# HG changeset patch # User kvn # Date 1405636804 25200 # Node ID b384ba33c9a0600a3adad24ad3409101e9208927 # Parent 99f0593d8c9fff6579d0b5e927f27cf9facefd3b 8050942: PPC64: implement template interpreter for ppc64le Reviewed-by: kvn, goetz Contributed-by: asmundak@google.com diff -r 99f0593d8c9f -r b384ba33c9a0 src/cpu/ppc/vm/assembler_ppc.hpp --- a/src/cpu/ppc/vm/assembler_ppc.hpp Mon Sep 08 14:13:01 2014 +0000 +++ b/src/cpu/ppc/vm/assembler_ppc.hpp Thu Jul 17 15:40:04 2014 -0700 @@ -298,6 +298,7 @@ LWZ_OPCODE = (32u << OPCODE_SHIFT), LWZX_OPCODE = (31u << OPCODE_SHIFT | 23u << 1), LWZU_OPCODE = (33u << OPCODE_SHIFT), + LWBRX_OPCODE = (31u << OPCODE_SHIFT | 534 << 1), LHA_OPCODE = (42u << OPCODE_SHIFT), LHAX_OPCODE = (31u << OPCODE_SHIFT | 343u << 1), @@ -306,6 +307,7 @@ LHZ_OPCODE = (40u << OPCODE_SHIFT), LHZX_OPCODE = (31u << OPCODE_SHIFT | 279u << 1), LHZU_OPCODE = (41u << OPCODE_SHIFT), + LHBRX_OPCODE = (31u << OPCODE_SHIFT | 790 << 1), LBZ_OPCODE = (34u << OPCODE_SHIFT), LBZX_OPCODE = (31u << OPCODE_SHIFT | 87u << 1), @@ -1364,11 +1366,17 @@ inline void lwax( Register d, Register s1, Register s2); inline void lwa( Register d, int si16, Register s1); + // 4 bytes reversed + inline void lwbrx( Register d, Register s1, Register s2); + // 2 bytes inline void lhzx( Register d, Register s1, Register s2); inline void lhz( Register d, int si16, Register s1); inline void lhzu( Register d, int si16, Register s1); + // 2 bytes reversed + inline void lhbrx( Register d, Register s1, Register s2); + // 2 bytes inline void lhax( Register d, Register s1, Register s2); inline void lha( Register d, int si16, Register s1); @@ -1858,10 +1866,12 @@ inline void lwz( Register d, int si16); inline void lwax( Register d, Register s2); inline void lwa( Register d, int si16); + inline void lwbrx(Register d, Register s2); inline void lhzx( Register d, Register s2); inline void lhz( Register d, int si16); inline void lhax( Register d, Register s2); inline void lha( Register d, int si16); + inline void lhbrx(Register d, Register s2); inline void lbzx( Register d, Register s2); inline void lbz( Register d, int si16); inline void ldx( Register d, Register s2); diff -r 99f0593d8c9f -r b384ba33c9a0 src/cpu/ppc/vm/assembler_ppc.inline.hpp --- a/src/cpu/ppc/vm/assembler_ppc.inline.hpp Mon Sep 08 14:13:01 2014 +0000 +++ b/src/cpu/ppc/vm/assembler_ppc.inline.hpp Thu Jul 17 15:40:04 2014 -0700 @@ -263,10 +263,14 @@ inline void Assembler::lwax( Register d, Register s1, Register s2) { emit_int32(LWAX_OPCODE | rt(d) | ra0mem(s1) | rb(s2));} inline void Assembler::lwa( Register d, int si16, Register s1) { emit_int32(LWA_OPCODE | rt(d) | ds(si16) | ra0mem(s1));} +inline void Assembler::lwbrx( Register d, Register s1, Register s2) { emit_int32(LWBRX_OPCODE | rt(d) | ra0mem(s1) | rb(s2));} + inline void Assembler::lhzx( Register d, Register s1, Register s2) { emit_int32(LHZX_OPCODE | rt(d) | ra0mem(s1) | rb(s2));} inline void