From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=0.7 required=3.0 tests=HEADER_FROM_DIFFERENT_DOMAINS, HTML_MESSAGE,MAILING_LIST_MULTI,SPF_HELO_NONE,SPF_PASS,UNPARSEABLE_RELAY, USER_AGENT_SANE_1 autolearn=no autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id B1867C3F2C6 for ; Fri, 28 Feb 2020 01:51:53 +0000 (UTC) Received: from lists.gnu.org (lists.gnu.org [209.51.188.17]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by mail.kernel.org (Postfix) with ESMTPS id 7705A24690 for ; Fri, 28 Feb 2020 01:51:53 +0000 (UTC) DMARC-Filter: OpenDMARC Filter v1.3.2 mail.kernel.org 7705A24690 Authentication-Results: mail.kernel.org; dmarc=none (p=none dis=none) header.from=c-sky.com Authentication-Results: mail.kernel.org; spf=pass smtp.mailfrom=qemu-devel-bounces+qemu-devel=archiver.kernel.org@nongnu.org Received: from localhost ([::1]:40588 helo=lists1p.gnu.org) by lists.gnu.org with esmtp (Exim 4.90_1) (envelope-from ) id 1j7Uoi-0007zd-Bu for qemu-devel@archiver.kernel.org; Thu, 27 Feb 2020 20:51:52 -0500 Received: from eggs.gnu.org ([2001:470:142:3::10]:50586) by lists.gnu.org with esmtp (Exim 4.90_1) (envelope-from ) id 1j7Unz-0007Rg-BH for qemu-devel@nongnu.org; Thu, 27 Feb 2020 20:51:09 -0500 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1j7Unw-0007Mz-QM for qemu-devel@nongnu.org; Thu, 27 Feb 2020 20:51:07 -0500 Received: from smtp2200-217.mail.aliyun.com ([121.197.200.217]:34959) by eggs.gnu.org with esmtps (TLS1.0:DHE_RSA_AES_256_CBC_SHA1:32) (Exim 4.71) (envelope-from ) id 1j7Unw-0007Db-15; Thu, 27 Feb 2020 20:51:04 -0500 X-Alimail-AntiSpam: AC=CONTINUE; BC=0.07436282|-1; CH=green; DM=CONTINUE|CONTINUE|true|0.450971-0.0173873-0.531642; DS=CONTINUE|ham_enroll_verification|0.0810743-0.000531321-0.918394; FP=0|0|0|0|0|-1|-1|-1; HT=e01a16378; MF=zhiwei_liu@c-sky.com; NM=1; PH=DS; RN=9; RT=9; SR=0; TI=SMTPD_---.GtT9WG3_1582854653; Received: from 192.168.3.18(mailfrom:zhiwei_liu@c-sky.com fp:SMTPD_---.GtT9WG3_1582854653) by smtp.aliyun-inc.com(10.147.41.143); Fri, 28 Feb 2020 09:50:53 +0800 Subject: Re: [PATCH v4 1/5] target/riscv: add vector unit stride load and store instructions To: Richard Henderson , alistair23@gmail.com, chihmin.chao@sifive.com, palmer@dabbelt.com References: <20200225103508.7651-1-zhiwei_liu@c-sky.com> <20200225103508.7651-2-zhiwei_liu@c-sky.com> <4cfb56d6-34a5-0e35-87a0-2aefaafa4221@linaro.org> From: LIU Zhiwei Message-ID: <287bde05-421c-f49c-2404-fdee183c9e12@c-sky.com> Date: Fri, 28 Feb 2020 09:50:52 +0800 User-Agent: Mozilla/5.0 (Windows NT 10.0; WOW64; rv:68.0) Gecko/20100101 Thunderbird/68.4.2 MIME-Version: 1.0 In-Reply-To: <4cfb56d6-34a5-0e35-87a0-2aefaafa4221@linaro.org> Content-Type: multipart/alternative; boundary="------------6091AF5C174EB126FE31EAF0" Content-Language: en-US X-detected-operating-system: by eggs.gnu.org: GNU/Linux 3.x [generic] [fuzzy] X-Received-From: 121.197.200.217 X-BeenThere: qemu-devel@nongnu.org X-Mailman-Version: 2.1.23 Precedence: list List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: wenmeng_zhang@c-sky.com, qemu-riscv@nongnu.org, linux-csky@vger.kernel.org, wxy194768@alibaba-inc.com, qemu-devel@nongnu.org Errors-To: qemu-devel-bounces+qemu-devel=archiver.kernel.org@nongnu.org Sender: "Qemu-devel" This is a multi-part message in MIME format. --------------6091AF5C174EB126FE31EAF0 Content-Type: text/plain; charset=utf-8; format=flowed Content-Transfer-Encoding: 8bit On 2020/2/28 3:17, Richard Henderson wrote: > On 2/25/20 2:35 AM, LIU Zhiwei wrote: >> +static bool vext_check_reg(DisasContext *s, uint32_t reg, bool widen) >> +{ >> + int legal = widen ? 2 << s->lmul : 1 << s->lmul; >> + >> + return !((s->lmul == 0x3 && widen) || (reg % legal)); >> +} >> + >> +static bool vext_check_overlap_mask(DisasContext *s, uint32_t vd, bool vm) >> +{ >> + return !