Doug Evans [Wed, 15 Jun 1994 07:49:12 +0000 (07:49 +0000)]
sparc.c (SKIP_CALLERS_UNIMP_P): Define.
* sparc/sparc.c (SKIP_CALLERS_UNIMP_P): Define.
(sparc_cpu_type, sparc_arg_count, sparc_n_named_args,
frame_base_offset, fpconv_stack_temp): New globals.
(leaf_reg_remap): Add additional registers for v9.
(sparc_override_options, sparc64_init_expanders,
sparc64_fpconv_stack_temp, intreg_operand, ccfp_reg_operand,
data_segment_operand, text_segment_operand, v9_regcmp_op,
arith11_operand, arith10_operand, arith11_double_operand,
arith10_double_operand, gen_v9_scc, emit_v9_brxx_insn,
sparc_init_modes, build_big_number, output_v9branch,
sparc_initialize_trampoline, sparc64_initialize_trampoline):
New functions.
(arith_double_operand, gen_compare_reg, finalize_pic,
emit_move_sequence, mem_aligned_8, output_move_double,
output_move_quad, output_fp_move_double, output_fp_move_quad,
output_block_move, save_regs, restore_regs): Add v9 support.
(sparc_mode_class): New enum.
(*_MODES): Redefine to use it.
(hard_32bit_mode_classes): Renamed from hard_regno_mode_ok.
(hard_regno_mode_classes, hard_64bit_mode_classes,
sparc_mode_class): New globals.
(num_gfregs): Renamed from num_fregs.
(compute_frame_size): Add v9 support. Simplify calculations.
(output_function_prologue): Call build_big_number to compute stack
size in %g1, then adjust %sp.
Fix saving of call saved registers. Handle new v9 registers.
(output_function_epilogue): Fix restoration of call saved registers.
Handle new v9 registers.
Use SKIP_CALLERS_UNIMP_P to see if unimp insn is at return address.
(sparc_builtin_saveregs): Define v9 version.
(output_cbranch): New argument fp_cond_reg. All callers changed.
Add v9 support.
(output_return): Use SKIP_CALLERS_UNIMP_P.
(print_operand): New codes '_', '@', 'C', 'D'.
(output_double_int): Handle LABEL_REF and MINUS for v9.
Use ASM_LONGLONG if assembler can handle it.
(encode, decode): Use 4 HOST_WIDE_INTs for encoded value with HOST_BITS_PER_WIDE_INT/2 bits in each.
(encode, decode): Use 4 HOST_WIDE_INTs for encoded
value with HOST_BITS_PER_WIDE_INT/2 bits in each.
(LOWPART, HIGHPART): New macros.
(BASE): Move definition outside of div_and_round_double.
(add_double, mul_double, lshift_double, rshift_double): Rewrite.
(lrotate_double): Use LOWPART, HIGHPART, and BASE.
(rrotate_double): Likewise.
(div_and_round_double): Major changes to code for general case.
Now it actually produces non-garbage results for large operands.
(div_and_round_double): Simplify condition for special code used when
divisor < BASE.
(const_binop): Delete special cases for multiplying by 0, 1, 2, 4, 8.
(fold, case *_DIV_EXPR): Don't try to optimize for overflow.
Richard Kenner [Tue, 14 Jun 1994 21:37:25 +0000 (17:37 -0400)]
(reload): Process and accumulate needs for RELOAD_FOR_OPADDR_ADDR reloads.
When determining if needs are nongroup needs, ignore inactive or optional
reloads.
(reload_reg_used_in_op_addr_reload): New variable to indicate when reloads of
class RELOAD_FOR_OPADDR_ADDR are in use.
(mark_reload_reg_in_use, clear_reload_reg_in_use): Process reloads of class
RELOAD_FOR_OPADDR_ADDR.
(reload_reg_free_p, reload_reg_free_before_p, reloads_conflict): Likewise.
(choose_reload_regs): Support RELOAD_FOR_OPADDR_ADDR reloads.
(emit_reload_insns): Output RELOAD_FOR_OPADDR_ADDR reloads.
