diff options
Diffstat (limited to 'arch/tile/lib')
| -rw-r--r-- | arch/tile/lib/Makefile | 2 | ||||
| -rw-r--r-- | arch/tile/lib/atomic_32.c | 90 | ||||
| -rw-r--r-- | arch/tile/lib/memcpy_32.S | 61 | ||||
| -rw-r--r-- | arch/tile/lib/memcpy_tile64.c | 280 | ||||
| -rw-r--r-- | arch/tile/lib/memset_32.c | 105 |
5 files changed, 1 insertions, 537 deletions
diff --git a/arch/tile/lib/Makefile b/arch/tile/lib/Makefile index 9adfd76fbdd8..c4211cbb2021 100644 --- a/arch/tile/lib/Makefile +++ b/arch/tile/lib/Makefile @@ -7,7 +7,7 @@ lib-y = cacheflush.o checksum.o cpumask.o delay.o uaccess.o \ strchr_$(BITS).o strlen_$(BITS).o strnlen_$(BITS).o lib-$(CONFIG_TILEGX) += memcpy_user_64.o -lib-$(CONFIG_TILEPRO) += atomic_32.o atomic_asm_32.o memcpy_tile64.o +lib-$(CONFIG_TILEPRO) += atomic_32.o atomic_asm_32.o lib-$(CONFIG_SMP) += spinlock_$(BITS).o usercopy_$(BITS).o obj-$(CONFIG_MODULES) += exports.o diff --git a/arch/tile/lib/atomic_32.c b/arch/tile/lib/atomic_32.c index 42eacb1f737a..5d91d1860640 100644 --- a/arch/tile/lib/atomic_32.c +++ b/arch/tile/lib/atomic_32.c @@ -20,50 +20,12 @@ #include <linux/atomic.h> #include <arch/chip.h> -/* See <asm/atomic_32.h> */ -#if ATOMIC_LOCKS_FOUND_VIA_TABLE() - -/* - * A block of memory containing locks for atomic ops. Each instance of this - * struct will be homed on a different CPU. - */ -struct atomic_locks_on_cpu { - int lock[ATOMIC_HASH_L2_SIZE]; -} __attribute__((aligned(ATOMIC_HASH_L2_SIZE * 4))); - -static DEFINE_PER_CPU(struct atomic_locks_on_cpu, atomic_lock_pool); - -/* The locks we'll use until __init_atomic_per_cpu is called. */ -static struct atomic_locks_on_cpu __initdata initial_atomic_locks; - -/* Hash into this vector to get a pointer to lock for the given atomic. */ -struct atomic_locks_on_cpu *atomic_lock_ptr[ATOMIC_HASH_L1_SIZE] - __write_once = { - [0 ... ATOMIC_HASH_L1_SIZE-1] (&initial_atomic_locks) -}; - -#else /* ATOMIC_LOCKS_FOUND_VIA_TABLE() */ - /* This page is remapped on startup to be hash-for-home. */ int atomic_locks[PAGE_SIZE / sizeof(int)] __page_aligned_bss; -#endif /* ATOMIC_LOCKS_FOUND_VIA_TABLE() */ - int *__atomic_hashed_lock(volatile void *v) { /* NOTE: this code must match "sys_cmpxchg" in kernel/intvec_32.S */ -#if ATOMIC_LOCKS_FOUND_VIA_TABLE() - unsigned long i = - (unsigned long) v & ((PAGE_SIZE-1) & -sizeof(long long)); - unsigned long n = __insn_crc32_32(0, i); - - /* Grab high bits for L1 index. */ - unsigned long l1_index = n >> ((sizeof(n) * 8) - ATOMIC_HASH_L1_SHIFT); - /* Grab low bits for L2 index. */ - unsigned long l2_index = n & (ATOMIC_HASH_L2_SIZE - 1); - - return &atomic_lock_ptr[l1_index]->lock[l2_index]; -#else /* * Use bits [3, 3 + ATOMIC_HASH_SHIFT) as the lock index. * Using mm works here because atomic_locks is page aligned. @@ -72,26 +34,13 @@ int *__atomic_hashed_lock(volatile void *v) (unsigned long)atomic_locks, 2, (ATOMIC_HASH_SHIFT + 2) - 1); return (int *)ptr; -#endif } #ifdef CONFIG_SMP /* Return whether the passed pointer is a valid atomic lock pointer. */ static int is_atomic_lock(int *p) { -#if ATOMIC_LOCKS_FOUND_VIA_TABLE() - int i; - for (i = 0; i < ATOMIC_HASH_L1_SIZE; ++i) { - - if (p >= &atomic_lock_ptr[i]->lock[0] && - p < &atomic_lock_ptr[i]->lock[ATOMIC_HASH_L2_SIZE]) { - return 1; - } - } - return 0; -#else return p >= &atomic_locks[0] && p < &atomic_locks[ATOMIC_HASH_SIZE]; -#endif } void __atomic_fault_unlock(int *irqlock_word) @@ -210,43 +159,6 @@ struct __get_user __atomic_bad_address(int __user *addr) void __init __init_atomic_per_cpu(void) { -#if ATOMIC_LOCKS_FOUND_VIA_TABLE() - - unsigned int i; - int actual_cpu; - - /* - * Before this is called from setup, we just have one lock for - * all atomic objects/operations. Here we replace the - * elements of atomic_lock_ptr so that they point at per_cpu - * integers. This seemingly over-complex approach stems from - * the fact that DEFINE_PER_CPU defines an entry for each cpu - * in the grid, not each cpu from 0..ATOMIC_HASH_SIZE-1. But - * for efficient hashing of atomics to their locks we want a - * compile time constant power of 2 for the size of this - * table, so we use ATOMIC_HASH_SIZE. - * - * Here we populate atomic_lock_ptr from the per cpu - * atomic_lock_pool, interspersing by actual cpu so that - * subsequent elements are homed on consecutive cpus. - */ - - actual_cpu = cpumask_first(cpu_possible_mask); - - for (i = 0; i < ATOMIC_HASH_L1_SIZE; ++i) { - /* - * Preincrement to slightly bias against using cpu 0, - * which has plenty of stuff homed on it already. - */ - actual_cpu = cpumask_next(actual_cpu, cpu_possible_mask); - if (actual_cpu >= nr_cpu_ids) - actual_cpu = cpumask_first(cpu_possible_mask); - - atomic_lock_ptr[i] = &per_cpu(atomic_lock_pool, actual_cpu); - } - -#else /* ATOMIC_LOCKS_FOUND_VIA_TABLE() */ - /* Validate power-of-two and "bigger than cpus" assumption */ BUILD_BUG_ON(ATOMIC_HASH_SIZE & (ATOMIC_HASH_SIZE-1)); BUG_ON(ATOMIC_HASH_SIZE < nr_cpu_ids); @@ -270,6 +182,4 @@ void __init __init_atomic_per_cpu(void) * That should not produce more indices than ATOMIC_HASH_SIZE. */ BUILD_BUG_ON((PAGE_SIZE >> 3) > ATOMIC_HASH_SIZE); - -#endif /* ATOMIC_LOCKS_FOUND_VIA_TABLE() */ } diff --git a/arch/tile/lib/memcpy_32.S b/arch/tile/lib/memcpy_32.S index 8ba7626cfeb1..a2771ae5da53 100644 --- a/arch/tile/lib/memcpy_32.S +++ b/arch/tile/lib/memcpy_32.S @@ -22,14 +22,6 @@ #include <linux/linkage.h> -/* On TILE64, we wrap these functions via arch/tile/lib/memcpy_tile64.c */ -#if !CHIP_HAS_COHERENT_LOCAL_CACHE() -#define memcpy __memcpy_asm -#define __copy_to_user_inatomic __copy_to_user_inatomic_asm -#define __copy_from_user_inatomic __copy_from_user_inatomic_asm -#define __copy_from_user_zeroing __copy_from_user_zeroing_asm -#endif - #define IS_MEMCPY 0 #define IS_COPY_FROM_USER 1 #define IS_COPY_FROM_USER_ZEROING 2 @@ -159,12 +151,9 @@ EX: { sw r0, r3; addi r0, r0, 4; addi r2, r2, -4 } { addi r3, r1, 60; andi r9, r9, -64 } -#if CHIP_HAS_WH64() /* No need to prefetch dst, we'll just do the wh64 * right before we copy a line. */ -#endif - EX: { lw r5, r3; addi r3, r3, 64; movei r4, 1 } /* Intentionally stall for a few cycles to leave L2 cache alone. */ { bnzt zero, .; move r27, lr } @@ -172,21 +161,6 @@ EX: { lw r6, r3; addi r3, r3, 64 } /* Intentionally stall for a few cycles to leave L2 cache alone. */ { bnzt zero, . } EX: { lw r7, r3; addi r3, r3, 64 } -#if !CHIP_HAS_WH64() - /* Prefetch the dest */ - /* Intentionally stall