void PPCXEmitter::Epilogue() {
u32 regSize = 8; // 4 in 32bit system
u32 stackFrameSize = 0x1F0;
//Break();
// Write Epilogue (restore stack frame, return)
// free stack
ADDI(R1, R1, stackFrameSize);
#if 0
ADDI(R12, R1, -0x98);
// Restore fpr
for(int i = 14; i < 32; i ++) {
LFD((PPCReg)i, R1, -((32 - i) * regSize));
}
#endif
// Restore gpr
for(int i = 14; i < 32; i ++) {
LD((PPCReg)i, R1, -((33 - i) * regSize));
}
// recover r12 (LR saved register)
LWZ (R12, R1, -0x8);
// Restore Lr
MTLR(R12);
#if 1
// load fpr buff
MOVI2R(R5, (u32)&_fprTmp);
// Load fpr
for(int i = 14; i < 32; i ++) {
LFD((PPCReg)i, R5, i * regSize);
}
#endif
}
void PPCXEmitter::Epilogue() {
u32 regSize = 8; // 4 in 32bit system
u32 stackFrameSize = 32*32;//(35 - 12) * regSize;
// Write Epilogue (restore stack frame, return)
// free stack
ADDI(R1, R1, stackFrameSize);
// Restore regs
for(int i = 14; i < 32; i ++) {
LD((PPCReg)i, R1, -((33 - i) * regSize));
}
// recover r12 (LR saved register)
LWZ (R12, R1, -0x8);
// Restore Lr
MTLR(R12);
}
void PPCXEmitter::Prologue() {
// Save regs
u32 regSize = 8; // 4 in 32bit system
u32 stackFrameSize = 0x1F0;
// Write Prologue (setup stack frame etc ...)
// Save Lr
MFLR(R12);
// Save gpr
for(int i = 14; i < 32; i ++) {
STD((PPCReg)i, R1, -((33 - i) * regSize));
}
// Save r12
STW(R12, R1, -0x8);
#if 0
// add fpr frame
ADDI(R12, R1, -0x98);
// Load fpr
for(int i = 14; i < 32; i ++) {
SFD((PPCReg)i, R1, -((32 - i) * regSize));
}
#endif
// allocate stack
STWU(R1, R1, -stackFrameSize);
#if 1
// load fpr buff
MOVI2R(R10, (u32)&_fprTmp);
// Save fpr
for(int i = 14; i < 32; i ++) {
SFD((PPCReg)i, R10, i * regSize);
}
#endif
}
/* Process 2OPI Integer instructions */
bool eval_2OPI_Int(struct lilith* vm, struct Instruction* c)
{
#ifdef DEBUG
char Name[20] = "ILLEGAL_2OPI";
#endif
/* 0x0E ... 0x2B */
/* 0xB0 ... 0xDF */
switch(c->raw2)
{
case 0x0E: /* ADDI */
{
#ifdef DEBUG
strncpy(Name, "ADDI", 19);
#elif TRACE
record_trace("ADDI");
#endif
ADDI(vm, c);
break;
}
case 0x0F: /* ADDUI */
{
#ifdef DEBUG
strncpy(Name, "ADDUI", 19);
#elif TRACE
record_trace("ADDUI");
#endif
ADDUI(vm, c);
break;
}
case 0x10: /* SUBI */
{
#ifdef DEBUG
strncpy(Name, "SUBI", 19);
#elif TRACE
record_trace("SUBI");
#endif
SUBI(vm, c);
break;
}
case 0x11: /* SUBUI */
{
#ifdef DEBUG
strncpy(Name, "SUBUI", 19);
#elif TRACE
record_trace("SUBUI");
#endif
SUBUI(vm, c);
break;
}
case 0x12: /* CMPI */
{
#ifdef DEBUG
strncpy(Name, "CMPI", 19);
#elif TRACE
record_trace("CMPI");
#endif
CMPI(vm, c);
break;
}
case 0x13: /* LOAD */
{
#ifdef DEBUG
strncpy(Name, "LOAD", 19);
#elif TRACE
record_trace("LOAD");
#endif
LOAD(vm, c);
break;
}
case 0x14: /* LOAD8 */
{
#ifdef DEBUG
strncpy(Name, "LOAD8", 19);
#elif TRACE
record_trace("LOAD8");
#endif
LOAD8(vm, c);
break;
}
case 0x15: /* LOADU8 */
{
#ifdef DEBUG
strncpy(Name, "LOADU8", 19);
#elif TRACE
record_trace("LOADU8");
#endif
LOADU8(vm, c);
break;
}
case 0x16: /* LOAD16 */
{
#ifdef DEBUG
strncpy(Name, "LOAD16", 19);
#elif TRACE
record_trace("LOAD16");
#endif
LOAD16(vm, c);
break;
}
case 0x17: /* LOADU16 */
{
#ifdef DEBUG
strncpy(Name, "LOADU16", 19);
#elif TRACE
record_trace("LOADU16");
#endif
LOADU16(vm, c);
break;
}
case 0x18: /* LOAD32 */
{
#ifdef DEBUG
strncpy(Name, "LOAD32", 19);
#elif TRACE
record_trace("LOAD32");
#endif
LOAD32(vm, c);
break;
}
case 0x19: /* LOADU32 */
{
#ifdef DEBUG
strncpy(Name, "LOADU32", 19);
#elif TRACE
record_trace("LOADU32");
#endif
LOADU32(vm, c);
break;
}
case 0x1F: /* CMPUI */
{
#ifdef DEBUG
strncpy(Name, "CMPUI", 19);
#elif TRACE
record_trace("CMPUI");
#endif
CMPUI(vm, c);
break;
}
case 0x20: /* STORE */
{
#ifdef DEBUG
strncpy(Name, "STORE", 19);
#elif TRACE
record_trace("STORE");
#endif
STORE(vm, c);
break;
}
case 0x21: /* STORE8 */
{
#ifdef DEBUG
strncpy(Name, "STORE8", 19);
#elif TRACE
record_trace("STORE8");
#endif
STORE8(vm, c);
break;
}
case 0x22: /* STORE16 */
{
#ifdef DEBUG
strncpy(Name, "STORE16", 19);
#elif TRACE
record_trace("STORE16");
#endif
STORE16(vm, c);
break;
}
