| Message ID | 20180126045742.5487-2-richard.henderson@linaro.org |
|---|---|
| State | Superseded |
| Headers | show |
| Series | tcg: generic vector operations | expand |
Richard Henderson <richard.henderson@linaro.org> writes: > This will be required for storing vector constants. > > Reviewed-by: Peter Maydell <peter.maydell@linaro.org> > Signed-off-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> > --- > tcg/tcg-pool.inc.c | 115 +++++++++++++++++++++++++++++++++++++++++++---------- > 1 file changed, 93 insertions(+), 22 deletions(-) > > diff --git a/tcg/tcg-pool.inc.c b/tcg/tcg-pool.inc.c > index 8a85131405..ecf7c9282b 100644 > --- a/tcg/tcg-pool.inc.c > +++ b/tcg/tcg-pool.inc.c > @@ -22,39 +22,110 @@ > > typedef struct TCGLabelPoolData { > struct TCGLabelPoolData *next; > - tcg_target_ulong data; > tcg_insn_unit *label; > - intptr_t addend; > - int type; > + int32_t addend; > + uint16_t rtype; > + uint16_t nlong; > + tcg_target_ulong data[]; > } TCGLabelPoolData; > > > -static void new_pool_label(TCGContext *s, tcg_target_ulong data, int type, > - tcg_insn_unit *label, intptr_t addend) > +static TCGLabelPoolData *new_pool_alloc(TCGContext *s, int nlong, int rtype, > + tcg_insn_unit *label, int addend) > { > - TCGLabelPoolData *n = tcg_malloc(sizeof(*n)); > - TCGLabelPoolData *i, **pp; > + TCGLabelPoolData *n = tcg_malloc(sizeof(TCGLabelPoolData) > + + sizeof(tcg_target_ulong) * nlong); > > - n->data = data; > n->label = label; > - n->type = type; > n->addend = addend; > + n->rtype = rtype; > + n->nlong = nlong; > + return n; > +} > + > +static void new_pool_insert(TCGContext *s, TCGLabelPoolData *n) > +{ > + TCGLabelPoolData *i, **pp; > + int nlong = n->nlong; > > /* Insertion sort on the pool. */ > - for (pp = &s->pool_labels; (i = *pp) && i->data < data; pp = &i->next) { > - continue; > + for (pp = &s->pool_labels; (i = *pp) != NULL; pp = &i->next) { > + if (nlong > i->nlong) { > + break; > + } > + if (nlong < i->nlong) { > + continue; > + } > + if (memcmp(n->data, i->data, sizeof(tcg_target_ulong) * nlong) >= 0) { > + break; > + } > } > n->next = *pp; > *pp = n; > } > > +/* The "usual" for generic integer code. */ > +static inline void new_pool_label(TCGContext *s, tcg_target_ulong d, int rtype, > + tcg_insn_unit *label, int addend) > +{ > + TCGLabelPoolData *n = new_pool_alloc(s, 1, rtype, label, addend); > + n->data[0] = d; > + new_pool_insert(s, n); > +} > + > +/* For v64 or v128, depending on the host. */ > +static inline void new_pool_l2(TCGContext *s, int rtype, tcg_insn_unit *label, > + int addend, tcg_target_ulong d0, > + tcg_target_ulong d1) > +{ > + TCGLabelPoolData *n = new_pool_alloc(s, 2, rtype, label, addend); > + n->data[0] = d0; > + n->data[1] = d1; > + new_pool_insert(s, n); > +} > + > +/* For v128 or v256, depending on the host. */ > +static inline void new_pool_l4(TCGContext *s, int rtype, tcg_insn_unit *label, > + int addend, tcg_target_ulong d0, > + tcg_target_ulong d1, tcg_target_ulong d2, > + tcg_target_ulong d3) > +{ > + TCGLabelPoolData *n = new_pool_alloc(s, 4, rtype, label, addend); > + n->data[0] = d0; > + n->data[1] = d1; > + n->data[2] = d2; > + n->data[3] = d3; > + new_pool_insert(s, n); > +} > + > +/* For