Hexagon: Add memcpy and memset accelerated functions

Signed-off-by: Richard Kuo <[e-mail]>
Acked-by: Arnd Bergmann <[e-mail]>
Signed-off-by: Linus Torvalds <[e-mail]>

arch/hexagon/lib/memcpy.S

···
+/*
+ * Copyright (c) 2010-2011, Code Aurora Forum. 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 version 2 and
+ * only version 2 as published by the Free Software Foundation.
+ *
+ * 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.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
+ * 02110-1301, USA.
+ */
+
+/*
+ * Description
+ *
+ *   library function for memcpy where length bytes are copied from
+ *   ptr_in to ptr_out. ptr_out is returned unchanged.
+ *   Allows any combination of alignment on input and output pointers
+ *   and length from 0 to 2^32-1
+ *
+ * Restrictions
+ *   The arrays should not overlap, the program will produce undefined output
+ *   if they do.
+ *   For blocks less than 16 bytes a byte by byte copy is performed. For
+ *   8byte alignments, and length multiples, a dword copy is performed up to
+ *   96bytes
+ * History
+ *
+ *   DJH  5/15/09 Initial version 1.0
+ *   DJH  6/ 1/09 Version 1.1 modified ABI to inlcude R16-R19
+ *   DJH  7/12/09 Version 1.2 optimized codesize down to 760 was 840
+ *   DJH 10/14/09 Version 1.3 added special loop for aligned case, was
+ *                            overreading bloated codesize back up to 892
+ *   DJH  4/20/10 Version 1.4 fixed Ldword_loop_epilog loop to prevent loads
+ *                            occuring if only 1 left outstanding, fixes bug
+ *                            # 3888, corrected for all alignments. Peeled off
+ *                            1 32byte chunk from kernel loop and extended 8byte
+ *                            loop at end to solve all combinations and prevent
+ *                            over read.  Fixed Ldword_loop_prolog to prevent
+ *                            overread for blocks less than 48bytes. Reduced
+ *                            codesize to 752 bytes
+ *   DJH  4/21/10 version 1.5 1.4 fix broke code for input block ends not
+ *                            aligned to dword boundaries,underwriting by 1
+ *                            byte, added detection for this and fixed. A
+ *                            little bloat.
+ *   DJH  4/23/10 version 1.6 corrected stack error, R20 was not being restored
+ *                            always, fixed the error of R20 being modified
+ *                            before it was being saved
+ * Natural c model
+ * ===============
+ * void * memcpy(char * ptr_out, char * ptr_in, int length) {
+ *   int i;
+ *   if(length) for(i=0; i < length; i++) { ptr_out[i] = ptr_in[i]; }
+ *   return(ptr_out);
+ * }
+ *
+ * Optimized memcpy function
+ * =========================
