tcg/s390x: Remove TCG_REG_TB

This reverts 829e1376d9 ("tcg/s390: Introduce TCG_REG_TB"), and
several follow-up patches.  The primary motivation is to reduce the
less-tested code paths, pre-z10.  Secondarily, this allows the
unconditional use of TCG_TARGET_HAS_direct_jump, which might be more
important for performance than any slight increase in code size.

Reviewed-by: Ilya Leoshkevich <iii@linux.ibm.com>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
v4: Do not simplify tgen_ori, tgen_xori.
This commit is contained in:
Richard Henderson 2022-11-30 00:17:50 +00:00
parent 4143f78dad
commit ccbecb441e

View File

@ -65,12 +65,6 @@
/* A scratch register that may be be used throughout the backend. */
#define TCG_TMP0 TCG_REG_R1
/* A scratch register that holds a pointer to the beginning of the TB.
We don't need this when we have pc-relative loads with the general
instructions extension facility. */
#define TCG_REG_TB TCG_REG_R12
#define USE_REG_TB (!HAVE_FACILITY(GEN_INST_EXT))
#ifndef CONFIG_SOFTMMU
#define TCG_GUEST_BASE_REG TCG_REG_R13
#endif
@ -813,8 +807,8 @@ static bool maybe_out_small_movi(TCGContext *s, TCGType type,
}
/* load a register with an immediate value */
static void tcg_out_movi_int(TCGContext *s, TCGType type, TCGReg ret,
tcg_target_long sval, bool in_prologue)
static void tcg_out_movi(TCGContext *s, TCGType type,
TCGReg ret, tcg_target_long sval)
{
tcg_target_ulong uval;
@ -853,14 +847,6 @@ static void tcg_out_movi_int(TCGContext *s, TCGType type, TCGReg ret,
tcg_out_insn(s, RIL, LARL, ret, off);
return;
}
} else if (USE_REG_TB && !in_prologue) {
ptrdiff_t off = tcg_tbrel_diff(s, (void *)sval);
if (off == sextract64(off, 0, 20)) {
/* This is certain to be an address within TB, and therefore
OFF will be negative; don't try RX_LA. */
tcg_out_insn(s, RXY, LAY, ret, TCG_REG_TB, TCG_REG_NONE, off);
return;
}
}
/* A 32-bit unsigned value can be loaded in 2 insns. And given
@ -876,10 +862,6 @@ static void tcg_out_movi_int(TCGContext *s, TCGType type, TCGReg ret,
if (HAVE_FACILITY(GEN_INST_EXT)) {
tcg_out_insn(s, RIL, LGRL, ret, 0);
new_pool_label(s, sval, R_390_PC32DBL, s->code_ptr - 2, 2);
} else if (USE_REG_TB && !in_prologue) {
tcg_out_insn(s, RXY, LG, ret, TCG_REG_TB, TCG_REG_NONE, 0);
new_pool_label(s, sval, R_390_20, s->code_ptr - 2,
tcg_tbrel_diff(s, NULL));
} else {
TCGReg base = ret ? ret : TCG_TMP0;
tcg_out_insn(s, RIL, LARL, base, 0);
@ -888,12 +870,6 @@ static void tcg_out_movi_int(TCGContext *s, TCGType type, TCGReg ret,
}
}
static void tcg_out_movi(TCGContext *s, TCGType type,
TCGReg ret, tcg_target_long sval)
{
tcg_out_movi_int(s, type, ret, sval, false);
}
/* Emit a load/store type instruction. Inputs are:
DATA: The register to be loaded or stored.
BASE+OFS: The effective address.
@ -1037,13 +1013,6 @@ static void tcg_out_ld_abs(TCGContext *s, TCGType type,
return;
}
}
if (USE_REG_TB) {
ptrdiff_t disp = tcg_tbrel_diff(s, abs);
if (disp == sextract64(disp, 0, 20)) {
tcg_out_ld(s, type, dest, TCG_REG_TB, disp);
return;
}
}
tcg_out_movi(s, TCG_TYPE_PTR, dest, addr & ~0xffff);
tcg_out_ld(s, type, dest, dest, addr & 0xffff);
@ -1243,17 +1212,7 @@ static void tgen_andi(TCGContext *s, TCGType type, TCGReg dest, uint64_t val)
return;
}
/* Use the constant pool if USE_REG_TB, but not for small constants. */
if (USE_REG_TB) {
if (!maybe_out_small_movi(s, type, TCG_TMP0, val)) {
tcg_out_insn(s, RXY, NG, dest, TCG_REG_TB, TCG_REG_NONE, 0);
new_pool_label(s, val & valid, R_390_20, s->code_ptr - 2,
tcg_tbrel_diff(s, NULL));
return;
}
} else {
tcg_out_movi(s, type, TCG_TMP0, val);
}
tcg_out_movi(s, type, TCG_TMP0, val);
if (type == TCG_TYPE_I32) {
tcg_out_insn(s, RR, NR, dest, TCG_TMP0);
} else {
