This can significantly reduce code size for generation of (some) 64-bit constants. With the side effect that we know for a fact that exit_tb can use the register to good effect.
Signed-off-by: Richard Henderson <r...@twiddle.net> --- tcg/ppc/tcg-target.c | 105 +++++++++++++++++++++++++++++++++++++++++++++------ 1 file changed, 93 insertions(+), 12 deletions(-) diff --git a/tcg/ppc/tcg-target.c b/tcg/ppc/tcg-target.c index e6283e1..c83fd9f 100644 --- a/tcg/ppc/tcg-target.c +++ b/tcg/ppc/tcg-target.c @@ -41,6 +41,30 @@ # define TCG_REG_TMP1 TCG_REG_R12 #endif +/* For the 64-bit target, we don't like the 5 insn sequence needed to build + full 64-bit addresses. Better to have a base register to which we can + apply a 32-bit displacement. + + There are generally three items of interest: + (1) helper functions in the main executable, + (2) TranslationBlock data structures, + (3) the return address in the epilogue. + + For user-only, we USE_STATIC_CODE_GEN_BUFFER, so the code_gen_buffer + will be inside the main executable, and thus near enough to make a + pointer to the epilogue be within 2GB of all helper functions. + + For softmmu, we'll let the kernel choose the address of code_gen_buffer, + and odds are it'll be somewhere close to the main malloc arena, and so + a pointer to the epilogue will be within 2GB of the TranslationBlocks. + + For --enable-pie, everything will be kinda near everything else, + somewhere in high memory. + + Thus we choose to keep the return address in a call-saved register. */ +#define TCG_REG_RA TCG_REG_R31 +#define USE_REG_RA (TCG_TARGET_REG_BITS == 64) + /* Shorthand for size of a pointer. Avoid promotion to unsigned. */ #define SZP ((int)sizeof(void *)) @@ -467,6 +491,8 @@ static int tcg_target_const_match(tcg_target_long val, TCGType type, #define TW XO31( 4) #define TRAP (TW | TO(31)) +#define NOP ORI /* ori 0,0,0 */ + #define RT(r) ((r)<<21) #define RS(r) ((r)<<21) #define RA(r) ((r)<<16) @@ -531,6 +557,9 @@ static const uint32_t tcg_to_isel[] = { [TCG_COND_GTU] = ISEL | BC_(7, CR_GT), }; +static void tcg_out_mem_long(TCGContext *s, int opi, int opx, TCGReg rt, + TCGReg base, tcg_target_long offset); + static void tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg) { tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || type == TCG_TYPE_I32); @@ -601,7 +630,17 @@ static void tcg_out_movi(TCGContext *s, TCGType type, TCGReg ret, tcg_out32(s, ADDI | TAI(ret, 0, arg)); tcg_out32(s, ORIS | SAI(ret, ret, arg >> 16)); } else { - int32_t high = arg >> 31 >> 1; + int32_t high; + + if (USE_REG_RA) { + intptr_t diff = arg - (intptr_t)tb_ret_addr; + if (diff == (int32_t)diff) { + tcg_out_mem_long(s, ADDI, ADD, ret, TCG_REG_RA, diff); + return; + } + } + + high = arg >> 31 >> 1; tcg_out_movi32(s, ret, high); if (high) { tcg_out_shli64(s, ret, ret, 32); @@ -1714,18 +1753,16 @@ static void tcg_target_qemu_prologue(TCGContext *s) { int i; - tcg_set_frame(s, TCG_REG_CALL_STACK, REG_SAVE_BOT - CPU_TEMP_BUF_SIZE, - CPU_TEMP_BUF_SIZE); - #ifdef _CALL_AIX - { - void **desc = (void **)s->code_ptr; - desc[0] = desc + 2; /* entry point */ - desc[1] = 0; /* environment pointer */ - s->code_ptr = (void *)(desc + 2); /* skip over descriptor */ - } + void **desc = (void **)s->code_ptr; + desc[0] = desc + 2; /* entry point */ + desc[1] = 0; /* environment pointer */ + s->code_ptr = (void *)(desc + 2); /* skip over descriptor */ #endif + tcg_set_frame(s, TCG_REG_CALL_STACK, REG_SAVE_BOT - CPU_TEMP_BUF_SIZE, + CPU_TEMP_BUF_SIZE); + /* Prologue */ tcg_out32(s, MFSPR | RT(TCG_REG_R0) | LR); tcg_out32(s, (SZR == 8 ? STDU : STWU) @@ -1746,10 +1783,36 @@ static void tcg_target_qemu_prologue(TCGContext *s) tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]); tcg_out32(s, MTSPR | RS(tcg_target_call_iarg_regs[1]) | CTR); - tcg_out32(s, BCCTR | BO_ALWAYS); + + if (USE_REG_RA) { +#ifdef _CALL_AIX + /* Make the caller load the value as the TOC into R2. */ + tb_ret_addr = s->code_ptr + 2; + desc[1] = tb_ret_addr; + tcg_out_mov(s, TCG_TYPE_PTR, TCG_REG_RA, TCG_REG_R2); + tcg_out32(s, BCCTR | BO_ALWAYS); +#elif defined(_CALL_ELF) && _CALL_ELF == 2 + /* Compute from the incoming R12 value. */ + tb_ret_addr = s->code_ptr + 2; + tcg_out32(s, ADDI | TAI(TCG_REG_RA, TCG_REG_R12, + tcg_ptr_byte_diff(tb_ret_addr, s->code_buf))); + tcg_out32(s, BCCTR | BO_ALWAYS); +#else + /* Reserve max 5 insns for the constant load. */ + tb_ret_addr = s->code_ptr + 6; + tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_RA, (intptr_t)tb_ret_addr); + tcg_out32(s, BCCTR | BO_ALWAYS); + while (s->code_ptr < tb_ret_addr) { + tcg_out32(s, NOP); + } +#endif + } else { + tcg_out32(s, BCCTR | BO_ALWAYS); + tb_ret_addr = s->code_ptr; + } /* Epilogue */ - tb_ret_addr = s->code_ptr; + assert(tb_ret_addr == s->code_ptr); tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_R0, TCG_REG_R1, FRAME_SIZE+LR_OFFSET); for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); ++i) { @@ -1769,6 +1832,21 @@ static void tcg_out_op(TCGContext *s, TCGOpcode opc, const TCGArg *args, switch (opc) { case INDEX_op_exit_tb: + if (USE_REG_RA) { + ptrdiff_t disp = tcg_pcrel_diff(s, tb_ret_addr); + + /* If we can use a direct branch, otherwise use the value in RA. + Note that the direct branch is always forward. If it's in + range now, it'll still be in range after the movi. Don't + bother about the 20 bytes where the test here fails but it + would succeed below. */ + if (!in_range_b(disp)) { + tcg_out32(s, MTSPR | RS(TCG_REG_RA) | CTR); + tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R3, args[0]); + tcg_out32(s, BCCTR | BO_ALWAYS); + break; + } + } tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R3, args[0]); tcg_out_b(s, 0, tb_ret_addr); break; @@ -2479,6 +2557,9 @@ static void tcg_target_init(TCGContext *s) tcg_regset_set_reg(s->reserved_regs, TCG_REG_R13); /* thread pointer */ #endif tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP1); /* mem temp */ + if (USE_REG_RA) { + tcg_regset_set_reg(s->reserved_regs, TCG_REG_RA); /* return addr */ + } tcg_add_target_add_op_defs(ppc_op_defs); } -- 1.9.3