case 0x23: /* STORE32 */
{
#ifdef DEBUG
strncpy(Name, "STORE32", 19);
#elif TRACE
record_trace("STORE32");
#endif
STORE32(vm, c);
break;
}
case 0xB0: /* ANDI */
{
#ifdef DEBUG
strncpy(Name, "ANDI", 19);
#elif TRACE
record_trace("ANDI");
#endif
ANDI(vm, c);
break;
}
case 0xB1: /* ORI */
{
#ifdef DEBUG
strncpy(Name, "ORI", 19);
#elif TRACE
record_trace("ORI");
#endif
ORI(vm, c);
break;
}
case 0xB2: /* XORI */
{
#ifdef DEBUG
strncpy(Name, "XORI", 19);
#elif TRACE
record_trace("XORI");
#endif
XORI(vm, c);
break;
}
case 0xB3: /* NANDI */
{
#ifdef DEBUG
strncpy(Name, "NANDI", 19);
#elif TRACE
record_trace("NANDI");
#endif
NANDI(vm, c);
break;
}
case 0xB4: /* NORI */
{
#ifdef DEBUG
strncpy(Name, "NORI", 19);
#elif TRACE
record_trace("NORI");
#endif
NORI(vm, c);
break;
}
case 0xB5: /* XNORI */
{
#ifdef DEBUG
strncpy(Name, "XNORI", 19);
#elif TRACE
record_trace("XNORI");
#endif
XNORI(vm, c);
break;
}
case 0xC0: /* CMPJUMPI.G */
{
#ifdef DEBUG
strncpy(Name, "CMPJUMPI.G", 19);
#elif TRACE
record_trace("CMPJUMPI.G");
#endif
CMPJUMPI_G(vm, c);
break;
}
case 0xC1: /* CMPJUMPI.GE */
{
#ifdef DEBUG
strncpy(Name, "CMPJUMPI.GE", 19);
#elif TRACE
record_trace("CMPJUMPI.GE");
#endif
CMPJUMPI_GE(vm, c);
break;
}
case 0xC2: /* CMPJUMPI.E */
{
#ifdef DEBUG
strncpy(Name, "CMPJUMPI.E", 19);
#elif TRACE
record_trace("CMPJUMPI.E");
#endif
CMPJUMPI_E(vm, c);
break;
}
case 0xC3: /* CMPJUMPI.NE */
{
#ifdef DEBUG
strncpy(Name, "CMPJUMPI.NE", 19);
#elif TRACE
record_trace("CMPJUMPI.NE");
#endif
CMPJUMPI_NE(vm, c);
break;
}
case 0xC4: /* CMPJUMPI.LE */
{
#ifdef DEBUG
strncpy(Name, "CMPJUMPI.LE", 19);
#elif TRACE
record_trace("CMPJUMPI.LE");
#endif
CMPJUMPI_LE(vm, c);
break;
}
case 0xC5: /* CMPJUMPI.L */
{
#ifdef DEBUG
strncpy(Name, "CMPJUMPI.L", 19);
#elif TRACE
record_trace("CMPJUMPI.L");
#endif
CMPJUMPI_L(vm, c);
break;
}
case 0xD0: /* CMPJUMPUI.G */
{
#ifdef DEBUG
strncpy(Name, "CMPJUMPUI.G", 19);
#elif TRACE
record_trace("CMPJUMPUI.G");
#endif
CMPJUMPUI_G(vm, c);
break;
}
case 0xD1: /* CMPJUMPUI.GE */
{
#ifdef DEBUG
strncpy(Name, "CMPJUMPUI.GE", 19);
#elif TRACE
record_trace("CMPJUMPUI.GE");
#endif
CMPJUMPUI_GE(vm, c);
break;
}
case 0xD4: /* CMPJUMPUI.LE */
{
#ifdef DEBUG
strncpy(Name, "CMPJUMPUI.LE", 19);
#elif TRACE
record_trace("CMPJUMPUI.LE");
#endif
CMPJUMPUI_LE(vm, c);
break;
}
case 0xD5: /* CMPJUMPUI.L */
{
#ifdef DEBUG
strncpy(Name, "CMPJUMPUI.L", 19);
#elif TRACE
record_trace("CMPJUMPUI.L");
#endif
CMPJUMPUI_L(vm, c);
break;
}
default:
{
illegal_instruction(vm, c);
break;
}
}
#ifdef DEBUG
fprintf(stdout, "# %s reg%u reg%u %i\n", Name, c->reg0, c->reg1, c->raw_Immediate);
#endif
return false;
}
{
// Read umd patches
{ psp_read, STDU(SP, 0xFF90, SP), &condition_psp_iso },
{ psp_read+4, MFLR(R0), &condition_psp_iso },
{ psp_read+8, STD(R0, 0x80, SP), &condition_psp_iso },
{ psp_read+0x0C, MR(R8, R7), &condition_psp_iso },
{ psp_read+0x10, MR(R7, R6), &condition_psp_iso },
{ psp_read+0x14, MR(R6, R5), &condition_psp_iso },
{ psp_read+0x18, MR(R5, R4), &condition_psp_iso },
{ psp_read+0x1C, MR(R4, R3), &condition_psp_iso },
{ psp_read+0x20, LI(R3, SYSCALL8_OPCODE_READ_PSP_UMD), &condition_psp_iso },
{ psp_read+0x24, LI(R11, 8), &condition_psp_iso },
{ psp_read+0x28, SC, &condition_psp_iso },
{ psp_read+0x2C, LD(R0, 0x80, SP), &condition_psp_iso },
{ psp_read+0x30, MTLR(R0), &condition_psp_iso },
{ psp_read+0x34, ADDI(SP, SP, 0x70), &condition_psp_iso },
{ psp_read+0x38, BLR, &condition_psp_iso },
// Read header patches
{ psp_read+0x3C, STDU(SP, 0xFF90, SP), &condition_psp_iso },
{ psp_read+0x40, MFLR(R0), &condition_psp_iso },
{ psp_read+0x44, STD(R0, 0x80, SP), &condition_psp_iso },
{ psp_read+0x48, MR(R7, R6), &condition_psp_iso },
{ psp_read+0x4C, MR(R6, R5), &condition_psp_iso },
{ psp_read+0x50, MR(R5, R4), &condition_psp_iso },
{ psp_read+0x54, MR(R4, R3), &condition_psp_iso },
{ psp_read+0x58, LI(R3, SYSCALL8_OPCODE_READ_PSP_HEADER), &condition_psp_iso },
{ psp_read+0x5C, LI(R11, 8), &condition_psp_iso },
{ psp_read+0x60, SC, &condition_psp_iso },