@ -1297,17 +1256,12 @@ static void tgen_ori(TCGContext *s, TCGType type, TCGReg dest, uint64_t val)
}
}
/* Use the constant pool if USE_REG_TB, but not for small constants. */
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 if (USE_REG_TB) {
tcg_out_insn(s, RXY, OG, dest, TCG_REG_TB, TCG_REG_NONE, 0);
new_pool_label(s, val, R_390_20, s->code_ptr - 2,
tcg_tbrel_diff(s, NULL));
} else {
/* Perform the OR via sequential modifications to the high and
low parts. Do this via recursion to handle 16-bit vs 32-bit
@ -1332,17 +1286,12 @@ static void tgen_xori(TCGContext *s, TCGType type, TCGReg dest, uint64_t val)
}
}
/* Use the constant pool if USE_REG_TB, but not for small constants. */
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 if (USE_REG_TB) {
tcg_out_insn(s, RXY, XG, dest, TCG_REG_TB, TCG_REG_NONE, 0);
new_pool_label(s, val, R_390_20, s->code_ptr - 2,
tcg_tbrel_diff(s, NULL));
} else {
/* Perform the xor by parts. */
tcg_debug_assert(HAVE_FACILITY(EXT_IMM));
@ -1395,19 +1344,6 @@ static int tgen_cmp(TCGContext *s, TCGType type, TCGCond c, TCGReg r1,
if (maybe_out_small_movi(s, type, TCG_TMP0, c2)) {
c2 = TCG_TMP0;
/* fall through to reg-reg */
} else if (USE_REG_TB) {
if (type == TCG_TYPE_I32) {
op = (is_unsigned ? RXY_CLY : RXY_CY);
tcg_out_insn_RXY(s, op, r1, TCG_REG_TB, TCG_REG_NONE, 0);
new_pool_label(s, (uint32_t)c2, R_390_20, s->code_ptr - 2,
4 - tcg_tbrel_diff(s, NULL));
} else {
op = (is_unsigned ? RXY_CLG : RXY_CG);
tcg_out_insn_RXY(s, op, r1, TCG_REG_TB, TCG_REG_NONE, 0);
new_pool_label(s, c2, R_390_20, s->code_ptr - 2,
tcg_tbrel_diff(s, NULL));
}
goto exit;
} else {
if (type == TCG_TYPE_I32) {
op = (is_unsigned ? RIL_CLRL : RIL_CRL);
@ -2109,35 +2045,21 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
if (!QEMU_PTR_IS_ALIGNED(s->code_ptr + 1, 4)) {
tcg_out16(s, NOP);
}
tcg_debug_assert(!USE_REG_TB);
tcg_out16(s, RIL_BRCL | (S390_CC_ALWAYS << 4));
s->tb_jmp_insn_offset[a0] = tcg_current_code_size(s);
s->code_ptr += 2;
} else {
/* load address stored at s->tb_jmp_target_addr + a0 */
tcg_out_ld_abs(s, TCG_TYPE_PTR, TCG_REG_TB,
tcg_out_ld_abs(s, TCG_TYPE_PTR, TCG_TMP0,
tcg_splitwx_to_rx(s->tb_jmp_target_addr + a0));
/* and go there */
tcg_out_insn(s, RR, BCR, S390_CC_ALWAYS, TCG_REG_TB);
tcg_out_insn(s, RR, BCR, S390_CC_ALWAYS, TCG_TMP0);
}
set_jmp_reset_offset(s, a0);
/* For the unlinked path of goto_tb, we need to reset
TCG_REG_TB to the beginning of this TB. */
if (USE_REG_TB) {
int ofs = -tcg_current_code_size(s);
/* All TB are restricted to 64KiB by unwind info. */
tcg_debug_assert(ofs == sextract64(ofs, 0, 20));
tcg_out_insn(s, RXY, LAY, TCG_REG_TB,
TCG_REG_TB, TCG_REG_NONE, ofs);
}
break;
case INDEX_op_goto_ptr:
a0 = args[0];
if (USE_REG_TB) {
tcg_out_mov(s, TCG_TYPE_PTR, TCG_REG_TB, a0);
}
tcg_out_insn(s, RR, BCR, S390_CC_ALWAYS, a0);
break;
@ -3405,9 +3327,6 @@ static void tcg_target_init(TCGContext *s)
/* XXX many insns can't be used with R0, so we better avoid it for now */
tcg_regset_set_reg(s->reserved_regs, TCG_REG_R0);
tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
if (USE_REG_TB) {
tcg_regset_set_reg(s->reserved_regs, TCG_REG_TB);
}
}
#define FRAME_SIZE ((int)(TCG_TARGET_CALL_STACK_OFFSET \
@ -3428,16 +3347,12 @@ static void tcg_target_qemu_prologue(TCGContext *s)
#ifndef CONFIG_SOFTMMU
if (guest_base >= 0x80000) {
tcg_out_movi_int(s, TCG_TYPE_PTR, TCG_GUEST_BASE_REG, guest_base, true);
tcg_out_movi(s, TCG_TYPE_PTR, TCG_GUEST_BASE_REG, guest_base);
tcg_regset_set_reg(s->reserved_regs, TCG_GUEST_BASE_REG);
}
#endif
tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
if (USE_REG_TB) {
tcg_out_mov(s, TCG_TYPE_PTR, TCG_REG_TB,
tcg_target_call_iarg_regs[1]);
}
/* br %r3 (go to TB) */
tcg_out_insn(s, RR, BCR, S390_CC_ALWAYS, tcg_target_call_iarg_regs[1]);