[PATCH v4 00/12] tcg/mips: Unaligned access and other cleanup

[PATCH v4 00/12] tcg/mips: Unaligned access and other cleanup

[Richard Henderson]

Based-on: <20220104021543.396571-1-richard.henderson@linaro.org>
("[PATCH v4 0/7] Unaligned access for user only")

Changes from v3:
  * Rebase on master, which has some patches applied.


r~


Richard Henderson (12):
  tcg/mips: Support unaligned access for user-only
  tcg/mips: Support unaligned access for softmmu
  tcg/mips: Move TCG_AREG0 to S8
  tcg/mips: Move TCG_GUEST_BASE_REG to S7
  tcg/mips: Unify TCG_GUEST_BASE_REG tests
  tcg/mips: Create and use TCG_REG_TB
  tcg/mips: Split out tcg_out_movi_one
  tcg/mips: Split out tcg_out_movi_two
  tcg/mips: Use the constant pool for 64-bit constants
  tcg/mips: Aggressively use the constant pool for n64 calls
  tcg/mips: Try tb-relative addresses in tcg_out_movi
  tcg/mips: Try three insns with shift and add in tcg_out_movi

 tcg/mips/tcg-target.h     |   5 +-
 tcg/mips/tcg-target.c.inc | 647 +++++++++++++++++++++++++++++++++-----
 2 files changed, 579 insertions(+), 73 deletions(-)

-- 
2.25.1

[PATCH v4 01/12] tcg/mips: Support unaligned access for user-only

[Richard Henderson]

This is kinda sorta the opposite of the other tcg hosts, where
we get (normal) alignment checks for free with host SIGBUS and
need to add code to support unaligned accesses.

Fortunately, the ISA contains pairs of instructions that are
used to implement unaligned memory accesses.  Use them.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/mips/tcg-target.h     |   2 -
 tcg/mips/tcg-target.c.inc | 334 +++++++++++++++++++++++++++++++++++++-
 2 files changed, 328 insertions(+), 8 deletions(-)

diff --git a/tcg/mips/tcg-target.h b/tcg/mips/tcg-target.h
index c366fdf74b..7669213175 100644
--- a/tcg/mips/tcg-target.h
+++ b/tcg/mips/tcg-target.h
@@ -207,8 +207,6 @@ extern bool use_mips32r2_instructions;
 void tb_target_set_jmp_target(uintptr_t, uintptr_t, uintptr_t, uintptr_t)
     QEMU_ERROR("code path is reachable");
 
-#ifdef CONFIG_SOFTMMU
 #define TCG_TARGET_NEED_LDST_LABELS
-#endif
 
 #endif
diff --git a/tcg/mips/tcg-target.c.inc b/tcg/mips/tcg-target.c.inc
index d8f6914f03..5737d8a269 100644
--- a/tcg/mips/tcg-target.c.inc
+++ b/tcg/mips/tcg-target.c.inc
@@ -24,6 +24,8 @@
  * THE SOFTWARE.
  */
 
+#include "../tcg-ldst.c.inc"
+
 #ifdef HOST_WORDS_BIGENDIAN
 # define MIPS_BE  1
 #else
@@ -230,16 +232,26 @@ typedef enum {
     OPC_ORI      = 015 << 26,
     OPC_XORI     = 016 << 26,
     OPC_LUI      = 017 << 26,
+    OPC_BNEL     = 025 << 26,
+    OPC_BNEZALC_R6 = 030 << 26,
     OPC_DADDIU   = 031 << 26,
+    OPC_LDL      = 032 << 26,
+    OPC_LDR      = 033 << 26,
     OPC_LB       = 040 << 26,
     OPC_LH       = 041 << 26,
+    OPC_LWL      = 042 << 26,
     OPC_LW       = 043 << 26,
     OPC_LBU      = 044 << 26,
     OPC_LHU      = 045 << 26,
+    OPC_LWR      = 046 << 26,
     OPC_LWU      = 047 << 26,
     OPC_SB       = 050 << 26,
     OPC_SH       = 051 << 26,
+    OPC_SWL      = 052 << 26,
     OPC_SW       = 053 << 26,
+    OPC_SDL      = 054 << 26,
+    OPC_SDR      = 055 << 26,
+    OPC_SWR      = 056 << 26,
     OPC_LD       = 067 << 26,
     OPC_SD       = 077 << 26,
 
@@ -1015,8 +1027,6 @@ static void tcg_out_call(TCGContext *s, const tcg_insn_unit *arg)
 }
 
 #if defined(CONFIG_SOFTMMU)
-#include "../tcg-ldst.c.inc"
-
 static void * const qemu_ld_helpers[(MO_SSIZE | MO_BSWAP) + 1] = {
     [MO_UB]   = helper_ret_ldub_mmu,
     [MO_SB]   = helper_ret_ldsb_mmu,
@@ -1324,7 +1334,82 @@ static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
     tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], TCG_AREG0);
     return true;
 }
-#endif
+
+#else
+
+static void tcg_out_test_alignment(TCGContext *s, bool is_ld, TCGReg addrlo,
+                                   TCGReg addrhi, unsigned a_bits)
+{
+    unsigned a_mask = (1 << a_bits) - 1;
+    TCGLabelQemuLdst *l = new_ldst_label(s);
+
+    l->is_ld = is_ld;
+    l->addrlo_reg = addrlo;
+    l->addrhi_reg = addrhi;
+
+    /* We are expecting a_bits to max out at 7, much lower than ANDI. */
+    tcg_debug_assert(a_bits < 16);
+    tcg_out_opc_imm(s, OPC_ANDI, TCG_TMP0, addrlo, a_mask);
+
+    l->label_ptr[0] = s->code_ptr;
+    if (use_mips32r6_instructions) {
+        tcg_out_opc_br(s, OPC_BNEZALC_R6, TCG_REG_ZERO, TCG_TMP0);
+    } else {
+        tcg_out_opc_br(s, OPC_BNEL, TCG_TMP0, TCG_REG_ZERO);
+        tcg_out_nop(s);
+    }
+
+    l->raddr = tcg_splitwx_to_rx(s->code_ptr);
+}
+
+static bool tcg_out_fail_alignment(TCGContext *s, TCGLabelQemuLdst *l)
+{
+    void *target;
+
+    if (!reloc_pc16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
+        return false;
+    }
+
+    if (TCG_TARGET_REG_BITS < TARGET_LONG_BITS) {
+        /* A0 is env, A1 is skipped, A2:A3 is the uint64_t address. */
+        TCGReg a2 = MIPS_BE ? l->addrhi_reg : l->addrlo_reg;
+        TCGReg a3 = MIPS_BE ? l->addrlo_reg : l->addrhi_reg;
+
+        if (a3 != TCG_REG_A2) {
+            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_A2, a2);
+            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_A3, a3);
+        } else if (a2 != TCG_REG_A3) {
+            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_A3, a3);
+            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_A2, a2);
+        } else {
+            tcg_out_mov(s, TCG_TYPE_I32, TCG_TMP0, TCG_REG_A2);
+            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_A2, TCG_REG_A3);
+            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_A3, TCG_TMP0);
+        }
+    } else {
+        tcg_out_mov(s, TCG_TYPE_TL, TCG_REG_A1, l->addrlo_reg);
+    }
+    tcg_out_mov(s, TCG_TYPE_PTR, TCG_REG_A0, TCG_AREG0);
+
+    /*
+     * Tail call to the helper, with the return address back inline.
+     * We have arrived here via BNEL, so $31 is already set.
+     */
+    target = (l->is_ld ? helper_unaligned_ld : helper_unaligned_st);
+    tcg_out_call_int(s, target, true);
+    return true;
+}
+
+static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
+{
+    return tcg_out_fail_alignment(s, l);
+}
+
+static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
+{
+    return tcg_out_fail_alignment(s, l);
+}
+#endif /* SOFTMMU */
 
 static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg lo, TCGReg hi,
                                    TCGReg base, MemOp opc, bool is_64)
@@ -1430,6 +1515,127 @@ static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg lo, TCGReg hi,
     }
 }
 
