MUL, args[3]);
@@ -3193,6 +3368,15 @@ static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
     case INDEX_op_qemu_st_a64_i64:
         return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I4(L, L, L, L);
 
+    case INDEX_op_qemu_ld_a32_i128:
+    case INDEX_op_qemu_ld_a64_i128:
+        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
+        return C_O2_I1(r, r, L);
+    case INDEX_op_qemu_st_a32_i128:
+    case INDEX_op_qemu_st_a64_i128:
+        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
+        return C_O0_I3(L, L, L);
+
     case INDEX_op_brcond2_i32:
         return C_O0_I4(r, r, ri, ri);
 
@@ -3962,6 +4146,7 @@ static void tcg_target_init(TCGContext *s)
 
     s->reserved_regs = 0;
     tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
+    tcg_regset_set_reg(s->reserved_regs, TCG_TMP_VEC);
 #ifdef _WIN64
     /* These are call saved, and we don't save them, so don't use them. */
     tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM6);
diff --git a/tcg/i386/tcg-target.h b/tcg/i386/tcg-target.h
index 0106946..b167f1e 100644
--- a/tcg/i386/tcg-target.h
+++ b/tcg/i386/tcg-target.h
@@ -118,7 +118,6 @@ typedef enum {
 #define have_avx1         (cpuinfo & CPUINFO_AVX1)
 #define have_avx2         (cpuinfo & CPUINFO_AVX2)
 #define have_movbe        (cpuinfo & CPUINFO_MOVBE)
-#define have_atomic16     (cpuinfo & CPUINFO_ATOMIC_VMOVDQA)
 
 /*
  * There are interesting instructions in AVX512, so long as we have AVX512VL,
@@ -202,7 +201,8 @@ typedef enum {
 #define TCG_TARGET_HAS_qemu_st8_i32     1
 #endif
 
-#define TCG_TARGET_HAS_qemu_ldst_i128   0
+#define TCG_TARGET_HAS_qemu_ldst_i128 \
+    (TCG_TARGET_REG_BITS == 64 && (cpuinfo & CPUINFO_ATOMIC_VMOVDQA))
 
 /* We do not support older SSE systems, only beginning with AVX1.  */
 #define TCG_TARGET_HAS_v64              have_avx1