+ * void * memcpy(char * ptr_out, char * ptr_in, int len) {
+ *   int i, prolog, kernel, epilog, mask;
+ *   u8 offset;
+ *   s64 data0, dataF8, data70;
+ *
+ *   s64 * ptr8_in;
+ *   s64 * ptr8_out;
+ *   s32 * ptr4;
+ *   s16 * ptr2;
+ *
+ *   offset = ((int) ptr_in) & 7;
+ *   ptr8_in = (s64 *) &ptr_in[-offset];   //read in the aligned pointers
+ *
+ *   data70 = *ptr8_in++;
+ *   dataF8 = *ptr8_in++;
+ *
+ *   data0 = HEXAGON_P_valignb_PPp(dataF8, data70, offset);
+ *
+ *   prolog = 32 - ((int) ptr_out);
+ *   mask  = 0x7fffffff >> HEXAGON_R_cl0_R(len);
+ *   prolog = prolog & mask;
+ *   kernel = len - prolog;
+ *   epilog = kernel & 0x1F;
+ *   kernel = kernel>>5;
+ *
+ *   if (prolog & 1) { ptr_out[0] = (u8) data0; data0 >>= 8; ptr_out += 1;}
+ *   ptr2 = (s16 *) &ptr_out[0];
+ *   if (prolog & 2) { ptr2[0] = (u16) data0;  data0 >>= 16; ptr_out += 2;}
+ *   ptr4 = (s32 *) &ptr_out[0];
+ *   if (prolog & 4) { ptr4[0] = (u32) data0;  data0 >>= 32; ptr_out += 4;}
+ *
+ *   offset = offset + (prolog & 7);
+ *   if (offset >= 8) {
+ *     data70 = dataF8;
+ *     dataF8 = *ptr8_in++;
+ *   }
+ *   offset = offset & 0x7;
+ *
+ *   prolog = prolog >> 3;
+ *   if (prolog) for (i=0; i < prolog; i++) {
+ *       data0 = HEXAGON_P_valignb_PPp(dataF8, data70, offset);
+ *       ptr8_out = (s64 *) &ptr_out[0]; *ptr8_out = data0; ptr_out += 8;
+ *       data70 = dataF8;
+ *       dataF8 = *ptr8_in++;
+ *   }
+ *   if(kernel) { kernel -= 1; epilog += 32; }
+ *   if(kernel) for(i=0; i < kernel; i++) {
+ *       data0 = HEXAGON_P_valignb_PPp(dataF8, data70, offset);
+ *       ptr8_out = (s64 *) &ptr_out[0]; *ptr8_out = data0; ptr_out += 8;
+ *       data70 = *ptr8_in++;
+ *
+ *       data0 = HEXAGON_P_valignb_PPp(data70, dataF8, offset);
+ *       ptr8_out = (s64 *) &ptr_out[0]; *ptr8_out = data0; ptr_out += 8;
+ *       dataF8 = *ptr8_in++;
+ *
+ *       data0 = HEXAGON_P_valignb_PPp(dataF8, data70, offset);
+ *       ptr8_out = (s64 *) &ptr_out[0]; *ptr8_out = data0; ptr_out += 8;
+ *       data70 = *ptr8_in++;
+ *
+ *       data0 = HEXAGON_P_valignb_PPp(data70, dataF8, offset);
+ *       ptr8_out = (s64 *) &ptr_out[0]; *ptr8_out = data0; ptr_out += 8;
+ *       dataF8 = *ptr8_in++;
+ *   }
+ *   epilogdws = epilog >> 3;
+ *   if (epilogdws) for (i=0; i < epilogdws; i++) {
+ *       data0 = HEXAGON_P_valignb_PPp(dataF8, data70, offset);
+ *       ptr8_out = (s64 *) &ptr_out[0]; *ptr8_out = data0; ptr_out += 8;
+ *       data70 = dataF8;
+ *       dataF8 = *ptr8_in++;
+ *   }
+ *   data0 = HEXAGON_P_valignb_PPp(dataF8, data70, offset);
+ *
+ *   ptr4 = (s32 *) &ptr_out[0];
+ *   if (epilog & 4) { ptr4[0] = (u32) data0; data0 >>= 32; ptr_out += 4;}
+ *   ptr2 = (s16 *) &ptr_out[0];
+ *   if (epilog & 2) { ptr2[0] = (u16) data0; data0 >>= 16; ptr_out += 2;}