+static void __attribute__((unused))
+tcg_out_qemu_ld_unalign(TCGContext *s, TCGReg lo, TCGReg hi,
+                                    TCGReg base, MemOp opc, bool is_64)
+{
+    const MIPSInsn lw1 = MIPS_BE ? OPC_LWL : OPC_LWR;
+    const MIPSInsn lw2 = MIPS_BE ? OPC_LWR : OPC_LWL;
+    const MIPSInsn ld1 = MIPS_BE ? OPC_LDL : OPC_LDR;
+    const MIPSInsn ld2 = MIPS_BE ? OPC_LDR : OPC_LDL;
+
+    bool sgn = (opc & MO_SIGN);
+
+    switch (opc & (MO_SSIZE | MO_BSWAP)) {
+    case MO_SW | MO_BE:
+    case MO_UW | MO_BE:
+        tcg_out_opc_imm(s, sgn ? OPC_LB : OPC_LBU, TCG_TMP0, base, 0);
+        tcg_out_opc_imm(s, OPC_LBU, lo, base, 1);
+        if (use_mips32r2_instructions) {
+            tcg_out_opc_bf(s, OPC_INS, lo, TCG_TMP0, 31, 8);
+        } else {
+            tcg_out_opc_sa(s, OPC_SLL, TCG_TMP0, TCG_TMP0, 8);
+            tcg_out_opc_reg(s, OPC_OR, lo, TCG_TMP0, TCG_TMP1);
+        }
+        break;
+
+    case MO_SW | MO_LE:
+    case MO_UW | MO_LE:
+        if (use_mips32r2_instructions && lo != base) {
+            tcg_out_opc_imm(s, OPC_LBU, lo, base, 0);
+            tcg_out_opc_imm(s, sgn ? OPC_LB : OPC_LBU, TCG_TMP0, base, 1);
+            tcg_out_opc_bf(s, OPC_INS, lo, TCG_TMP0, 31, 8);
+        } else {
+            tcg_out_opc_imm(s, OPC_LBU, TCG_TMP0, base, 0);
+            tcg_out_opc_imm(s, sgn ? OPC_LB : OPC_LBU, TCG_TMP1, base, 1);
+            tcg_out_opc_sa(s, OPC_SLL, TCG_TMP1, TCG_TMP1, 8);
+            tcg_out_opc_reg(s, OPC_OR, lo, TCG_TMP0, TCG_TMP1);
+        }
+        break;
+
+    case MO_SL:
+    case MO_UL:
+        tcg_out_opc_imm(s, lw1, lo, base, 0);
+        tcg_out_opc_imm(s, lw2, lo, base, 3);
+        if (TCG_TARGET_REG_BITS == 64 && is_64 && !sgn) {
+            tcg_out_ext32u(s, lo, lo);
+        }
+        break;
+
+    case MO_UL | MO_BSWAP:
+    case MO_SL | MO_BSWAP:
+        if (use_mips32r2_instructions) {
+            tcg_out_opc_imm(s, lw1, lo, base, 0);
+            tcg_out_opc_imm(s, lw2, lo, base, 3);
+            tcg_out_bswap32(s, lo, lo,
+                            TCG_TARGET_REG_BITS == 64 && is_64
+                            ? (sgn ? TCG_BSWAP_OS : TCG_BSWAP_OZ) : 0);
+        } else {
+            const tcg_insn_unit *subr =
+                (TCG_TARGET_REG_BITS == 64 && is_64 && !sgn
+                 ? bswap32u_addr : bswap32_addr);
+
+            tcg_out_opc_imm(s, lw1, TCG_TMP0, base, 0);
+            tcg_out_bswap_subr(s, subr);
+            /* delay slot */
+            tcg_out_opc_imm(s, lw2, TCG_TMP0, base, 3);
+            tcg_out_mov(s, is_64 ? TCG_TYPE_I64 : TCG_TYPE_I32, lo, TCG_TMP3);
+        }
+        break;
+
+    case MO_Q:
+        if (TCG_TARGET_REG_BITS == 64) {
+            tcg_out_opc_imm(s, ld1, lo, base, 0);
+            tcg_out_opc_imm(s, ld2, lo, base, 7);
+        } else {
+            tcg_out_opc_imm(s, lw1, MIPS_BE ? hi : lo, base, 0 + 0);
+            tcg_out_opc_imm(s, lw2, MIPS_BE ? hi : lo, base, 0 + 3);
+            tcg_out_opc_imm(s, lw1, MIPS_BE ? lo : hi, base, 4 + 0);
+            tcg_out_opc_imm(s, lw2, MIPS_BE ? lo : hi, base, 4 + 3);
+        }
+        break;
+
+    case MO_Q | MO_BSWAP:
+        if (TCG_TARGET_REG_BITS == 64) {
+            if (use_mips32r2_instructions) {
+                tcg_out_opc_imm(s, ld1, lo, base, 0);
+                tcg_out_opc_imm(s, ld2, lo, base, 7);
+                tcg_out_bswap64(s, lo, lo);
+            } else {
+                tcg_out_opc_imm(s, ld1, TCG_TMP0, base, 0);
+                tcg_out_bswap_subr(s, bswap64_addr);
+                /* delay slot */
+                tcg_out_opc_imm(s, ld2, TCG_TMP0, base, 7);
+                tcg_out_mov(s, TCG_TYPE_I64, lo, TCG_TMP3);
+            }
+        } else if (use_mips32r2_instructions) {
+            tcg_out_opc_imm(s, lw1, TCG_TMP0, base, 0 + 0);
+            tcg_out_opc_imm(s, lw2, TCG_TMP0, base, 0 + 3);
+            tcg_out_opc_imm(s, lw1, TCG_TMP1, base, 4 + 0);
+            tcg_out_opc_imm(s, lw2, TCG_TMP1, base, 4 + 3);
+            tcg_out_opc_reg(s, OPC_WSBH, TCG_TMP0, 0, TCG_TMP0);
+            tcg_out_opc_reg(s, OPC_WSBH, TCG_TMP1, 0, TCG_TMP1);
+            tcg_out_opc_sa(s, OPC_ROTR, MIPS_BE ? lo : hi, TCG_TMP0, 16);
+            tcg_out_opc_sa(s, OPC_ROTR, MIPS_BE ? hi : lo, TCG_TMP1, 16);
+        } else {
+            tcg_out_opc_imm(s, lw1, TCG_TMP0, base, 0 + 0);
+            tcg_out_bswap_subr(s, bswap32_addr);
+            /* delay slot */
+            tcg_out_opc_imm(s, lw2, TCG_TMP0, base, 0 + 3);
+            tcg_out_opc_imm(s, lw1, TCG_TMP0, base, 4 + 0);
+            tcg_out_mov(s, TCG_TYPE_I32, MIPS_BE ? lo : hi, TCG_TMP3);
+            tcg_out_bswap_subr(s, bswap32_addr);
+            /* delay slot */
+            tcg_out_opc_imm(s, lw2, TCG_TMP0, base, 4 + 3);
+            tcg_out_mov(s, TCG_TYPE_I32, MIPS_BE ? hi : lo, TCG_TMP3);
+        }
+        break;
+
+    default:
+        g_assert_not_reached();
+    }
+}
+
 static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, bool is_64)
 {
     TCGReg addr_regl, addr_regh __attribute__((unused));
@@ -1438,6 +1644,8 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, bool is_64)
     MemOp opc;
 #if defined(CONFIG_SOFTMMU)
     tcg_insn_unit *label_ptr[2];
+#else
+    unsigned a_bits, s_bits;
 #endif
     TCGReg base = TCG_REG_A0;
 
