diff options
Diffstat (limited to 'src/sljit/sljitNativePPC_common.c')
| -rw-r--r-- | src/sljit/sljitNativePPC_common.c | 168 |
1 files changed, 138 insertions, 30 deletions
diff --git a/src/sljit/sljitNativePPC_common.c b/src/sljit/sljitNativePPC_common.c index 21cdb0cf..f3871147 100644 --- a/src/sljit/sljitNativePPC_common.c +++ b/src/sljit/sljitNativePPC_common.c @@ -203,8 +203,13 @@ static const sljit_u8 freg_map[SLJIT_NUMBER_OF_FLOAT_REGISTERS + 3] = { #define OR (HI(31) | LO(444)) #define ORI (HI(24)) #define ORIS (HI(25)) -#define RLDICL (HI(30)) +#define RLDCL (HI(30) | LO(8)) +#define RLDICL (HI(30) | LO(0 << 1)) +#define RLDICR (HI(30) | LO(1 << 1)) +#define RLDIMI (HI(30) | LO(3 << 1)) +#define RLWIMI (HI(20)) #define RLWINM (HI(21)) +#define RLWNM (HI(23)) #define SLD (HI(31) | LO(27)) #define SLW (HI(31) | LO(24)) #define SRAD (HI(31) | LO(794)) @@ -233,9 +238,24 @@ static const sljit_u8 freg_map[SLJIT_NUMBER_OF_FLOAT_REGISTERS + 3] = { #define SIMM_MIN (-0x8000) #define UIMM_MAX (0xffff) -#define RLDI(dst, src, sh, mb, type) \ - (HI(30) | S(src) | A(dst) | ((sljit_ins)(type) << 2) | (((sljit_ins)(sh) & 0x1f) << 11) \ - | (((sljit_ins)(sh) & 0x20) >> 4) | (((sljit_ins)(mb) & 0x1f) << 6) | ((sljit_ins)(mb) & 0x20)) +/* Shift helpers. */ +#define RLWI_SH(sh) ((sljit_ins)(sh) << 11) +#define RLWI_MBE(mb, me) (((sljit_ins)(mb) << 6) | ((sljit_ins)(me) << 1)) +#define RLDI_SH(sh) ((((sljit_ins)(sh) & 0x1f) << 11) | (((sljit_ins)(sh) & 0x20) >> 4)) +#define RLDI_MB(mb) ((((sljit_ins)(mb) & 0x1f) << 6) | ((sljit_ins)(mb) & 0x20)) +#define RLDI_ME(me) RLDI_MB(me) + +#define SLWI(shift) (RLWINM | RLWI_SH(shift) | RLWI_MBE(0, 31 - (shift))) +#define SLDI(shift) (RLDICR | RLDI_SH(shift) | RLDI_ME(63 - (shift))) +/* shift > 0 */ +#define SRWI(shift) (RLWINM | RLWI_SH(32 - (shift)) | RLWI_MBE((shift), 31)) +#define SRDI(shift) (RLDICL | RLDI_SH(64 - (shift)) | RLDI_MB(shift)) + +#if (defined SLJIT_CONFIG_PPC_32 && SLJIT_CONFIG_PPC_32) +#define SLWI_W(shift) SLWI(shift) +#else /* !SLJIT_CONFIG_PPC_32 */ +#define SLWI_W(shift) SLDI(shift) +#endif /* SLJIT_CONFIG_PPC_32 */ #if (defined SLJIT_INDIRECT_CALL && SLJIT_INDIRECT_CALL) SLJIT_API_FUNC_ATTRIBUTE void sljit_set_function_context(void** func_ptr, struct sljit_function_context* context, sljit_uw addr, void* func) @@ -371,7 +391,7 @@ static SLJIT_INLINE void put_label_set(struct sljit_put_label *put_label) inst++; } - inst[1] = RLDI(reg, reg, 32, 31, 1); + inst[1] = SLDI(32) | S(reg) | A(reg); inst[2] = ORIS | S(reg) | A(reg) | IMM((addr >> 16) & 0xffff); inst += 2; } @@ -379,7 +399,7 @@ static SLJIT_INLINE void put_label_set(struct sljit_put_label *put_label) inst[1] = ORI | S(reg) | A(reg) | IMM(addr & 0xffff); } -#endif +#endif /* SLJIT_CONFIG_PPC_64 */ SLJIT_API_FUNC_ATTRIBUTE void* sljit_generate_code(struct sljit_compiler *compiler) { @@ -641,9 +661,13 