{ psp_read+0x64, LD(R0, 0x80, SP), &condition_psp_iso },
{ psp_read+0x68, MTLR(R0), &condition_psp_iso },
{ psp_read+0x6C, ADDI(SP, SP, 0x70), &condition_psp_iso },
void ffts_generate_func_code(ffts_plan_t *p, size_t N, size_t leafN, int sign) {
int count = tree_count(N, leafN, 0) + 1;
size_t *ps = malloc(count * 2 * sizeof(size_t));
size_t *pps = ps;
#ifdef __x86_64__
if(sign < 0) p->constants = sse_constants;
else p->constants = sse_constants_inv;
#endif
elaborate_tree(&pps, N, leafN, 0);
pps[0] = 0;
pps[1] = 0;
pps = ps;
#ifdef __arm__
if(N < 8192) p->transform_size = 8192;
else p->transform_size = N;
#else
if(N < 2048) p->transform_size = 16384;
else p->transform_size = 16384 + 2*N/8 * __builtin_ctzl(N);
#endif
#ifdef __APPLE__
p->transform_base = mmap(NULL, p->transform_size, PROT_WRITE | PROT_READ, MAP_ANON | MAP_SHARED, -1, 0);
#else
#define MAP_ANONYMOUS 0x20
p->transform_base = mmap(NULL, p->transform_size, PROT_WRITE | PROT_READ, MAP_ANONYMOUS | MAP_SHARED, -1, 0);
#endif
/*
if(p->transform_base == MAP_FAILED) {
fprintf(stderr, "MAP FAILED\n");
exit(1);
}*/
insns_t *func = p->transform_base;//valloc(8192);
insns_t *fp = func;
//fprintf(stderr, "Allocating %d bytes \n", p->transform_size);
//fprintf(stderr, "Base address = %016p\n", func);
if(!func) {
fprintf(stderr, "NOMEM\n");
exit(1);
}
insns_t *x_8_addr = fp;
#ifdef __arm__
#ifdef HAVE_NEON
memcpy(fp, neon_x8, neon_x8_t - neon_x8);
if(sign < 0) {
fp[31] ^= 0x00200000; fp[32] ^= 0x00200000; fp[33] ^= 0x00200000; fp[34] ^= 0x00200000;
fp[65] ^= 0x00200000; fp[66] ^= 0x00200000; fp[70] ^= 0x00200000; fp[74] ^= 0x00200000;
fp[97] ^= 0x00200000; fp[98] ^= 0x00200000; fp[102] ^= 0x00200000; fp[104] ^= 0x00200000;
}
fp += (neon_x8_t - neon_x8) / 4;
#else
memcpy(fp, vfp_x8, vfp_end - vfp_x8);
if(sign > 0) {
fp[65] ^= 0x00000040;
fp[66] ^= 0x00000040;
fp[68] ^= 0x00000040;
fp[70] ^= 0x00000040;
fp[103] ^= 0x00000040;
fp[104] ^= 0x00000040;
fp[105] ^= 0x00000040;
fp[108] ^= 0x00000040;
fp[113] ^= 0x00000040;
fp[114] ^= 0x00000040;
fp[117] ^= 0x00000040;
fp[118] ^= 0x00000040;
}
fp += (vfp_end - vfp_x8) / 4;
#endif
#else
align_mem16(&fp, 0);
x_8_addr = fp;
align_mem16(&fp, 5);
memcpy(fp, x8_soft, x8_hard - x8_soft);
fp += (x8_hard - x8_soft);
//fprintf(stderr, "X8 start address = %016p\n", x_8_addr);
#endif
//uint32_t *x_8_t_addr = fp;
//memcpy(fp, neon_x8_t, neon_end - neon_x8_t);
//fp += (neon_end - neon_x8_t) / 4;
insns_t *x_4_addr = fp;
#ifdef __arm__
#ifdef HAVE_NEON
memcpy(fp, neon_x4, neon_x8 - neon_x4);
if(sign < 0) {
fp[26] ^= 0x00200000; fp[28] ^= 0x00200000; fp[31] ^= 0x00200000; fp[32] ^= 0x00200000;
}
fp += (neon_x8 - neon_x4) / 4;
#else
memcpy(fp, vfp_x4, vfp_x8 - vfp_x4);
if(sign > 0) {
fp[36] ^= 0x00000040;
fp[38] ^= 0x00000040;
fp[43] ^= 0x00000040;
fp[44] ^= 0x00000040;
}
fp += (vfp_x8 - vfp_x4) / 4;
#endif
#else
align_mem16(&fp, 0);
x_4_addr = fp;
memcpy(fp, x4, x8_soft - x4);
fp += (x8_soft - x4);
#endif
insns_t *start = fp;
#ifdef __arm__
*fp = PUSH_LR(); fp++;
*fp = 0xed2d8b10; fp++;
ADDI(&fp, 3, 1, 0);
ADDI(&fp, 7, 1, N);
ADDI(&fp, 5, 1, 2*N);
ADDI(&fp, 10, 7, 2*N);
ADDI(&fp, 4, 5, 2*N);
ADDI(&fp, 8, 10, 2*N);
ADDI(&fp, 6, 4, 2*N);
ADDI(&fp, 9, 8, 2*N);
*fp = LDRI(12, 0, ((uint32_t)&p->offsets) - ((uint32_t)p)); fp++; // load offsets into r12
// *fp++ = LDRI(1, 0, 4); // load ws into r1
ADDI(&fp, 1, 0, 0);
ADDI(&fp, 0, 2, 0), // mov out into r0
#endif
#ifdef __arm__
*fp = LDRI(2, 1, ((uint32_t)&p->ee_ws) - ((uint32_t)p)); fp++;
#ifdef HAVE_NEON
MOVI(&fp, 11, p->i0);
#else
MOVI(&fp, 11, p->i0);
#endif
#else
align_mem16(&fp, 0);
start = fp;
*fp++ = 0x4c;
*fp++ = 0x8b;
*fp++ = 0x07;
uint32_t lp_cnt = p->i0 * 4;
MOVI(&fp, RCX, lp_cnt);
//LEA(&fp, R8, RDI, ((uint32_t)&p->offsets) - ((uint32_t)p));
#endif
//fp++;
#ifdef __arm__
#ifdef HAVE_NEON
memcpy(fp, neon_ee, neon_oo - neon_ee);
if(sign < 0) {
fp[33] ^= 0x00200000; fp[37] ^= 0x00200000; fp[38] ^= 0x00200000; fp[39] ^= 0x00200000;
fp[40] ^= 0x00200000; fp[41] ^= 0x00200000; fp[44] ^= 0x00200000; fp[45] ^= 0x00200000;