@@ -1467,7 +1675,27 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, bool is_64)
     } else {
         tcg_out_opc_reg(s, ALIAS_PADD, base, TCG_GUEST_BASE_REG, addr_regl);
     }
-    tcg_out_qemu_ld_direct(s, data_regl, data_regh, base, opc, is_64);
+    a_bits = get_alignment_bits(opc);
+    s_bits = opc & MO_SIZE;
+    /*
+     * R6 removes the left/right instructions but requires the
+     * system to support misaligned memory accesses.
+     */
+    if (use_mips32r6_instructions) {
+        if (a_bits) {
+            tcg_out_test_alignment(s, true, addr_regl, addr_regh, a_bits);
+        }
+        tcg_out_qemu_ld_direct(s, data_regl, data_regh, base, opc, is_64);
+    } else {
+        if (a_bits && a_bits != s_bits) {
+            tcg_out_test_alignment(s, true, addr_regl, addr_regh, a_bits);
+        }
+        if (a_bits >= s_bits) {
+            tcg_out_qemu_ld_direct(s, data_regl, data_regh, base, opc, is_64);
+        } else {
+            tcg_out_qemu_ld_unalign(s, data_regl, data_regh, base, opc, is_64);
+        }
+    }
 #endif
 }
 
@@ -1532,6 +1760,79 @@ static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg lo, TCGReg hi,
     }
 }
 
+static void __attribute__((unused))
+tcg_out_qemu_st_unalign(TCGContext *s, TCGReg lo, TCGReg hi,
+                                    TCGReg base, MemOp opc)
+{
+    const MIPSInsn sw1 = MIPS_BE ? OPC_SWL : OPC_SWR;
+    const MIPSInsn sw2 = MIPS_BE ? OPC_SWR : OPC_SWL;
+    const MIPSInsn sd1 = MIPS_BE ? OPC_SDL : OPC_SDR;
+    const MIPSInsn sd2 = MIPS_BE ? OPC_SDR : OPC_SDL;
+
+    /* Don't clutter the code below with checks to avoid bswapping ZERO.  */
+    if ((lo | hi) == 0) {
+        opc &= ~MO_BSWAP;
+    }
+
+    switch (opc & (MO_SIZE | MO_BSWAP)) {
+    case MO_16 | MO_BE:
+        tcg_out_opc_sa(s, OPC_SRL, TCG_TMP0, lo, 8);
+        tcg_out_opc_imm(s, OPC_SB, TCG_TMP0, base, 0);
+        tcg_out_opc_imm(s, OPC_SB, lo, base, 1);
+        break;
+
+    case MO_16 | MO_LE:
+        tcg_out_opc_sa(s, OPC_SRL, TCG_TMP0, lo, 8);
+        tcg_out_opc_imm(s, OPC_SB, lo, base, 0);
+        tcg_out_opc_imm(s, OPC_SB, TCG_TMP0, base, 1);
+        break;
+
+    case MO_32 | MO_BSWAP:
+        tcg_out_bswap32(s, TCG_TMP3, lo, 0);
+        lo = TCG_TMP3;
+        /* fall through */
+    case MO_32:
+        tcg_out_opc_imm(s, sw1, lo, base, 0);
+        tcg_out_opc_imm(s, sw2, lo, base, 3);
+        break;
+
+    case MO_64 | MO_BSWAP:
+        if (TCG_TARGET_REG_BITS == 64) {
+            tcg_out_bswap64(s, TCG_TMP3, lo);
+            lo = TCG_TMP3;
+        } else if (use_mips32r2_instructions) {
+            tcg_out_opc_reg(s, OPC_WSBH, TCG_TMP0, 0, MIPS_BE ? hi : lo);
+            tcg_out_opc_reg(s, OPC_WSBH, TCG_TMP1, 0, MIPS_BE ? lo : hi);
+            tcg_out_opc_sa(s, OPC_ROTR, TCG_TMP0, TCG_TMP0, 16);
+            tcg_out_opc_sa(s, OPC_ROTR, TCG_TMP1, TCG_TMP1, 16);
+            hi = MIPS_BE ? TCG_TMP0 : TCG_TMP1;
+            lo = MIPS_BE ? TCG_TMP1 : TCG_TMP0;
+        } else {
+            tcg_out_bswap32(s, TCG_TMP3, MIPS_BE ? lo : hi, 0);
+            tcg_out_opc_imm(s, sw1, TCG_TMP3, base, 0);
+            tcg_out_opc_imm(s, sw2, TCG_TMP3, base, 3);
+            tcg_out_bswap32(s, TCG_TMP3, MIPS_BE ? hi : lo, 0);
+            tcg_out_opc_imm(s, sw1, TCG_TMP3, base, 4);
+            tcg_out_opc_imm(s, sw2, TCG_TMP3, base, 7);
+            break;
+        }
+        /* fall through */
+    case MO_64:
+        if (TCG_TARGET_REG_BITS == 64) {
+            tcg_out_opc_imm(s, sd1, lo, base, 0);
+            tcg_out_opc_imm(s, sd2, lo, base, 7);
+        } else {
+            tcg_out_opc_imm(s, sw1, MIPS_BE ? hi : lo, base, 0);
+            tcg_out_opc_imm(s, sw2, MIPS_BE ? hi : lo, base, 3);
+            tcg_out_opc_imm(s, sw1, MIPS_BE ? lo : hi, base, 4);
+            tcg_out_opc_imm(s, sw2, MIPS_BE ? lo : hi, base, 7);
+        }
+        break;
+
+    default:
+        tcg_abort();
+    }
+}
 static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is_64)
 {
     TCGReg addr_regl, addr_regh __attribute__((unused));
@@ -1540,6 +1841,8 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is_64)
     MemOp opc;
 #if defined(CONFIG_SOFTMMU)
     tcg_insn_unit *label_ptr[2];
+#else
+    unsigned a_bits, s_bits;
 #endif
     TCGReg base = TCG_REG_A0;
 
@@ -1558,7 +1861,6 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is_64)
                         data_regl, data_regh, addr_regl, addr_regh,
                         s->code_ptr, label_ptr);
 #else
-    base = TCG_REG_A0;
     if (TCG_TARGET_REG_BITS > TARGET_LONG_BITS) {
         tcg_out_ext32u(s, base, addr_regl);
         addr_regl = base;
@@ -1570,7 +1872,27 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is_64)
     } else {
         tcg_out_opc_reg(s, ALIAS_PADD, base, TCG_GUEST_BASE_REG, addr_regl);
     }
-    tcg_out_qemu_st_direct(s, data_regl, data_regh, base, opc);
+    a_bits = get_alignment_bits(opc);
+    s_bits = opc & MO_SIZE;
+    /*
+     * R6 removes the left/right instructions but requires the
+     * system to support misaligned memory accesses.
+     */
+    if (use_mips32r6_instructions) {
+        if (a_bits) {
+            tcg_out_test_alignment(s, true, addr_regl, addr_regh, a_bits);
+        }
+        tcg_out_qemu_st_direct(s, data_regl, data_regh, base, opc);
+    } else {
+        if (a_bits && a_bits != s_bits) {
+            tcg_out_test_alignment(s, true, addr_regl, addr_regh, a_bits);
+        }
+        if (a_bits >= s_bits) {
+            tcg_out_qemu_st_direct(s, data_regl, data_regh, base, opc);
+        } else {
+            tcg_out_qemu_st_unalign(s, data_regl, data_regh, base, opc);
+        }
+    }
 #endif
 }
 
-- 
2.25.1

[PATCH v4 02/12] tcg/mips: Support unaligned access for softmmu

[Richard Henderson]

We can use the routines just added for user-only to emit
unaligned accesses in softmmu mode too.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/mips/tcg-target.c.inc | 91 ++++++++++++++++++++++-----------------
 1 file changed, 51 insertions(+), 40 deletions(-)

diff --git a/tcg/mips/tcg-target.c.inc b/tcg/mips/tcg-target.c.inc
index 5737d8a269..7682059d92 100644
--- a/tcg/mips/tcg-target.c.inc
+++ b/tcg/mips/tcg-target.c.inc
@@ -1134,8 +1134,10 @@ static void tcg_out_tlb_load(TCGContext *s, TCGReg base, TCGReg addrl,
                              tcg_insn_unit *label_ptr[2], bool is_load)
 {
     MemOp opc = get_memop(oi);
-    unsigned s_bits = opc & MO_SIZE;
     unsigned a_bits = get_alignment_bits(opc);
+    unsigned s_bits = opc & MO_SIZE;
+    unsigned a_mask = (1 << a_bits) - 1;
+    unsigned s_mask = (1 << s_bits) - 1;
     int mem_index = get_mmuidx(oi);
     int fast_off = TLB_MASK_TABLE_OFS(mem_index);
     int mask_off = fast_off + offsetof(CPUTLBDescFast, mask);
@@ -1143,7 +1145,7 @@ static void tcg_out_tlb_load(TCGContext *s, TCGReg base, TCGReg addrl,
     int add_off = offsetof(CPUTLBEntry, addend);
     int cmp_off = (is_load ? offsetof(CPUTLBEntry, addr_read)
                    : offsetof(CPUTLBEntry, addr_write));
-    target_ulong mask;
+    target_ulong tlb_mask;
 
