Projects
openEuler:24.03:SP1:Everything
gcc
_service:tar_scm:0156-Backport-SME-aarch64-Add-...
Sign Up
Log In
Username
Password
Overview
Repositories
Revisions
Requests
Users
Attributes
Meta
File _service:tar_scm:0156-Backport-SME-aarch64-Add-backend-support-for-DFP.patch of Package gcc
From 8394394bd26c7be6129b9a4e673d2a3530d9efde Mon Sep 17 00:00:00 2001 From: Christophe Lyon <christophe.lyon@arm.com> Date: Fri, 11 Mar 2022 16:21:02 +0000 Subject: [PATCH 057/157] [Backport][SME] aarch64: Add backend support for DFP Reference: https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=0dc8e1e7026d9b8ec8b669c051786d426a52cd22 This patch updates the aarch64 backend as needed to support DFP modes (SD, DD and TD). Changes v1->v2: * Drop support for DFP modes in aarch64_gen_{load||store}[wb]_pair as these are only used in prologue/epilogue where DFP modes are not used. Drop the changes to the corresponding patterns in aarch64.md, and useless GPF_PAIR iterator. * In aarch64_reinterpret_float_as_int, handle DDmode the same way as DFmode (needed in case the representation of the floating-point value can be loaded using mov/movk. * In aarch64_float_const_zero_rtx_p, reject constants with DFP mode: when X is zero, the callers want to emit either '0' or 'zr' depending on the context, which is not the way 0.0 is represented in DFP mode (in particular fmov d0, #0 is not right for DFP). * In aarch64_legitimate_constant_p, accept DFP 2022-03-31 Christophe Lyon <christophe.lyon@arm.com> gcc/ * config/aarch64/aarch64.cc (aarch64_split_128bit_move): Handle DFP modes. (aarch64_mode_valid_for_sched_fusion_p): Likewise. (aarch64_classify_address): Likewise. (aarch64_legitimize_address_displacement): Likewise. (aarch64_reinterpret_float_as_int): Likewise. (aarch64_float_const_zero_rtx_p): Likewise. (aarch64_can_const_movi_rtx_p): Likewise. (aarch64_anchor_offset): Likewise. (aarch64_secondary_reload): Likewise. (aarch64_rtx_costs): Likewise. (aarch64_legitimate_constant_p): Likewise. (aarch64_gimplify_va_arg_expr): Likewise. (aapcs_vfp_sub_candidate): Likewise. (aarch64_vfp_is_call_or_return_candidate): Likewise. (aarch64_output_scalar_simd_mov_immediate): Likewise. (aarch64_gen_adjusted_ldpstp): Likewise. (aarch64_scalar_mode_supported_p): Accept DFP modes if enabled. * config/aarch64/aarch64.md (movsf_aarch64): Use SFD iterator and rename into mov<mode>_aarch64. (movdf_aarch64): Use DFD iterator and rename into mov<mode>_aarch64. (movtf_aarch64): Use TFD iterator and rename into mov<mode>_aarch64. (split pattern for move TF mode): Use TFD iterator. * config/aarch64/iterators.md (GPF_TF_F16_MOV): Add DFP modes. (SFD, DFD, TFD): New iterators. (GPF_TF): Add DFP modes. (TX, DX, DX2): Likewise. --- gcc/config/aarch64/aarch64.cc | 82 ++++++++++++++++++++++----------- gcc/config/aarch64/aarch64.md | 34 +++++++------- gcc/config/aarch64/iterators.md | 24 +++++++--- 3 files changed, 89 insertions(+), 51 deletions(-) diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc index 055b436b1..02210ed13 100644 --- a/gcc/config/aarch64/aarch64.cc +++ b/gcc/config/aarch64/aarch64.cc @@ -5068,7 +5068,7 @@ aarch64_split_128bit_move (rtx dst, rtx src) machine_mode mode = GET_MODE (dst); - gcc_assert (mode == TImode || mode == TFmode); + gcc_assert (mode == TImode || mode == TFmode || mode == TDmode); gcc_assert (!