fp[46] ^= 0x00200000; fp[47] ^= 0x00200000; fp[48] ^= 0x00200000; fp[57] ^= 0x00200000;
}
fp += (neon_oo - neon_ee) / 4;
#else
memcpy(fp, vfp_e, vfp_o - vfp_e);
if(sign > 0) {
fp[64] ^= 0x00000040; fp[65] ^= 0x00000040; fp[68] ^= 0x00000040; fp[75] ^= 0x00000040;
fp[76] ^= 0x00000040; fp[79] ^= 0x00000040; fp[80] ^= 0x00000040; fp[83] ^= 0x00000040;
fp[84] ^= 0x00000040; fp[87] ^= 0x00000040; fp[91] ^= 0x00000040; fp[93] ^= 0x00000040;
}
fp += (vfp_o - vfp_e) / 4;
#endif
#else
//fprintf(stderr, "Body start address = %016p\n", start);
PUSH(&fp, RBP);
PUSH(&fp, RBX);
PUSH(&fp, R10);
PUSH(&fp, R11);
PUSH(&fp, R12);
PUSH(&fp, R13);
PUSH(&fp, R14);
PUSH(&fp, R15);
int i;
memcpy(fp, leaf_ee_init, leaf_ee - leaf_ee_init);
//fprintf(stderr, "Leaf ee init address = %016p\n", leaf_ee_init);
//fprintf(stderr, "Constants address = %016p\n", sse_constants);
//fprintf(stderr, "Constants address = %016p\n", p->constants);
//int32_t val = READ_IMM32(fp + 3);
//fprintf(stderr, "diff = 0x%x\n", ((uint32_t)&p->constants) - ((uint32_t)p));
//int64_t v2 = val + (int64_t)((void *)leaf_ee_init - (void *)fp );
//fprintf(stderr, "IMM = 0x%llx\n", v2);
//IMM32_NI(fp + 3, ((int64_t) READ_IMM32(fp + 3)) + ((void *)leaf_ee_init - (void *)fp ));
fp += (leaf_ee - leaf_ee_init);
//fprintf(stderr, "Leaf start address = %016p\n", fp);
align_mem16(&fp, 9);
memcpy(fp, leaf_ee, leaf_oo - leaf_ee);
uint32_t offsets[8] = {0, N, N/2, 3*N/2, N/4, 5*N/4, 7*N/4, 3*N/4};
uint32_t offsets_o[8] = {0, N, N/2, 3*N/2, 7*N/4, 3*N/4, N/4, 5*N/4};
uint32_t offsets_oe[8] = {7*N/4, 3*N/4, N/4, 5*N/4, 0, N, 3*N/2, N/2};
for(i=0;i<8;i++) IMM32_NI(fp + sse_leaf_ee_offsets[i], offsets[i]*4);
fp += (leaf_oo - leaf_ee);
if(__builtin_ctzl(N) & 1){
if(p->i1) {
lp_cnt += p->i1 * 4;
MOVI(&fp, RCX, lp_cnt);
align_mem16(&fp, 4);
memcpy(fp, leaf_oo, leaf_eo - leaf_oo);
for(i=0;i<8;i++) IMM32_NI(fp + sse_leaf_oo_offsets[i], offsets_o[i]*4);
fp += (leaf_eo - leaf_oo);
}
memcpy(fp, leaf_oe, leaf_end - leaf_oe);
lp_cnt += 4;
for(i=0;i<8;i++) IMM32_NI(fp + sse_leaf_oe_offsets[i], offsets_o[i]*4);
fp += (leaf_end - leaf_oe);
}else{
memcpy(fp, leaf_eo, leaf_oe - leaf_eo);
lp_cnt += 4;
for(i=0;i<8;i++) IMM32_NI(fp + sse_leaf_eo_offsets[i], offsets[i]*4);
fp += (leaf_oe - leaf_eo);
if(p->i1) {
lp_cnt += p->i1 * 4;
MOVI(&fp, RCX, lp_cnt);
align_mem16(&fp, 4);
memcpy(fp, leaf_oo, leaf_eo - leaf_oo);
for(i=0;i<8;i++) IMM32_NI(fp + sse_leaf_oo_offsets[i], offsets_o[i]*4);
fp += (leaf_eo - leaf_oo);
}
}
if(p->i1) {
lp_cnt += p->i1 * 4;
MOVI(&fp, RCX, lp_cnt);
align_mem16(&fp, 9);
memcpy(fp, leaf_ee, leaf_oo - leaf_ee);
for(i=0;i<8;i++) IMM32_NI(fp + sse_leaf_ee_offsets[i], offsets_oe[i]*4);
fp += (leaf_oo - leaf_ee);
}
//fprintf(stderr, "Body start address = %016p\n", fp);
//LEA(&fp, R8, RDI, ((uint32_t)&p->ws) - ((uint32_t)p));
memcpy(fp, x_init, x4 - x_init);
//IMM32_NI(fp + 3, ((int64_t)READ_IMM32(fp + 3)) + ((void *)x_init - (void *)fp ));
fp += (x4 - x_init);
int32_t pAddr = 0;
int32_t pN = 0;
int32_t pLUT = 0;
count = 2;
while(pps[0]) {
if(!pN) {
MOVI(&fp, RCX, pps[0] / 4);
}else{
if((pps[1]*4)-pAddr) ADDI(&fp, RDX, (pps[1] * 4)- pAddr);
if(pps[0] > leafN && pps[0] - pN) {
int diff = __builtin_ctzl(pps[0]) - __builtin_ctzl(pN);
*fp++ = 0xc1;
if(diff > 0) {
*fp++ = 0xe1;
*fp++ = (diff & 0xff);
}else{
*fp++ = 0xe9;
*fp++ = ((-diff) & 0xff);
}
}
}
if(p->ws_is[__builtin_ctzl(pps[0]/leafN)-1]*8 - pLUT)
ADDI(&fp, R8, p->ws_is[__builtin_ctzl(pps[0]/leafN)-1]*8 - pLUT);
if(pps[0] == 2*leafN) {
CALL(&fp, x_4_addr);
// }else if(!pps[2]){
// //uint32_t *x_8_t_addr = fp;
// memcpy(fp, neon_x8_t, neon_ee - neon_x8_t);
// fp += (neon_ee - neon_x8_t) / 4;
// //*fp++ = BL(fp+2, x_8_t_addr);
}else{
CALL(&fp, x_8_addr);
}
pAddr = pps[1] * 4;
if(pps[0] > leafN)
pN = pps[0];
pLUT = p->ws_is[__builtin_ctzl(pps[0]/leafN)-1]*8;//LUT_offset(pps[0], leafN);
// fprintf(stderr, "LUT offset for %d is %d\n", pN, pLUT);
count += 4;
pps += 2;
}
#endif
#ifdef __arm__