     /* Load tlb_mask[mmu_idx] and tlb_table[mmu_idx].  */
     tcg_out_ld(s, TCG_TYPE_PTR, TCG_TMP0, TCG_AREG0, mask_off);
@@ -1157,27 +1159,13 @@ static void tcg_out_tlb_load(TCGContext *s, TCGReg base, TCGReg addrl,
     /* Add the tlb_table pointer, creating the CPUTLBEntry address in TMP3.  */
     tcg_out_opc_reg(s, ALIAS_PADD, TCG_TMP3, TCG_TMP3, TCG_TMP1);
 
-    /* We don't currently support unaligned accesses.
-       We could do so with mips32r6.  */
-    if (a_bits < s_bits) {
-        a_bits = s_bits;
-    }
-
-    /* Mask the page bits, keeping the alignment bits to compare against.  */
-    mask = (target_ulong)TARGET_PAGE_MASK | ((1 << a_bits) - 1);
-
     /* Load the (low-half) tlb comparator.  */
     if (TCG_TARGET_REG_BITS < TARGET_LONG_BITS) {
-        tcg_out_ld(s, TCG_TYPE_I32, TCG_TMP0, TCG_TMP3, cmp_off + LO_OFF);
-        tcg_out_movi(s, TCG_TYPE_I32, TCG_TMP1, mask);
+        tcg_out_ldst(s, OPC_LW, TCG_TMP0, TCG_TMP3, cmp_off + LO_OFF);
     } else {
         tcg_out_ldst(s, (TARGET_LONG_BITS == 64 ? OPC_LD
                          : TCG_TARGET_REG_BITS == 64 ? OPC_LWU : OPC_LW),
                      TCG_TMP0, TCG_TMP3, cmp_off);
-        tcg_out_movi(s, TCG_TYPE_TL, TCG_TMP1, mask);
-        /* No second compare is required here;
-           load the tlb addend for the fast path.  */
-        tcg_out_ld(s, TCG_TYPE_PTR, TCG_TMP2, TCG_TMP3, add_off);
     }
 
     /* Zero extend a 32-bit guest address for a 64-bit host. */
@@ -1185,7 +1173,25 @@ static void tcg_out_tlb_load(TCGContext *s, TCGReg base, TCGReg addrl,
         tcg_out_ext32u(s, base, addrl);
         addrl = base;
     }
-    tcg_out_opc_reg(s, OPC_AND, TCG_TMP1, TCG_TMP1, addrl);
+
+    /*
+     * Mask the page bits, keeping the alignment bits to compare against.
+     * For unaligned accesses, compare against the end of the access to
+     * verify that it does not cross a page boundary.
+     */
+    tlb_mask = (target_ulong)TARGET_PAGE_MASK | a_mask;
+    tcg_out_movi(s, TCG_TYPE_I32, TCG_TMP1, tlb_mask);
+    if (a_mask >= s_mask) {
+        tcg_out_opc_reg(s, OPC_AND, TCG_TMP1, TCG_TMP1, addrl);
+    } else {
+        tcg_out_opc_imm(s, ALIAS_PADDI, TCG_TMP2, addrl, s_mask - a_mask);
+        tcg_out_opc_reg(s, OPC_AND, TCG_TMP1, TCG_TMP1, TCG_TMP2);
+    }
+
+    if (TCG_TARGET_REG_BITS >= TARGET_LONG_BITS) {
+        /* Load the tlb addend for the fast path.  */
+        tcg_out_ld(s, TCG_TYPE_PTR, TCG_TMP2, TCG_TMP3, add_off);
+    }
 
     label_ptr[0] = s->code_ptr;
     tcg_out_opc_br(s, OPC_BNE, TCG_TMP1, TCG_TMP0);
@@ -1193,7 +1199,7 @@ static void tcg_out_tlb_load(TCGContext *s, TCGReg base, TCGReg addrl,
     /* Load and test the high half tlb comparator.  */
     if (TCG_TARGET_REG_BITS < TARGET_LONG_BITS) {
         /* delay slot */
-        tcg_out_ld(s, TCG_TYPE_I32, TCG_TMP0, TCG_TMP3, cmp_off + HI_OFF);
+        tcg_out_ldst(s, OPC_LW, TCG_TMP0, TCG_TMP3, cmp_off + HI_OFF);
 
         /* Load the tlb addend for the fast path.  */
         tcg_out_ld(s, TCG_TYPE_PTR, TCG_TMP2, TCG_TMP3, add_off);
@@ -1515,8 +1521,7 @@ static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg lo, TCGReg hi,
     }
 }
 
-static void __attribute__((unused))
-tcg_out_qemu_ld_unalign(TCGContext *s, TCGReg lo, TCGReg hi,
+static void tcg_out_qemu_ld_unalign(TCGContext *s, TCGReg lo, TCGReg hi,
                                     TCGReg base, MemOp opc, bool is_64)
 {
     const MIPSInsn lw1 = MIPS_BE ? OPC_LWL : OPC_LWR;
@@ -1645,8 +1650,8 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, bool is_64)
 #if defined(CONFIG_SOFTMMU)
     tcg_insn_unit *label_ptr[2];
 #else
-    unsigned a_bits, s_bits;
 #endif
+    unsigned a_bits, s_bits;
     TCGReg base = TCG_REG_A0;
 
     data_regl = *args++;
@@ -1655,10 +1660,20 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, bool is_64)
     addr_regh = (TCG_TARGET_REG_BITS < TARGET_LONG_BITS ? *args++ : 0);
     oi = *args++;
     opc = get_memop(oi);
+    a_bits = get_alignment_bits(opc);
+    s_bits = opc & MO_SIZE;
 
+    /*
+     * R6 removes the left/right instructions but requires the
+     * system to support misaligned memory accesses.
+     */
 #if defined(CONFIG_SOFTMMU)
     tcg_out_tlb_load(s, base, addr_regl, addr_regh, oi, label_ptr, 1);
-    tcg_out_qemu_ld_direct(s, data_regl, data_regh, base, opc, is_64);
+    if (use_mips32r6_instructions || a_bits >= s_bits) {
+        tcg_out_qemu_ld_direct(s, data_regl, data_regh, base, opc, is_64);
+    } else {
+        tcg_out_qemu_ld_unalign(s, data_regl, data_regh, base, opc, is_64);
+    }
     add_qemu_ldst_label(s, 1, oi,
                         (is_64 ? TCG_TYPE_I64 : TCG_TYPE_I32),
                         data_regl, data_regh, addr_regl, addr_regh,
@@ -1675,12 +1690,6 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, bool is_64)
     } else {
         tcg_out_opc_reg(s, ALIAS_PADD, base, TCG_GUEST_BASE_REG, addr_regl);
     }
-    a_bits = get_alignment_bits(opc);
-    s_bits = opc & MO_SIZE;
-    /*
-     * R6 removes the left/right instructions but requires the
-     * system to support misaligned memory accesses.
-     */
     if (use_mips32r6_instructions) {
         if (a_bits) {
             tcg_out_test_alignment(s, true, addr_regl, addr_regh, a_bits);
@@ -1760,8 +1769,7 @@ static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg lo, TCGReg hi,
     }
 }
 