(side_effects_p (src) || side_effects_p (dst))); gcc_assert (mode == GET_MODE (src) || GET_MODE (src) == VOIDmode); @@ -10834,6 +10834,7 @@ aarch64_mode_valid_for_sched_fusion_p (machine_mode mode) { return mode == SImode || mode == DImode || mode == SFmode || mode == DFmode + || mode == SDmode || mode == DDmode || (aarch64_vector_mode_supported_p (mode) && (known_eq (GET_MODE_SIZE (mode), 8) || (known_eq (GET_MODE_SIZE (mode), 16) @@ -10876,12 +10877,13 @@ aarch64_classify_address (struct aarch64_address_info *info, vec_flags &= ~VEC_PARTIAL; /* On BE, we use load/store pair for all large int mode load/stores. - TI/TFmode may also use a load/store pair. */ + TI/TF/TDmode may also use a load/store pair. */ bool advsimd_struct_p = (vec_flags == (VEC_ADVSIMD | VEC_STRUCT)); bool load_store_pair_p = (type == ADDR_QUERY_LDP_STP || type == ADDR_QUERY_LDP_STP_N || mode == TImode || mode == TFmode + || mode == TDmode || (BYTES_BIG_ENDIAN && advsimd_struct_p)); /* If we are dealing with ADDR_QUERY_LDP_STP_N that means the incoming mode corresponds to the actual size of the memory being loaded/stored and the @@ -10955,7 +10957,7 @@ aarch64_classify_address (struct aarch64_address_info *info, info->offset = op1; info->const_offset = offset; - /* TImode and TFmode values are allowed in both pairs of X + /* TImode, TFmode and TDmode values are allowed in both pairs of X registers and individual Q registers. The available address modes are: X,X: 7-bit signed scaled offset @@ -10964,7 +10966,7 @@ aarch64_classify_address (struct aarch64_address_info *info, When performing the check for pairs of X registers i.e. LDP/STP pass down DImode since that is the natural size of the LDP/STP instruction memory accesses. */ - if (mode == TImode || mode == TFmode) + if (mode == TImode || mode == TFmode || mode == TDmode) return (aarch64_offset_7bit_signed_scaled_p (DImode, offset) && (aarch64_offset_9bit_signed_unscaled_p (mode, offset) || offset_12bit_unsigned_scaled_p (mode, offset))); @@ -11087,14 +11089,14 @@ aarch64_classify_address (struct aarch64_address_info *info, info->offset = XEXP (XEXP (x, 1), 1); info->const_offset = offset; - /* TImode and TFmode values are allowed in both pairs of X + /* TImode, TFmode and TDmode values are allowed in both pairs of X registers and individual Q registers. The available address modes are: X,X: 7-bit signed scaled offset Q: 9-bit signed offset We conservatively require an offset representable in either mode. */ - if (mode == TImode || mode == TFmode) + if (mode == TImode || mode == TFmode || mode == TDmode) return (aarch64_offset_7bit_signed_scaled_p (mode, offset) && aarch64_offset_9bit_signed_unscaled_p (mode, offset)); @@ -11256,9 +11258,9 @@ aarch64_legitimize_address_displacement (rtx *offset1, rtx *offset2, offset. Use 4KB range for 1- and 2-byte accesses and a 16KB range otherwise to increase opportunities for sharing the base address of different sizes. Unaligned accesses use the signed - 9-bit range, TImode/TFmode use the intersection of signed + 9-bit range, TImode/TFmode/TDmode use the intersection of signed scaled 7-bit and signed 9-bit offset. */ - if (mode == TImode || mode == TFmode) + if (mode == TImode || mode == TFmode || mode == TDmode) second_offset = ((const_offset + 0x100) & 0x1f8) - 0x100; else if ((const_offset & (size - 1)) != 0) second_offset = ((const_offset + 0x100) & 0x1ff) - 0x100; @@ -11339,7 +11341,7 @@ aarch64_reinterpret_float_as_int (rtx value, unsigned HOST_WIDE_INT *intval) CONST_DOUBLE_REAL_VALUE (value), REAL_MODE_FORMAT (mode)); - if (mode == DFmode) + if (mode == DFmode || mode == DDmode) { int order = BYTES_BIG_ENDIAN ? 1 : 0; ival = zext_hwi (res[order], 32); @@ -11380,11 +11382,15 @@ aarch64_float_const_rtx_p (rtx x) return false; } -/* Return TRUE if rtx X is immediate constant 0.0 */ +/* Return TRUE if rtx X is immediate constant 0.0 (but not in Decimal + Floating Point). */ bool aarch64_float_const_zero_rtx_p (rtx x) { - if (GET_MODE (x) == VOIDmode) + /* 0.0 in Decimal Floating Point cannot be represented by #0 or + zr as our callers expect, so no need to check the actual + value if X is of Decimal Floating Point type. */ + if (GET_MODE_CLASS (GET_MODE (x)) == MODE_DECIMAL_FLOAT) return false; if (REAL_VALUE_MINUS_ZERO (*CONST_DOUBLE_REAL_VALUE (x))) @@ -11422,7 +11428,7 @@ aarch64_can_const_movi_rtx_p (rtx x, machine_mode mode) else return false; - /* use a 64 bit mode for everything except for DI/DF mode, where we use + /* use a 64 bit mode for everything except for DI/DF/DD mode, where we use a 128 bit vector mode. */ int width = GET_MODE_BITSIZE (imode) == 64 ? 128 : 64; @@ -12628,7 +12634,7 @@ aarch64_anchor_offset (HOST_WIDE_INT offset, HOST_WIDE_INT size, if (IN_RANGE (offset, -256, 0)) return 0; - if (mode == TImode || mode == TFmode) + if (mode == TImode || mode == TFmode || mode == TDmode) return (offset + 0x100) & ~0x1ff; /* Use 12-bit offset by access size. */ @@ -12737,7 +12743,9 @@ aarch64_secondary_reload (bool in_p ATTRIBUTE_UNUSED, rtx x, /* Without the TARGET_SIMD instructions we cannot move a Q register to a Q register directly. We need a scratch. */ - if (REG_P (x) && (mode == TFmode || mode == TImode) && mode == GET_MODE (x) + if (REG_P (x) + && (mode == TFmode || mode == TImode || mode == TDmode) + && mode == GET_MODE (x) && FP_REGNUM_P (REGNO (x)) && !TARGET_SIMD && reg_class_subset_p (rclass, FP_REGS)) { @@ -12745,14 +12753,16 @@ aarch64_secondary_reload (bool in_p ATTRIBUTE_UNUSED, rtx x, return NO_REGS; } - /* A TFmode or TImode memory access should be handled via an FP_REGS + /* A TFmode, TImode or TDmode memory access should be handled via an FP_REGS because AArch64 has richer addressing modes for LDR/STR instructions than LDP/STP instructions. */ if (TARGET_FLOAT && rclass == GENERAL_REGS && known_eq (GET_MODE_SIZE (mode), 16) && MEM_P (x)) return FP_REGS; - if (rclass == FP_REGS && (mode == TImode || mode == TFmode) && CONSTANT_P(x)) + if (rclass == FP_REGS + && (mode == TImode || mode == TFmode || mode == TDmode) + && CONSTANT_P(x)) return GENERAL_REGS; return NO_REGS; @@ -13883,9 +13893,9 @@ aarch64_rtx_costs (rtx x, machine_mode mode, int outer ATTRIBUTE_UNUSED, *cost += extra_cost->ldst.storev; else if (GET_MODE_CLASS (mode) == MODE_INT) *cost += extra_cost->ldst.store; - else if (mode == SFmode) + else if (mode == SFmode || mode == SDmode) *cost += extra_cost->ldst.storef; - else if (mode == DFmode) + else if (mode == DFmode || mode == DDmode) *cost += extra_cost->ldst.stored; *cost += @@ -14009,11 +14019,11 @@ aarch64_rtx_costs (rtx x, machine_mode mode, int outer ATTRIBUTE_UNUSED, /* mov[df,sf]_aarch64. */ if (aarch64_float_const_representable_p (x)) /* FMOV (scalar immediate). */ - *cost += extra_cost->fp[mode == DFmode].fpconst; + *cost += extra_cost->fp[mode == DFmode || mode == DDmode].fpconst; else if (!aarch64_float_const_zero_rtx_p (x)) { /* This will be a load from memory. */ - if (mode == DFmode) + if (mode == DFmode || mode == DDmode) *cost += extra_cost->ldst.loadd; else *cost += extra_cost->ldst.loadf; @@ -14039,9 +14049,9 @@ aarch64_rtx_costs (rtx x, machine_mode mode, int outer ATTRIBUTE_UNUSED, *cost += extra_cost->ldst.loadv; else if (GET_MODE_CLASS (mode) == MODE_INT) *cost += extra_cost->ldst.load; - else if (mode == SFmode) + else if (mode == SFmode || mode == SDmode) *cost += extra_cost->ldst.loadf; - else if (mode == DFmode) + else if (mode == DFmode || mode == DDmode) *cost += extra_cost->ldst.loadd; *cost += @@ -19623,7 +19633,7 @@ aarch64_legitimate_constant_p (machine_mode mode, rtx x) { /* Support CSE and rematerialization of common constants. */ if (CONST_INT_P (x) - || (CONST_DOUBLE_P (x) && GET_MODE_CLASS (mode) == MODE_FLOAT)) + || CONST_DOUBLE_P (x)) return true; /* Only accept variable-length vector constants if they can be @@ -20064,6 +20074,18 @@ aarch64_gimplify_va_arg_expr (tree valist, tree type, gimple_seq *pre_p, field_t = long_double_type_node; field_ptr_t = long_double_ptr_type_node; break; + case SDmode: + field_t = dfloat32_type_node; + field_ptr_t = build_pointer_type (dfloat32_type_node); + break; + case DDmode: + field_t = dfloat64_type_node; + field_ptr_t = build_pointer_type (dfloat64_type_node); + break; + case TDmode: + field_t = dfloat128_type_node; + field_ptr_t = build_pointer_type (dfloat128_type_node); + break; case E_HFmode: field_t = aarch64_fp16_type_node; field_ptr_t = aarch64_fp16_ptr_type_node; @@ -20315,7 +20337,8 @@ aapcs_vfp_sub_candidate (const_tree type, machine_mode *modep, case REAL_TYPE: mode = TYPE_MODE (type); if (mode != DFmode && mode != SFmode - && mode != TFmode && mode != HFmode) + && mode != TFmode && mode != HFmode + && mode != SDmode && mode != DDmode && mode != TDmode) return -1; if (*modep == VOIDmode) @@ -20631,7 +20654,9 @@ aarch64_vfp_is_call_or_return_candidate (machine_mode mode, machine_mode new_mode = VOIDmode; bool composite_p = aarch64_composite_type_p (type, mode); - if ((!composite_p && GET_MODE_CLASS (mode) == MODE_FLOAT) + if ((!composite_p + && (GET_MODE_CLASS (mode) == MODE_FLOAT + || GET_MODE_CLASS (mode) == MODE_DECIMAL_FLOAT)) || aarch64_short_vector_p (type, mode)) { *count = 1; @@ -23565,7 +23590,7 @@ aarch64_output_scalar_simd_mov_immediate (rtx immediate, scalar_int_mode mode) } machine_mode vmode; - /* use a 64 bit mode for everything except for DI/DF mode, where we use + /* use a 64 bit mode for everything except for DI/DF/DD mode, where we use a 128 bit vector mode. */ int width = GET_MODE_BITSIZE (mode) == 64 ? 128 : 64; @@ -26417,7 +26442,7 @@ aarch64_gen_adjusted_ldpstp (rtx *operands, bool load, base_off = (off_val_1 + off_val_3) / 2; else /* However, due to issues with negative LDP/STP offset generation for - larger modes, for DF, DI and vector modes. we must not use negative + larger modes, for DF, DD, DI and vector modes. we must not use negative addresses smaller than 9 signed unadjusted bits can store. This provides the most range in this case. */ base_off = off_val_1; @@ -26695,6 +26720,9 @@ aarch64_libgcc_floating_mode_supported_p (scalar_float_mode mode) static bool aarch64_scalar_mode_supported_p (scalar_mode mode) { + if (DECIMAL_FLOAT_MODE_P (mode)) + return default_decimal_float_supported_p (); + return (mode == HFmode ? true : default_scalar_mode_supported_p (mode)); diff --git a/gcc/config/aarch64/aarch64.md b/gcc/config/aarch64/aarch64.md index a78476c8a..8757a962f 100644 --- a/gcc/config/aarch64/aarch64.md +++ b/gcc/config/aarch64/aarch64.md @@ -1476,11 +1476,11 @@ (set_attr "arch" "simd,fp16,simd,*,simd,*,simd,*,fp16,simd,*,*,*,*,*")] ) -(define_insn "*movsf_aarch64" - [(set (match_operand:SF 0 "nonimmediate_operand" "=w,w ,?r,w,w ,w ,w,m,r,m ,r,r") - (match_operand:SF 1 "general_operand" "Y ,?rY, w,w,Ufc,Uvi,m,w,m,rY,r,M"))] - "TARGET_FLOAT && (register_operand (operands[0], SFmode) - || aarch64_reg_or_fp_zero (operands[1], SFmode))" +(define_insn "*mov<mode>_aarch64" + [(set (match_operand:SFD 0 "nonimmediate_operand" "=w,w ,?r,w,w ,w ,w,m,r,m ,r,r") + (match_operand:SFD 1 "general_operand" "Y ,?rY, w,w,Ufc,Uvi,m,w,m,rY,r,M"))] + "TARGET_FLOAT && (register_operand (operands[0], <MODE>mode) + || aarch64_reg_or_fp_zero (operands[1], <MODE>mode))" "@ movi\\t%0.2s, #0 fmov\\t%s0, %w1 @@ -1500,11 +1500,11 @@ (set_attr "arch" "simd,*,*,*,*,simd,*,*,*,*,*,*")] ) -(define_insn "*movdf_aarch64" - [(set (match_operand:DF 0 "nonimmediate_operand" "=w, w ,?r,w,w ,w ,w,m,r,m ,r,r") - (match_operand:DF 1 "general_operand" "Y , ?rY, w,w,Ufc,Uvi,m,w,m,rY,r,N"))] - "TARGET_FLOAT && (register_operand (operands[0], DFmode) - || aarch64_reg_or_fp_zero (operands[1], DFmode))" +(define_insn "*mov<mode>_aarch64" + [(set (match_operand:DFD 0 "nonimmediate_operand" "=w, w ,?r,w,w ,w ,w,m,r,m ,r,r") + (match_operand:DFD 1 "general_operand" "Y , ?rY, w,w,Ufc,Uvi,m,w,m,rY,r,N"))] + "TARGET_FLOAT && (register_operand (operands[0], <MODE>mode) + || aarch64_reg_or_fp_zero (operands[1], <MODE>mode))" "@ movi\\t%d0, #0 fmov\\t%d0, %x1 @@ -1545,13 +1545,13 @@ } ) -(define_insn "*movtf_aarch64" - [(set (match_operand:TF 0 +(define_insn "*mov<mode>_aarch64" + [(set (match_operand:TFD 0 "nonimmediate_operand" "=w,?r ,w ,?r,w,?w,w,m,?r,m ,m") - (match_operand:TF 1 + (match_operand:TFD 1 "general_operand" " w,?rY,?r,w ,Y,Y ,m,w,m ,?r,Y"))] - "TARGET_FLOAT && (register_operand (operands[0], TFmode) - || aarch64_reg_or_fp_zero (operands[1], TFmode))" + "TARGET_FLOAT && (register_operand (operands[0], <MODE>mode) + || aarch64_reg_or_fp_zero (operands[1], <MODE>mode))" "@ mov\\t%0.16b, %1.16b # @@ -1571,8 +1571,8 @@ ) (define_split - [(set (match_operand:TF 0 "register_operand" "") - (match_operand:TF 1 "nonmemory_operand" ""))] + [(set (match_operand:TFD 0 "register_operand" "") + (match_operand:TFD 1 "nonmemory_operand" ""))] "reload_completed && aarch64_split_128bit_move_p (operands[0], operands[1])" [(const_int 0)] { diff --git a/gcc/config/aarch64/iterators.md b/gcc/config/aarch64/iterators.md index 967e6b0b1..d0cd1b788 100644 --- a/gcc/config/aarch64/iterators.md +++ b/gcc/config/aarch64/iterators.md @@ -67,14 +67,24 @@ (define_mode_iterator GPF_TF_F16 [HF SF DF TF]) ;; Iterator for all scalar floating point modes suitable for moving, including -;; special BF type (HF, SF, DF, TF and BF) -(define_mode_iterator GPF_TF_F16_MOV [HF BF SF DF TF]) +;; special BF type and decimal floating point types (HF, SF, DF, TF, BF, +;; SD, DD and TD) +(define_mode_iterator GPF_TF_F16_MOV [HF BF SF DF TF SD DD TD]) + +;; Iterator for scalar 32bit fp modes (SF, SD) +(define_mode_iterator SFD [SD SF]) + +;; Iterator for scalar 64bit fp modes (DF, DD) +(define_mode_iterator DFD [DD DF]) + +;; Iterator for scalar 128bit fp modes (TF, TD) +(define_mode_iterator TFD [TD TF]) ;; Double vector modes. (define_mode_iterator VDF [V2SF V4HF]) -;; Iterator for all scalar floating point modes (SF, DF and TF) -(define_mode_iterator GPF_TF [SF DF TF]) +;; Iterator for all scalar floating point modes (SF, DF, TF, SD, DD, and TD) +(define_mode_iterator GPF_TF [SF DF TF SD DD TD]) ;; Integer Advanced SIMD modes. (define_mode_iterator VDQ_I [V8QI V16QI V4HI V8HI V2SI V4SI V2DI]) @@ -301,7 +311,7 @@ ;; 2 and 4 lane SI modes. (define_mode_iterator VS [V2SI V4SI]) -(define_mode_iterator TX [TI TF]) +(define_mode_iterator TX [TI TF TD]) ;; Advanced SIMD opaque structure modes. (define_mode_iterator VSTRUCT [OI CI XI]) @@ -403,10 +413,10 @@ V4x8HF V4x4SF V4x2DF V4x8BF]) ;; Double scalar modes -(define_mode_iterator DX [DI DF]) +(define_mode_iterator DX [DI DF DD]) ;; Duplicate of the above -(define_mode_iterator DX2 [DI DF]) +(define_mode_iterator DX2 [DI DF DD]) ;; Single scalar modes (define_mode_iterator SX [SI SF]) -- 2.33.0
Locations
Projects
Search
Status Monitor
Help
Open Build Service
OBS Manuals
API Documentation
OBS Portal
Reporting a Bug
Contact
Mailing List
Forums
Chat (IRC)
Twitter
Open Build Service (OBS)
is an
openSUSE project
.
浙ICP备2022010568号-2