#ifdef HAVE_NEON
if(__builtin_ctzl(N) & 1){
ADDI(&fp, 2, 7, 0);
ADDI(&fp, 7, 9, 0);
ADDI(&fp, 9, 2, 0);
ADDI(&fp, 2, 8, 0);
ADDI(&fp, 8, 10, 0);
ADDI(&fp, 10, 2, 0);
if(p->i1) {
MOVI(&fp, 11, p->i1);
memcpy(fp, neon_oo, neon_eo - neon_oo);
if(sign < 0) {
fp[12] ^= 0x00200000; fp[13] ^= 0x00200000; fp[14] ^= 0x00200000; fp[15] ^= 0x00200000;
fp[27] ^= 0x00200000; fp[29] ^= 0x00200000; fp[30] ^= 0x00200000; fp[31] ^= 0x00200000;
fp[46] ^= 0x00200000; fp[47] ^= 0x00200000; fp[48] ^= 0x00200000; fp[57] ^= 0x00200000;
}
fp += (neon_eo - neon_oo) / 4;
}
*fp = LDRI(11, 1, ((uint32_t)&p->oe_ws) - ((uint32_t)p)); fp++;
memcpy(fp, neon_oe, neon_end - neon_oe);
if(sign < 0) {
fp[19] ^= 0x00200000; fp[20] ^= 0x00200000; fp[22] ^= 0x00200000; fp[23] ^= 0x00200000;
fp[37] ^= 0x00200000; fp[38] ^= 0x00200000; fp[40] ^= 0x00200000; fp[41] ^= 0x00200000;
fp[64] ^= 0x00200000; fp[65] ^= 0x00200000; fp[66] ^= 0x00200000; fp[67] ^= 0x00200000;
}
fp += (neon_end - neon_oe) / 4;
}else{
*fp = LDRI(11, 1, ((uint32_t)&p->eo_ws) - ((uint32_t)p)); fp++;
memcpy(fp, neon_eo, neon_oe - neon_eo);
if(sign < 0) {
fp[10] ^= 0x00200000; fp[11] ^= 0x00200000; fp[13] ^= 0x00200000; fp[14] ^= 0x00200000;
fp[31] ^= 0x00200000; fp[33] ^= 0x00200000; fp[34] ^= 0x00200000; fp[35] ^= 0x00200000;
fp[59] ^= 0x00200000; fp[60] ^= 0x00200000; fp[61] ^= 0x00200000; fp[62] ^= 0x00200000;
}
fp += (neon_oe - neon_eo) / 4;
ADDI(&fp, 2, 7, 0);
ADDI(&fp, 7, 9, 0);
ADDI(&fp, 9, 2, 0);
ADDI(&fp, 2, 8, 0);
ADDI(&fp, 8, 10, 0);
ADDI(&fp, 10, 2, 0);
if(p->i1) {
MOVI(&fp, 11, p->i1);
memcpy(fp, neon_oo, neon_eo - neon_oo);
if(sign < 0) {
fp[12] ^= 0x00200000; fp[13] ^= 0x00200000; fp[14] ^= 0x00200000; fp[15] ^= 0x00200000;
fp[27] ^= 0x00200000; fp[29] ^= 0x00200000; fp[30] ^= 0x00200000; fp[31] ^= 0x00200000;
fp[46] ^= 0x00200000; fp[47] ^= 0x00200000; fp[48] ^= 0x00200000; fp[57] ^= 0x00200000;
}
fp += (neon_eo - neon_oo) / 4;
}
}
if(p->i1) {
ADDI(&fp, 2, 3, 0);
ADDI(&fp, 3, 7, 0);
ADDI(&fp, 7, 2, 0);
ADDI(&fp, 2, 4, 0);
ADDI(&fp, 4, 8, 0);
ADDI(&fp, 8, 2, 0);
ADDI(&fp, 2, 5, 0);
ADDI(&fp, 5, 9, 0);
ADDI(&fp, 9, 2, 0);
ADDI(&fp, 2, 6, 0);
ADDI(&fp, 6, 10, 0);
ADDI(&fp, 10, 2, 0);
ADDI(&fp, 2, 9, 0);
ADDI(&fp, 9, 10, 0);
ADDI(&fp, 10, 2, 0);
*fp = LDRI(2, 1, ((uint32_t)&p->ee_ws) - ((uint32_t)p)); fp++;
MOVI(&fp, 11, p->i1);
memcpy(fp, neon_ee, neon_oo - neon_ee);
if(sign < 0) {
fp[33] ^= 0x00200000; fp[37] ^= 0x00200000; fp[38] ^= 0x00200000; fp[39] ^= 0x00200000;
fp[40] ^= 0x00200000; fp[41] ^= 0x00200000; fp[44] ^= 0x00200000; fp[45] ^= 0x00200000;
fp[46] ^= 0x00200000; fp[47] ^= 0x00200000; fp[48] ^= 0x00200000; fp[57] ^= 0x00200000;
}
fp += (neon_oo - neon_ee) / 4;
}
#else
ADDI(&fp, 2, 7, 0);
ADDI(&fp, 7, 9, 0);
ADDI(&fp, 9, 2, 0);
ADDI(&fp, 2, 8, 0);
ADDI(&fp, 8, 10, 0);
ADDI(&fp, 10, 2, 0);
MOVI(&fp, 11, (p->i1>0) ? p->i1 : 1);
memcpy(fp, vfp_o, vfp_x4 - vfp_o);
if(sign > 0) {
fp[22] ^= 0x00000040; fp[24] ^= 0x00000040; fp[25] ^= 0x00000040; fp[26] ^= 0x00000040;
fp[62] ^= 0x00000040; fp[64] ^= 0x00000040; fp[65] ^= 0x00000040; fp[66] ^= 0x00000040;
}
fp += (vfp_x4 - vfp_o) / 4;
ADDI(&fp, 2, 3, 0);
ADDI(&fp, 3, 7, 0);
ADDI(&fp, 7, 2, 0);
ADDI(&fp, 2, 4, 0);
ADDI(&fp, 4, 8, 0);
ADDI(&fp, 8, 2, 0);
ADDI(&fp, 2, 5, 0);
ADDI(&fp, 5, 9, 0);
ADDI(&fp, 9, 2, 0);
ADDI(&fp, 2, 6, 0);
ADDI(&fp, 6, 10, 0);
ADDI(&fp, 10, 2, 0);
ADDI(&fp, 2, 9, 0);
ADDI(&fp, 9, 10, 0);
ADDI(&fp, 10, 2, 0);
*fp = LDRI(2, 1, ((uint32_t)&p->ee_ws) - ((uint32_t)p)); fp++;
MOVI(&fp, 11, (p->i2>0) ? p->i2 : 1);
memcpy(fp, vfp_e, vfp_o - vfp_e);
if(sign > 0) {
fp[64] ^= 0x00000040; fp[65] ^= 0x00000040; fp[68] ^= 0x00000040; fp[75] ^= 0x00000040;
fp[76] ^= 0x00000040; fp[79] ^= 0x00000040; fp[80] ^= 0x00000040; fp[83] ^= 0x00000040;
fp[84] ^= 0x00000040; fp[87] ^= 0x00000040; fp[91] ^= 0x00000040; fp[93] ^= 0x00000040;
}
fp += (vfp_o - vfp_e) / 4;
#endif
*fp = LDRI(2, 1, ((uint32_t)&p->ws) - ((uint32_t)p)); fp++; // load offsets into r12
//ADDI(&fp, 2, 1, 0);
MOVI(&fp, 1, 0);
// args: r0 - out
// r1 - N