(s->lmul > 1 && vm == 0 && vd == 0); >> +} >> + >> +static bool vext_check_nf(DisasContext *s, uint32_t nf) >> +{ >> + return s->lmul * (nf + 1) <= 8; >> +} > Some commentary would be good here, quoting the rule being applied. E.g. "The > destination vector register group for a masked vector instruction can only > overlap the source mask regis- > ter (v0) when LMUL=1. (Section 5.3)" Good idea. >> +static bool ld_us_op(DisasContext *s, arg_r2nfvm *a, uint8_t seq) >> +{ >> + uint8_t nf = a->nf + 1; > Perhaps NF should have the +1 done during decode, so that it cannot be > forgotten here or elsewhere. Perhaps not. It will  not be used elsewhere. And it will need one more bit in FIELD(). > E.g. > > %nf 31:3 !function=ex_plus_1 > @r2_nfvm ... ... vm:1 ..... ..... ... ..... ....... \ > &r2nfvm %nf %rs1 %rd > > Where ex_plus_1 is the obvious modification of ex_shift_1(). >> +static inline uint32_t vext_nf(uint32_t desc) >> +{ >> + return (simd_data(desc) >> 11) & 0xf; >> +} >> + >> +static inline uint32_t vext_mlen(uint32_t desc) >> +{ >> + return simd_data(desc) & 0xff; >> +} >> + >> +static inline uint32_t vext_vm(uint32_t desc) >> +{ >> + return (simd_data(desc) >> 8) & 0x1; >> +} >> + >> +static inline uint32_t vext_lmul(uint32_t desc) >> +{ >> + return (simd_data(desc) >> 9) & 0x3; >> +} > You should use FIELD() to define the fields, and then use FIELD_EX32 and > FIELD_DP32 to reference them. Nice, I will find some place to define the fields. >> +/* >> + * This function checks watchpoint before real load operation. >> + * >> + * In softmmu mode, the TLB API probe_access is enough for watchpoint check. >> + * In user mode, there is no watchpoint support now. >> + * >> + * It will triggle an exception if there is no mapping in TLB > trigger. Yes. >> + * and page table walk can't fill the TLB entry. Then the guest >> + * software can return here after process the exception or never return. >> + */ >> +static void probe_read_access(CPURISCVState *env, target_ulong addr, >> + target_ulong len, uintptr_t ra) >> +{ >> + while (len) { >> + const target_ulong pagelen = -(addr | TARGET_PAGE_MASK); >> + const target_ulong curlen = MIN(pagelen, len); >> + >> + probe_read(env, addr, curlen, cpu_mmu_index(env, false), ra); >> + addr += curlen; >> + len -= curlen; >> + } >> +} >> + >> +static void probe_write_access(CPURISCVState *env, target_ulong addr, >> + target_ulong len, uintptr_t ra) >> +{ >> + while (len) { >> + const target_ulong pagelen = -(addr | TARGET_PAGE_MASK); >> + const target_ulong curlen = MIN(pagelen, len); >> + >> + probe_write(env, addr, curlen, cpu_mmu_index(env, false), ra); >> + addr += curlen; >> + len -= curlen; >> + } >> +} > A loop is overkill -- the access cannot span to 3 pages. Yes, I will just do as you suggest! In the unit stride load, without mask,  the max access len is checked . It is 512 in bytes. And current target page is 4096 in bytes. #define TARGET_PAGE_BITS 12 /* 4 KiB Pages */ > These two functions > can be merged using probe_access and MMU_DATA_{LOAD,STORE}. > >> + >> +#ifdef HOST_WORDS_BIGENDIAN >> +static void vext_clear(void *tail, uint32_t cnt, uint32_t tot) >> +{ >> + /* >> + * Split the remaining range to two parts. >> + * The first part is in the last uint64_t unit. >> + * The second part start from the next uint64_t unit. >> + */ >> + int part1 = 0, part2 = tot - cnt; >> + if (cnt % 64) { >> + part1 = 64 - (cnt % 64); >> + part2 = tot - cnt - part1; >> + memset(tail & ~(63ULL), 0, part1); >> + memset((tail + 64) & ~(63ULL), 0, part2); > You're confusing bit and byte offsets -- cnt and tot are both byte offsets. Yes, I will fix it. >> +static inline int vext_elem_mask(void *v0, int mlen, int index) >> +{ >> + >> + int idx = (index * mlen) / 8; >> + int pos = (index * mlen) % 8; >> + >> + switch (mlen) { >> + case 8: >> + return *((uint8_t *)v0 + H1(index)) & 0x1; >> + case 16: >> + return *((uint16_t *)v0 + H2(index)) & 0x1; >> + case 32: >> + return *((uint32_t *)v0 + H4(index)) & 0x1; >> + case 64: >> + return *((uint64_t *)v0 + index) & 0x1; >> + default: >> >> >> >> >> >> >> >> >> >> + return (*((uint8_t *)v0 + H1(idx)) >> pos) & 0x1; >> + } > This is not what I had in mind, and looks wrong as well. > > int idx = (index * mlen) / 64; > int pos = (index * mlen) % 64; > return (((uint64_t *)v0)[idx] >> pos) & 1; > > You also might consider passing log2(mlen), so the multiplication could be > strength-reduced to a shift. I don't think so. For example, when mlen is 8 bits and index is 0, it will reduce to return (((uint64_t *)v0)[0]) & 1 And it's not right. The right bit is first bit in vector register 0. And in host big endianess, it will be  the first bit of the seventh byte. > >> +#define GEN_VEXT_LD_ELEM(NAME, MTYPE, ETYPE, H, LDSUF) \ >> +static void vext_##NAME##_ld_elem(CPURISCVState *env, abi_ptr addr, \ >> + uint32_t idx, void *vd, uintptr_t retaddr) \ >> +{ \ >> + int mmu_idx = cpu_mmu_index(env, false); \ >> + MTYPE data; \ >> + ETYPE *cur = ((ETYPE *)vd + H(idx)); \ >> + data = cpu_##LDSUF##_mmuidx_ra(env, addr, mmu_idx, retaddr); \ >> + *cur = data; \ >> +} \ > If you're going to use cpu_mmu_index, you might as well use cpu_SUFF_data_ra(), > which does not require the mmu_idx parameter. Good. >> +#define GEN_VEXT_ST_ELEM(NAME, ETYPE, H, STSUF) \ >> +static void vext_##NAME##_st_elem(CPURISCVState *env, abi_ptr addr, \ >> + uint32_t idx, void *vd, uintptr_t retaddr) \ >> +{ \ >> + int mmu_idx = cpu_mmu_index(env, false); \ >> + ETYPE data = *((ETYPE *)vd + H(idx)); \ >> + cpu_##STSUF##_mmuidx_ra(env, addr, data, mmu_idx, retaddr); \ >> +} > Likewise. > >> +/* >> + *** unit-stride: load vector element from continuous guest memory >> + */ >> +static inline void vext_ld_us_mask(void *vd, void *v0, target_ulong base, >> + CPURISCVState *env, uint32_t desc, >> + vext_ld_elem_fn ld_elem, >> + vext_ld_clear_elem clear_elem, >> + uint32_t esz, uint32_t msz, uintptr_t ra) >> +{ >> + uint32_t i, k; >> + uint32_t mlen = vext_mlen(desc); > You don't need to pass mlen, since it's Yes. >> +/* unit-stride: store vector element to guest memory */ >> +static void vext_st_us_mask(void *vd, void *v0, target_ulong base, >> + CPURISCVState *env, uint32_t desc, >> + vext_st_elem_fn st_elem, >> + uint32_t esz, uint32_t msz, uintptr_t ra) >> +{ >> + uint32_t i, k; >> + uint32_t nf = vext_nf(desc); >> + uint32_t mlen = vext_mlen(desc); >> + uint32_t vlmax = vext_maxsz(desc) / esz; >> + >> + /* probe every access*/ >> + for (i = 0; i < env->vl; i++) { >> + if (!vext_elem_mask(v0, mlen, i)) { >> + continue; >> + } >> + probe_write_access(env, base + nf * i * msz, nf * msz, ra); >> + } >> + /* store bytes to guest memory */ >> + for (i = 0; i < env->vl; i++) { >> + k = 0; >> + if (!vext_elem_mask(v0, mlen, i)) { >> + continue; >> + } >> + while (k < nf) { >> + target_ulong addr = base + (i * nf + k) * msz; >> + st_elem(env, addr, i + k * vlmax, vd, ra); >> + k++; >> + } >> + } >> +} > I'll note that vext_ld_us_mask and vext_st_us_mask are identical, except for: > > 1) probe_read/write_access (which I already suggested merging, using > MMUAccessType), > > 2) the name of the ld_elem/st_elem variable (the function types are already > identical), and > > 3) the clear loop at the end of the load (which could be conditional on > clear_elem != NULL; after inlining, this should be optimized away). Good idea. Thanks. Zhiwei >> +static void vext_st_us(void *vd, target_ulong base, >> + CPURISCVState *env, uint32_t desc, >> + vext_st_elem_fn st_elem, >> + uint32_t esz, uint32_t msz, uintptr_t ra) > Similarly. > > > r~ --------------6091AF5C174EB126FE31EAF0 Content-Type: text/html; charset=utf-8 Content-Transfer-Encoding: 8bit

