Message ID | 55508D02.4050007@imgtec.com |
---|---|
State | New |
Headers | show |
Hi Sameera, Sameera Deshpande <Sameera.Deshpande@imgtec.com> writes: > Changelog: > gcc/ > * config/mips/mips.md (JOIN_MODE): New mode iterator. > (join2_load_Store<JOIN_MODE:mode>): New pattern. > (join2_loadhi): Likewise. > (define_peehole2): Add peephole2 patterns to join 2 > HI/SI/SF/DF-mode > load-load and store-stores. > * config/mips/mips.opt (mload-store-pairs): New option. > (TARGET_LOAD_STORE_PAIRS): New macro. > * config/mips/mips.h (ENABLE_LD_ST_PAIRS): Likewise. > * config/mips/mips-protos.h (mips_load_store_bonding_p): New > prototype. > * config/mips/mips.c (mips_load_store_bonding_p): New function. > > gcc/testsuite/ > * gcc.target/mips/p5600-bonding.c : New testcase to test > bonding. Just 'New file.' is fine for the changelog. >diff --git a/gcc/testsuite/gcc.target/mips/p5600-bonding.c b/gcc/testsuite/gcc.target/mips/p5600-bonding.c >new file mode 100644 >index 0000000..122b9f8 >--- /dev/null >+++ b/gcc/testsuite/gcc.target/mips/p5600-bonding.c >@@ -0,0 +1,19 @@ >+/* { dg-do compile } */ >+/* { dg-options "-dp -mtune=p5600 -mno-micromips -mno-mips16" } */ >+/* { dg-skip-if "Bonding needs peephole optimization." { *-*-* } { "-O0" "-O1" } { "" } } */ >+typedef int VINT32 __attribute__ ((vector_size((16)))); >+ >+void memory_operation_fun2_si(void * __restrict src, void * __restrict dest, int num) Code style applies for testcases too, return type on line above, space after function name, line length. >+{ >+ VINT32 *vsrc = (VINT32 *)src; Indentation. >+ VINT32 *vdest = (VINT32 *)dest; >+ int i; >+ >+ for (i = 0; i < num - 1; i+=2) >+ { Indentation >+ vdest[i] = (vdest[i] + vsrc[i]); Unnecessary brackets. >+ vdest[i + 1] = vdest[i + 1] + vsrc[i + 1]; >+ } >+} >+/* { dg-final { scan-assembler "join2_" } } */ >+ OK with those changes. Thanks, Matthew
On May 11, 2015, at 4:05 AM, sameera <sameera.deshpande@imgtec.com> wrote: >>> +(define_insn "*join2_loadhi" >>> + [(set (match_operand:SI 0 "register_operand" "=r") >>> + (any_extend:SI (match_operand:HI 1 "non_volatile_mem_operand" "m"))) >>> + (set (match_operand:SI 2 "register_operand" "=r") >>> + (any_extend:SI (match_operand:HI 3 "non_volatile_mem_operand" "m")))] >>> + "ENABLE_LD_ST_PAIRS && reload_completed" >>> + { >>> + /* Reg-renaming pass reuses base register if it is dead after bonded loads. >>> + Hardware does not bond those loads, even when they are consecutive. >>> + However, order of the loads need to be checked for correctness. */ >>> + if (!reg_overlap_mentioned_p (operands[0], operands[1])) >>> + { >>> + output_asm_insn ("lh<u>\t%0,%1", operands); >>> + output_asm_insn ("lh<u>\t%2,%3", operands); >>> + } >>> + else >>> + { >>> + output_asm_insn ("lh<u>\t%2,%3", operands); >>> + output_asm_insn ("lh<u>\t%0,%1", operands); >>> + } >>> + >>> + return ""; >>> + } >>> + [(set_attr "move_type" "load") >>> + (set_attr "insn_count" "2")]) > However, unlike other architectures, we do not generate single instruction for bonded pair, Actually, you do. The above is 1 instruction pattern. Doesn’t matter much what it prints as or what the CPU thinks of it. > because of which it is difficult to check if bonding is happening or not. Hence, an assembly file is generated with debug dumps, and the bonded loads/stores are identified by their pattern names. Nothing wrong with that approach. Also, in the assembly, one can look for sequences of instruction if they way. See gcc/testsuite/gcc.target/aarch64/fuse_adrp_add_1.c: /* { dg-final { scan-assembler "adrp\tx.*, fixed_regs\n\tadd\tx.*, x.*fixed_regs" } } */ in the test suite for example. > I am trying FUSION for MIPS as suggested by Mike, and testing the perf impact of it along with other mips specific options. I think you will discover it is virtually what you have now, and works better. The fusion just can peephole over greater distances, that’s the only real difference.