// r2 - ws
// ADDI(&fp, 3, 1, 0); // put N into r3 for counter
int32_t pAddr = 0;
int32_t pN = 0;
int32_t pLUT = 0;
count = 2;
while(pps[0]) {
// fprintf(stderr, "size %zu at %zu - diff %zu\n", pps[0], pps[1]*4, (pps[1]*4) - pAddr);
if(!pN) {
MOVI(&fp, 1, pps[0]);
}else{
if((pps[1]*4)-pAddr) ADDI(&fp, 0, 0, (pps[1] * 4)- pAddr);
if(pps[0] - pN) ADDI(&fp, 1, 1, pps[0] - pN);
}
if(p->ws_is[__builtin_ctzl(pps[0]/leafN)-1]*8 - pLUT)
ADDI(&fp, 2, 2, p->ws_is[__builtin_ctzl(pps[0]/leafN)-1]*8 - pLUT);
if(pps[0] == 2*leafN) {
*fp = BL(fp+2, x_4_addr); fp++;
}else if(!pps[2]){
//uint32_t *x_8_t_addr = fp;
#ifdef HAVE_NEON
memcpy(fp, neon_x8_t, neon_ee - neon_x8_t);
if(sign < 0) {
fp[31] ^= 0x00200000; fp[32] ^= 0x00200000; fp[33] ^= 0x00200000; fp[34] ^= 0x00200000;
fp[65] ^= 0x00200000; fp[66] ^= 0x00200000; fp[70] ^= 0x00200000; fp[74] ^= 0x00200000;
fp[97] ^= 0x00200000; fp[98] ^= 0x00200000; fp[102] ^= 0x00200000; fp[104] ^= 0x00200000;
}
fp += (neon_ee - neon_x8_t) / 4;
//*fp++ = BL(fp+2, x_8_t_addr);
#else
*fp = BL(fp+2, x_8_addr); fp++;
#endif
}else{
*fp = BL(fp+2, x_8_addr); fp++;
}
pAddr = pps[1] * 4;
pN = pps[0];
pLUT = p->ws_is[__builtin_ctzl(pps[0]/leafN)-1]*8;//LUT_offset(pps[0], leafN);
// fprintf(stderr, "LUT offset for %d is %d\n", pN, pLUT);
count += 4;
pps += 2;
}
*fp++ = 0xecbd8b10;
*fp++ = POP_LR(); count++;
#else
POP(&fp, R15);
POP(&fp, R14);
POP(&fp, R13);
POP(&fp, R12);
POP(&fp, R11);
POP(&fp, R10);
POP(&fp, RBX);
POP(&fp, RBP);
RET(&fp);
//uint8_t *pp = func;
//int counter = 0;
//do{
// printf("%02x ", *pp);
// if(counter++ % 16 == 15) printf("\n");
//} while(++pp < fp);
//printf("\n");
#endif
// *fp++ = B(14); count++;
//for(int i=0;i<(neon_x8 - neon_x4)/4;i++)
// fprintf(stderr, "%08x\n", x_4_addr[i]);
//fprintf(stderr, "\n");
//for(int i=0;i<count;i++)
free(ps);
if (mprotect(func, p->transform_size, PROT_READ | PROT_EXEC)) {
perror("Couldn't mprotect");
exit(1);
}
#ifdef __APPLE__
sys_icache_invalidate(func, p->transform_size);
#elif __ANDROID__
cacheflush((long)(func), (long)(func) + p->transform_size, 0);
#elif __linux__
#ifdef __GNUC__
__clear_cache((long)(func), (long)(func) + p->transform_size);
#endif
#endif
//fprintf(stderr, "size of transform %zu = %d\n", N, (fp-func)*4);
p->transform = (void *) (start);
}
transform_func_t ffts_generate_func_code(ffts_plan_t *p, size_t N, size_t leaf_N, int sign)
{
uint32_t offsets[8] = {0, 4*N, 2*N, 6*N, N, 5*N, 7*N, 3*N};
uint32_t offsets_o[8] = {0, 4*N, 2*N, 6*N, 7*N, 3*N, N, 5*N};
int32_t pAddr = 0;
int32_t pN = 0;
int32_t pLUT = 0;
insns_t *fp;
insns_t *start;
insns_t *x_4_addr;
insns_t *x_8_addr;
uint32_t loop_count;
int count;
ptrdiff_t len;
size_t *ps;
size_t *pps;
count = ffts_tree_count(N, leaf_N, 0) + 1;
ps = pps = malloc(2 * count * sizeof(*ps));
if (!ps) {
return NULL;
}
ffts_elaborate_tree(&pps, N, leaf_N, 0);
pps[0] = 0;
pps[1] = 0;
pps = ps;
#ifdef HAVE_SSE
if (sign < 0) {
p->constants = (const void*) sse_constants;
} else {
p->constants = (const void*) sse_constants_inv;
}
#endif
fp = (insns_t*) p->transform_base;
/* generate base cases */
x_4_addr = generate_size4_base_case(&fp, sign);
x_8_addr = generate_size8_base_case(&fp, sign);
#ifdef __arm__
start = generate_prologue(&fp, p);
#ifdef HAVE_NEON
memcpy(fp, neon_ee, neon_oo - neon_ee);
if (sign < 0) {
fp[33] ^= 0x00200000;
fp[37] ^= 0x00200000;
fp[38] ^= 0x00200000;
fp[39] ^= 0x00200000;
fp[40] ^= 0x00200000;
fp[41] ^= 0x00200000;
fp[44] ^= 0x00200000;
fp[45] ^= 0x00200000;
fp[46] ^= 0x00200000;
fp[47] ^= 0x00200000;
fp[48] ^= 0x00200000;
fp[57] ^= 0x00200000;
}
fp += (neon_oo - neon_ee) / 4;
#else
memcpy(fp, vfp_e, vfp_o - vfp_e);
if (sign > 0) {
fp[64] ^= 0x00000040;
fp[65] ^= 0x00000040;
fp[68] ^= 0x00000040;
fp[75] ^= 0x00000040;
fp[76] ^= 0x00000040;
fp[79] ^= 0x00000040;
fp[80] ^= 0x00000040;
fp[83] ^= 0x00000040;
fp[84] ^= 0x00000040;
fp[87] ^= 0x00000040;
fp[91] ^= 0x00000040;
fp[93] ^= 0x00000040;
}
fp += (vfp_o - vfp_e) / 4;
#endif
#else
/* generate functions */