for a few cycles to leave L2 cache alone. */ - { bnzt zero, . } - /* Use a real load to cause a TLB miss if necessary. We aren't using - * r28, so this should be fine. - */ -EX: { lw r28, r9; addi r9, r9, 64 } - /* Intentionally stall for a few cycles to leave L2 cache alone. */ - { bnzt zero, . } - { prefetch r9; addi r9, r9, 64 } - /* Intentionally stall for a few cycles to leave L2 cache alone. */ - { bnzt zero, . } - { prefetch r9; addi r9, r9, 64 } -#endif /* Intentionally stall for a few cycles to leave L2 cache alone. */ { bz zero, .Lbig_loop2 } @@ -287,13 +261,8 @@ EX: { lw r7, r3; addi r3, r3, 64 } /* Fill second L1D line. */ EX: { lw r17, r17; addi r1, r1, 48; mvz r3, r13, r1 } /* r17 = WORD_4 */ -#if CHIP_HAS_WH64() /* Prepare destination line for writing. */ EX: { wh64 r9; addi r9, r9, 64 } -#else - /* Prefetch dest line */ - { prefetch r9; addi r9, r9, 64 } -#endif /* Load seven words that are L1D hits to cover wh64 L2 usage. */ /* Load the three remaining words from the last L1D line, which @@ -331,16 +300,7 @@ EX: { lw r18, r1; addi r1, r1, 4 } /* r18 = WORD_8 */ EX: { sw r0, r16; addi r0, r0, 4; add r16, r0, r2 } /* store(WORD_0) */ EX: { sw r0, r13; addi r0, r0, 4; andi r16, r16, -64 } /* store(WORD_1) */ EX: { sw r0, r14; addi r0, r0, 4; slt_u r16, r9, r16 } /* store(WORD_2) */ -#if CHIP_HAS_WH64() EX: { sw r0, r15; addi r0, r0, 4; addi r13, sp, -64 } /* store(WORD_3) */ -#else - /* Back up the r9 to a cache line we are already storing to - * if it gets past the end of the dest vector. Strictly speaking, - * we don't need to back up to the start of a cache line, but it's free - * and tidy, so why not? - */ -EX: { sw r0, r15; addi r0, r0, 4; andi r13, r0, -64 } /* store(WORD_3) */ -#endif /* Store second L1D line. */ EX: { sw r0, r17; addi r0, r0, 4; mvz r9, r16, r13 }/* store(WORD_4) */ EX: { sw r0, r19; addi r0, r0, 4 } /* store(WORD_5) */ @@ -404,7 +364,6 @@ EX: { sb r0, r3; addi r0, r0, 1; addi r2, r2, -1 } .Ldest_is_word_aligned: -#if CHIP_HAS_DWORD_ALIGN() EX: { andi r8, r0, 63; lwadd_na r6, r1, 4} { slti_u r9, r2, 64; bz r8, .Ldest_is_L2_line_aligned } @@ -512,26 +471,6 @@ EX: { swadd r0, r13, 4; addi r2, r2, -32 } /* Move r1 back to the point where it corresponds to r0. */ { addi r1, r1, -4 } -#else /* !CHIP_HAS_DWORD_ALIGN() */ - - /* Compute right/left shift counts and load initial source words. */ - { andi r5, r1, -4; andi r3, r1, 3 } -EX: { lw r6, r5; addi r5, r5, 4; shli r3, r3, 3 } -EX: { lw r7, r5; addi r5, r5, 4; sub r4, zero, r3 } - - /* Load and store one word at a time, using shifts and ORs - * to correct for the misaligned src. - */ -.Lcopy_unaligned_src_loop: - { shr r6, r6, r3; shl r8, r7, r4 } -EX: { lw r7, r5; or r8, r8, r6; move r6, r7 } -EX: { sw r0, r8; addi r0, r0, 4; addi r2, r2, -4 } - { addi r5, r5, 4; slti_u r8, r2, 8 } - { bzt r8, .Lcopy_unaligned_src_loop; addi r1, r1, 4 } - - { bz r2, .Lcopy_unaligned_done } -#endif /* !CHIP_HAS_DWORD_ALIGN() */ - /* Fall through */ /* diff --git a/arch/tile/lib/memcpy_tile64.c b/arch/tile/lib/memcpy_tile64.c deleted file mode 100644 index 0290c222847b..000000000000 --- a/arch/tile/lib/memcpy_tile64.c +++ /dev/null @@ -1,280 +0,0 @@ -/* - * Copyright 2010 Tilera