+ *   if (epilog & 1) { *ptr_out++ = (u8) data0; }
+ *
+ *   return(ptr_out - length);
+ * }
+ *
+ * Codesize : 784 bytes
+ */
+
+
+#define ptr_out		R0	/*  destination  pounter  */
+#define ptr_in		R1	/*  source pointer  */
+#define len		R2	/*  length of copy in bytes  */
+
+#define data70		R13:12	/*  lo 8 bytes of non-aligned transfer  */
+#define dataF8		R11:10	/*  hi 8 bytes of non-aligned transfer  */
+#define ldata0		R7:6	/*  even 8 bytes chunks  */
+#define ldata1		R25:24	/*  odd 8 bytes chunks  */
+#define data1		R7	/*  lower 8 bytes of ldata1  */
+#define data0		R6	/*  lower 8 bytes of ldata0  */
+
+#define ifbyte		p0	/*  if transfer has bytes in epilog/prolog  */
+#define ifhword		p0	/*  if transfer has shorts in epilog/prolog  */
+#define ifword		p0	/*  if transfer has words in epilog/prolog  */
+#define noprolog	p0	/*  no prolog, xfer starts at 32byte  */
+#define nokernel	p1	/*  no 32byte multiple block in the transfer  */
+#define noepilog	p0	/*  no epilog, xfer ends on 32byte boundary  */
+#define align		p2	/*  alignment of input rel to 8byte boundary  */
+#define kernel1		p0	/*  kernel count == 1  */
+
+#define dalign		R25	/*  rel alignment of input to output data  */
+#define star3		R16	/*  number bytes in prolog - dwords  */
+#define rest		R8	/*  length - prolog bytes  */
+#define back		R7	/*  nr bytes > dword boundary in src block  */
+#define epilog		R3	/*  bytes in epilog  */
+#define inc		R15:14	/*  inc kernel by -1 and defetch ptr by 32  */
+#define kernel		R4	/*  number of 32byte chunks in kernel  */
+#define ptr_in_p_128	R5	/*  pointer for prefetch of input data  */
+#define mask		R8	/*  mask used to determine prolog size  */
+#define shift		R8	/*  used to work a shifter to extract bytes  */
+#define shift2		R5	/*  in epilog to workshifter to extract bytes */
+#define prolog		R15	/*  bytes in  prolog  */
+#define epilogdws	R15	/*  number dwords in epilog  */
+#define shiftb		R14	/*  used to extract bytes  */
+#define offset		R9	/*  same as align in reg  */
+#define ptr_out_p_32	R17	/*  pointer to output dczero  */
+#define align888	R14	/*  if simple dword loop can be used  */
+#define len8		R9	/*  number of dwords in length  */
+#define over		R20	/*  nr of bytes > last inp buf dword boundary */
+
+#define ptr_in_p_128kernel	R5:4	/*  packed fetch pointer & kernel cnt */
+
+	.section .text
+	.p2align 4
+        .global memcpy
+        .type memcpy, @function
+memcpy:
+{
+	p2 = cmp.eq(len, #0);		/*  =0 */
+	align888 = or(ptr_in, ptr_out);	/*  %8 < 97 */
+	p0 = cmp.gtu(len, #23);		/*  %1, <24 */
+	p1 = cmp.eq(ptr_in, ptr_out);	/*  attempt to overwrite self */
+}
+{
+	p1 = or(p2, p1);
+	p3 = cmp.gtu(len, #95);		/*  %8 < 97 */
+	align888 = or(align888, len);	/*  %8 < 97 */