Assembler::lhz( Register d, int si16, Register s1) { emit_int32(LHZ_OPCODE | rt(d) | d1(si16) | ra0mem(s1));} inline void Assembler::lhzu( Register d, int si16, Register s1) { assert(d != s1, "according to ibm manual"); emit_int32(LHZU_OPCODE | rt(d) | d1(si16) | rta0mem(s1));} +inline void Assembler::lhbrx( Register d, Register s1, Register s2) { emit_int32(LHBRX_OPCODE | rt(d) | ra0mem(s1) | rb(s2));} + inline void Assembler::lhax( Register d, Register s1, Register s2) { emit_int32(LHAX_OPCODE | rt(d) | ra0mem(s1) | rb(s2));} inline void Assembler::lha( Register d, int si16, Register s1) { emit_int32(LHA_OPCODE | rt(d) | d1(si16) | ra0mem(s1));} inline void Assembler::lhau( Register d, int si16, Register s1) { assert(d != s1, "according to ibm manual"); emit_int32(LHAU_OPCODE | rt(d) | d1(si16) | rta0mem(s1));} @@ -736,10 +740,12 @@ inline void Assembler::lwz( Register d, int si16 ) { emit_int32( LWZ_OPCODE | rt(d) | d1(si16));} inline void Assembler::lwax( Register d, Register s2) { emit_int32( LWAX_OPCODE | rt(d) | rb(s2));} inline void Assembler::lwa( Register d, int si16 ) { emit_int32( LWA_OPCODE | rt(d) | ds(si16));} +inline void Assembler::lwbrx(Register d, Register s2) { emit_int32( LWBRX_OPCODE| rt(d) | rb(s2));} inline void Assembler::lhzx( Register d, Register s2) { emit_int32( LHZX_OPCODE | rt(d) | rb(s2));} inline void Assembler::lhz( Register d, int si16 ) { emit_int32( LHZ_OPCODE | rt(d) | d1(si16));} inline void Assembler::lhax( Register d, Register s2) { emit_int32( LHAX_OPCODE | rt(d) | rb(s2));} inline void Assembler::lha( Register d, int si16 ) { emit_int32( LHA_OPCODE | rt(d) | d1(si16));} +inline void Assembler::lhbrx(Register d, Register s2) { emit_int32( LHBRX_OPCODE| rt(d) | rb(s2));} inline void Assembler::lbzx( Register d, Register s2) { emit_int32( LBZX_OPCODE | rt(d) | rb(s2));} inline void Assembler::lbz( Register d, int si16 ) { emit_int32( LBZ_OPCODE | rt(d) | d1(si16));} inline void Assembler::ld( Register d, int si16 ) { emit_int32( LD_OPCODE | rt(d) | ds(si16));} diff -r 99f0593d8c9f -r b384ba33c9a0 src/cpu/ppc/vm/interp_masm_ppc_64.cpp --- a/src/cpu/ppc/vm/interp_masm_ppc_64.cpp Mon Sep 08 14:13:01 2014 +0000 +++ b/src/cpu/ppc/vm/interp_masm_ppc_64.cpp Thu Jul 17 15:40:04 2014 -0700 @@ -119,9 +119,15 @@ // Call the Interpreter::remove_activation_preserving_args_entry() // func to get the address of the same-named entrypoint in the // generated interpreter code. +#if defined(ABI_ELFv2) + call_c(CAST_FROM_FN_PTR(address, + Interpreter::remove_activation_preserving_args_entry), + relocInfo::none); +#else call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, Interpreter::remove_activation_preserving_args_entry), relocInfo::none); +#endif // Jump to Interpreter::_remove_activation_preserving_args_entry. mtctr(R3_RET); @@ -331,29 +337,40 @@ void InterpreterMacroAssembler::get_2_byte_integer_at_bcp(int bcp_offset, Register Rdst, signedOrNot is_signed) { +#if