Hi Mike, Thanks for your comments. Please find my comments inlined. - Thanks and regards, Sameera D. On Monday 11 May 2015 10:09 PM, Mike Stump wrote: > On May 11, 2015, at 4:05 AM, sameera <sameera.deshpande@imgtec.com> wrote: >>>> +(define_insn "*join2_loadhi" >>>> + [(set (match_operand:SI 0 "register_operand" "=r") >>>> + (any_extend:SI (match_operand:HI 1 "non_volatile_mem_operand" "m"))) >>>> + (set (match_operand:SI 2 "register_operand" "=r") >>>> + (any_extend:SI (match_operand:HI 3 "non_volatile_mem_operand" "m")))] >>>> + "ENABLE_LD_ST_PAIRS && reload_completed" >>>> + { >>>> + /* Reg-renaming pass reuses base register if it is dead after bonded loads. >>>> + Hardware does not bond those loads, even when they are consecutive. >>>> + However, order of the loads need to be checked for correctness. */ >>>> + if (!reg_overlap_mentioned_p (operands[0], operands[1])) >>>> + { >>>> + output_asm_insn ("lh<u>\t%0,%1", operands); >>>> + output_asm_insn ("lh<u>\t%2,%3", operands); >>>> + } >>>> + else >>>> + { >>>> + output_asm_insn ("lh<u>\t%2,%3", operands); >>>> + output_asm_insn ("lh<u>\t%0,%1", operands); >>>> + } >>>> + >>>> + return ""; >>>> + } >>>> + [(set_attr "move_type" "load") >>>> + (set_attr "insn_count" "2")]) > >> However, unlike other architectures, we do not generate single instruction for bonded pair, > > Actually, you do. The above is 1 instruction pattern. Doesn’t matter much what it prints as or what the CPU thinks of it. The pattern is single, however, the asm code will have multiple instructions generated for the pattern. > >> because of which it is difficult to check if bonding is happening or not. Hence, an assembly file is generated with debug dumps, and the bonded loads/stores are identified by their pattern names. > > Nothing wrong with that approach. Also, in the assembly, one can look for sequences of instruction if they way. Load/store bonding is not just contiguous load/store instructions, but they also need to have same base register and offset with specific difference. Hence, The way you suggested might not be useful always. Hence, I am comparing the pattern name instead. > See gcc/testsuite/gcc.target/aarch64/fuse_adrp_add_1.c: > > /* { dg-final { scan-assembler "adrp\tx.*, fixed_regs\n\tadd\tx.*, x.*fixed_regs" } } */ > > in the test suite for example. > >> I am trying FUSION for MIPS as suggested by Mike, and testing the perf impact of it along with other mips specific options. > > I think you will discover it is virtually what you have now, and works better. The fusion just can peephole over greater distances, that’s the only real difference. Yes, in many cases I see clear improvement. However, it also tries to bring loads/stores together, which were split intentionally by msched-weight option, introduced for MIPS. I need to measure performance and do perf tuning (if needed) for that option before sending it for review. >
diff --git a/gcc/config/mips/mips-protos.h b/gcc/config/mips/mips-protos.h index b48e04f..244eb8d 100644 --- a/gcc/config/mips/mips-protos.h +++ b/gcc/config/mips/mips-protos.h @@ -360,6 +360,7 @@ extern bool mips_epilogue_uses (unsigned int); extern void mips_final_prescan_insn (rtx_insn *, rtx *, int); extern int mips_trampoline_code_size (void); extern void mips_function_profiler (FILE *); +extern bool mips_load_store_bonding_p (rtx *, machine_mode, bool); typedef rtx (*mulsidi3_gen_fn) (rtx, rtx, rtx); #ifdef RTX_CODE diff --git a/gcc/config/mips/mips.c b/gcc/config/mips/mips.c index bf69850..4fc15c4 100644 --- a/gcc/config/mips/mips.c +++ b/gcc/config/mips/mips.c @@ -18241,6 +18241,66 @@ umips_load_store_pair_p_1 (bool load_p, bool swap_p, return true; } +bool +mips_load_store_bonding_p (rtx *operands, machine_mode mode, bool load_p) +{ + rtx reg1, reg2, mem1, mem2, base1, base2; + enum reg_class rc1, rc2; + HOST_WIDE_INT