+	len8 = lsr(len, #3);		/*  %8 < 97 */
+}
+{
+	dcfetch(ptr_in);		/*  zero/ptrin=ptrout causes fetch */
+	p2 = bitsclr(align888, #7);	/*  %8 < 97  */
+	if(p1) jumpr r31;		/*  =0  */
+}
+{
+	p2 = and(p2,!p3);			/*  %8 < 97  */
+	if (p2.new) len = add(len, #-8);	/*  %8 < 97  */
+	if (p2.new) jump:NT .Ldwordaligned; 	/*  %8 < 97  */
+}
+{
+	if(!p0) jump .Lbytes23orless;	/*  %1, <24  */
+	mask.l = #LO(0x7fffffff);
+	/*  all bytes before line multiples of data  */
+	prolog = sub(#0, ptr_out);
+}
+{
+	/*  save r31 on stack, decrement sp by 16  */
+	allocframe(#24);
+	mask.h = #HI(0x7fffffff);
+	ptr_in_p_128 = add(ptr_in, #32);
+	back = cl0(len);
+}
+{
+	memd(sp+#0) = R17:16;		/*  save r16,r17 on stack6  */
+	r31.l = #LO(.Lmemcpy_return);	/*  set up final return pointer  */
+	prolog &= lsr(mask, back);
+	offset = and(ptr_in, #7);
+}
+{
+	memd(sp+#8) = R25:24;		/*  save r25,r24 on stack  */
+	dalign = sub(ptr_out, ptr_in);
+	r31.h = #HI(.Lmemcpy_return);	/*  set up final return pointer  */
+}
+{
+	/*  see if there if input buffer end if aligned  */
+	over = add(len, ptr_in);
+	back = add(len, offset);
+	memd(sp+#16) = R21:20;		/*  save r20,r21 on stack  */
+}
+{
+	noprolog = bitsclr(prolog, #7);
+	prolog = and(prolog, #31);
+	dcfetch(ptr_in_p_128);
+	ptr_in_p_128 = add(ptr_in_p_128, #32);
+}
+{
+	kernel = sub(len, prolog);
+	shift = asl(prolog, #3);
+	star3 = and(prolog, #7);
+	ptr_in = and(ptr_in, #-8);
+}
+{
+	prolog = lsr(prolog, #3);
+	epilog = and(kernel, #31);
+	ptr_out_p_32 = add(ptr_out, prolog);
+	over = and(over, #7);
+}
+{
+	p3 = cmp.gtu(back, #8);
+	kernel = lsr(kernel, #5);
+	dcfetch(ptr_in_p_128);
+	ptr_in_p_128 = add(ptr_in_p_128, #32);
+}
+{
+	p1 = cmp.eq(prolog, #0);
+	if(!p1.new) prolog = add(prolog, #1);
+	dcfetch(ptr_in_p_128);	/*  reserve the line 64bytes on  */
+	ptr_in_p_128 = add(ptr_in_p_128, #32);
+}
+{
+	nokernel = cmp.eq(kernel,#0);
+	dcfetch(ptr_in_p_128);	/* reserve the line 64bytes on  */
+	ptr_in_p_128 = add(ptr_in_p_128, #32);
+	shiftb = and(shift, #8);
+}
+{
+	dcfetch(ptr_in_p_128);		/*  reserve the line 64bytes on  */
+	ptr_in_p_128 = add(ptr_in_p_128, #32);
+	if(nokernel) jump .Lskip64;
+	p2 = cmp.eq(kernel, #1);	/*  skip ovr if kernel == 0  */
+}
+{
+	dczeroa(ptr_out_p_32);
+	/*  don't advance pointer  */
+	if(!p2) ptr_out_p_32 = add(ptr_out_p_32, #32);
+}
+{
+	dalign = and(dalign, #31);
+	dczeroa(ptr_out_p_32);
+}
+.Lskip64:
+{
+	data70 = memd(ptr_in++#16);
+	if(p3) dataF8 = memd(ptr_in+#8);
+	if(noprolog) jump .Lnoprolog32;
+	align = offset;
+}
+/*  upto initial 7 bytes  */
+{
+	ldata0 = valignb(dataF8, data70, align);
+	ifbyte = tstbit(shift,#3);
+	offset = add(offset, star3);
+}
+{
+	if(ifbyte) memb(ptr_out++#1) = data0;
+	ldata0 = lsr(ldata0, shiftb);
+	shiftb = and(shift, #16);
+	ifhword = tstbit(shift,#4);