On 2020/2/28 3:17, Richard Henderson wrote:
On 2/25/20 2:35 AM, LIU Zhiwei wrote:

+static bool vext_check_reg(DisasContext *s, uint32_t reg, bool widen)
+{
+    int legal = widen ? 2 << s->lmul : 1 << s->lmul;
+
+    return !((s->lmul == 0x3 && widen) || (reg % legal));
+}
+
+static bool vext_check_overlap_mask(DisasContext *s, uint32_t vd, bool vm)
+{
+    return !(s->lmul > 1 && vm == 0 && vd == 0);
+}
+
+static bool vext_check_nf(DisasContext *s, uint32_t nf)
+{
+    return s->lmul * (nf + 1) <= 8;
+}

Some commentary would be good here, quoting the rule being applied.  E.g. "The
destination vector register group for a masked vector instruction can only
overlap the source mask regis-
ter (v0) when LMUL=1. (Section 5.3)"
Good idea.

      

        
+static bool ld_us_op(DisasContext *s, arg_r2nfvm *a, uint8_t seq)
+{
+    uint8_t nf = a->nf + 1;


      

      
Perhaps NF should have the +1 done during decode, so that it cannot be
forgotten here or elsewhere. 

    

    

    Perhaps not. It will  not be used elsewhere. And it will need one
    more bit in FIELD().

    

      
 E.g.

%nf      31:3  !function=ex_plus_1
@r2_nfvm ... ... vm:1 ..... ..... ... ..... ....... \
         &r2nfvm %nf %rs1 %rd

Where ex_plus_1 is the obvious modification of ex_shift_1().


    

    

      

      

        
+static inline uint32_t vext_nf(uint32_t desc)
+{
+    return (simd_data(desc) >> 11) & 0xf;
+}
+
+static inline uint32_t vext_mlen(uint32_t desc)
+{
+    return simd_data(desc) & 0xff;
+}
+
+static inline uint32_t vext_vm(uint32_t desc)
+{
+    return (simd_data(desc) >> 8) & 0x1;
+}
+
+static inline uint32_t vext_lmul(uint32_t desc)
+{
+    return (simd_data(desc) >> 9) & 0x3;
+}


      

      
You should use FIELD() to define the fields, and then use FIELD_EX32 and
FIELD_DP32 to reference them.