defined(VM_LITTLE_ENDIAN) + if (bcp_offset) { + load_const_optimized(Rdst, bcp_offset); + lhbrx(Rdst, R14_bcp, Rdst); + } else { + lhbrx(Rdst, R14_bcp); + } + if (is_signed == Signed) { + extsh(Rdst, Rdst); + } +#else // Read Java big endian format. if (is_signed == Signed) { lha(Rdst, bcp_offset, R14_bcp); } else { lhz(Rdst, bcp_offset, R14_bcp); } -#if 0 - assert(Rtmp != Rdst, "need separate temp register"); - Register Rfirst = Rtmp; - lbz(Rfirst, bcp_offset, R14_bcp); // first byte - lbz(Rdst, bcp_offset+1, R14_bcp); // second byte - - // Rdst = ((Rfirst<<8) & 0xFF00) | (Rdst &~ 0xFF00) - rldimi(/*RA=*/Rdst, /*RS=*/Rfirst, /*sh=*/8, /*mb=*/48); - if (is_signed == Signed) { - extsh(Rdst, Rdst); - } #endif } void InterpreterMacroAssembler::get_4_byte_integer_at_bcp(int bcp_offset, Register Rdst, signedOrNot is_signed) { +#if defined(VM_LITTLE_ENDIAN) + if (bcp_offset) { + load_const_optimized(Rdst, bcp_offset); + lwbrx(Rdst, R14_bcp, Rdst); + } else { + lwbrx(Rdst, R14_bcp); + } + if (is_signed == Signed) { + extsw(Rdst, Rdst); + } +#else // Read Java big endian format. if (bcp_offset & 3) { // Offset unaligned? load_const_optimized(Rdst, bcp_offset); @@ -369,19 +386,27 @@ lwz(Rdst, bcp_offset, R14_bcp); } } +#endif } + // Load the constant pool cache index from the bytecode stream. // // Kills / writes: // - Rdst, Rscratch void InterpreterMacroAssembler::get_cache_index_at_bcp(Register Rdst, int bcp_offset, size_t index_size) { assert(bcp_offset > 0, "bcp is still pointing to start of bytecode"); + // Cache index is always in the native format, courtesy of Rewriter. if (index_size == sizeof(u2)) { - get_2_byte_integer_at_bcp(bcp_offset, Rdst, Unsigned); + lhz(Rdst, bcp_offset, R14_bcp); } else if (index_size == sizeof(u4)) { assert(EnableInvokeDynamic, "giant index used only for JSR 292"); - get_4_byte_integer_at_bcp(bcp_offset, Rdst, Signed); + if (bcp_offset & 3) { + load_const_optimized(Rdst, bcp_offset); + lwax(Rdst, R14_bcp, Rdst); + } else { + lwa(Rdst, bcp_offset, R14_bcp); + } assert(ConstantPool::decode_invokedynamic_index(~123) == 123, "else change next line"); nand(Rdst, Rdst, Rdst); // convert to plain index } else if (index_size == sizeof(u1)) { @@ -398,6 +423,29 @@ add(cache, R27_constPoolCache, cache); } +// Load 4-byte signed or unsigned integer in Java format (that is, big-endian format) +// from (Rsrc)+offset. +void InterpreterMacroAssembler::get_u4(Register Rdst, Register Rsrc, int offset, + signedOrNot is_signed) { +#if defined(VM_LITTLE_ENDIAN) + if (offset) { + load_const_optimized(Rdst, offset); + lwbrx(Rdst, Rdst, Rsrc); + } else { + lwbrx(Rdst, Rsrc); + } + if (is_signed == Signed) { + extsw(Rdst, Rdst); + } +#else + if (is_signed == Signed) { + lwa(Rdst, offset, Rsrc); + } else { + lwz(Rdst, offset, Rsrc); + } +#endif +} + // Load object from cpool->resolved_references(index). void InterpreterMacroAssembler::load_resolved_reference_at_index(Register result, Register