+}
+{
+	if(ifhword) memh(ptr_out++#2) = data0;
+	ldata0 = lsr(ldata0, shiftb);
+	ifword = tstbit(shift,#5);
+	p2 = cmp.gtu(offset, #7);
+}
+{
+	if(ifword) memw(ptr_out++#4) = data0;
+	if(p2) data70 = dataF8;
+	if(p2) dataF8 = memd(ptr_in++#8);	/*  another 8 bytes  */
+	align = offset;
+}
+.Lnoprolog32:
+{
+	p3 = sp1loop0(.Ldword_loop_prolog, prolog)
+	rest = sub(len, star3);	/*  whats left after the loop  */
+	p0 = cmp.gt(over, #0);
+}
+	if(p0) rest = add(rest, #16);
+.Ldword_loop_prolog:
+{
+	if(p3) memd(ptr_out++#8) = ldata0;
+	ldata0 = valignb(dataF8, data70, align);
+	p0 = cmp.gt(rest, #16);
+}
+{
+	data70 = dataF8;
+	if(p0) dataF8 = memd(ptr_in++#8);
+	rest = add(rest, #-8);
+}:endloop0
+.Lkernel:
+{
+	/*  kernel is at least 32bytes  */
+	p3 = cmp.gtu(kernel, #0);
+	/*  last itn. remove edge effects  */
+	if(p3.new) kernel = add(kernel, #-1);
+	/*  dealt with in last dword loop  */
+	if(p3.new) epilog = add(epilog, #32);
+}
+{
+	nokernel = cmp.eq(kernel, #0);		/*  after adjustment, recheck */
+	if(nokernel.new) jump:NT .Lepilog;	/*  likely not taken  */
+	inc = combine(#32, #-1);
+	p3 = cmp.gtu(dalign, #24);
+}
+{
+	if(p3) jump .Lodd_alignment;
+}
+{
+	loop0(.Loword_loop_25to31, kernel);
+	kernel1 = cmp.gtu(kernel, #1);
+	rest = kernel;
+}
+	.falign
+.Loword_loop_25to31:
+{
+	dcfetch(ptr_in_p_128);	/*  prefetch 4 lines ahead  */
+	if(kernel1) ptr_out_p_32 = add(ptr_out_p_32, #32);
+}
+{
+	dczeroa(ptr_out_p_32);	/*  reserve the next 32bytes in cache  */
+	p3 = cmp.eq(kernel, rest);
+}
+{
+	/*  kernel -= 1  */
+	ptr_in_p_128kernel = vaddw(ptr_in_p_128kernel, inc);
+	/*  kill write on first iteration  */
+	if(!p3) memd(ptr_out++#8) = ldata1;
+	ldata1 = valignb(dataF8, data70, align);
+	data70 = memd(ptr_in++#8);
+}
+{
+	memd(ptr_out++#8) = ldata0;
+	ldata0 = valignb(data70, dataF8, align);
+	dataF8 = memd(ptr_in++#8);
+}
+{
+	memd(ptr_out++#8) = ldata1;
+	ldata1 = valignb(dataF8, data70, align);
+	data70 = memd(ptr_in++#8);
+}
+{
+	memd(ptr_out++#8) = ldata0;
+	ldata0 = valignb(data70, dataF8, align);
+	dataF8 = memd(ptr_in++#8);
+	kernel1 = cmp.gtu(kernel, #1);
+}:endloop0
+{
+	memd(ptr_out++#8) = ldata1;
+	jump .Lepilog;
+}
+.Lodd_alignment:
+{
+	loop0(.Loword_loop_00to24, kernel);
+	kernel1 = cmp.gtu(kernel, #1);
+	rest = add(kernel, #-1);
+}
+	.falign
+.Loword_loop_00to24:
+{
+	dcfetch(ptr_in_p_128);	/*  prefetch 4 lines ahead  */
+	ptr_in_p_128kernel = vaddw(ptr_in_p_128kernel, inc);
+	if(kernel1) ptr_out_p_32 = add(ptr_out_p_32, #32);
+}
+{
+	dczeroa(ptr_out_p_32);	/*  reserve the next 32bytes in cache  */
+}
+{
+	memd(ptr_out++#8) = ldata0;
+	ldata0 = valignb(dataF8, data70, align);
+	data70 = memd(ptr_in++#8);
+}
+{
+	memd(ptr_out++#8) = ldata0;
+	ldata0 = valignb(data70, dataF8, align);
+	dataF8 = memd(ptr_in++#8);