Corporation. All Rights Reserved. - * - * This program is free software; you can redistribute it and/or - * modify it under the terms of the GNU General Public License - * as published by the Free Software Foundation, version 2. - * - * This program is distributed in the hope that it will be useful, but - * WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or - * NON INFRINGEMENT. See the GNU General Public License for - * more details. - */ - -#include <linux/string.h> -#include <linux/smp.h> -#include <linux/module.h> -#include <linux/uaccess.h> -#include <asm/fixmap.h> -#include <asm/kmap_types.h> -#include <asm/tlbflush.h> -#include <hv/hypervisor.h> -#include <arch/chip.h> - - -#if !CHIP_HAS_COHERENT_LOCAL_CACHE() - -/* Defined in memcpy.S */ -extern unsigned long __memcpy_asm(void *to, const void *from, unsigned long n); -extern unsigned long __copy_to_user_inatomic_asm( - void __user *to, const void *from, unsigned long n); -extern unsigned long __copy_from_user_inatomic_asm( - void *to, const void __user *from, unsigned long n); -extern unsigned long __copy_from_user_zeroing_asm( - void *to, const void __user *from, unsigned long n); - -typedef unsigned long (*memcpy_t)(void *, const void *, unsigned long); - -/* Size above which to consider TLB games for performance */ -#define LARGE_COPY_CUTOFF 2048 - -/* Communicate to the simulator what we are trying to do. */ -#define sim_allow_multiple_caching(b) \ - __insn_mtspr(SPR_SIM_CONTROL, \ - SIM_CONTROL_ALLOW_MULTIPLE_CACHING | ((b) << _SIM_CONTROL_OPERATOR_BITS)) - -/* - * Copy memory by briefly enabling incoherent cacheline-at-a-time mode. - * - * We set up our own source and destination PTEs that we fully control. - * This is the only way to guarantee that we don't race with another - * thread that is modifying the PTE; we can't afford to try the - * copy_{to,from}_user() technique of catching the interrupt, since - * we must run with interrupts disabled to avoid the risk of some - * other code seeing the incoherent data in our cache. (Recall that - * our cache is indexed by PA, so even if the other code doesn't use - * our kmap_atomic virtual addresses, they'll still hit in cache using - * the normal VAs that aren't supposed to hit in cache.) - */ -static void memcpy_multicache(void *dest, const void *source, - pte_t dst_pte, pte_t src_pte, int len) -{ - int idx; - unsigned long flags, newsrc, newdst; - pmd_t *pmdp; - pte_t *ptep; - int type0, type1; - int cpu = smp_processor_id(); - - /* - * Disable interrupts so that we don't recurse into memcpy() - * in an interrupt handler, nor accidentally reference - * the PA of the source from an interrupt routine. Also - * notify the simulator that we're playing games so we don't - * generate spurious coherency warnings. - */ - local_irq_save(flags); - sim_allow_multiple_caching(1); - - /* Set up the new dest mapping */ - type0 = kmap_atomic_idx_push(); - idx = FIX_KMAP_BEGIN + (KM_TYPE_NR * cpu) + type0; - newdst = __fix_to_virt(idx) + ((unsigned long)dest & (PAGE_SIZE-1)); - pmdp = pmd_offset(pud_offset(pgd_offset_k(newdst), newdst), newdst); - ptep = pte_offset_kernel(pmdp, newdst); - if (pte_val(*ptep) != pte_val(dst_pte)) { - set_pte(ptep, dst_pte); - local_flush_tlb_page(NULL, newdst, PAGE_SIZE); - } - - /* Set