diff options
Diffstat (limited to 'src/gallium/auxiliary/rtasm')
-rw-r--r-- | src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c | 382 | ||||
-rw-r--r-- | src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h | 76 | ||||
-rw-r--r-- | src/gallium/auxiliary/rtasm/rtasm_x86sse.c | 44 | ||||
-rw-r--r-- | src/gallium/auxiliary/rtasm/rtasm_x86sse.h | 7 |
4 files changed, 453 insertions, 56 deletions
diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c index 491141f190..dea1aed032 100644 --- a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c +++ b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c @@ -164,6 +164,27 @@ rem_prefix(const char *longname) } +static const char * +reg_name(int reg) +{ + switch (reg) { + case SPE_REG_SP: + return "$sp"; + case SPE_REG_RA: + return "$lr"; + default: + { + /* cycle through four buffers to handle multiple calls per printf */ + static char buf[4][10]; + static int b = 0; + b = (b + 1) % 4; + sprintf(buf[b], "$%d", reg); + return buf[b]; + } + } +} + + static void emit_RR(struct spe_function *p, unsigned op, unsigned rT, unsigned rA, unsigned rB, const char *name) { @@ -176,7 +197,8 @@ static void emit_RR(struct spe_function *p, unsigned op, unsigned rT, assert(p->num_inst <= p->max_inst); if (p->print) { indent(p); - printf("%s\t$%d, $%d, $%d\n", rem_prefix(name), rT, rA, rB); + printf("%s\t%s, %s, %s\n", + rem_prefix(name), reg_name(rT), reg_name(rA), reg_name(rB)); } } @@ -194,7 +216,8 @@ static void emit_RRR(struct spe_function *p, unsigned op, unsigned rT, assert(p->num_inst <= p->max_inst); if (p->print) { indent(p); - printf("%s\t$%d, $%d, $%d, $%d\n", rem_prefix(name), rT, rA, rB, rC); + printf("%s\t%s, %s, %s, %s\n", rem_prefix(name), reg_name(rT), + reg_name(rA), reg_name(rB), reg_name(rC)); } } @@ -211,7 +234,8 @@ static void emit_RI7(struct spe_function *p, unsigned op, unsigned rT, assert(p->num_inst <= p->max_inst); if (p->print) { indent(p); - printf("%s\t$%d, $%d, 0x%x\n", rem_prefix(name), rT, rA, imm); + printf("%s\t%s, %s, 0x%x\n", + rem_prefix(name), reg_name(rT), reg_name(rA), imm); } } @@ -229,7 +253,8 @@ static void emit_RI8(struct spe_function *p, unsigned op, unsigned rT, assert(p->num_inst <= p->max_inst); if (p->print) { indent(p); - printf("%s\t$%d, $%d, 0x%x\n", rem_prefix(name), rT, rA, imm); + printf("%s\t%s, %s, 0x%x\n", + rem_prefix(name), reg_name(rT), reg_name(rA), imm); } } @@ -247,15 +272,22 @@ static void emit_RI10(struct spe_function *p, unsigned op, unsigned rT, assert(p->num_inst <= p->max_inst); if (p->print) { indent(p); - if (strcmp(name, "spe_lqd") == 0 || - strcmp(name, "spe_stqd") == 0) - printf("%s\t$%d, 0x%x($%d)\n", rem_prefix(name), rT, imm, rA); - else - printf("%s\t$%d, $%d, 0x%x\n", rem_prefix(name), rT, rA, imm); + printf("%s\t%s, %s, 0x%x\n", + rem_prefix(name), reg_name(rT), reg_name(rA), imm); } } +/** As above, but do range checking on signed immediate value */ +static void emit_RI10s(struct spe_function *p, unsigned op, unsigned rT, + unsigned rA, int imm, const char *name) +{ + assert(imm <= 511); + assert(imm >= -512); + emit_RI10(p, op, rT, rA, imm, name); +} + + static void emit_RI16(struct spe_function *p, unsigned op, unsigned rT, int imm, const char *name) { @@ -267,7 +299,7 @@ static void emit_RI16(struct spe_function *p, unsigned op, unsigned rT, assert(p->num_inst <= p->max_inst); if (p->print) { indent(p); - printf("%s\t$%d, 0x%x\n", rem_prefix(name), rT, imm); + printf("%s\t%s, 0x%x\n", rem_prefix(name), reg_name(rT), imm); } } @@ -283,7 +315,7 @@ static void emit_RI18(struct spe_function *p, unsigned op, unsigned rT, assert(p->num_inst <= p->max_inst); if (p->print) { indent(p); - printf("%s\t$%d, 0x%x\n", rem_prefix(name), rT, imm); + printf("%s\t%s, 0x%x\n", rem_prefix(name), reg_name(rT), imm); } } @@ -332,6 +364,12 @@ void _name (struct spe_function *p, unsigned rT, unsigned rA, int imm) \ emit_RI10(p, _op, rT, rA, imm, __FUNCTION__); \ } +#define EMIT_RI10s(_name, _op) \ +void _name (struct spe_function *p, unsigned rT, unsigned rA, int imm) \ +{ \ + emit_RI10s(p, _op, rT, rA, imm, __FUNCTION__); \ +} + #define EMIT_RI16(_name, _op) \ void _name (struct spe_function *p, unsigned rT, int imm) \ { \ @@ -353,20 +391,28 @@ void _name (struct spe_function *p, int imm) \ #include "rtasm_ppc_spe.h" + /** * Initialize an spe_function. * \param code_size size of instruction buffer to allocate, in bytes. */ void spe_init_func(struct spe_function *p, unsigned code_size) { + unsigned int i; + p->store = align_malloc(code_size, 16); p->num_inst = 0; p->max_inst = code_size / SPE_INST_SIZE; + p->set_count = 0; + memset(p->regs, 0, SPE_NUM_REGS * sizeof(p->regs[0])); + /* Conservatively treat R0 - R2 and R80 - R127 as non-volatile. */ - p->regs[0] = ~7; - p->regs[1] = (1U << (80 - 64)) - 1; + p->regs[0] = p->regs[1] = p->regs[2] = 1; + for (i = 80; i <= 127; i++) { + p->regs[i] = 1; + } p->print = false; p->indent = 0; @@ -398,12 +444,8 @@ int spe_allocate_available_register(struct spe_function *p) { unsigned i; for (i = 0; i < SPE_NUM_REGS; i++) { - const uint64_t mask = (1ULL << (i % 64)); - const unsigned idx = i / 64; - - assert(idx < 2); - if ((p->regs[idx] & mask) != 0) { - p->regs[idx] &= ~mask; + if (p->regs[i] == 0) { + p->regs[i] = 1; return i; } } @@ -417,31 +459,84 @@ int spe_allocate_available_register(struct spe_function *p) */ int spe_allocate_register(struct spe_function *p, int reg) { - const unsigned idx = reg / 64; - const unsigned bit = reg % 64; - assert(reg < SPE_NUM_REGS); - assert((p->regs[idx] & (1ULL << bit)) != 0); - - p->regs[idx] &= ~(1ULL << bit); + assert(p->regs[reg] == 0); + p->regs[reg] = 1; return reg; } /** - * Mark the given SPE register as "unallocated". + * Mark the given SPE register as "unallocated". Note that this should + * only be used on registers allocated in the current register set; an + * assertion will fail if an attempt is made to deallocate a register + * allocated in an earlier register set. */ void spe_release_register(struct spe_function *p, int reg) { - const unsigned idx = reg / 64; - const unsigned bit = reg % 64; + assert(reg < SPE_NUM_REGS); + assert(p->regs[reg] == 1); - assert(idx < 2); + p->regs[reg] = 0; +} - assert(reg < SPE_NUM_REGS); - assert((p->regs[idx] & (1ULL << bit)) == 0); +/** + * Start a new set of registers. This can be called if + * it will be difficult later to determine exactly what + * registers were actually allocated during a code generation + * sequence, and you really just want to deallocate all of them. + */ +void spe_allocate_register_set(struct spe_function *p) +{ + unsigned int i; - p->regs[idx] |= (1ULL << bit); + /* Keep track of the set count. If it ever wraps around to 0, + * we're in trouble. + */ + p->set_count++; + assert(p->set_count > 0); + + /* Increment the allocation count of all registers currently + * allocated. Then any registers that are allocated in this set + * will be the only ones with a count of 1; they'll all be released + * when the register set is released. + */ + for (i = 0; i < SPE_NUM_REGS; i++) { + if (p->regs[i] > 0) + p->regs[i]++; + } +} + +void spe_release_register_set(struct spe_function *p) +{ + unsigned int i; + + /* If the set count drops below zero, we're in trouble. */ + assert(p->set_count > 0); + p->set_count--; + + /* Drop the allocation level of all registers. Any allocated + * during this register set will drop to 0 and then become + * available. + */ + for (i = 0; i < SPE_NUM_REGS; i++) { + if (p->regs[i] > 0) + p->regs[i]--; + } +} + + +unsigned +spe_get_registers_used(const struct spe_function *p, ubyte used[]) +{ + unsigned i, num = 0; + /* only count registers in the range available to callers */ + for (i = 2; i < 80; i++) { + if (p->regs[i]) { + used[num++] = i; + } + } + return num; } @@ -459,7 +554,7 @@ spe_indent(struct spe_function *p, int spaces) } -extern void +void spe_comment(struct spe_function *p, int rel_indent, const char *s) { if (p->print) { @@ -472,6 +567,56 @@ spe_comment(struct spe_function *p, int rel_indent, const char *s) /** + * Load quad word. + * NOTE: offset is in bytes and the least significant 4 bits must be zero! + */ +void spe_lqd(struct spe_function *p, unsigned rT, unsigned rA, int offset) +{ + const boolean pSave = p->print; + + /* offset must be a multiple of 16 */ + assert(offset % 16 == 0); + /* offset must fit in 10-bit signed int field, after shifting */ + assert((offset >> 4) <= 511); + assert((offset >> 4) >= -512); + + p->print = FALSE; + emit_RI10(p, 0x034, rT, rA, offset >> 4, "spe_lqd"); + p->print = pSave; + + if (p->print) { + indent(p); + printf("lqd\t%s, %d(%s)\n", reg_name(rT), offset, reg_name(rA)); + } +} + + +/** + * Store quad word. + * NOTE: offset is in bytes and the least significant 4 bits must be zero! + */ +void spe_stqd(struct spe_function *p, unsigned rT, unsigned rA, int offset) +{ + const boolean pSave = p->print; + + /* offset must be a multiple of 16 */ + assert(offset % 16 == 0); + /* offset must fit in 10-bit signed int field, after shifting */ + assert((offset >> 4) <= 511); + assert((offset >> 4) >= -512); + + p->print = FALSE; + emit_RI10(p, 0x024, rT, rA, offset >> 4, "spe_stqd"); + p->print = pSave; + + if (p->print) { + indent(p); + printf("stqd\t%s, %d(%s)\n", reg_name(rT), offset, reg_name(rA)); + } +} + + +/** * For branch instructions: * \param d if 1, disable interupts if branch is taken * \param e if 1, enable interupts if branch is taken @@ -603,22 +748,187 @@ void spe_load_uint(struct spe_function *p, unsigned rT, unsigned int ui) { /* If the whole value is in the lower 18 bits, use ila, which * doesn't sign-extend. Otherwise, if the two halfwords of - * the constant are identical, use ilh. Otherwise, we have - * to use ilhu followed by iohl. + * the constant are identical, use ilh. Otherwise, if every byte of + * the desired value is 0x00 or 0xff, we can use Form Select Mask for + * Bytes Immediate (fsmbi) to load the value in a single instruction. + * Otherwise, in the general case, we have to use ilhu followed by iohl. */ - if ((ui & 0xfffc0000) == ui) { + if ((ui & 0x0003ffff) == ui) { spe_ila(p, rT, ui); } else if ((ui >> 