    

    Nice, I will find some place to define the fields.

    

      

      

        
+/*
+ * This function checks watchpoint before real load operation.
+ *
+ * In softmmu mode, the TLB API probe_access is enough for watchpoint check.
+ * In user mode, there is no watchpoint support now.
+ *
+ * It will triggle an exception if there is no mapping in TLB


      

      
trigger.


    

    Yes.

    

      

      

        
+ * and page table walk can't fill the TLB entry. Then the guest
+ * software can return here after process the exception or never return.
+ */
+static void probe_read_access(CPURISCVState *env, target_ulong addr,
+        target_ulong len, uintptr_t ra)
+{
+    while (len) {
+        const target_ulong pagelen = -(addr | TARGET_PAGE_MASK);
+        const target_ulong curlen = MIN(pagelen, len);
+
+        probe_read(env, addr, curlen, cpu_mmu_index(env, false), ra);
+        addr += curlen;
+        len -= curlen;
+    }
+}
+
+static void probe_write_access(CPURISCVState *env, target_ulong addr,
+        target_ulong len, uintptr_t ra)
+{
+    while (len) {
+        const target_ulong pagelen = -(addr | TARGET_PAGE_MASK);
+        const target_ulong curlen = MIN(pagelen, len);
+
+        probe_write(env, addr, curlen, cpu_mmu_index(env, false), ra);
+        addr += curlen;
+        len -= curlen;
+    }
+}


      

      
A loop is overkill -- the access cannot span to 3 pages.  

    

    Yes, I will just do as you suggest!

    

    In the unit stride load, without mask,  the max access len is
    checked . It is 512 in bytes.

    And current target page is 4096 in bytes.

    

    #define TARGET_PAGE_BITS 12 /* 4 KiB Pages */

    

    

      
These two functions
can be merged using probe_access and MMU_DATA_{LOAD,STORE}.



      

        
+
+#ifdef HOST_WORDS_BIGENDIAN
+static void vext_clear(void *tail, uint32_t cnt, uint32_t tot)
+{
+    /*
+     * Split the remaining range to two parts.
+     * The first part is in the last uint64_t unit.
+     * The second part start from the next uint64_t unit.
+     */
+    int part1 = 0, part2 = tot - cnt;
+    if (cnt % 64) {
+        part1 = 64 - (cnt % 64);
+        part2 = tot - cnt - part1;
+        memset(tail & ~(63ULL), 0, part1);
+        memset((tail + 64) & ~(63ULL), 0, part2);


      

      
You're confusing bit and byte offsets -- cnt and tot are both byte offsets.


    

    Yes, I will fix it.

    

      

      

        
+static inline int vext_elem_mask(void *v0, int mlen, int index)
+{
+
+    int idx = (index * mlen) / 8;
+    int pos = (index * mlen) % 8;
+
+    switch (mlen) {
+    case 8:
+        return *((uint8_t *)v0 + H1(index)) & 0x1;
+    case 16:
+        return *((uint16_t *)v0 + H2(index)) & 0x1;
+    case 32:
+        return *((uint32_t *)v0 + H4(index)) & 0x1;
+    case 64:
+        return *((uint64_t *)v0 + index) & 0x1;
+    default:

        

          
            

              

              		
              

              		
              

              		
              

              		
              

              		
              

              		
              

              		
              

              		
            

          
        

        
+        return (*((uint8_t *)v0 + H1(idx)) >> pos) & 0x1;
+    }


      

      
This is not what I had in mind, and looks wrong as well.

    int idx = (index * mlen) / 64;
    int pos = (index * mlen) % 64;
    return (((uint64_t *)v0)[idx] >> pos) & 1;

You also might consider passing log2(mlen), so the multiplication could be
strength-reduced to a shift.

    

    I don't think so. For example, when mlen is 8 bits and index is 0,
    it will reduce to

    
return (((uint64_t *)v0)[0]) & 1

    And it's not right. 