up the new source mapping */ - type1 = kmap_atomic_idx_push(); - idx += (type0 - type1); - src_pte = hv_pte_set_nc(src_pte); - src_pte = hv_pte_clear_writable(src_pte); /* be paranoid */ - newsrc = __fix_to_virt(idx) + ((unsigned long)source & (PAGE_SIZE-1)); - pmdp = pmd_offset(pud_offset(pgd_offset_k(newsrc), newsrc), newsrc); - ptep = pte_offset_kernel(pmdp, newsrc); - __set_pte(ptep, src_pte); /* set_pte() would be confused by this */ - local_flush_tlb_page(NULL, newsrc, PAGE_SIZE); - - /* Actually move the data. */ - __memcpy_asm((void *)newdst, (const void *)newsrc, len); - - /* - * Remap the source as locally-cached and not OLOC'ed so that - * we can inval without also invaling the remote cpu's cache. - * This also avoids known errata with inv'ing cacheable oloc data. - */ - src_pte = hv_pte_set_mode(src_pte, HV_PTE_MODE_CACHE_NO_L3); - src_pte = hv_pte_set_writable(src_pte); /* need write access for inv */ - __set_pte(ptep, src_pte); /* set_pte() would be confused by this */ - local_flush_tlb_page(NULL, newsrc, PAGE_SIZE); - - /* - * Do the actual invalidation, covering the full L2 cache line - * at the end since __memcpy_asm() is somewhat aggressive. - */ - __inv_buffer((void *)newsrc, len); - - /* - * We're done: notify the simulator that all is back to normal, - * and re-enable interrupts and pre-emption. - */ - kmap_atomic_idx_pop(); - kmap_atomic_idx_pop(); - sim_allow_multiple_caching(0); - local_irq_restore(flags); -} - -/* - * Identify large copies from remotely-cached memory, and copy them - * via memcpy_multicache() if they look good, otherwise fall back - * to the particular kind of copying passed as the memcpy_t function. - */ -static unsigned long fast_copy(void *dest, const void *source, int len, - memcpy_t func) -{ - int cpu = get_cpu(); - unsigned long retval; - - /* - * Check if it's big enough to bother with. We may end up doing a - * small copy via TLB manipulation if we're near a page boundary, - * but presumably we'll make it up when we hit the second page. - */ - while (len >= LARGE_COPY_CUTOFF) { - int copy_size, bytes_left_on_page; - pte_t *src_ptep, *dst_ptep; - pte_t src_pte, dst_pte; - struct page *src_page, *dst_page; - - /* Is the source page oloc'ed to a remote cpu? */ -retry_source: - src_ptep = virt_to_pte(current->mm, (unsigned long)source); - if (src_ptep == NULL) - break; - src_pte = *src_ptep; - if (!hv_pte_get_present(src_pte) || - !hv_pte_get_readable(src_pte) || - hv_pte_get_mode(src_pte) != HV_PTE_MODE_CACHE_TILE_L3) - break; - if (get_remote_cache_cpu(src_pte) == cpu) - break; - src_page = pfn_to_page(pte_pfn(src_pte)); - get_page(src_page); - if (pte_val(src_pte) != pte_val(*src_ptep)) { - put_page(src_page); - goto retry_source; - } - if (pte_huge(src_pte)) { - /* Adjust the PTE to correspond to a small page */ - int pfn = pte_pfn(src_pte); - pfn += (((unsigned long)source & (HPAGE_SIZE-1)) - >> PAGE_SHIFT); - src_pte = pfn_pte(pfn, src_pte); - src_pte = pte_mksmall(src_pte); - } - - /* Is the destination page writable? */ -retry_dest: - dst_ptep = virt_to_pte(current->mm, (unsigned long)dest); - if (dst_ptep == NULL) { - put_page(src_page); - break; - } - dst_pte = *dst_ptep; - if (!hv_pte_get_present(dst_pte) || - !hv_pte_get_writable(dst_pte)) { - put_page(src_page); - break; - } - dst_page = pfn_to_page(pte_pfn(dst_pte)); - if (dst_page == src_page) { - /* - * Source and dest are on the same page; this - * potentially exposes us to incoherence if any - * part of src and dest overlap on a cache line. - * Just give up rather than trying to be precise. - */ - put_page(src_page); - break; - } - get_page(dst_page); - if (pte_val(dst_pte) != pte_val(*dst_ptep)) { - put_page(dst_page); - goto retry_dest; - } - if (pte_huge(dst_pte)) { - /* Adjust the PTE to correspond to a small page */ - int pfn = pte_pfn(dst_pte); - pfn += (((unsigned long)dest & (HPAGE_SIZE-1)) - >> PAGE_SHIFT); - dst_pte = pfn_pte(pfn, dst_pte); - dst_pte = pte_mksmall(dst_pte); - } - - /* All looks good: create a cachable PTE and copy from it */ - copy_size = len; - bytes_left_on_page = - PAGE_SIZE - (((int)source) & (PAGE_SIZE-1)); - if (copy_size > bytes_left_on_page) - copy_size = bytes_left_on_page; - bytes_left_on_page = - PAGE_SIZE - (((int)dest) & (PAGE_SIZE-1)); - if (copy_size > bytes_left_on_page) - copy_size = bytes_left_on_page; - memcpy_multicache(dest, source, dst_pte, src_pte, copy_size); - - /* Release the pages */ - put_page(dst_page); - put_page(src_page); - - /* Continue on the next page */ - dest += copy_size; - source += copy_size; - len -= copy_size; - } - - retval = func(dest, source, len); - put_cpu(); - return retval; -} - -void *memcpy(void *to, const void *from, __kernel_size_t n) -{ - if (n < LARGE_COPY_CUTOFF) - return (void *)__memcpy_asm(to, from, n); - else - return (void *)fast_copy(to, from, n, __memcpy_asm); -} - -unsigned long __copy_to_user_inatomic(void __user *to, const void *from, - unsigned long n) -{ - if (n < LARGE_COPY_CUTOFF) - return __copy_to_user_inatomic_asm(to, from, n); - else - return fast_copy(to, from, n, __copy_to_user_inatomic_asm); -} - -unsigned long __copy_from_user_inatomic(void *to, const void __user *from, - unsigned long n) -{ - if (n < LARGE_COPY_CUTOFF) - return __copy_from_user_inatomic_asm(to, from, n); - else - return fast_copy(to, from, n, __copy_from_user_inatomic_asm); -} - -unsigned long __copy_from_user_zeroing(void *to, const void __user *from, - unsigned long n) -{ - if (n < LARGE_COPY_CUTOFF) - return __copy_from_user_zeroing_asm(to, from, n); - else - return fast_copy(to, from, n, __copy_from_user_zeroing_asm); -} - -#endif /* !CHIP_HAS_COHERENT_LOCAL_CACHE() */ diff --git a/arch/tile/lib/memset_32.c b/arch/tile/lib/memset_32.c index 9a7837d11f7d..2042bfe6595f 100644 --- a/arch/tile/lib/memset_32.c +++ b/arch/tile/lib/memset_32.c @@ -23,11 +23,7 @@ void *memset(void *s, int c, size_t n) int n32; uint32_t v16, v32; uint8_t *out8 = s; -#if !CHIP_HAS_WH64() - int ahead32; -#else int to_align32; -#endif /* Experimentation shows that a trivial tight loop is a win up until * around a size of 20, where writing a word at a time starts to win. @@ -58,21 +54,6 @@ void *memset(void *s, int c, size_t n) return s; } -#if !CHIP_HAS_WH64() - /* Use a spare issue slot to start prefetching the first cache - * line early. This instruction is free as the store can be buried - * in otherwise idle issue slots doing ALU ops. - */ - __insn_prefetch(out8); - - /* We prefetch the end so that a short memset that spans two cache - * lines gets some prefetching benefit. Again we believe this is free - * to issue. - */ - __insn_prefetch(&out8[n - 1]); -#endif /* !CHIP_HAS_WH64() */ - - /* Align 'out8'. We know n >= 3 so this won't write past the end. */ while (((uintptr_t) out8 & 3) != 0) { *out8++ = c; @@ -93,90 +74,6 @@ void *memset(void *s, int c, size_t n) /* This must be at least 8 or the following loop doesn't work. */ #define CACHE_LINE_SIZE_IN_WORDS (CHIP_L2_LINE_SIZE() / 4) -#if !CHIP_HAS_WH64() - - ahead32 = CACHE_LINE_SIZE_IN_WORDS; - - /* We already prefetched the first and last cache lines, so - * we only need to do more prefetching if we are storing - * to more than two cache lines. - */ - if (n32 > CACHE_LINE_SIZE_IN_WORDS * 2) { - int i; - - /* Prefetch the next several cache lines. - * This is the setup code for the software-pipelined - * loop below. - */ -#define MAX_PREFETCH 5 - ahead32 = n32 & -CACHE_LINE_SIZE_IN_WORDS; - if (ahead32 > MAX_PREFETCH * CACHE_LINE_SIZE_IN_WORDS) - ahead32 = MAX_PREFETCH * CACHE_LINE_SIZE_IN_WORDS; - - for (i = CACHE_LINE_SIZE_IN_WORDS; - i < ahead32; i += CACHE_LINE_SIZE_IN_WORDS) - __insn_prefetch(&out32[i]); - } - - if (n32 > ahead32) { - while (1) { - int j; - - /* Prefetch by reading one word several cache lines - * ahead. Since loads are non-blocking this will - * cause the full cache line to be read while we are - * finishing earlier cache lines. Using a store - * here causes microarchitectural performance - * problems where a victimizing store miss goes to - * the head of the retry FIFO and locks the pipe for - * a few cycles. So a few subsequent stores in this - * loop go into the retry FIFO, and then later - * stores see other stores to the same cache line - * are already in the retry FIFO and themselves go - * into the retry FIFO, filling it up and grinding - * to a halt waiting for the original miss to be - * satisfied. - */ - __insn_prefetch(&out32[ahead32]); - -#if CACHE_LINE_SIZE_IN_WORDS % 4 != 0 -#error "Unhandled CACHE_LINE_SIZE_IN_WORDS" -#endif - - n32 -= CACHE_LINE_SIZE_IN_WORDS; - - /* Save icache space by only partially unrolling - * this loop. - */ - for (j = CACHE_LINE_SIZE_IN_WORDS / 4; j > 0; j--) { - *out32++ = v32; - *out32++ = v32; - *out32++ = v32; - *out32++ = v32; - } - - /* To save compiled code size, reuse this loop even - * when we run out of prefetching to do by dropping - * ahead32 down. - */ - if (n32 <= ahead32) { - /* Not even a full cache line left, - * so stop now. - */ - if (n32 < CACHE_LINE_SIZE_IN_WORDS) - break; - - /* Choose a small enough value that we don't - * prefetch past the end. There's no sense - * in touching cache lines we don't have to. - */ - ahead32 = CACHE_LINE_SIZE_IN_WORDS - 1; - } - } - } - -#else /* CHIP_HAS_WH64() */ - /* Determine how many words we need to emit before the 'out32' * pointer becomes aligned modulo the cache line size. */ @@ -233,8 +130,6 @@ void *memset(void *s, int c, size_t n) n32 &= CACHE_LINE_SIZE_IN_WORDS - 1; } -#endif /* CHIP_HAS_WH64() */ - /* Now handle any leftover values. */ if (n32 != 0) { do { |