16) == (ui & 0xffff)) { spe_ilh(p, rT, ui & 0xffff); } + else if ( + ((ui & 0x000000ff) == 0 || (ui & 0x000000ff) == 0x000000ff) && + ((ui & 0x0000ff00) == 0 || (ui & 0x0000ff00) == 0x0000ff00) && + ((ui & 0x00ff0000) == 0 || (ui & 0x00ff0000) == 0x00ff0000) && + ((ui & 0xff000000) == 0 || (ui & 0xff000000) == 0xff000000) + ) { + unsigned int mask = 0; + /* fsmbi duplicates each bit in the given mask eight times, + * using a 16-bit value to initialize a 16-byte quadword. + * Each 4-bit nybble of the mask corresponds to a full word + * of the result; look at the value and figure out the mask + * (replicated for each word in the quadword), and then + * form the "select mask" to get the value. + */ + if ((ui & 0x000000ff) == 0x000000ff) mask |= 0x1111; + if ((ui & 0x0000ff00) == 0x0000ff00) mask |= 0x2222; + if ((ui & 0x00ff0000) == 0x00ff0000) mask |= 0x4444; + if ((ui & 0xff000000) == 0xff000000) mask |= 0x8888; + spe_fsmbi(p, rT, mask); + } else { + /* The general case: this usually uses two instructions, but + * may use only one if the low-order 16 bits of each word are 0. + */ spe_ilhu(p, rT, ui >> 16); if (ui & 0xffff) spe_iohl(p, rT, ui & 0xffff); } } +/** + * This function is constructed identically to spe_xor_uint() below. + * Changes to one should be made in the other. + */ +void +spe_and_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui) +{ + /* If we can, emit a single instruction, either And Byte Immediate + * (which uses the same constant across each byte), And Halfword Immediate + * (which sign-extends a 10-bit immediate to 16 bits and uses that + * across each halfword), or And Word Immediate (which sign-extends + * a 10-bit immediate to 32 bits). + * + * Otherwise, we'll need to use a temporary register. + */ + unsigned int tmp; + + /* If the upper 23 bits are all 0s or all 1s, sign extension + * will work and we can use And Word Immediate + */ + tmp = ui & 0xfffffe00; + if (tmp == 0xfffffe00 || tmp == 0) { + spe_andi(p, rT, rA, ui & 0x000003ff); + return; + } + + /* If the ui field is symmetric along halfword boundaries and + * the upper 7 bits of each halfword are all 0s or 1s, we + * can use And Halfword Immediate + */ + tmp = ui & 0xfe00fe00; + if ((tmp == 0xfe00fe00 || tmp == 0) && ((ui >> 16) == (ui & 0x0000ffff))) { + spe_andhi(p, rT, rA, ui & 0x000003ff); + return; + } + + /* If the ui field is symmetric in each byte, then we can use + * the And Byte Immediate instruction. + */ + tmp = ui & 0x000000ff; + if ((ui >> 24) == tmp && ((ui >> 16) & 0xff) == tmp && ((ui >> 8) & 0xff) == tmp) { + spe_andbi(p, rT, rA, tmp); + return; + } + + /* Otherwise, we'll have to use a temporary register. */ + unsigned int tmp_reg = spe_allocate_available_register(p); + spe_load_uint(p, tmp_reg, ui); + spe_and(p, rT, rA, tmp_reg); + spe_release_register(p, tmp_reg); +} + + +/** + * This function is constructed identically to spe_and_uint() above. + * Changes to one should be made in the other. + */ +void +spe_xor_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui) +{ + /* If we can, emit a single instruction, either Exclusive Or Byte + * Immediate (which uses the same constant across each byte), Exclusive + * Or Halfword Immediate (which sign-extends a 10-bit immediate to + * 16 bits and uses that across each halfword), or Exclusive Or Word + * Immediate (which sign-extends a 10-bit immediate to 32 