-static void __attribute__((unused))
-tcg_out_qemu_st_unalign(TCGContext *s, TCGReg lo, TCGReg hi,
+static void tcg_out_qemu_st_unalign(TCGContext *s, TCGReg lo, TCGReg hi,
                                     TCGReg base, MemOp opc)
 {
     const MIPSInsn sw1 = MIPS_BE ? OPC_SWL : OPC_SWR;
@@ -1841,9 +1849,8 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is_64)
     MemOp opc;
 #if defined(CONFIG_SOFTMMU)
     tcg_insn_unit *label_ptr[2];
-#else
-    unsigned a_bits, s_bits;
 #endif
+    unsigned a_bits, s_bits;
     TCGReg base = TCG_REG_A0;
 
     data_regl = *args++;
@@ -1852,10 +1859,20 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is_64)
     addr_regh = (TCG_TARGET_REG_BITS < TARGET_LONG_BITS ? *args++ : 0);
     oi = *args++;
     opc = get_memop(oi);
+    a_bits = get_alignment_bits(opc);
+    s_bits = opc & MO_SIZE;
 
+    /*
+     * R6 removes the left/right instructions but requires the
+     * system to support misaligned memory accesses.
+     */
 #if defined(CONFIG_SOFTMMU)
     tcg_out_tlb_load(s, base, addr_regl, addr_regh, oi, label_ptr, 0);
-    tcg_out_qemu_st_direct(s, data_regl, data_regh, base, opc);
+    if (use_mips32r6_instructions || a_bits >= s_bits) {
+        tcg_out_qemu_st_direct(s, data_regl, data_regh, base, opc);
+    } else {
+        tcg_out_qemu_st_unalign(s, data_regl, data_regh, base, opc);
+    }
     add_qemu_ldst_label(s, 0, oi,
                         (is_64 ? TCG_TYPE_I64 : TCG_TYPE_I32),
                         data_regl, data_regh, addr_regl, addr_regh,
@@ -1872,12 +1889,6 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is_64)
     } else {
         tcg_out_opc_reg(s, ALIAS_PADD, base, TCG_GUEST_BASE_REG, addr_regl);
     }
-    a_bits = get_alignment_bits(opc);
-    s_bits = opc & MO_SIZE;
-    /*
-     * R6 removes the left/right instructions but requires the
-     * system to support misaligned memory accesses.
-     */
     if (use_mips32r6_instructions) {
         if (a_bits) {
             tcg_out_test_alignment(s, true, addr_regl, addr_regh, a_bits);
-- 
2.25.1

[PATCH v4 03/12] tcg/mips: Move TCG_AREG0 to S8

[Richard Henderson]

No functional change; just moving the saved reserved regs to the end.

Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/mips/tcg-target.h     | 2 +-
 tcg/mips/tcg-target.c.inc | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/tcg/mips/tcg-target.h b/tcg/mips/tcg-target.h
index 7669213175..28c42e23e1 100644
--- a/tcg/mips/tcg-target.h
+++ b/tcg/mips/tcg-target.h
@@ -76,7 +76,7 @@ typedef enum {
     TCG_REG_RA,
 
     TCG_REG_CALL_STACK = TCG_REG_SP,
-    TCG_AREG0 = TCG_REG_S0,
+    TCG_AREG0 = TCG_REG_S8,
 } TCGReg;
 
 /* used for function call generation */
diff --git a/tcg/mips/tcg-target.c.inc b/tcg/mips/tcg-target.c.inc
index 7682059d92..5702a6ad92 100644
--- a/tcg/mips/tcg-target.c.inc
+++ b/tcg/mips/tcg-target.c.inc
@@ -2548,7 +2548,7 @@ static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
 }
 
 static const int tcg_target_callee_save_regs[] = {
-    TCG_REG_S0,       /* used for the global env (TCG_AREG0) */
+    TCG_REG_S0,
     TCG_REG_S1,
     TCG_REG_S2,
     TCG_REG_S3,
@@ -2556,7 +2556,7 @@ static const int tcg_target_callee_save_regs[] = {
     TCG_REG_S5,
     TCG_REG_S6,
     TCG_REG_S7,
-    TCG_REG_S8,
+    TCG_REG_S8,       /* used for the global env (TCG_AREG0) */
     TCG_REG_RA,       /* should be last for ABI compliance */
 };
 
-- 
2.25.1

[PATCH v4 04/12] tcg/mips: Move TCG_GUEST_BASE_REG to S7

[Richard Henderson]

No functional change; just moving the saved reserved regs to the end.

Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/mips/tcg-target.c.inc | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/tcg/mips/tcg-target.c.inc b/tcg/mips/tcg-target.c.inc
index 5702a6ad92..1bfe6aea0e 100644
--- a/tcg/mips/tcg-target.c.inc
+++ b/tcg/mips/tcg-target.c.inc
@@ -86,7 +86,7 @@ static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
 #define TCG_TMP3  TCG_REG_T7
 
 #ifndef CONFIG_SOFTMMU
-#define TCG_GUEST_BASE_REG TCG_REG_S1
+#define TCG_GUEST_BASE_REG TCG_REG_S7
 #endif
 
 /* check if we really need so many registers :P */
@@ -2555,7 +2555,7 @@ static const int tcg_target_callee_save_regs[] = {
     TCG_REG_S4,
     TCG_REG_S5,
     TCG_REG_S6,
-    TCG_REG_S7,
+    TCG_REG_S7,       /* used for guest_base */
     TCG_REG_S8,       /* used for the global env (TCG_AREG0) */
     TCG_REG_RA,       /* should be last for ABI compliance */
 };
-- 
2.25.1

[PATCH v4 05/12] tcg/mips: Unify TCG_GUEST_BASE_REG tests

[Richard Henderson]

In tcg_out_qemu_ld/st, we already check for guest_base matching int16_t.
Mirror that when setting up TCG_GUEST_BASE_REG in the prologue.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/mips/tcg-target.c.inc | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tcg/mips/tcg-target.c.inc b/tcg/mips/tcg-target.c.inc
index 1bfe6aea0e..46616784f8 100644
--- a/tcg/mips/tcg-target.c.inc
+++ b/tcg/mips/tcg-target.c.inc
@@ -2677,7 +2677,7 @@ static void tcg_target_qemu_prologue(TCGContext *s)
     }
 
 #ifndef CONFIG_SOFTMMU
-    if (guest_base) {
+    if (guest_base != (int16_t)guest_base) {
         tcg_out_movi(s, TCG_TYPE_PTR, TCG_GUEST_BASE_REG, guest_base);
         tcg_regset_set_reg(s->reserved_regs, TCG_GUEST_BASE_REG);
     }
-- 
2.25.1

[PATCH v4 06/12] tcg/mips: Create and use TCG_REG_TB

[Richard Henderson]

This vastly reduces the size of code generated for 64-bit addresses.
The code for exit_tb, for instance, where we load a (tagged) pointer
to the current TB, goes from

0x400aa9725c:  li       v0,64
0x400aa97260:  dsll     v0,v0,0x10
0x400aa97264:  ori      v0,v0,0xaa9
0x400aa97268:  dsll     v0,v0,0x10
0x400aa9726c:  j        0x400aa9703c
0x400aa97270:  ori      v0,v0,0x7083

to

0x400aa97240:  j        0x400aa97040
0x400aa97244:  daddiu   v0,s6,-189

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/mips/tcg-target.c.inc | 75 ++++++++++++++++++++++++++++++++-------
 1 file changed, 62 insertions(+), 13 deletions(-)

diff --git a/tcg/mips/tcg-target.c.inc b/tcg/mips/tcg-target.c.inc
index 46616784f8..76fb1dada0 100644
--- a/tcg/mips/tcg-target.c.inc
+++ b/tcg/mips/tcg-target.c.inc
@@ -88,6 +88,11 @@ static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
 #ifndef CONFIG_SOFTMMU
 #define TCG_GUEST_BASE_REG TCG_REG_S7
 #endif
+#if TCG_TARGET_REG_BITS == 64
+#define TCG_REG_TB         TCG_REG_S6
+#else
+#define TCG_REG_TB         (qemu_build_not_reached(), TCG_REG_ZERO)
+#endif
 