Jeff Law [Tue, 14 Jun 1994 19:18:43 +0000 (13:18 -0600)]
jump.c (condjump_in_parallel_p): New function to detect conditional jumps within PARALLEL insns.
* jump.c (condjump_in_parallel_p): New function to detect
conditional jumps within PARALLEL insns.
(jump_optimize): Allow for some simple optimizations involving
conditional jumps within PARALLEL insns.
* reorg.c (get_jump_flags): Handle conditional jumps in PARALLEL
insns.
(get_branch_condition, fill_simple_delay_slots): Likewise.
(fill_eager_delay_slots, relax_delay_slots, dbr_schedule): Likewise.
(storeinthi): Rework to generate better code, add big-endian support.
(movhi): Add support for processors running in big-endian mode.
(movhi_bigend): New expansion pattern.
(matcher for movhi): Remove incorrect alternative for storing HImode to
memory. Only match when little endian.
(matcher for big-endian movhi): New pattern.
(matchers/peepholes for extended pre/post increment HImode loads): Only
match when little-endian.
(matcher for operating on adjacent memory locations): Load operation into
arith operands.
(peepholes for extended pre-increment byte loads): New patterns.
(FUNCTION_PROFILER, TRAMPOLINE_TEMPLATE, ASM_FILE_START,
ASM_OUTPUT_REG_{PUSH,POP}, PRINT_OPERAND_ADDRESS): Output register
prefix before any explicit register name.
(ASM_OUTPUT_LONG_DOUBLE, ASM_OUTPUT_DOUBLE, ASM_OUTPUT_FLOAT,
ASM_OUTPUT_COMMON): Output the target dependent comment char.
(ASM_FILE_START): Output register alias for r10.
(REGISTER_NAMES): Output "r9" now for r9, not "rfp".
(ADDITIONAL_REGISTER_NAMES): New macro.
(all patterns explicitly mentioning regs): Output the target dependent prefix before the register name.
(all patterns explicitly mentioning regs): Output the target dependent
prefix before the register name.
(all patterns with asm comments): Output the target dependent comment
character.
(REGNO_OK_FOR_FP_P): Use FP_REGNO_P.
(FUNCTION_VALUE_REGNO_P): Revert last change.
(CONDITIONAL_REGISTER_USAGE): Partially revert last changes;
Make non-existing 1.0 fp regs fixed, including odd-numbered regs.
(enum reg_class): Get rid of SNAKE_FP_REGS, GENERAL_OR_SNAKE_FP_REGS,
and NON_SHIFT_REGS.
(REG_CLASS_NAMES): Likewise.
(REG_CLASS_CONTENTS): Likewise.
(REGNO_REG_CLASS): Likewise.
(HARD_REGNO_MODE_OK): Corresponding changes. Don't force 1.0 fpregs
to even regno here; fixed_regs does the job better.
(FP_REG_CLASS_P): Get rid of SNAKE_FP_REGS.
(REG_CLASS_FROM_LETTER, case 'z'): Deleted.
(HARD_REGNO_NREGS): Partially revert last change.
(CLASS_MAX_NREGS): Partially revert last change.
(all patterns): Update instruction generation to reflect new meanings of constraints 'K' and 'L'.
(all patterns): Update instruction generation to reflect new meanings of
constraints 'K' and 'L'. Use output type 'B' to generate an inverted
constant. Use output type 'N' to generate a negated floating point
constant.
Jeff Law [Fri, 3 Jun 1994 03:06:45 +0000 (21:06 -0600)]
pa.h (TARGET_SWITCHES): Enable TARGET_LONG_CALLS when TARGET_PORTABLE_RUNTIME is enabled.
* pa.h (TARGET_SWITCHES): Enable TARGET_LONG_CALLS when
TARGET_PORTABLE_RUNTIME is enabled.
* pa.c (output_call): If TARGET_LONG_CALLS is enabled, then emit
an inline long-call sequence.
* pa.md (millicode define_delay): Disable delay slots if
TARGET_LONG_CALLS.
(call_internal_reg, call_value_internal_reg): If TARGET_LONG_CALLS
is enabled, then emit an inline long-call sequence. Fix length
computation for TARGET_LONG_CALLS.