bits). + * + * Otherwise, we'll need to use a temporary register. + */ + unsigned int tmp; + + /* If the upper 23 bits are all 0s or all 1s, sign extension + * will work and we can use Exclusive Or Word Immediate + */ + tmp = ui & 0xfffffe00; + if (tmp == 0xfffffe00 || tmp == 0) { + spe_xori(p, rT, rA, ui & 0x000003ff); + return; + } + + /* If the ui field is symmetric along halfword boundaries and + * the upper 7 bits of each halfword are all 0s or 1s, we + * can use Exclusive Or Halfword Immediate + */ + tmp = ui & 0xfe00fe00; + if ((tmp == 0xfe00fe00 || tmp == 0) && ((ui >> 16) == (ui & 0x0000ffff))) { + spe_xorhi(p, rT, rA, ui & 0x000003ff); + return; + } + + /* If the ui field is symmetric in each byte, then we can use + * the Exclusive Or Byte Immediate instruction. + */ + tmp = ui & 0x000000ff; + if ((ui >> 24) == tmp && ((ui >> 16) & 0xff) == tmp && ((ui >> 8) & 0xff) == tmp) { + spe_xorbi(p, rT, rA, tmp); + return; + } + + /* Otherwise, we'll have to use a temporary register. */ + unsigned int tmp_reg = spe_allocate_available_register(p); + spe_load_uint(p, tmp_reg, ui); + spe_xor(p, rT, rA, tmp_reg); + spe_release_register(p, tmp_reg); +} + +void +spe_compare_equal_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui) +{ + /* If the comparison value is 9 bits or less, it fits inside a + * Compare Equal Word Immediate instruction. + */ + if ((ui & 0x000001ff) == ui) { + spe_ceqi(p, rT, rA, ui); + } + /* Otherwise, we're going to have to load a word first. */ + else { + unsigned int tmp_reg = spe_allocate_available_register(p); + spe_load_uint(p, tmp_reg, ui); + spe_ceq(p, rT, rA, tmp_reg); + spe_release_register(p, tmp_reg); + } +} + +void +spe_compare_greater_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui) +{ + /* If the comparison value is 10 bits or less, it fits inside a + * Compare Logical Greater Than Word Immediate instruction. + */ + if ((ui & 0x000003ff) == ui) { + spe_clgti(p, rT, rA, ui); + } + /* Otherwise, we're going to have to load a word first. */ + else { + unsigned int tmp_reg = spe_allocate_available_register(p); + spe_load_uint(p, tmp_reg, ui); + spe_clgt(p, rT, rA, tmp_reg); + spe_release_register(p, tmp_reg); + } +} void spe_splat(struct spe_function *p, unsigned rT, unsigned rA) diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h index 61c7edeb60..d6a3c02f20 100644 --- a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h +++ b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h @@ -53,17 +53,26 @@ struct spe_function uint num_inst; uint max_inst; - /** - * Mask of used / unused registers - * - * Each set bit corresponds to an available register. Each cleared bit - * corresponds to an allocated register. + /** + * The "set count" reflects the number of nested register sets + * are allowed. In the unlikely case that we exceed the set count, + * register allocation will start to be confused, which is critical + * enough that we check for it. + */ + unsigned char set_count; + + /** + * Flags for used and unused registers. Each byte corresponds to a + * register; a 0 in that byte means that the register is available. + * A value of 1 means that the register was allocated in the current + * register set. Any other value N means that the register was allocated + * N register sets ago. * * \sa * spe_allocate_register, spe_allocate_available_register, - * spe_release_register + * spe_allocate_register_set, spe_release_register_set, spe_release_register, */ - uint64_t regs[SPE_NUM_REGS / 64]; + unsigned char regs[SPE_NUM_REGS]; boolean print; /**< print/dump instructions as they're emitted? */ int indent; /**< number of spaces to indent */ @@ -77,6 +86,11 @@ extern unsigned spe_code_size(const struct spe_function *p); extern int spe_allocate_available_register(struct spe_function *p); extern int spe_allocate_register(struct spe_function *p, int reg); extern void spe_release_register(struct spe_function *p, int reg); +extern void spe_allocate_register_set(struct spe_function *p); +extern void spe_release_register_set(struct spe_function *p); + +extern unsigned +spe_get_registers_used(const struct spe_function *p, ubyte used[]); extern void spe_print_code(struct spe_function *p, boolean enable); extern void spe_indent(struct spe_function *p, int spaces); @@ -105,6 +119,9 @@ extern void spe_comment(struct spe_function *p, int rel_indent, const char *s); #define EMIT_RI10(_name, _op) \ extern void _name (struct spe_function *p, unsigned rT, unsigned rA, \ int imm) +#define EMIT_RI10s(_name, _op) \ + extern void _name (struct spe_function *p, unsigned rT, unsigned rA, \ + int imm) #define EMIT_RI16(_name, _op) \ extern void _name (struct spe_function *p, unsigned rT, int imm) #define EMIT_RI18(_name, _op) \ @@ -117,11 +134,9 @@ extern void spe_comment(struct spe_function *p, int rel_indent, const char *s); /* Memory load / store instructions */ -EMIT_RI10(spe_lqd, 0x034); EMIT_RR (spe_lqx, 0x1c4); EMIT_RI16(spe_lqa, 0x061); EMIT_RI16(spe_lqr, 0x067); -EMIT_RI10(spe_stqd, 0x024); EMIT_RR (spe_stqx, 0x144); EMIT_RI16(spe_stqa, 0x041); EMIT_RI16(spe_stqr, 0x047); @@ -151,7 +166,7 @@ EMIT_RI16(spe_fsmbi, 0x065); EMIT_RR (spe_ah, 0x0c8); EMIT_RI10(spe_ahi, 0x01d); EMIT_RR (spe_a, 0x0c0); -EMIT_RI10(spe_ai, 0x01c); +EMIT_RI10s(spe_ai, 0x01c); EMIT_RR (spe_sfh, 0x048); EMIT_RI10(spe_sfhi, 0x00d); EMIT_RR (spe_sf, 0x040); @@ -189,19 +204,19 @@ EMIT_R (spe_xshw, 0x2ae); EMIT_R (spe_xswd, 0x2a6); EMIT_RR (spe_and, 0x0c1); EMIT_RR (spe_andc, 0x2c1); -EMIT_RI10(spe_andbi, 0x016); -EMIT_RI10(spe_andhi, 0x015); -EMIT_RI10(spe_andi, 0x014); +EMIT_RI10s(spe_andbi, 0x016); +EMIT_RI10s(spe_andhi, 0x015); +EMIT_RI10s(spe_andi, 0x014); EMIT_RR (spe_or, 0x041); EMIT_RR (spe_orc, 0x2c9); -EMIT_RI10(spe_orbi, 0x006); -EMIT_RI10(spe_orhi, 0x005); -EMIT_RI10(spe_ori, 0x004); +EMIT_RI10s(spe_orbi, 0x006); +EMIT_RI10s(spe_orhi, 0x005); +EMIT_RI10s(spe_ori, 0x004); EMIT_R (spe_orx, 0x1f0); EMIT_RR (spe_xor, 0x241); -EMIT_RI10(spe_xorbi, 0x026); -EMIT_RI10(spe_xorhi, 0x025); -EMIT_RI10(spe_xori, 0x024); +EMIT_RI10s(spe_xorbi, 0x026); +EMIT_RI10s(spe_xorhi, 0x025); +EMIT_RI10s(spe_xori, 0x024); EMIT_RR (spe_nand, 0x0c9); EMIT_RR (spe_nor, 0x049); EMIT_RR (spe_eqv, 0x249); @@ -279,6 +294,12 @@ EMIT_RI16(spe_brz, 0x040); EMIT_RI16(spe_brhnz, 0x046); EMIT_RI16(spe_brhz, 0x044); +extern void +spe_lqd(struct spe_function *p, unsigned rT, unsigned rA, int offset); + +extern void +spe_stqd(struct spe_function *p, unsigned rT, unsigned rA, int offset); + extern void spe_bi(struct spe_function *p, unsigned rA, int d, int e); extern void spe_iret(struct spe_function *p, unsigned rA, int d, int e); extern void spe_bisled(struct spe_function *p, unsigned rT, unsigned rA, @@ -307,6 +328,22 @@ spe_load_int(struct spe_function *p, unsigned rT, int i); extern void