+}
+{
+	memd(ptr_out++#8) = ldata0;
+	ldata0 = valignb(dataF8, data70, align);
+	data70 = memd(ptr_in++#8);
+}
+{
+	memd(ptr_out++#8) = ldata0;
+	ldata0 = valignb(data70, dataF8, align);
+	dataF8 = memd(ptr_in++#8);
+	kernel1 = cmp.gtu(kernel, #1);
+}:endloop0
+.Lepilog:
+{
+	noepilog = cmp.eq(epilog,#0);
+	epilogdws = lsr(epilog, #3);
+	kernel = and(epilog, #7);
+}
+{
+	if(noepilog) jumpr r31;
+	if(noepilog) ptr_out = sub(ptr_out, len);
+	p3 = cmp.eq(epilogdws, #0);
+	shift2 = asl(epilog, #3);
+}
+{
+	shiftb = and(shift2, #32);
+	ifword = tstbit(epilog,#2);
+	if(p3) jump .Lepilog60;
+	if(!p3) epilog = add(epilog, #-16);
+}
+{
+	loop0(.Ldword_loop_epilog, epilogdws);
+	/*  stop criteria is lsbs unless = 0 then its 8  */
+	p3 = cmp.eq(kernel, #0);
+	if(p3.new) kernel= #8;
+	p1 = cmp.gt(over, #0);
+}
+	/*  if not aligned to end of buffer execute 1 more iteration  */
+	if(p1) kernel= #0;
+.Ldword_loop_epilog:
+{
+	memd(ptr_out++#8) = ldata0;
+	ldata0 = valignb(dataF8, data70, align);
+	p3 = cmp.gt(epilog, kernel);
+}
+{
+	data70 = dataF8;
+	if(p3) dataF8 = memd(ptr_in++#8);
+	epilog = add(epilog, #-8);
+}:endloop0
+/* copy last 7 bytes */
+.Lepilog60:
+{
+	if(ifword) memw(ptr_out++#4) = data0;
+	ldata0 = lsr(ldata0, shiftb);
+	ifhword = tstbit(epilog,#1);
+	shiftb = and(shift2, #16);
+}
+{
+	if(ifhword) memh(ptr_out++#2) = data0;
+	ldata0 = lsr(ldata0, shiftb);
+	ifbyte = tstbit(epilog,#0);
+	if(ifbyte.new) len = add(len, #-1);
+}
+{
+	if(ifbyte) memb(ptr_out) = data0;
+	ptr_out = sub(ptr_out, len);	/*  return dest pointer  */
+        jumpr r31;
+}
+/*  do byte copy for small n  */
+.Lbytes23orless:
+{
+	p3 = sp1loop0(.Lbyte_copy, len);
+	len = add(len, #-1);
+}
+.Lbyte_copy:
+{
+	data0 = memb(ptr_in++#1);
+	if(p3) memb(ptr_out++#1) = data0;
+}:endloop0
+{
+	memb(ptr_out) = data0;
+	ptr_out = sub(ptr_out, len);
+	jumpr r31;
+}
+/*  do dword copies for aligned in, out and length  */
+.Ldwordaligned:
+{
+	p3 = sp1loop0(.Ldword_copy, len8);
+}
+.Ldword_copy:
+{
+	if(p3) memd(ptr_out++#8) = ldata0;
+	ldata0 = memd(ptr_in++#8);
+}:endloop0
+{
+	memd(ptr_out) = ldata0;
+	ptr_out = sub(ptr_out, len);
+	jumpr r31;	/*  return to function caller  */
+}
+.Lmemcpy_return:
+	r21:20 = memd(sp+#16);	/*  restore r20+r21  */
+{
+	r25:24 = memd(sp+#8);	/*  restore r24+r25  */
+	r17:16 = memd(sp+#0);	/*  restore r16+r17  */
+}
+	deallocframe;	/*  restore r31 and incrment stack by 16  */
+	jumpr r31

arch/hexagon/lib/memset.S

···
+/*
+ * Copyright (c) 2011 Code Aurora Forum. 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 version 2 and
+ * only version 2 as published by the Free Software Foundation.
+ *
+ * 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.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
+ * 02110-1301, USA.