(millicode calls): Fix length computation for TARGET_LONG_CALLS.
output_func_{prologue,epilogue}): Change declarations from output_{prologue,epilogue}.
output_func_{prologue,epilogue}): Change declarations from
output_{prologue,epilogue}.
(arm_output_llc, arithmetic_instr, output_arithmetic_with_immediate_mulitply,
output_arithmetic_with_shift, shift_instr): Delete function definition.
(ASM_OUTPUT_OPCODE): Delete.
(PRINT_OPERAND): Call arm_print_operand, instead of doing it directly.
(ARM_COMMENT_CHAR): New macro.
(ARM_REG_PREFIX): New macro.
(ARM_SIGN_EXTEND): New macro.
(PRINT_OPERAND_ADDRESS): Delete variable 'shift'. Change type of
variable 'offset' to HOST_WIDE_INT. Use arm_print_operand to handle
shifts in addresses.
(const_pool_offset): Correct typo.
(fp_const_from_val): New function.
(all instruction generators): mark position where conditionalization
should be with '%?'.
(output_move_double): Use new '%m' output type to simplify load/store
multiple generation.
(shift_instr): Delete.
(shift_op): New function.
(output_arithmetic): Delete.
(output_arithmetic_with_shift): Delete.
(output_arithmetic_with_immediate_multiply): Delete.
(output_shifted_move): Delete.
(output_shift_compare): Delete.
(arm_print_operand): New function.
(FUNCTION_{PROLOGUE,EPILOGUE}): Call output_func_{prologue,epilogue}.
(INITIAL_ELIMINATION_OFFSET): Inspect all registers, but only add those that
are not in call_used_regs.
(GO_IF_LEGITIMATE_INDEX): Change type of variable 'range' to HOST_WIDE_INT;
change type of variable 'code' to enum rtx_code.
(arm_output_asm_insn): Delete declaration.
(EXTRA_CONSTAINT): New constraint 'R' to match a constant pool label.
Make constraint 'S' never match when not optimizing.
(CONSTANT_ADDRESS_P): Only accept symbols marked with SYMBOL_REF_FLAG when
optimizing.
(ENCODE_SECTION_INFO): Don't mark any symbols when not optimizing.
(ASM_OUTPUT_REG_{PUSH,POP}: Don't call arm_increase_location.
(ASM_OUTPUT_ADDR_VEC_ELT): Likewise.
#include <string.h>.
Declare some prototypes.
(output_memory_reference_mode): Change type to enum machine_mode.
(arm_const_nmoves): Delete.
(adjacent_mem_locations, {load,store}_multiple_operation): Explicitly
declare to return int.
({load,store}_multiple_operation): Change type of 'count', 'base' and 'i' to
HOST_WIDE_INT.
(output_add_immediate): Change type of 'n' to HOST_WIDE_INT.
(output_multi_immediate): Change type of Argument 'n' to HOST_WIDE_INT, if
wider than 32 bits, mask out the high bits.
(output_arithmetic_with_immediate_multiply): Change type of 'shift'
to HOST_WIDE_INT.
(output_func_{prologue,epilogue}): Renamed from output_{prologue,epilogue}.
Check all registers to see if they are live, but only push/pop them if they
are not in call_used_regs.
(const_pool_offset): New function.
(get_prologue_size): New function.
(output_func_prologue): Eliminate variable code_size.
(output_func_epilogue): Only call arm_increase_location when optimizing; also
add the size of the function just compiled, and the size of the prologue.
(arm_output_asm_insn): Delete. All callers changed to use output_asm_insn.
(arm_output_llc): Delete.
(output_load_symbol): New first parameter 'insn'. Rewrite so that assembler
can detect whether we made a mistake.
(emit_move_sequence): Add missing argument to symbolic_operand.
(output_block_move): Prefer addl to add.
(print_operand): Print fp regs with L suffix when appropriate.
(output_arg_descriptor): Rewrite to accomodate new convention for
USE information.
(compute_frame_size): Update for changed register numbers.
Simplify loops over general registers to stop at register 4.
(hppa_expand_prologue): Likewise.
(hppa_expand_epilogue): Likewise.