 /* check if we really need so many registers :P */
 static const int tcg_target_reg_alloc_order[] = {
@@ -1971,34 +1976,72 @@ static void tcg_out_op(TCGContext *s, TCGOpcode opc,
     switch (opc) {
     case INDEX_op_exit_tb:
         {
-            TCGReg b0 = TCG_REG_ZERO;
+            TCGReg base = TCG_REG_ZERO;
+            int16_t lo = 0;
 
-            a0 = (intptr_t)a0;
-            if (a0 & ~0xffff) {
-                tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_V0, a0 & ~0xffff);
-                b0 = TCG_REG_V0;
+            if (a0) {
+                intptr_t ofs;
+                if (TCG_TARGET_REG_BITS == 64) {
+                    ofs = tcg_tbrel_diff(s, (void *)a0);
+                    lo = ofs;
+                    if (ofs == lo) {
+                        base = TCG_REG_TB;
+                    } else {
+                        base = TCG_REG_V0;
+                        tcg_out_movi(s, TCG_TYPE_PTR, base, ofs - lo);
+                        tcg_out_opc_reg(s, ALIAS_PADD, base, base, TCG_REG_TB);
+                    }
+                } else {
+                    ofs = a0;
+                    lo = ofs;
+                    base = TCG_REG_V0;
+                    tcg_out_movi(s, TCG_TYPE_PTR, base, ofs - lo);
+                }
             }
             if (!tcg_out_opc_jmp(s, OPC_J, tb_ret_addr)) {
                 tcg_out_movi(s, TCG_TYPE_PTR, TCG_TMP0,
                              (uintptr_t)tb_ret_addr);
                 tcg_out_opc_reg(s, OPC_JR, 0, TCG_TMP0, 0);
             }
-            tcg_out_opc_imm(s, OPC_ORI, TCG_REG_V0, b0, a0 & 0xffff);
+            tcg_out_opc_imm(s, ALIAS_PADDI, TCG_REG_V0, base, lo);
         }
         break;
     case INDEX_op_goto_tb:
         /* indirect jump method */
         tcg_debug_assert(s->tb_jmp_insn_offset == 0);
-        tcg_out_ld(s, TCG_TYPE_PTR, TCG_TMP0, TCG_REG_ZERO,
-                   (uintptr_t)(s->tb_jmp_target_addr + a0));
-        tcg_out_opc_reg(s, OPC_JR, 0, TCG_TMP0, 0);
-        tcg_out_nop(s);
-        set_jmp_reset_offset(s, a0);
+        {
+            TCGReg base, dest;
+            intptr_t ofs;
+
+            if (TCG_TARGET_REG_BITS == 64) {
+                dest = base = TCG_REG_TB;
+                ofs = tcg_tbrel_diff(s, s->tb_jmp_target_addr + a0);
+            } else {
+                dest = TCG_TMP0;
+                base = TCG_REG_ZERO;
+                ofs = (intptr_t)(s->tb_jmp_target_addr + a0);
+            }
+            tcg_out_ld(s, TCG_TYPE_PTR, dest, base, ofs);
+            tcg_out_opc_reg(s, OPC_JR, 0, dest, 0);
+            /* delay slot */
+            tcg_out_nop(s);
+
+            set_jmp_reset_offset(s, args[0]);
+            if (TCG_TARGET_REG_BITS == 64) {
+                /* For the unlinked case, need to reset TCG_REG_TB. */
+                tcg_out_ldst(s, ALIAS_PADDI, TCG_REG_TB, TCG_REG_TB,
+                             -tcg_current_code_size(s));
+            }
+        }
         break;
     case INDEX_op_goto_ptr:
         /* jmp to the given host address (could be epilogue) */
         tcg_out_opc_reg(s, OPC_JR, 0, a0, 0);
-        tcg_out_nop(s);
+        if (TCG_TARGET_REG_BITS == 64) {
+            tcg_out_mov(s, TCG_TYPE_PTR, TCG_REG_TB, a0);
+        } else {
+            tcg_out_nop(s);
+        }
         break;
     case INDEX_op_br:
         tcg_out_brcond(s, TCG_COND_EQ, TCG_REG_ZERO, TCG_REG_ZERO,
@@ -2554,7 +2597,7 @@ static const int tcg_target_callee_save_regs[] = {
     TCG_REG_S3,
     TCG_REG_S4,
     TCG_REG_S5,
-    TCG_REG_S6,
+    TCG_REG_S6,       /* used for the tb base (TCG_REG_TB) */
     TCG_REG_S7,       /* used for guest_base */
     TCG_REG_S8,       /* used for the global env (TCG_AREG0) */
     TCG_REG_RA,       /* should be last for ABI compliance */
@@ -2682,6 +2725,9 @@ static void tcg_target_qemu_prologue(TCGContext *s)
         tcg_regset_set_reg(s->reserved_regs, TCG_GUEST_BASE_REG);
     }
 #endif
+    if (TCG_TARGET_REG_BITS == 64) {
+        tcg_out_mov(s, TCG_TYPE_PTR, TCG_REG_TB, tcg_target_call_iarg_regs[1]);
+    }
 
     /* Call generated code */
     tcg_out_opc_reg(s, OPC_JR, 0, tcg_target_call_iarg_regs[1], 0);
@@ -2863,6 +2909,9 @@ static void tcg_target_init(TCGContext *s)
     tcg_regset_set_reg(s->reserved_regs, TCG_REG_RA);   /* return address */
     tcg_regset_set_reg(s->reserved_regs, TCG_REG_SP);   /* stack pointer */
     tcg_regset_set_reg(s->reserved_regs, TCG_REG_GP);   /* global pointer */
+    if (TCG_TARGET_REG_BITS == 64) {
+        tcg_regset_set_reg(s->reserved_regs, TCG_REG_TB); /* tc->tc_ptr */
+    }
 }
 
 typedef struct {
-- 
2.25.1

[PATCH v4 07/12] tcg/mips: Split out tcg_out_movi_one

[Richard Henderson]

Emit all constants that can be loaded in exactly one insn.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/mips/tcg-target.c.inc | 26 ++++++++++++++++++++------
 1 file changed, 20 insertions(+), 6 deletions(-)

diff --git a/tcg/mips/tcg-target.c.inc b/tcg/mips/tcg-target.c.inc
index 76fb1dada0..8741fdd49c 100644
--- a/tcg/mips/tcg-target.c.inc
+++ b/tcg/mips/tcg-target.c.inc
@@ -524,20 +524,34 @@ static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
     return true;
 }
 
+static bool tcg_out_movi_one(TCGContext *s, TCGReg ret, tcg_target_long arg)
+{
+    if (arg == (int16_t)arg) {
+        tcg_out_opc_imm(s, OPC_ADDIU, ret, TCG_REG_ZERO, arg);
+        return true;
+    }
+    if (arg == (uint16_t)arg) {
+        tcg_out_opc_imm(s, OPC_ORI, ret, TCG_REG_ZERO, arg);
+        return true;
+    }
+    if (arg == (int32_t)arg && (arg & 0xffff) == 0) {
+        tcg_out_opc_imm(s, OPC_LUI, ret, TCG_REG_ZERO, arg >> 16);
+        return true;
+    }
+    return false;
+}
+
 static void tcg_out_movi(TCGContext *s, TCGType type,
                          TCGReg ret, tcg_target_long arg)
 {
     if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I32) {
         arg = (int32_t)arg;
     }
-    if (arg == (int16_t)arg) {
-        tcg_out_opc_imm(s, OPC_ADDIU, ret, TCG_REG_ZERO, arg);
-        return;
-    }
-    if (arg == (uint16_t)arg) {
-        tcg_out_opc_imm(s, OPC_ORI, ret, TCG_REG_ZERO, arg);
+
+    if (tcg_out_movi_one(s, ret, arg)) {
         return;
     }
+
     if (TCG_TARGET_REG_BITS == 32 || arg == (int32_t)arg) {
         tcg_out_opc_imm(s, OPC_LUI, ret, TCG_REG_ZERO, arg >> 16);
     } else {
-- 
2.25.1