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_has_cpu_feature(sljit_s32 feature_type) /* A saved register is set to a zero value. */ case SLJIT_HAS_ZERO_REGISTER: case SLJIT_HAS_CLZ: + case SLJIT_HAS_ROT: case SLJIT_HAS_PREFETCH: return 1; + case SLJIT_HAS_CTZ: + return 2; + default: return 0; } @@ -1125,11 +1149,7 @@ static sljit_s32 emit_op_mem(struct sljit_compiler *compiler, sljit_s32 inp_flag offs_reg = OFFS_REG(arg); if (argw != 0) { -#if (defined SLJIT_CONFIG_PPC_32 && SLJIT_CONFIG_PPC_32) - FAIL_IF(push_inst(compiler, RLWINM | S(OFFS_REG(arg)) | A(tmp_reg) | ((sljit_ins)argw << 11) | ((31 - (sljit_ins)argw) << 1))); -#else /* !SLJIT_CONFIG_PPC_32 */ - FAIL_IF(push_inst(compiler, RLDI(tmp_reg, OFFS_REG(arg), argw, 63 - argw, 1))); -#endif /* SLJIT_CONFIG_PPC_32 */ + FAIL_IF(push_inst(compiler, SLWI_W(argw) | S(OFFS_REG(arg)) | A(tmp_reg))); offs_reg = tmp_reg; } @@ -1309,11 +1329,7 @@ static sljit_s32 emit_prefetch(struct sljit_compiler *compiler, if (srcw == 0) return push_inst(compiler, DCBT | A(src & REG_MASK) | B(OFFS_REG(src))); -#if (defined SLJIT_CONFIG_PPC_32 && SLJIT_CONFIG_PPC_32) - FAIL_IF(push_inst(compiler, RLWINM | S(OFFS_REG(src)) | A(TMP_REG1) | ((sljit_ins)srcw << 11) | ((31 - (sljit_ins)srcw) << 1))); -#else - FAIL_IF(push_inst(compiler, RLDI(TMP_REG1, OFFS_REG(src), srcw, 63 - srcw, 1))); -#endif + FAIL_IF(push_inst(compiler, SLWI_W(srcw) | S(OFFS_REG(src)) | A(TMP_REG1))); return push_inst(compiler, DCBT | A(src & REG_MASK) | B(TMP_REG1)); } @@ -1398,10 +1414,11 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op1(struct sljit_compiler *compile return emit_op(compiler, SLJIT_NOT, flags, dst, dstw, TMP_REG1, 0, src, srcw); case SLJIT_CLZ: + case SLJIT_CTZ: #if (defined SLJIT_CONFIG_PPC_64 && SLJIT_CONFIG_PPC_64) - return emit_op(compiler, SLJIT_CLZ, flags | (!(op_flags & SLJIT_32) ? 0 : ALT_FORM1), dst, dstw, TMP_REG1, 0, src, srcw); + return emit_op(compiler, op, flags | (!(op_flags & SLJIT_32) ? 0 : ALT_FORM1), dst, dstw, TMP_REG1, 0, src, srcw); #else - return emit_op(compiler, SLJIT_CLZ, flags, dst, dstw, TMP_REG1, 0, src, srcw); + return emit_op(compiler, op, flags, dst, dstw, TMP_REG1, 0, src, srcw); #endif } @@ -1681,6 +1698,8 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op2(struct sljit_compiler *compile case SLJIT_MLSHR: case SLJIT_ASHR: case SLJIT_MASHR: + case SLJIT_ROTL: + case SLJIT_ROTR: #if (defined SLJIT_CONFIG_PPC_64 && SLJIT_CONFIG_PPC_64) if (op & SLJIT_32) flags |= ALT_FORM2; @@ -1710,6 +1729,102 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op2u(struct sljit_compiler *compil #undef TEST_SUB_FORM2 #undef TEST_SUB_FORM3 +SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_shift_into(struct sljit_compiler *compiler, sljit_s32 op, + sljit_s32 src_dst, + sljit_s32 src1, sljit_sw src1w, + sljit_s32 src2, sljit_sw src2w) +{ + sljit_s32 is_right; +#if (defined SLJIT_CONFIG_PPC_64 && SLJIT_CONFIG_PPC_64) + sljit_s32 inp_flags = ((op & SLJIT_32) ? INT_DATA : WORD_DATA) | LOAD_DATA; + sljit_sw bit_length = (op & SLJIT_32) ? 