    

    The right bit is first bit in vector register 0. And in host big
    endianess, 

    it will be  the first bit of the seventh byte. 

    

      



      

        
+#define GEN_VEXT_LD_ELEM(NAME, MTYPE, ETYPE, H, LDSUF)              \
+static void vext_##NAME##_ld_elem(CPURISCVState *env, abi_ptr addr, \
+        uint32_t idx, void *vd, uintptr_t retaddr)                  \
+{                                                                   \
+    int mmu_idx = cpu_mmu_index(env, false);                        \
+    MTYPE data;                                                     \
+    ETYPE *cur = ((ETYPE *)vd + H(idx));                            \
+    data = cpu_##LDSUF##_mmuidx_ra(env, addr, mmu_idx, retaddr);    \
+    *cur = data;                                                    \
+}                                                                   \


      

      
If you're going to use cpu_mmu_index, you might as well use cpu_SUFF_data_ra(),
which does not require the mmu_idx parameter.


    

    Good.

    

      

      

        
+#define GEN_VEXT_ST_ELEM(NAME, ETYPE, H, STSUF)                       \
+static void vext_##NAME##_st_elem(CPURISCVState *env, abi_ptr addr,   \
+        uint32_t idx, void *vd, uintptr_t retaddr)                    \
+{                                                                     \
+    int mmu_idx = cpu_mmu_index(env, false);                          \
+    ETYPE data = *((ETYPE *)vd + H(idx));                             \
+    cpu_##STSUF##_mmuidx_ra(env, addr, data, mmu_idx, retaddr);       \
+}


      

      
Likewise.



      

        
+/*
+ *** unit-stride: load vector element from continuous guest memory
+ */
+static inline void vext_ld_us_mask(void *vd, void *v0, target_ulong base,
+        CPURISCVState *env, uint32_t desc,
+        vext_ld_elem_fn ld_elem,
+        vext_ld_clear_elem clear_elem,
+        uint32_t esz, uint32_t msz, uintptr_t ra)
+{
+    uint32_t i, k;
+    uint32_t mlen = vext_mlen(desc);


      

      
You don't need to pass mlen, since it's


    

    Yes.

    

      

      

        
+/* unit-stride: store vector element to guest memory */
+static void vext_st_us_mask(void *vd, void *v0, target_ulong base,
+        CPURISCVState *env, uint32_t desc,
+        vext_st_elem_fn st_elem,
+        uint32_t esz, uint32_t msz, uintptr_t ra)
+{
+    uint32_t i, k;
+    uint32_t nf = vext_nf(desc);
+    uint32_t mlen = vext_mlen(desc);
+    uint32_t vlmax = vext_maxsz(desc) / esz;
+
+    /* probe every access*/
+    for (i = 0; i < env->vl; i++) {
+        if (!vext_elem_mask(v0, mlen, i)) {
+            continue;
+        }
+        probe_write_access(env, base + nf * i * msz, nf * msz, ra);
+    }
+    /* store bytes to guest memory */
+    for (i = 0; i < env->vl; i++) {
+        k = 0;
+        if (!vext_elem_mask(v0, mlen, i)) {
+            continue;
+        }
+        while (k < nf) {
+            target_ulong addr = base + (i * nf + k) * msz;
+            st_elem(env, addr, i + k * vlmax, vd, ra);
+            k++;
+        }
+    }
+}


      

      
I'll note that vext_ld_us_mask and vext_st_us_mask are identical, except for:

1) probe_read/write_access (which I already suggested merging, using
MMUAccessType),

2) the name of the ld_elem/st_elem variable (the function types are already
identical), and

3) the clear loop at the end of the load (which could be conditional on
clear_elem != NULL; after inlining, this should be optimized away).


    

    Good idea. Thanks.

    

    Zhiwei

    

      

      

        
+static void vext_st_us(void *vd, target_ulong base,
+        CPURISCVState *env, uint32_t desc,
+        vext_st_elem_fn st_elem,
+        uint32_t esz, uint32_t msz, uintptr_t ra)


      

      
Similarly.


r~


    

    

  


--------------6091AF5C174EB126FE31EAF0--