[PATCH v4 08/12] tcg/mips: Split out tcg_out_movi_two

[Richard Henderson]

Emit all 32-bit signed constants, which can be loaded in two insns.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/mips/tcg-target.c.inc | 35 ++++++++++++++++++++++++-----------
 1 file changed, 24 insertions(+), 11 deletions(-)

diff --git a/tcg/mips/tcg-target.c.inc b/tcg/mips/tcg-target.c.inc
index 8741fdd49c..142583b613 100644
--- a/tcg/mips/tcg-target.c.inc
+++ b/tcg/mips/tcg-target.c.inc
@@ -541,6 +541,22 @@ static bool tcg_out_movi_one(TCGContext *s, TCGReg ret, tcg_target_long arg)
     return false;
 }
 
+static bool tcg_out_movi_two(TCGContext *s, TCGReg ret, tcg_target_long arg)
+{
+    /*
+     * All signed 32-bit constants are loadable with two immediates,
+     * and everything else requires more work.
+     */
+    if (arg == (int32_t)arg) {
+        if (!tcg_out_movi_one(s, ret, arg)) {
+            tcg_out_opc_imm(s, OPC_LUI, ret, TCG_REG_ZERO, arg >> 16);
+            tcg_out_opc_imm(s, OPC_ORI, ret, ret, arg & 0xffff);
+        }
+        return true;
+    }
+    return false;
+}
+
 static void tcg_out_movi(TCGContext *s, TCGType type,
                          TCGReg ret, tcg_target_long arg)
 {
@@ -548,21 +564,18 @@ static void tcg_out_movi(TCGContext *s, TCGType type,
         arg = (int32_t)arg;
     }
 
-    if (tcg_out_movi_one(s, ret, arg)) {
+    /* Load all 32-bit constants. */
+    if (tcg_out_movi_two(s, ret, arg)) {
         return;
     }
 
-    if (TCG_TARGET_REG_BITS == 32 || arg == (int32_t)arg) {
-        tcg_out_opc_imm(s, OPC_LUI, ret, TCG_REG_ZERO, arg >> 16);
+    tcg_out_movi(s, TCG_TYPE_I32, ret, arg >> 31 >> 1);
+    if (arg & 0xffff0000ull) {
+        tcg_out_dsll(s, ret, ret, 16);
+        tcg_out_opc_imm(s, OPC_ORI, ret, ret, arg >> 16);
+        tcg_out_dsll(s, ret, ret, 16);
     } else {
-        tcg_out_movi(s, TCG_TYPE_I32, ret, arg >> 31 >> 1);
-        if (arg & 0xffff0000ull) {
-            tcg_out_dsll(s, ret, ret, 16);
-            tcg_out_opc_imm(s, OPC_ORI, ret, ret, arg >> 16);
-            tcg_out_dsll(s, ret, ret, 16);
-        } else {
-            tcg_out_dsll(s, ret, ret, 32);
-        }
+        tcg_out_dsll(s, ret, ret, 32);
     }
     if (arg & 0xffff) {
         tcg_out_opc_imm(s, OPC_ORI, ret, ret, arg & 0xffff);
-- 
2.25.1

[PATCH v4 09/12] tcg/mips: Use the constant pool for 64-bit constants

[Richard Henderson]

During normal processing, the constant pool is accessible via
TCG_REG_TB.  During the prologue, it is accessible via TCG_REG_T9.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/mips/tcg-target.h     |  1 +
 tcg/mips/tcg-target.c.inc | 65 +++++++++++++++++++++++++++++----------
 2 files changed, 49 insertions(+), 17 deletions(-)

diff --git a/tcg/mips/tcg-target.h b/tcg/mips/tcg-target.h
index 28c42e23e1..839364b493 100644
--- a/tcg/mips/tcg-target.h
+++ b/tcg/mips/tcg-target.h
@@ -208,5 +208,6 @@ void tb_target_set_jmp_target(uintptr_t, uintptr_t, uintptr_t, uintptr_t)
     QEMU_ERROR("code path is reachable");
 
 #define TCG_TARGET_NEED_LDST_LABELS
+#define TCG_TARGET_NEED_POOL_LABELS
 
 #endif
diff --git a/tcg/mips/tcg-target.c.inc b/tcg/mips/tcg-target.c.inc
index 142583b613..41cb155eb0 100644
--- a/tcg/mips/tcg-target.c.inc
+++ b/tcg/mips/tcg-target.c.inc
@@ -25,6 +25,7 @@
  */
 
 #include "../tcg-ldst.c.inc"
+#include "../tcg-pool.c.inc"
 
 #ifdef HOST_WORDS_BIGENDIAN
 # define MIPS_BE  1
@@ -166,9 +167,18 @@ static bool reloc_pc16(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
 static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
                         intptr_t value, intptr_t addend)
 {
-    tcg_debug_assert(type == R_MIPS_PC16);
-    tcg_debug_assert(addend == 0);
-    return reloc_pc16(code_ptr, (const tcg_insn_unit *)value);
+    value += addend;
+    switch (type) {
+    case R_MIPS_PC16:
+        return reloc_pc16(code_ptr, (const tcg_insn_unit *)value);
+    case R_MIPS_16:
+        if (value != (int16_t)value) {
+            return false;
+        }
+        *code_ptr = deposit32(*code_ptr, 0, 16, value);
+        return true;
+    }
+    g_assert_not_reached();
 }
 
 #define TCG_CT_CONST_ZERO 0x100
@@ -500,6 +510,11 @@ static void tcg_out_nop(TCGContext *s)
     tcg_out32(s, 0);
 }
 
+static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
+{
+    memset(p, 0, count * sizeof(tcg_insn_unit));
+}
+
 static void tcg_out_dsll(TCGContext *s, TCGReg rd, TCGReg rt, TCGArg sa)
 {
     tcg_out_opc_sa64(s, OPC_DSLL, OPC_DSLL32, rd, rt, sa);
@@ -557,8 +572,15 @@ static bool tcg_out_movi_two(TCGContext *s, TCGReg ret, tcg_target_long arg)
     return false;
 }
 
-static void tcg_out_movi(TCGContext *s, TCGType type,
-                         TCGReg ret, tcg_target_long arg)
+static void tcg_out_movi_pool(TCGContext *s, TCGReg ret,
+                              tcg_target_long arg, TCGReg tbreg)
+{
+    new_pool_label(s, arg, R_MIPS_16, s->code_ptr, tcg_tbrel_diff(s, NULL));
+    tcg_out_opc_imm(s, OPC_LD, ret, tbreg, 0);
+}
+
+static void tcg_out_movi_int(TCGContext *s, TCGType type, TCGReg ret,
+                             tcg_target_long arg, TCGReg tbreg)
 {
     if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I32) {
         arg = (int32_t)arg;
@@ -568,18 +590,17 @@ static void tcg_out_movi(TCGContext *s, TCGType type,
     if (tcg_out_movi_two(s, ret, arg)) {
         return;
     }
+    assert(TCG_TARGET_REG_BITS == 64);
 
-    tcg_out_movi(s, TCG_TYPE_I32, ret, arg >> 31 >> 1);
-    if (arg & 0xffff0000ull) {
-        tcg_out_dsll(s, ret, ret, 16);
-        tcg_out_opc_imm(s, OPC_ORI, ret, ret, arg >> 16);
-        tcg_out_dsll(s, ret, ret, 16);
-    } else {
-        tcg_out_dsll(s, ret, ret, 32);
-    }
-    if (arg & 0xffff) {
-        tcg_out_opc_imm(s, OPC_ORI, ret, ret, arg & 0xffff);
-    }
+    /* Otherwise, put 64-bit constants into the constant pool. */
+    tcg_out_movi_pool(s, ret, arg, tbreg);
+}
+
+static void tcg_out_movi(TCGContext *s, TCGType type,
+                         TCGReg ret, tcg_target_long arg)
+{
+    TCGReg tbreg = TCG_TARGET_REG_BITS == 64 ? TCG_REG_TB : 0;
+    tcg_out_movi_int(s, type, ret, arg, tbreg);
 }
 
 static void tcg_out_bswap16(TCGContext *s, TCGReg ret, TCGReg arg, int flags)
@@ -2748,10 +2769,20 @@ static void tcg_target_qemu_prologue(TCGContext *s)
 