v256, for 32-bit host. */ > +static inline void new_pool_l8(TCGContext *s, int rtype, tcg_insn_unit *label, > + int addend, tcg_target_ulong d0, > + tcg_target_ulong d1, tcg_target_ulong d2, > + tcg_target_ulong d3, tcg_target_ulong d4, > + tcg_target_ulong d5, tcg_target_ulong d6, > + tcg_target_ulong d7) > +{ > + TCGLabelPoolData *n = new_pool_alloc(s, 8, rtype, label, addend); > + n->data[0] = d0; > + n->data[1] = d1; > + n->data[2] = d2; > + n->data[3] = d3; > + n->data[4] = d4; > + n->data[5] = d5; > + n->data[6] = d6; > + n->data[7] = d7; > + new_pool_insert(s, n); > +} > + > /* To be provided by cpu/tcg-target.inc.c. */ > static void tcg_out_nop_fill(tcg_insn_unit *p, int count); > > static bool tcg_out_pool_finalize(TCGContext *s) > { > TCGLabelPoolData *p = s->pool_labels; > - tcg_target_ulong d, *a; > + TCGLabelPoolData *l = NULL; > + void *a; > > if (p == NULL) { > return true; > @@ -62,24 +133,24 @@ static bool tcg_out_pool_finalize(TCGContext *s) > > /* ??? Round up to qemu_icache_linesize, but then do not round > again when allocating the next TranslationBlock structure. */ > - a = (void *)ROUND_UP((uintptr_t)s->code_ptr, sizeof(tcg_target_ulong)); > + a = (void *)ROUND_UP((uintptr_t)s->code_ptr, > + sizeof(tcg_target_ulong) * p->nlong); > tcg_out_nop_fill(s->code_ptr, (tcg_insn_unit *)a - s->code_ptr); > s->data_gen_ptr = a; > > - /* Ensure the first comparison fails. */ > - d = p->data + 1; > - > for (; p != NULL; p = p->next) { > - if (p->data != d) { > - d = p->data; > - if (unlikely((void *)a > s->code_gen_highwater)) { > + size_t size = sizeof(tcg_target_ulong) * p->nlong; > + if (!l || l->nlong != p->nlong || memcmp(l->data, p->data, size)) { > + if (unlikely(a > s->code_gen_highwater)) { > return false; > } > - *a++ = d; > + memcpy(a, p->data, size); > + a += size; > + l = p; > } > - patch_reloc(p->label, p->type, (intptr_t)(a - 1), p->addend); > + patch_reloc(p->label, p->rtype, (intptr_t)a - size, p->addend); > } > > - s->code_ptr = (void *)a; > + s->code_ptr = a; > return true; > } -- Alex Bennée
diff --git a/tcg/tcg-pool.inc.c b/tcg/tcg-pool.inc.c index 8a85131405..ecf7c9282b 100644 --- a/tcg/tcg-pool.inc.c +++ b/tcg/tcg-pool.inc.c @@ -22,39 +22,110 @@ typedef struct TCGLabelPoolData { struct TCGLabelPoolData *next; - tcg_target_ulong data; tcg_insn_unit *label; - intptr_t addend; - int type; + int32_t addend; + uint16_t rtype; + uint16_t nlong; + tcg_target_ulong data[]; } TCGLabelPoolData; -static void new_pool_label(TCGContext *s, tcg_target_ulong data, int type, - tcg_insn_unit *label, intptr_t addend) +static TCGLabelPoolData *new_pool_alloc(TCGContext *s, int nlong, int rtype, + tcg_insn_unit *label, int addend) { - TCGLabelPoolData *n = tcg_malloc(sizeof(*n)); - TCGLabelPoolData *i, **pp; + TCGLabelPoolData *n = tcg_malloc(sizeof(TCGLabelPoolData) + + sizeof(tcg_target_ulong) * nlong); - n->data = data; n->label = label; - n->type = type; n->addend = addend; + n->rtype = rtype; + n->nlong = nlong; + return n; +} + +static void new_pool_insert(TCGContext *s, TCGLabelPoolData *n) +{ + TCGLabelPoolData *i, **pp; + int nlong = n->nlong; /* Insertion sort on the pool. */ - for (pp = &s->pool_labels; (i = *pp) && i->data < data; pp = &i->next) { - continue; + for (pp = &s->pool_labels; (i = *pp) != NULL; pp = &i->next) { + if (nlong > i->nlong) { + break; + } + if (nlong < i->nlong) { + continue; + } + if (memcmp(n->data, i->data, sizeof(tcg_target_ulong) * nlong) >= 0) { + break; + } } n->next = *pp; *pp = n; } +/* The "usual" for generic integer code. */ +static inline void new_pool_label(TCGContext *s, tcg_target_ulong d, int rtype, + tcg_insn_unit *label, int addend) +{ + TCGLabelPoolData *n = new_pool_alloc(s, 1, rtype, label, addend); + n->data[0] = d; + new_pool_insert(s, n); +} + +/* For v64 or v128, depending on the host. */ +static inline void new_pool_l2(TCGContext *s, int rtype, tcg_insn_unit *label, + int addend, tcg_target_ulong d0, + tcg_target_ulong d1) +{ + TCGLabelPoolData *n = new_pool_alloc(s, 2, rtype, label, addend); + n->data[0] = d0; + n->data[1] = d1; + new_pool_insert(s, n); +} + +/* For v128 or v256, depending on the host. */ +static inline void new_pool_l4(TCGContext *s, int rtype, tcg_insn_unit *label, + int addend, tcg_target_ulong d0, + tcg_target_ulong d1, tcg_target_ulong d2, + tcg_target_ulong d3) +{ + TCGLabelPoolData *n = new_pool_alloc(s, 4, rtype, label, addend); + n->data[0] = d0; + n->data[1] = d1; + n->data[2] = d2; + n->data[3] = d3; + new_pool_insert(s, n); +} + +/* For v256, for 32-bit host. */ +static inline void new_pool_l8(TCGContext *s, int rtype, tcg_insn_unit *label, + int addend, tcg_target_ulong d0, + tcg_target_ulong d1, tcg_target_ulong d2, + tcg_target_ulong d3, tcg_target_ulong d4, + tcg_target_ulong d5, tcg_target_ulong d6, + tcg_target_ulong d7) +{ + TCGLabelPoolData *n = new_pool_alloc(s, 8, rtype, label, addend); + n->data[0] = d0; + n->data[1] = d1; + n->data[2] = d2; + n->data[3] = d3; + n->data[4] = d4; + n->data[5] = d5; + n->data[6] = d6; + n->data[7] = d7; + new_pool_insert(s, n); +} + /* To be provided by cpu/tcg-target.inc.c. */ static void tcg_out_nop_fill(tcg_insn_unit *p, int count); static bool tcg_out_pool_finalize(TCGContext *s) { TCGLabelPoolData *p = s->pool_labels; - tcg_target_ulong d, *a; + TCGLabelPoolData *l = NULL; + void *a; if (p == NULL) { return true; @@ -62,24 +133,24 @@ static bool tcg_out_pool_finalize(TCGContext *s) /* ??? Round up to qemu_icache_linesize, but then do not round again when allocating the next TranslationBlock structure. */ - a = (void *)ROUND_UP((uintptr_t)s->code_ptr, sizeof(tcg_target_ulong)); + a = (void *)ROUND_UP((uintptr_t)s->code_ptr, + sizeof(tcg_target_ulong) * p->nlong); tcg_out_nop_fill(s->code_ptr, (tcg_insn_unit *)a - s->code_ptr); s->data_gen_ptr = a; - /* Ensure the first comparison fails. */ - d = p->data + 1; - for (; p != NULL; p = p->next) { - if (p->data != d) { - d = p->data; - if (unlikely((void *)a > s->code_gen_highwater)) { + size_t size = sizeof(tcg_target_ulong) * p->nlong; + if (!l || l->nlong != p->nlong || memcmp(l->data, p->data, size)) { + if (unlikely(a > s->code_gen_highwater)) { return false; } - *a++ = d; + memcpy(a, p->data, size); + a += size; + l = p; } - patch_reloc(p->label, p->type, (intptr_t)(a - 1), p->addend); + patch_reloc(p->label, p->rtype, (intptr_t)a - size, p->addend); } - s->code_ptr = (void *)a; + s->code_ptr = a; return true; }