+ */
+
+
+/* HEXAGON assembly optimized memset */
+/* Replaces the standard library function memset */
+
+
+        .macro HEXAGON_OPT_FUNC_BEGIN name
+	.text
+	.p2align 4
+	.globl \name
+	.type  \name, @function
+\name:
+	.endm
+
+	.macro HEXAGON_OPT_FUNC_FINISH name
+	.size  \name, . - \name
+	.endm
+
+/* FUNCTION: memset (v2 version) */
+#if __HEXAGON_ARCH__ < 3
+HEXAGON_OPT_FUNC_BEGIN memset
+	{
+		r6 = #8
+		r7 = extractu(r0, #3 , #0)
+		p0 = cmp.eq(r2, #0)
+		p1 = cmp.gtu(r2, #7)
+	}
+	{
+		r4 = vsplatb(r1)
+		r8 = r0           /* leave r0 intact for return val  */
+		r9 = sub(r6, r7)  /* bytes until double alignment  */
+		if p0 jumpr r31   /* count == 0, so return  */
+	}
+	{
+		r3 = #0
+		r7 = #0
+		p0 = tstbit(r9, #0)
+		if p1 jump 2f /* skip byte loop */
+	}
+
+/* less than 8 bytes to set, so just set a byte at a time and return  */
+
+		loop0(1f, r2) /* byte loop */
+	.falign
+1: /* byte loop */
+	{
+		memb(r8++#1) = r4
+	}:endloop0
+		jumpr r31
+	.falign
+2: /* skip byte loop */
+	{
+		r6 = #1
+		p0 = tstbit(r9, #1)
+		p1 = cmp.eq(r2, #1)
+		if !p0 jump 3f /* skip initial byte store */
+	}
+	{
+		memb(r8++#1) = r4
+		r3:2 = sub(r3:2, r7:6)
+		if p1 jumpr r31
+	}
+	.falign
+3: /* skip initial byte store */
+	{
+		r6 = #2
+		p0 = tstbit(r9, #2)
+		p1 = cmp.eq(r2, #2)
+		if !p0 jump 4f /* skip initial half store */
+	}
+	{
+		memh(r8++#2) = r4
+		r3:2 = sub(r3:2, r7:6)
+		if p1 jumpr r31
+	}
+	.falign
+4: /* skip initial half store */
+	{
+		r6 = #4
+		p0 = cmp.gtu(r2, #7)
+		p1 = cmp.eq(r2, #4)
+		if !p0 jump 5f /* skip initial word store */
+	}
+	{
+		memw(r8++#4) = r4
+		r3:2 = sub(r3:2, r7:6)
+		p0 = cmp.gtu(r2, #11)
+		if p1 jumpr r31
+	}
+	.falign
+5: /* skip initial word store */
+	{
+		r10 = lsr(r2, #3)
+		p1 = cmp.eq(r3, #1)
+		if !p0 jump 7f /* skip double loop */
+	}
+	{
+		r5 = r4
+		r6 = #8
+		loop0(6f, r10) /* double loop */
+	}
+
+/* set bytes a double word at a time  */
+
+	.falign
+6: /* double loop */
+	{
+		memd(r8++#8) = r5:4
+		r3:2 = sub(r3:2, r7:6)
+		p1 = cmp.eq(r2, #8)
+	}:endloop0
+	.falign
+7: /* skip double loop */
+	{
+		p0 = tstbit(r2, #2)
+		if p1 jumpr r31
+	}
+	{
+		r6 = #4
+		p0 = tstbit(r2, #1)
+		p1 = cmp.eq(r2, #4)
+		if !p0 jump 8f /* skip final word store */
+	}
+	{
+		memw(r8++#4) = r4
+		r3:2 = sub(r3:2, r7:6)
+		if p1 jumpr r31
+	}
+	.falign
+8: /* skip final word store */
+	{
+		p1 = cmp.eq(r2, #2)
+		if !p0 jump 9f /* skip final half store */