spe_load_uint(struct spe_function *p, unsigned rT, unsigned int ui); +/** And immediate value into rT. */ +extern void +spe_and_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui); + +/** Xor immediate value into rT. */ +extern void +spe_xor_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui); + +/** Compare equal with immediate value. */ +extern void +spe_compare_equal_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui); + +/** Compare greater with immediate value. */ +extern void +spe_compare_greater_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui); + /** Replicate word 0 of rA across rT. */ extern void spe_splat(struct spe_function *p, unsigned rT, unsigned rA); @@ -388,6 +425,7 @@ EMIT_R (spe_wrch, 0x10d); #undef EMIT_RI7 #undef EMIT_RI8 #undef EMIT_RI10 +#undef EMIT_RI10s #undef EMIT_RI16 #undef EMIT_RI18 #undef EMIT_I16 diff --git a/src/gallium/auxiliary/rtasm/rtasm_x86sse.c b/src/gallium/auxiliary/rtasm/rtasm_x86sse.c index 3bba9dcc07..99ee74cf14 100644 --- a/src/gallium/auxiliary/rtasm/rtasm_x86sse.c +++ b/src/gallium/auxiliary/rtasm/rtasm_x86sse.c @@ -371,7 +371,11 @@ void x86_jcc( struct x86_function *p, DUMP_I(cc); if (offset < 0) { - assert(p->csr - p->store > -offset); + /*assert(p->csr - p->store > -offset);*/ + if (p->csr - p->store <= -offset) { + /* probably out of memory (using the error_overflow buffer) */ + return; + } } if (offset <= 127 && offset >= -128) { @@ -675,6 +679,44 @@ void x86_and( struct x86_function *p, * SSE instructions */ +void sse_prefetchnta( struct x86_function *p, struct x86_reg ptr) +{ + DUMP_R( ptr ); + assert(ptr.mod != mod_REG); + emit_2ub(p, 0x0f, 0x18); + emit_modrm_noreg(p, 0, ptr); +} + +void sse_prefetch0( struct x86_function *p, struct x86_reg ptr) +{ + DUMP_R( ptr ); + assert(ptr.mod != mod_REG); + emit_2ub(p, 0x0f, 0x18); + emit_modrm_noreg(p, 1, ptr); +} + +void sse_prefetch1( struct x86_function *p, struct x86_reg ptr) +{ + DUMP_R( ptr ); + assert(ptr.mod != mod_REG); + emit_2ub(p, 0x0f, 0x18); + emit_modrm_noreg(p, 2, ptr); +} + +void sse_movntps( struct x86_function *p, + struct x86_reg dst, + struct x86_reg src) +{ + DUMP_RR( dst, src ); + + assert(dst.mod != mod_REG); + assert(src.mod == mod_REG); + emit_2ub(p, 0x0f, 0x2b); + emit_modrm(p, src, dst); +} + + + void sse_movss( struct x86_function *p, struct x86_reg dst, diff --git a/src/gallium/auxiliary/rtasm/rtasm_x86sse.h b/src/gallium/auxiliary/rtasm/rtasm_x86sse.h index 510aa1b0de..1b5eaaca85 100644 --- a/src/gallium/auxiliary/rtasm/rtasm_x86sse.h +++ b/src/gallium/auxiliary/rtasm/rtasm_x86sse.h @@ -185,6 +185,13 @@ void sse2_pshufd( struct x86_function *p, struct x86_reg dest, struct x86_reg ar void sse2_rcpps( struct x86_function *p, struct x86_reg dst, struct x86_reg src ); void sse2_rcpss( struct x86_function *p, struct x86_reg dst, struct x86_reg src ); + +void sse_prefetchnta( struct x86_function *p, struct x86_reg ptr); +void sse_prefetch0( struct x86_function *p, struct x86_reg ptr); +void sse_prefetch1( struct x86_function *p, struct x86_reg ptr); + +void sse_movntps( struct x86_function *p, struct x86_reg dst, struct x86_reg src); + void sse_addps( struct x86_function *p, struct x86_reg dst, struct x86_reg src ); void sse_addss( struct x86_function *p, struct x86_reg dst, struct x86_reg src ); void sse_cvtps2pi( struct x86_function *p, struct x86_reg dst, struct x86_reg src ); |