32 : 64; +#else /* !SLJIT_CONFIG_PPC_64 */ + sljit_s32 inp_flags = WORD_DATA | LOAD_DATA; + sljit_sw bit_length = 32; +#endif /* SLJIT_CONFIG_PPC_64 */ + + CHECK_ERROR(); + CHECK(check_sljit_emit_shift_into(compiler, op, src_dst, src1, src1w, src2, src2w)); + + is_right = (GET_OPCODE(op) == SLJIT_LSHR || GET_OPCODE(op) == SLJIT_MLSHR); + + if (src_dst == src1) { + SLJIT_SKIP_CHECKS(compiler); + return sljit_emit_op2(compiler, (is_right ? SLJIT_ROTR : SLJIT_ROTL) | (op & SLJIT_32), src_dst, 0, src_dst, 0, src2, src2w); + } + + ADJUST_LOCAL_OFFSET(src1, src1w); + ADJUST_LOCAL_OFFSET(src2, src2w); + + if (src2 & SLJIT_IMM) { + src2w &= bit_length - 1; + + if (src2w == 0) + return SLJIT_SUCCESS; + } else if (src2 & SLJIT_MEM) { + FAIL_IF(emit_op_mem(compiler, inp_flags, TMP_REG2, src2, src2w, TMP_REG2)); + src2 = TMP_REG2; + } + + if (src1 & SLJIT_MEM) { + FAIL_IF(emit_op_mem(compiler, inp_flags, TMP_REG1, src1, src1w, TMP_REG1)); + src1 = TMP_REG1; + } else if (src1 & SLJIT_IMM) { + FAIL_IF(load_immediate(compiler, TMP_REG1, src1w)); + src1 = TMP_REG1; + } + + if (src2 & SLJIT_IMM) { +#if (defined SLJIT_CONFIG_PPC_64 && SLJIT_CONFIG_PPC_64) + if (!(op & SLJIT_32)) { + if (is_right) { + FAIL_IF(push_inst(compiler, SRDI(src2w) | S(src_dst) | A(src_dst))); + return push_inst(compiler, RLDIMI | S(src1) | A(src_dst) | RLDI_SH(64 - src2w) | RLDI_MB(0)); + } + + FAIL_IF(push_inst(compiler, SLDI(src2w) | S(src_dst) | A(src_dst))); + /* Computes SRDI(64 - src2w). */ + FAIL_IF(push_inst(compiler, RLDICL | S(src1) | A(TMP_REG1) | RLDI_SH(src2w) | RLDI_MB(64 - src2w))); + return push_inst(compiler, OR | S(src_dst) | A(src_dst) | B(TMP_REG1)); + } +#endif /* SLJIT_CONFIG_PPC_64 */ + + if (is_right) { + FAIL_IF(push_inst(compiler, SRWI(src2w) | S(src_dst) | A(src_dst))); + return push_inst(compiler, RLWIMI | S(src1) | A(src_dst) | RLWI_SH(32 - src2w) | RLWI_MBE(0, src2w - 1)); + } + + FAIL_IF(push_inst(compiler, SLWI(src2w) | S(src_dst) | A(src_dst))); + return push_inst(compiler, RLWIMI | S(src1) | A(src_dst) | RLWI_SH(src2w) | RLWI_MBE(32 - src2w, 31)); + } + +#if (defined SLJIT_CONFIG_PPC_64 && SLJIT_CONFIG_PPC_64) + if (!(op & SLJIT_32)) { + if (GET_OPCODE(op) == SLJIT_MSHL || GET_OPCODE(op) == SLJIT_MLSHR) { + FAIL_IF(push_inst(compiler, ANDI | S(src2) | A(TMP_REG2) | 0x3f)); + src2 = TMP_REG2; + } + + FAIL_IF(push_inst(compiler, (is_right ? SRD : SLD) | S(src_dst) | A(src_dst) | B(src2))); + FAIL_IF(push_inst(compiler, (is_right ? SLDI(1) : SRDI(1)) | S(src1) | A(TMP_REG1))); + FAIL_IF(push_inst(compiler, XORI | S(src2) | A(TMP_REG2) | 0x3f)); + FAIL_IF(push_inst(compiler, (is_right ? SLD : SRD) | S(TMP_REG1) | A(TMP_REG1) | B(TMP_REG2))); + return push_inst(compiler, OR | S(src_dst) | A(src_dst) | B(TMP_REG1)); + } +#endif /* SLJIT_CONFIG_PPC_64 */ + + if (GET_OPCODE(op) == SLJIT_MSHL || GET_OPCODE(op) == SLJIT_MLSHR) { + FAIL_IF(push_inst(compiler, ANDI | S(src2) | A(TMP_REG2) | 0x1f)); + src2 = TMP_REG2; + } + + FAIL_IF(push_inst(compiler, (is_right ? SRW : SLW) | S(src_dst) | A(src_dst) | B(src2))); + FAIL_IF(push_inst(compiler, (is_right ? SLWI(1) : SRWI(1)) | S(src1) | A(TMP_REG1))); + FAIL_IF(push_inst(compiler, XORI | S(src2) | A(TMP_REG2) | 0x1f)); + FAIL_IF(push_inst(compiler, (is_right ? SLW : SRW) | S(TMP_REG1) | A(TMP_REG1) | B(TMP_REG2))); + return push_inst(compiler, OR | S(src_dst) | A(src_dst) | B(TMP_REG1)); +} + SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op_src(struct sljit_compiler *compiler, sljit_s32 op, sljit_s32 src, sljit_sw srcw) { @@ -1722,7 +1837,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op_src(struct sljit_compiler *comp if (FAST_IS_REG(src)) FAIL_IF(push_inst(compiler, MTLR | S(src))); else { - FAIL_IF(emit_op(compiler, SLJIT_MOV, WORD_DATA, TMP_REG2, 0, TMP_REG1, 0, src, srcw)); + FAIL_IF(emit_op_mem(compiler, WORD_DATA | LOAD_DATA, TMP_REG2, src, srcw, TMP_REG2)); FAIL_IF(push_inst(compiler, MTLR | S(TMP_REG2))); } @@ -1818,11 +1933,7 @@ static SLJIT_INLINE sljit_s32 sljit_emit_fop1_conv_sw_from_f64(struct sljit_comp if (dst & OFFS_REG_MASK) { dstw &= 0x3; if (dstw) { -#if (defined SLJIT_CONFIG_PPC_32 && SLJIT_CONFIG_PPC_32) - FAIL_IF(push_inst(compiler, RLWINM | S(OFFS_REG(dst)) | A(TMP_REG1) | ((sljit_ins)dstw << 11) | ((31 - (sljit_ins)dstw) << 1))); -#else - FAIL_IF(push_inst(compiler, RLDI(TMP_REG1, OFFS_REG(dst), dstw, 63 - dstw, 1))); -#endif + FAIL_IF(push_inst(compiler, SLWI_W(dstw) | S(OFFS_REG(dst)) | A(TMP_REG1))); dstw = TMP_REG1; } else @@ -2442,7 +2553,8 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op_flags(struct sljit_compiler *co } FAIL_IF(push_inst(compiler, (from_xer ? MFXER : MFCR) | D(reg))); - FAIL_IF(push_inst(compiler, RLWINM | S(reg) | A(reg) | ((1 + bit) << 11) | (31 << 6) | (31 << 1))); + /* Simplified mnemonics: extrwi. */ + FAIL_IF(push_inst(compiler, RLWINM | S(reg) | A(reg) | RLWI_SH(1 + bit) | RLWI_MBE(31, 31))); if (invert) FAIL_IF(push_inst(compiler, XORI | S(reg) | A(reg) | 0x1)); @@ -2504,11 +2616,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_mem(struct sljit_compiler *compile memw &= 0x3; if (memw != 0) { -#if (defined SLJIT_CONFIG_PPC_32 && SLJIT_CONFIG_PPC_32) - FAIL_IF(push_inst(compiler, RLWINM | S(OFFS_REG(mem)) | A(TMP_REG1) | ((sljit_ins)memw << 11) | ((31 - (sljit_ins)memw) << 1))); -#else /* !SLJIT_CONFIG_PPC_32 */ - FAIL_IF(push_inst(compiler, RLDI(TMP_REG1, OFFS_REG(mem), memw, 63 - memw, 1))); -#endif /* SLJIT_CONFIG_PPC_32 */ + FAIL_IF(push_inst(compiler, SLWI_W(memw) | S(OFFS_REG(mem)) | A(TMP_REG1))); FAIL_IF(push_inst(compiler, ADD | D(TMP_REG1) | A(TMP_REG1) | B(mem & REG_MASK))); } else FAIL_IF(push_inst(compiler, ADD | D(TMP_REG1) | A(mem & REG_MASK) | B(OFFS_REG(mem)))); |