 #ifndef CONFIG_SOFTMMU
     if (guest_base != (int16_t)guest_base) {
-        tcg_out_movi(s, TCG_TYPE_PTR, TCG_GUEST_BASE_REG, guest_base);
+        /*
+         * The function call abi for n32 and n64 will have loaded $25 (t9)
+         * with the address of the prologue, so we can use that instead
+         * of TCG_REG_TB.
+         */
+#if TCG_TARGET_REG_BITS == 64 && !defined(__mips_abicalls)
+# error "Unknown mips abi"
+#endif
+        tcg_out_movi_int(s, TCG_TYPE_PTR, TCG_GUEST_BASE_REG, guest_base,
+                         TCG_TARGET_REG_BITS == 64 ? TCG_REG_T9 : 0);
         tcg_regset_set_reg(s->reserved_regs, TCG_GUEST_BASE_REG);
     }
 #endif
+
     if (TCG_TARGET_REG_BITS == 64) {
         tcg_out_mov(s, TCG_TYPE_PTR, TCG_REG_TB, tcg_target_call_iarg_regs[1]);
     }
-- 
2.25.1

[PATCH v4 10/12] tcg/mips: Aggressively use the constant pool for n64 calls

[Richard Henderson]

Repeated calls to a single helper are common -- especially
the ones for softmmu memory access.  Prefer the constant pool
to longer sequences to increase sharing.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/mips/tcg-target.c.inc | 16 +++++++++++++---
 1 file changed, 13 insertions(+), 3 deletions(-)

diff --git a/tcg/mips/tcg-target.c.inc b/tcg/mips/tcg-target.c.inc
index 41cb155eb0..e967f62869 100644
--- a/tcg/mips/tcg-target.c.inc
+++ b/tcg/mips/tcg-target.c.inc
@@ -1057,9 +1057,19 @@ static void tcg_out_movcond(TCGContext *s, TCGCond cond, TCGReg ret,
 
 static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *arg, bool tail)
 {
-    /* Note that the ABI requires the called function's address to be
-       loaded into T9, even if a direct branch is in range.  */
-    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_T9, (uintptr_t)arg);
+    /*
+     * Note that __mips_abicalls requires the called function's address
+     * to be loaded into $25 (t9), even if a direct branch is in range.
+     *
+     * For n64, always drop the pointer into the constant pool.
+     * We can re-use helper addresses often and do not want any
+     * of the longer sequences tcg_out_movi may try.
+     */
+    if (sizeof(uintptr_t) == 8) {
+        tcg_out_movi_pool(s, TCG_REG_T9, (uintptr_t)arg, TCG_REG_TB);
+    } else {
+        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_T9, (uintptr_t)arg);
+    }
 
     /* But do try a direct branch, allowing the cpu better insn prefetch.  */
     if (tail) {
-- 
2.25.1

[PATCH v4 11/12] tcg/mips: Try tb-relative addresses in tcg_out_movi

[Richard Henderson]

These addresses are often loaded by the qemu_ld/st slow path,
for loading the retaddr value.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/mips/tcg-target.c.inc | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/tcg/mips/tcg-target.c.inc b/tcg/mips/tcg-target.c.inc
index e967f62869..a128c70154 100644
--- a/tcg/mips/tcg-target.c.inc
+++ b/tcg/mips/tcg-target.c.inc
@@ -582,6 +582,8 @@ static void tcg_out_movi_pool(TCGContext *s, TCGReg ret,
 static void tcg_out_movi_int(TCGContext *s, TCGType type, TCGReg ret,
                              tcg_target_long arg, TCGReg tbreg)
 {
+    tcg_target_long tmp;
+
     if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I32) {
         arg = (int32_t)arg;
     }
@@ -592,6 +594,17 @@ static void tcg_out_movi_int(TCGContext *s, TCGType type, TCGReg ret,
     }
     assert(TCG_TARGET_REG_BITS == 64);
 
+    /* Load addresses within 2GB of TB with 1 or 3 insns. */
+    tmp = tcg_tbrel_diff(s, (void *)arg);
+    if (tmp == (int16_t)tmp) {
+        tcg_out_opc_imm(s, OPC_DADDIU, ret, tbreg, tmp);
+        return;
+    }
+    if (tcg_out_movi_two(s, ret, tmp)) {
+        tcg_out_opc_reg(s, OPC_DADDU, ret, ret, tbreg);
+        return;
+    }
+
     /* Otherwise, put 64-bit constants into the constant pool. */
     tcg_out_movi_pool(s, ret, arg, tbreg);
 }
-- 
2.25.1

[PATCH v4 12/12] tcg/mips: Try three insns with shift and add in tcg_out_movi

[Richard Henderson]

These sequences are inexpensive to test.  Maxing out at three insns
results in the same space as a load plus the constant pool entry.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/mips/tcg-target.c.inc | 44 +++++++++++++++++++++++++++++++++++++++
 1 file changed, 44 insertions(+)

diff --git a/tcg/mips/tcg-target.c.inc b/tcg/mips/tcg-target.c.inc
index a128c70154..185241da17 100644
--- a/tcg/mips/tcg-target.c.inc
+++ b/tcg/mips/tcg-target.c.inc
@@ -583,6 +583,7 @@ static void tcg_out_movi_int(TCGContext *s, TCGType type, TCGReg ret,
                              tcg_target_long arg, TCGReg tbreg)
 {
     tcg_target_long tmp;
+    int sh, lo;
 
     if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I32) {
         arg = (int32_t)arg;
@@ -605,6 +606,49 @@ static void tcg_out_movi_int(TCGContext *s, TCGType type, TCGReg ret,
         return;
     }
 
+    /*
+     * Load bitmasks with a right-shift.  This is good for things
+     * like 0x0fff_ffff_ffff_fff0: ADDUI r,0xff00 + DSRL r,r,4.
+     * or similarly using LUI.  For this to work, bit 31 must be set.
+     */
+    if (arg > 0 && (int32_t)arg < 0) {
+        sh = clz64(arg);
+        if (tcg_out_movi_one(s, ret, arg << sh)) {
+            tcg_out_dsrl(s, ret, ret, sh);
+            return;
+        }
+    }
+
+    /*
+     * Load slightly larger constants using left-shift.
+     * Limit this sequence to 3 insns to avoid too much expansion.
+     */
+    sh = ctz64(arg);
+    if (sh && tcg_out_movi_two(s, ret, arg >> sh)) {
+        tcg_out_dsll(s, ret, ret, sh);
+        return;
+    }
+
+    /*
+     * Load slightly larger constants using left-shift and add/or.
+     * Prefer addi with a negative immediate when that would produce
+     * a larger shift.  For this to work, bits 15 and 16 must be set.
+     */
+    lo = arg & 0xffff;
+    if (lo) {
+        if ((arg & 0x18000) == 0x18000) {
+            lo = (int16_t)arg;
+        }
+        tmp = arg - lo;
+        sh = ctz64(tmp);
+        tmp >>= sh;
+        if (tcg_out_movi_one(s, ret, tmp)) {
+            tcg_out_dsll(s, ret, ret, sh);
+            tcg_out_opc_imm(s, lo < 0 ? OPC_DADDIU : OPC_ORI, ret, ret, lo);
+            return;
+        }
+    }
+
     /* Otherwise, put 64-bit constants into the constant pool. */
     tcg_out_movi_pool(s, ret, arg, tbreg);
 }
-- 
2.25.1

Re: [PATCH v4 05/12] tcg/mips: Unify TCG_GUEST_BASE_REG tests

[Philippe Mathieu-Daudé]

On 1/8/22 07:36, Richard Henderson wrote:
> In tcg_out_qemu_ld/st, we already check for guest_base matching int16_t.
> Mirror that when setting up TCG_GUEST_BASE_REG in the prologue.
> 
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
>  tcg/mips/tcg-target.c.inc | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)

Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>