start = generate_prologue(&fp, p);
loop_count = 4 * p->i0;
generate_leaf_init(&fp, loop_count);
if (ffts_ctzl(N) & 1) {
generate_leaf_ee(&fp, offsets, p->i1 ? 6 : 0);
if (p->i1) {
loop_count += 4 * p->i1;
generate_leaf_oo(&fp, loop_count, offsets_o, 7);
}
loop_count += 4;
generate_leaf_oe(&fp, offsets_o);
} else {
generate_leaf_ee(&fp, offsets, N >= 256 ? 2 : 8);
loop_count += 4;
generate_leaf_eo(&fp, offsets);
if (p->i1) {
loop_count += 4 * p->i1;
generate_leaf_oo(&fp, loop_count, offsets_o, N >= 256 ? 4 : 7);
}
}
if (p->i1) {
uint32_t offsets_oe[8] = {7*N, 3*N, N, 5*N, 0, 4*N, 6*N, 2*N};
loop_count += 4 * p->i1;
/* align loop/jump destination */
#ifdef _M_X64
x86_mov_reg_imm(fp, X86_EBX, loop_count);
#else
x86_mov_reg_imm(fp, X86_ECX, loop_count);
ffts_align_mem16(&fp, 9);
#endif
generate_leaf_ee(&fp, offsets_oe, 0);
}
generate_transform_init(&fp);
/* generate subtransform calls */
count = 2;
while (pps[0]) {
size_t ws_is;
if (!pN) {
#ifdef _M_X64
x86_mov_reg_imm(fp, X86_EBX, pps[0]);
#else
x86_mov_reg_imm(fp, X86_ECX, pps[0] / 4);
#endif
} else {
int offset = (4 * pps[1]) - pAddr;
if (offset) {
#ifdef _M_X64
x64_alu_reg_imm_size(fp, X86_ADD, X64_R8, offset, 8);
#else
x64_alu_reg_imm_size(fp, X86_ADD, X64_RDX, offset, 8);
#endif
}
if (pps[0] > leaf_N && pps[0] - pN) {
int factor = ffts_ctzl(pps[0]) - ffts_ctzl(pN);
#ifdef _M_X64
if (factor > 0) {
x86_shift_reg_imm(fp, X86_SHL, X86_EBX, factor);
} else {
x86_shift_reg_imm(fp, X86_SHR, X86_EBX, -factor);
}
#else
if (factor > 0) {
x86_shift_reg_imm(fp, X86_SHL, X86_ECX, factor);
} else {
x86_shift_reg_imm(fp, X86_SHR, X86_ECX, -factor);
}
#endif
}
}
ws_is = 8 * p->ws_is[ffts_ctzl(pps[0] / leaf_N) - 1];
if (ws_is != pLUT) {
int offset = (int) (ws_is - pLUT);
#ifdef _M_X64
x64_alu_reg_imm_size(fp, X86_ADD, X64_R9, offset, 8);
#else
x64_alu_reg_imm_size(fp, X86_ADD, X64_R8, offset, 8);
#endif
}
if (pps[0] == 2 * leaf_N) {
x64_call_code(fp, x_4_addr);
} else {
x64_call_code(fp, x_8_addr);
}
pAddr = 4 * pps[1];
if (pps[0] > leaf_N) {
pN = pps[0];
}
pLUT = ws_is;//LUT_offset(pps[0], leafN);
//fprintf(stderr, "LUT offset for %d is %d\n", pN, pLUT);
count += 4;
pps += 2;
}
#endif
#ifdef __arm__
#ifdef HAVE_NEON
if (ffts_ctzl(N) & 1) {
ADDI(&fp, 2, 7, 0);
ADDI(&fp, 7, 9, 0);
ADDI(&fp, 9, 2, 0);
ADDI(&fp, 2, 8, 0);
ADDI(&fp, 8, 10, 0);
ADDI(&fp, 10, 2, 0);
if(p->i1) {
MOVI(&fp, 11, p->i1);
memcpy(fp, neon_oo, neon_eo - neon_oo);
if(sign < 0) {
fp[12] ^= 0x00200000;
fp[13] ^= 0x00200000;
fp[14] ^= 0x00200000;
fp[15] ^= 0x00200000;
fp[27] ^= 0x00200000;
fp[29] ^= 0x00200000;
fp[30] ^= 0x00200000;
fp[31] ^= 0x00200000;
fp[46] ^= 0x00200000;
fp[47] ^= 0x00200000;
fp[48] ^= 0x00200000;
fp[57] ^= 0x00200000;
}
fp += (neon_eo - neon_oo) / 4;
}
*fp = LDRI(11, 1, ((uint32_t)&p->oe_ws) - ((uint32_t)p));
fp++;
memcpy(fp, neon_oe, neon_end - neon_oe);
if(sign < 0) {
fp[19] ^= 0x00200000;
fp[20] ^= 0x00200000;
fp[22] ^= 0x00200000;
fp[23] ^= 0x00200000;
fp[37] ^= 0x00200000;
fp[38] ^= 0x00200000;
fp[40] ^= 0x00200000;
fp[41] ^= 0x00200000;
fp[64] ^= 0x00200000;
fp[65] ^= 0x00200000;
fp[66] ^= 0x00200000;
fp[67] ^= 0x00200000;
}
fp += (neon_end - neon_oe) / 4;
} else {
*fp = LDRI(11, 1, ((uint32_t)&p->eo_ws) - ((uint32_t)p));
fp++;
memcpy(fp, neon_eo, neon_oe - neon_eo);
if(sign < 0) {
fp[10] ^= 0x00200000;
fp[11] ^= 0x00200000;
fp[13] ^= 0x00200000;
fp[14] ^= 0x00200000;
fp[31] ^= 0x00200000;
fp[33] ^= 0x00200000;
fp[34] ^= 0x00200000;
fp[35] ^= 0x00200000;
fp[59] ^= 0x00200000;
fp[60] ^= 0x00200000;
fp[61] ^= 0x00200000;
fp[62] ^= 0x00200000;
}
fp += (neon_oe - neon_eo) / 4;
ADDI(&fp, 2, 7, 0);
ADDI(&fp, 7, 9, 0);
ADDI(&fp, 9, 2, 0);
ADDI(&fp, 2, 8, 0);
ADDI(&fp, 8, 10, 0);
ADDI(&fp, 10, 2, 0);
if(p->i1) {
MOVI(&fp, 11, p->i1);
memcpy(fp, neon_oo, neon_eo - neon_oo);
if(sign < 0) {
fp[12] ^= 0x00200000;
fp[13] ^= 0x00200000;
fp[14] ^= 0x00200000;
fp[15] ^= 0x00200000;
fp[27] ^= 0x00200000;
fp[29] ^= 0x00200000;
fp[30] ^= 0x00200000;
fp[31] ^= 0x00200000;
fp[46] ^= 0x00200000;
fp[47] ^= 0x00200000;