offset1, offset2; + + if (load_p) + { + reg1 = operands[0]; + reg2 = operands[2]; + mem1 = operands[1]; + mem2 = operands[3]; + } + else + { + reg1 = operands[1]; + reg2 = operands[3]; + mem1 = operands[0]; + mem2 = operands[2]; + } + + if (mips_address_insns (XEXP (mem1, 0), mode, false) == 0 + || mips_address_insns (XEXP (mem2, 0), mode, false) == 0) + return false; + + mips_split_plus (XEXP (mem1, 0), &base1, &offset1); + mips_split_plus (XEXP (mem2, 0), &base2, &offset2); + + /* Base regs do not match. */ + if (!REG_P (base1) || !rtx_equal_p (base1, base2)) + return false; + + /* Either of the loads is clobbering base register. It is legitimate to bond + loads if second load clobbers base register. However, hardware does not + support such bonding. */ + if (load_p + && (REGNO (reg1) == REGNO (base1) + || (REGNO (reg2) == REGNO (base1)))) + return false; + + /* Loading in same registers. */ + if (load_p + && REGNO (reg1) == REGNO (reg2)) + return false; + + /* The loads/stores are not of same type. */ + rc1 = REGNO_REG_CLASS (REGNO (reg1)); + rc2 = REGNO_REG_CLASS (REGNO (reg2)); + if (rc1 != rc2 + && !reg_class_subset_p (rc1, rc2) + && !reg_class_subset_p (rc2, rc1)) + return false; + + if (abs (offset1 - offset2) != GET_MODE_SIZE (mode)) + return false; + + return true; +} + /* OPERANDS describes the operands to a pair of SETs, in the order dest1, src1, dest2, src2. Return true if the operands can be used in an LWP or SWP instruction; LOAD_P says which. */ diff --git a/gcc/config/mips/mips.h b/gcc/config/mips/mips.h index 4bd83f5..c9accd1 100644 --- a/gcc/config/mips/mips.h +++ b/gcc/config/mips/mips.h @@ -3162,3 +3162,10 @@ extern GTY(()) struct target_globals *mips16_globals; #define STANDARD_STARTFILE_PREFIX_1 "/lib64/" #define STANDARD_STARTFILE_PREFIX_2 "/usr/lib64/" #endif + +/* Load store bonding is not supported by micromips and fix_24k. The + performance can be degraded for those targets. Hence, do not bond for + micromips or fix_24k. */ +#define ENABLE_LD_ST_PAIRS \ + (TARGET_LOAD_STORE_PAIRS && TUNE_P5600 \ + && !TARGET_MICROMIPS && !TARGET_FIX_24K) diff --git a/gcc/config/mips/mips.md b/gcc/config/mips/mips.md index ed4c0ba..0e2b172 100644 --- a/gcc/config/mips/mips.md +++ b/gcc/config/mips/mips.md @@ -754,6 +754,11 @@ (define_mode_iterator MOVEP1 [SI SF]) (define_mode_iterator MOVEP2 [SI SF]) +(define_mode_iterator JOIN_MODE [HI + SI + (SF "TARGET_HARD_FLOAT") + (DF "TARGET_HARD_FLOAT + && TARGET_DOUBLE_FLOAT")]) ;; This mode iterator allows :HILO to be used as the mode of the ;; concatenated HI and LO registers. @@ -7407,6 +7412,112 @@ { return MIPS_CALL ("jal", operands, 0, -1); } [(set_attr "type" "call") (set_attr "insn_count" "3")]) + +;; Match paired HI/SI/SF/DFmode load/stores. +(define_insn "*join2_load_store<JOIN_MODE:mode>" + [(set (match_operand:JOIN_MODE 0 "nonimmediate_operand" "=d,f,m,m") + (match_operand:JOIN_MODE 1 "nonimmediate_operand" "m,m,d,f")) + (set (match_operand:JOIN_MODE 2 "nonimmediate_operand" "=d,f,m,m") + (match_operand:JOIN_MODE 3 "nonimmediate_operand" "m,m,d,f"))] + "ENABLE_LD_ST_PAIRS && reload_completed" + { + bool load_p = (which_alternative == 0 || which_alternative == 1); + /* Reg-renaming pass reuses base register if it is dead after bonded loads. + Hardware does not bond those loads, even when they are consecutive. + However, order of the loads need to be checked for correctness. */ + if (!load_p || !reg_overlap_mentioned_p (operands[0], operands[1])) + { + output_asm_insn (mips_output_move (operands[0], operands[1]), + operands); + output_asm_insn (mips_output_move (operands[2], operands[3]), + &operands[2]); + } + else + { + output_asm_insn (mips_output_move (operands[2], operands[3]), + &operands[2]); + output_asm_insn (mips_output_move (operands[0], operands[1]), + operands); + } + return ""; + } + [(set_attr "move_type" "load,fpload,store,fpstore") + (set_attr "insn_count" "2,2,2,2")]) + +;; 2 HI/SI/SF/DF loads are joined. +;; P5600 does not support bonding of two LBs, hence QI mode is not included. +;; The loads must be non-volatile as they might be reordered at the time of asm +;; generation. +(define_peephole2 + [(set (match_operand:JOIN_MODE 0 "register_operand") + (match_operand:JOIN_MODE 1 "non_volatile_mem_operand")) + (set (match_operand:JOIN_MODE 2 "register_operand") + (match_operand:JOIN_MODE 3 "non_volatile_mem_operand"))] + "ENABLE_LD_ST_PAIRS + && mips_load_store_bonding_p (operands, <JOIN_MODE:MODE>mode, true)" + [(parallel [(set (match_dup 0) + (match_dup 1)) + (set (match_dup 2) + (match_dup 3))])] + "") + +;; 2 HI/SI/SF/DF stores are joined. +;; P5600 does not support bonding of two SBs, hence QI mode is not included. +(define_peephole2 + [(set (match_operand:JOIN_MODE 0 "memory_operand") + (match_operand:JOIN_MODE 1 "register_operand")) + (set (match_operand:JOIN_MODE 2 "memory_operand") + (match_operand:JOIN_MODE 3 "register_operand"))] + "ENABLE_LD_ST_PAIRS + && mips_load_store_bonding_p (operands, <JOIN_MODE:MODE>mode, false)" + [(parallel [(set (match_dup 0) + (match_dup 1)) + (set (match_dup 2) + (match_dup 3))])] + "") + +;; Match paired HImode loads. +(define_insn "*join2_loadhi" + [(set (match_operand:SI 0 "register_operand" "=r") + (any_extend:SI (match_operand:HI 1 "non_volatile_mem_operand" "m"))) + (set (match_operand:SI 2 "register_operand" "=r") + (any_extend:SI (match_operand:HI 3 "non_volatile_mem_operand" "m")))] + "ENABLE_LD_ST_PAIRS && reload_completed" + { + /* Reg-renaming pass reuses base register if it is dead after bonded loads. + Hardware does not bond those loads, even when they are consecutive. + However, order of the loads need to be checked for correctness. */ + if (!reg_overlap_mentioned_p (operands[0], operands[1])) + { + output_asm_insn ("lh<u>\t%0,%1", operands); + output_asm_insn ("lh<u>\t%2,%3", operands); + } + else + { + output_asm_insn ("lh<u>\t%2,%3", operands); + output_asm_insn ("lh<u>\t%0,%1", operands); + } + + return ""; + } + [(set_attr "move_type" "load") + (set_attr "insn_count" "2")]) + + +;; 2 HI loads are joined. +(define_peephole2 + [(set (match_operand:SI 0 "register_operand") + (any_extend:SI (match_operand:HI 1 "non_volatile_mem_operand"))) + (set (match_operand:SI 2 "register_operand") + (any_extend:SI (match_operand:HI 3 "non_volatile_mem_operand")))] + "ENABLE_LD_ST_PAIRS + && mips_load_store_bonding_p (operands, HImode, true)" + [(parallel [(set (match_dup 0) + (any_extend:SI (match_dup 1))) + (set (match_dup 2) + (any_extend:SI (match_dup 3)))])] + "") + ;; Synchronization instructions. diff --git a/gcc/config/mips/mips.opt b/gcc/config/mips/mips.opt index 9e89aa9..a9baebe 100644 --- a/gcc/config/mips/mips.opt +++ b/gcc/config/mips/mips.opt @@ -418,3 +418,7 @@ Enable use of odd-numbered single-precision registers noasmopt Driver + +mload-store-pairs +Target Report Var(TARGET_LOAD_STORE_PAIRS) Init(1) +Enable load/store bonding. diff --git a/gcc/testsuite/gcc.target/mips/p5600-bonding.c b/gcc/testsuite/gcc.target/mips/p5600-bonding.c new file mode 100644 index 0000000..122b9f8 --- /dev/null +++ b/gcc/testsuite/gcc.target/mips/p5600-bonding.c @@ -0,0 +1,19 @@ +/* { dg-do compile } */ +/* { dg-options "-dp -mtune=p5600 -mno-micromips -mno-mips16" } */ +/* { dg-skip-if "Bonding needs peephole optimization." { *-*-* } { "-O0" "-O1" } { "" } } */ +typedef int VINT32 __attribute__ ((vector_size((16)))); + +void memory_operation_fun2_si(void * __restrict src, void * __restrict dest, int num) +{ + VINT32 *vsrc = (VINT32 *)src; + VINT32 *vdest = (VINT32 *)dest; + int i; + + for (i = 0; i < num - 1; i+=2) + { + vdest[i] = (vdest[i] + vsrc[i]); + vdest[i + 1] = vdest[i + 1] + vsrc[i + 1]; + } +} +/* { dg-final { scan-assembler "join2_" } } */ +