index) { assert_different_registers(result, index); diff -r 99f0593d8c9f -r b384ba33c9a0 src/cpu/ppc/vm/interp_masm_ppc_64.hpp --- a/src/cpu/ppc/vm/interp_masm_ppc_64.hpp Mon Sep 08 14:13:01 2014 +0000 +++ b/src/cpu/ppc/vm/interp_masm_ppc_64.hpp Thu Jul 17 15:40:04 2014 -0700 @@ -130,6 +130,7 @@ void get_cache_and_index_at_bcp(Register cache, int bcp_offset, size_t index_size = sizeof(u2)); + void get_u4(Register Rdst, Register Rsrc, int offset, signedOrNot is_signed); // common code diff -r 99f0593d8c9f -r b384ba33c9a0 src/cpu/ppc/vm/templateInterpreter_ppc.cpp --- a/src/cpu/ppc/vm/templateInterpreter_ppc.cpp Mon Sep 08 14:13:01 2014 +0000 +++ b/src/cpu/ppc/vm/templateInterpreter_ppc.cpp Thu Jul 17 15:40:04 2014 -0700 @@ -175,8 +175,12 @@ const Register size = R12_scratch2; __ get_cache_and_index_at_bcp(cache, 1, index_size); - // Big Endian (get least significant byte of 64 bit value): + // Get least significant byte of 64 bit value: +#if defined(VM_LITTLE_ENDIAN) + __ lbz(size, in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset()), cache); +#else __ lbz(size, in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset()) + 7, cache); +#endif __ sldi(size, size, Interpreter::logStackElementSize); __ add(R15_esp, R15_esp, size); __ dispatch_next(state, step); @@ -857,7 +861,9 @@ // Our signature handlers copy required arguments to the C stack // (outgoing C args), R3_ARG1 to R10_ARG8, and FARG1 to FARG13. __ mr(R3_ARG1, R18_locals); +#if !defined(ABI_ELFv2) __ ld(signature_handler_fd, 0, signature_handler_fd); +#endif __ call_stub(signature_handler_fd); @@ -1019,8 +1025,13 @@ // native result across the call. No oop is present. __ mr(R3_ARG1, R16_thread); +#if defined(ABI_ELFv2) + __ call_c(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans), + relocInfo::none); +#else __ call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, JavaThread::check_special_condition_for_native_trans), relocInfo::none); +#endif __ bind(sync_check_done); diff -r 99f0593d8c9f -r b384ba33c9a0 src/cpu/ppc/vm/templateTable_ppc_64.cpp --- a/src/cpu/ppc/vm/templateTable_ppc_64.cpp Mon Sep 08 14:13:01 2014 +0000 +++ b/src/cpu/ppc/vm/templateTable_ppc_64.cpp Thu Jul 17 15:40:04 2014 -0700 @@ -188,8 +188,12 @@ assert(byte_no == f1_byte || byte_no == f2_byte, "byte_no out of range"); assert(load_bc_into_bc_reg, "we use bc_reg as temp"); __ get_cache_and_index_at_bcp(Rtemp /* dst = cache */, 1); - // Big Endian: ((*(cache+indices))>>((1+byte_no)*8))&0xFF + // ((*(cache+indices))>>((1+byte_no)*8))&0xFF: +#if defined(VM_LITTLE_ENDIAN) + __ lbz(Rnew_bc, in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::indices_offset()) + 1 + byte_no, Rtemp); +#else __ lbz(Rnew_bc, in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::indices_offset()) + 7 - (1 + byte_no), Rtemp); +#endif __ cmpwi(CCR0, Rnew_bc, 0); __ li(Rnew_bc, (unsigned int)(unsigned char)new_bc); __ beq(CCR0, L_patch_done); @@ -1838,8 +1842,8 @@ __ clrrdi(Rdef_offset_addr, Rdef_offset_addr, log2_long((jlong)BytesPerInt)); // Load lo & hi. - __ lwz(Rlow_byte, BytesPerInt, Rdef_offset_addr); - __ lwz(Rhigh_byte, BytesPerInt * 2, Rdef_offset_addr); + __ get_u4(Rlow_byte, Rdef_offset_addr, BytesPerInt, InterpreterMacroAssembler::Unsigned); + __ get_u4(Rhigh_byte, Rdef_offset_addr, 2 *BytesPerInt, InterpreterMacroAssembler::Unsigned); // Check for default case (=index outside [low,high]). __ cmpw(CCR0, R17_tos, Rlow_byte); @@ -1853,12 +1857,17 @@ __ profile_switch_case(Rindex, Rhigh_byte /* scratch */, Rscratch1, Rscratch2); __ sldi(Rindex, Rindex, LogBytesPerInt); __ addi(Rindex, Rindex, 3 * BytesPerInt); +#if defined(VM_LITTLE_ENDIAN) + __ lwbrx(Roffset, Rdef_offset_addr, Rindex); + __ extsw(Roffset, Roffset); +#else __ lwax(Roffset, Rdef_offset_addr, Rindex); +#endif __ b(Ldispatch); __ bind(Ldefault_case); __ profile_switch_default(Rhigh_byte, Rscratch1); - __ lwa(Roffset, 0, Rdef_offset_addr); + __ get_u4(Roffset, Rdef_offset_addr, 0, InterpreterMacroAssembler::Signed); __ bind(Ldispatch); @@ -1874,12 +1883,11 @@ // Table switch using linear search through cases. // Bytecode stream format: // Bytecode (1) | 4-byte padding | default offset (4) | count (4) | value/offset pair1 (8) | value/offset pair2 (8) | ... -// Note: Everything is big-endian format here. So on little endian machines, we have to revers offset and count and cmp value. +// Note: Everything is big-endian format here. void TemplateTable::fast_linearswitch() { transition(itos, vtos); - Label Lloop_entry, Lsearch_loop, Lfound, Lcontinue_execution, Ldefault_case; - + Label Lloop_entry, Lsearch_loop, Lcontinue_execution, Ldefault_case; Register Rcount = R3_ARG1, Rcurrent_pair = R4_ARG2, Rdef_offset_addr = R5_ARG3, // Is going to contain address of default offset. @@ -1893,47 +1901,40 @@ __ clrrdi(Rdef_offset_addr, Rdef_offset_addr, log2_long((jlong)BytesPerInt)); // Setup loop counter and limit. - __ lwz(Rcount, BytesPerInt, Rdef_offset_addr); // Load count. + __ get_u4(Rcount, Rdef_offset_addr, BytesPerInt, InterpreterMacroAssembler::Unsigned); __ addi(Rcurrent_pair, Rdef_offset_addr, 2 * BytesPerInt); // Rcurrent_pair now points to first pair. - // Set up search loop. - __ cmpwi(CCR0, Rcount, 0); - __ beq(CCR0, Ldefault_case); - __ mtctr(Rcount); - - // linear table search - __ bind(Lsearch_loop); - - __ lwz(Rvalue, 0, Rcurrent_pair); - __ lwa(Roffset, 1 * BytesPerInt, Rcurrent_pair); - - __ cmpw(CCR0, Rvalue, Rcmp_value); - __ beq(CCR0, Lfound); - - __ addi(Rcurrent_pair, Rcurrent_pair, 2 * BytesPerInt); - __ bdnz(Lsearch_loop); - - // default case + __ cmpwi(CCR0, Rcount, 0); + __ bne(CCR0, Lloop_entry); + + // Default case __ bind(Ldefault_case); - - __ lwa(Roffset, 0, Rdef_offset_addr); + __ get_u4(Roffset, Rdef_offset_addr, 0, InterpreterMacroAssembler::Signed); if (ProfileInterpreter) { __ profile_switch_default(Rdef_offset_addr, Rcount/* scratch */); - __ b(Lcontinue_execution); } - - // Entry found, skip Roffset bytecodes and continue. - __ bind(Lfound); + __ b(Lcontinue_execution); + + // Next iteration + __ bind(Lsearch_loop); + __ bdz(Ldefault_case); + __ addi(Rcurrent_pair, Rcurrent_pair, 2 * BytesPerInt); + __ bind(Lloop_entry); + __ get_u4(Rvalue, Rcurrent_pair, 0, InterpreterMacroAssembler::Unsigned); + __ cmpw(CCR0, Rvalue, Rcmp_value); + __ bne(CCR0, Lsearch_loop); + + // Found, load offset. + __ get_u4(Roffset, Rcurrent_pair, BytesPerInt, InterpreterMacroAssembler::Signed); + // Calculate case index and profile + __ mfctr(Rcurrent_pair); if (ProfileInterpreter) { - // Calc the num of the pair we hit. Careful, Rcurrent_pair points 2 ints - // beyond the actual current pair due to the auto update load above! - __ sub(Rcurrent_pair, Rcurrent_pair, Rdef_offset_addr); - __ addi(Rcurrent_pair, Rcurrent_pair, - 2 * BytesPerInt); - __ srdi(Rcurrent_pair, Rcurrent_pair, LogBytesPerInt + 1); + __ sub(Rcurrent_pair, Rcount, Rcurrent_pair); __ profile_switch_case(Rcurrent_pair, Rcount /*scratch*/, Rdef_offset_addr/*scratch*/, Rscratch); - __ bind(Lcontinue_execution); } + + __ bind(Lcontinue_execution); __ add(R14_bcp, Roffset, R14_bcp); __ dispatch_next(vtos); } @@ -1989,7 +1990,7 @@ // initialize i & j __ li(Ri,0); - __ lwz(Rj, -BytesPerInt, Rarray); + __ get_u4(Rj, Rarray, -BytesPerInt, InterpreterMacroAssembler::Unsigned); // and start. Label entry; @@ -2006,7 +2007,11 @@ // i = h; // } __ sldi(Rscratch, Rh, log_entry_size); +#if defined(VM_LITTLE_ENDIAN) + __ lwbrx(Rscratch, Rscratch, Rarray); +#else __ lwzx(Rscratch, Rscratch, Rarray); +#endif // if (key < current value) // Rh = Rj @@ -2038,20 +2043,20 @@ // Ri = value offset __ sldi(Ri, Ri, log_entry_size); __ add(Ri, Ri, Rarray); - __ lwz(Rscratch, 0, Ri); + __ get_u4(Rscratch, Ri, 0, InterpreterMacroAssembler::Unsigned); Label not_found; // Ri = offset offset __ cmpw(CCR0, Rkey, Rscratch); __ beq(CCR0, not_found); // entry not found -> j = default offset - __ lwz(Rj, -2 * BytesPerInt, Rarray); + __ get_u4(Rj, Rarray, -2 * BytesPerInt, InterpreterMacroAssembler::Unsigned); __ b(default_case); __ bind(not_found); // entry found -> j = offset __ profile_switch_case(Rh, Rj, Rscratch, Rkey); - __ lwz(Rj, BytesPerInt, Ri); + __ get_u4(Rj, Ri, BytesPerInt, InterpreterMacroAssembler::Unsigned); if (ProfileInterpreter) { __ b(continue_execution); @@ -2146,8 +2151,11 @@ assert(byte_no == f1_byte || byte_no == f2_byte, "byte_no out of range"); // We are resolved if the indices offset contains the current bytecode. - // Big Endian: +#if defined(VM_LITTLE_ENDIAN) + __ lbz(Rscratch, in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::indices_offset()) + byte_no + 1, Rcache); +#else __ lbz(Rscratch, in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::indices_offset()) + 7 - (byte_no + 1), Rcache); +#endif // Acquire by cmp-br-isync (see below). __ cmpdi(CCR0, Rscratch, (int)bytecode()); __ beq(CCR0, Lresolved);