fp[48] ^= 0x00200000;
fp[57] ^= 0x00200000;
}
fp += (neon_eo - neon_oo) / 4;
}
}
if(p->i1) {
ADDI(&fp, 2, 3, 0);
ADDI(&fp, 3, 7, 0);
ADDI(&fp, 7, 2, 0);
ADDI(&fp, 2, 4, 0);
ADDI(&fp, 4, 8, 0);
ADDI(&fp, 8, 2, 0);
ADDI(&fp, 2, 5, 0);
ADDI(&fp, 5, 9, 0);
ADDI(&fp, 9, 2, 0);
ADDI(&fp, 2, 6, 0);
ADDI(&fp, 6, 10, 0);
ADDI(&fp, 10, 2, 0);
ADDI(&fp, 2, 9, 0);
ADDI(&fp, 9, 10, 0);
ADDI(&fp, 10, 2, 0);
*fp = LDRI(2, 1, ((uint32_t)&p->ee_ws) - ((uint32_t)p));
fp++;
MOVI(&fp, 11, p->i1);
memcpy(fp, neon_ee, neon_oo - neon_ee);
if(sign < 0) {
fp[33] ^= 0x00200000;
fp[37] ^= 0x00200000;
fp[38] ^= 0x00200000;
fp[39] ^= 0x00200000;
fp[40] ^= 0x00200000;
fp[41] ^= 0x00200000;
fp[44] ^= 0x00200000;
fp[45] ^= 0x00200000;
fp[46] ^= 0x00200000;
fp[47] ^= 0x00200000;
fp[48] ^= 0x00200000;
fp[57] ^= 0x00200000;
}
fp += (neon_oo - neon_ee) / 4;
}
#else
ADDI(&fp, 2, 7, 0);
ADDI(&fp, 7, 9, 0);
ADDI(&fp, 9, 2, 0);
ADDI(&fp, 2, 8, 0);
ADDI(&fp, 8, 10, 0);
ADDI(&fp, 10, 2, 0);
MOVI(&fp, 11, (p->i1>0) ? p->i1 : 1);
memcpy(fp, vfp_o, vfp_x4 - vfp_o);
if(sign > 0) {
fp[22] ^= 0x00000040;
fp[24] ^= 0x00000040;
fp[25] ^= 0x00000040;
fp[26] ^= 0x00000040;
fp[62] ^= 0x00000040;
fp[64] ^= 0x00000040;
fp[65] ^= 0x00000040;
fp[66] ^= 0x00000040;
}
fp += (vfp_x4 - vfp_o) / 4;
ADDI(&fp, 2, 3, 0);
ADDI(&fp, 3, 7, 0);
ADDI(&fp, 7, 2, 0);
ADDI(&fp, 2, 4, 0);
ADDI(&fp, 4, 8, 0);
ADDI(&fp, 8, 2, 0);
ADDI(&fp, 2, 5, 0);
ADDI(&fp, 5, 9, 0);
ADDI(&fp, 9, 2, 0);
ADDI(&fp, 2, 6, 0);
ADDI(&fp, 6, 10, 0);
ADDI(&fp, 10, 2, 0);
ADDI(&fp, 2, 9, 0);
ADDI(&fp, 9, 10, 0);
ADDI(&fp, 10, 2, 0);
*fp = LDRI(2, 1, ((uint32_t)&p->ee_ws) - ((uint32_t)p));
fp++;
MOVI(&fp, 11, (p->i2>0) ? p->i2 : 1);
memcpy(fp, vfp_e, vfp_o - vfp_e);
if(sign > 0) {
fp[64] ^= 0x00000040;
fp[65] ^= 0x00000040;
fp[68] ^= 0x00000040;
fp[75] ^= 0x00000040;
fp[76] ^= 0x00000040;
fp[79] ^= 0x00000040;
fp[80] ^= 0x00000040;
fp[83] ^= 0x00000040;
fp[84] ^= 0x00000040;
fp[87] ^= 0x00000040;
fp[91] ^= 0x00000040;
fp[93] ^= 0x00000040;
}
fp += (vfp_o - vfp_e) / 4;
#endif
*fp = LDRI(2, 1, ((uint32_t)&p->ws) - ((uint32_t)p));
fp++; // load offsets into r12
//ADDI(&fp, 2, 1, 0);
MOVI(&fp, 1, 0);
// args: r0 - out
// r1 - N
// r2 - ws
// ADDI(&fp, 3, 1, 0); // put N into r3 for counter
count = 2;
while(pps[0]) {
// fprintf(stderr, "size %zu at %zu - diff %zu\n", pps[0], pps[1]*4, (pps[1]*4) - pAddr);
if(!pN) {
MOVI(&fp, 1, pps[0]);
} else {
if((pps[1]*4)-pAddr) ADDI(&fp, 0, 0, (pps[1] * 4)- pAddr);
if(pps[0] - pN) ADDI(&fp, 1, 1, pps[0] - pN);
}
if (p->ws_is[ffts_ctzl(pps[0]/leaf_N)-1]*8 - pLUT) {
ADDI(&fp, 2, 2, p->ws_is[ffts_ctzl(pps[0]/leaf_N)-1]*8 - pLUT);
}
if(pps[0] == 2 * leaf_N) {
*fp = BL(fp+2, x_4_addr);
fp++;
} else if(!pps[2]) {
//uint32_t *x_8_t_addr = fp;
#ifdef HAVE_NEON
memcpy(fp, neon_x8_t, neon_ee - neon_x8_t);
if(sign < 0) {
fp[31] ^= 0x00200000;
fp[32] ^= 0x00200000;
fp[33] ^= 0x00200000;
fp[34] ^= 0x00200000;
fp[65] ^= 0x00200000;
fp[66] ^= 0x00200000;
fp[70] ^= 0x00200000;
fp[74] ^= 0x00200000;
fp[97] ^= 0x00200000;
fp[98] ^= 0x00200000;
fp[102] ^= 0x00200000;
fp[104] ^= 0x00200000;
}
fp += (neon_ee - neon_x8_t) / 4;
//*fp++ = BL(fp+2, x_8_t_addr);
#else
*fp = BL(fp+2, x_8_addr);
fp++;
#endif
} else {
*fp = BL(fp+2, x_8_addr);
fp++;
}
pAddr = pps[1] * 4;
pN = pps[0];
pLUT = p->ws_is[ffts_ctzl(pps[0]/leaf_N)-1]*8;//LUT_offset(pps[0], leafN);
// fprintf(stderr, "LUT offset for %d is %d\n", pN, pLUT);
count += 4;
pps += 2;
}
*fp++ = 0xecbd8b10;
*fp++ = POP_LR();
count++;
#else
generate_epilogue(&fp);
#endif
// *fp++ = B(14); count++;
//for(int i=0;i<(neon_x8 - neon_x4)/4;i++)
// fprintf(stderr, "%08x\n", x_4_addr[i]);
//fprintf(stderr, "\n");
//for(int i=0;i<count;i++)
//fprintf(stderr, "size of transform %u = %d\n", N, (fp - x_8_addr) * sizeof(*fp));
free(ps);
#if defined(_MSC_VER)
#pragma warning(push)
/* disable type cast warning from data pointer to function pointer */
#pragma warning(disable : 4055)
#endif
return (transform_func_t) start;
#if defined(_MSC_VER)
#pragma warning(pop)
#endif
}