+	}
+	{
+		memh(r8++#2) = r4
+		if p1 jumpr r31
+	}
+	.falign
+9: /* skip final half store */
+	{
+		memb(r8++#1) = r4
+		jumpr r31
+	}
+HEXAGON_OPT_FUNC_FINISH memset
+#endif
+
+
+/*  FUNCTION: memset (v3 and higher version)  */
+#if __HEXAGON_ARCH__ >= 3
+HEXAGON_OPT_FUNC_BEGIN memset
+	{
+		r7=vsplatb(r1)
+		r6 = r0
+		if (r2==#0) jump:nt .L1
+	}
+	{
+		r5:4=combine(r7,r7)
+		p0 = cmp.gtu(r2,#8)
+		if (p0.new) jump:nt .L3
+	}
+	{
+		r3 = r0
+		loop0(.L47,r2)
+	}
+	.falign
+.L47:
+	{
+		memb(r3++#1) = r1
+	}:endloop0 /* start=.L47 */
+		jumpr r31
+.L3:
+	{
+		p0 = tstbit(r0,#0)
+		if (!p0.new) jump:nt .L8
+		p1 = cmp.eq(r2, #1)
+	}
+	{
+		r6 = add(r0, #1)
+		r2 = add(r2,#-1)
+		memb(r0) = r1
+		if (p1) jump .L1
+	}
+.L8:
+	{
+		p0 = tstbit(r6,#1)
+		if (!p0.new) jump:nt .L10
+	}
+	{
+		r2 = add(r2,#-2)
+		memh(r6++#2) = r7
+		p0 = cmp.eq(r2, #2)
+		if (p0.new) jump:nt .L1
+	}
+.L10:
+	{
+		p0 = tstbit(r6,#2)
+		if (!p0.new) jump:nt .L12
+	}
+	{
+		r2 = add(r2,#-4)
+		memw(r6++#4) = r7
+		p0 = cmp.eq(r2, #4)
+		if (p0.new) jump:nt .L1
+	}
+.L12:
+	{
+		p0 = cmp.gtu(r2,#127)
+		if (!p0.new) jump:nt .L14
+	}
+		r3 = and(r6,#31)
+		if (r3==#0) jump:nt .L17
+	{
+		memd(r6++#8) = r5:4
+		r2 = add(r2,#-8)
+	}
+		r3 = and(r6,#31)
+		if (r3==#0) jump:nt .L17
+	{
+		memd(r6++#8) = r5:4
+		r2 = add(r2,#-8)
+	}
+		r3 = and(r6,#31)
+		if (r3==#0) jump:nt .L17
+	{
+		memd(r6++#8) = r5:4
+		r2 = add(r2,#-8)
+	}
+.L17:
+	{
+		r3 = lsr(r2,#5)
+		if (r1!=#0) jump:nt .L18
+	}
+	{
+		r8 = r3
+		r3 = r6
+		loop0(.L46,r3)
+	}
+	.falign
+.L46:
+	{
+		dczeroa(r6)
+		r6 = add(r6,#32)
+		r2 = add(r2,#-32)
+	}:endloop0 /* start=.L46 */
+.L14:
+	{
+		p0 = cmp.gtu(r2,#7)
+		if (!p0.new) jump:nt .L28
+		r8 = lsr(r2,#3)
+	}
+		loop0(.L44,r8)
+	.falign
+.L44:
+	{
+		memd(r6++#8) = r5:4
+		r2 = add(r2,#-8)
+	}:endloop0 /* start=.L44 */
+.L28:
+	{
+		p0 = tstbit(r2,#2)
+		if (!p0.new) jump:nt .L33
+	}
+	{
+		r2 = add(r2,#-4)
+		memw(r6++#4) = r7
+	}
+.L33:
+	{
+		p0 = tstbit(r2,#1)
+		if (!p0.new) jump:nt .L35
+	}
+	{
+		r2 = add(r2,#-2)
+		memh(r6++#2) = r7
+	}
+.L35:
+		p0 = cmp.eq(r2,#1)
+		if (p0) memb(r6) = r1
+.L1:
+		jumpr r31
+.L18:
+		loop0(.L45,r3)
+	.falign
+.L45:
+		dczeroa(r6)
+	{
+		memd(r6++#8) = r5:4
+		r2 = add(r2,#-32)
+	}
+		memd(r6++#8) = r5:4
+		memd(r6++#8) = r5:4
+	{
+		memd(r6++#8) = r5:4
+	}:endloop0 /* start=.L45  */
+		jump .L14
+HEXAGON_OPT_FUNC_FINISH memset
+#endif