nvfx: use relocations array for vp constants

diff --git a/src/gallium/drivers/nvfx/nvfx_fragprog.c b/src/gallium/drivers/nvfx/nvfx_fragprog.c
index db33ecd78cfdcd0078734b2d10542fb7360d1f28..025989ac5bb29f8e3b7ecaacbe2efb9c5d27e8c0 100644 (file)
--- a/src/gallium/drivers/nvfx/nvfx_fragprog.c
+++ b/src/gallium/drivers/nvfx/nvfx_fragprog.c
@@ -270,7 +270,7 @@ nv40_fp_if(struct nvfx_fpc *fpc, struct nvfx_src src)
 static void
 nv40_fp_cal(struct nvfx_fpc *fpc, unsigned target)
 {
-        struct nvfx_label_relocation reloc;
+        struct nvfx_relocation reloc;
         uint32_t *hw;
         fpc->inst_offset = fpc->fp->insn_len;
         grow_insns(fpc, 4);
@@ -284,7 +284,7 @@ nv40_fp_cal(struct nvfx_fpc *fpc, unsigned target)
         hw[3] = 0;
         reloc.target = target;
         reloc.location = fpc->inst_offset + 2;
-        util_dynarray_append(&fpc->label_relocs, struct nvfx_label_relocation, reloc);
+        util_dynarray_append(&fpc->label_relocs, struct nvfx_relocation, reloc);
 }
 
 static void
@@ -306,7 +306,7 @@ nv40_fp_ret(struct nvfx_fpc *fpc)
 static void
 nv40_fp_rep(struct nvfx_fpc *fpc, unsigned count, unsigned target)
 {
-        struct nvfx_label_relocation reloc;
+        struct nvfx_relocation reloc;
         uint32_t *hw;
         fpc->inst_offset = fpc->fp->insn_len;
         grow_insns(fpc, 4);
@@ -325,7 +325,7 @@ nv40_fp_rep(struct nvfx_fpc *fpc, unsigned count, unsigned target)
         hw[3] = 0; /* | end_offset */
         reloc.target = target;
         reloc.location = fpc->inst_offset + 3;
-        util_dynarray_append(&fpc->label_relocs, struct nvfx_label_relocation, reloc);
+        util_dynarray_append(&fpc->label_relocs, struct nvfx_relocation, reloc);
         //util_dynarray_append(&fpc->loop_stack, unsigned, target);
 }
 
@@ -333,7 +333,7 @@ nv40_fp_rep(struct nvfx_fpc *fpc, unsigned count, unsigned target)
 static void
 nv40_fp_bra(struct nvfx_fpc *fpc, unsigned target)
 {
-        struct nvfx_label_relocation reloc;
+        struct nvfx_relocation reloc;
         uint32_t *hw;
         fpc->inst_offset = fpc->fp->insn_len;
         grow_insns(fpc, 4);
@@ -349,10 +349,10 @@ nv40_fp_bra(struct nvfx_fpc *fpc, unsigned target)
         hw[3] = 0; /* | endif_offset */
         reloc.target = target;
         reloc.location = fpc->inst_offset + 2;
-        util_dynarray_append(&fpc->label_relocs, struct nvfx_label_relocation, reloc);
+        util_dynarray_append(&fpc->label_relocs, struct nvfx_relocation, reloc);
         reloc.target = target;
         reloc.location = fpc->inst_offset + 3;
-        util_dynarray_append(&fpc->label_relocs, struct nvfx_label_relocation, reloc);
+        util_dynarray_append(&fpc->label_relocs, struct nvfx_relocation, reloc);
 }
 
 static void
@@ -1041,9 +1041,9 @@ nvfx_fragprog_translate(struct nvfx_context *nvfx,
        }
        util_dynarray_append(&insns, unsigned, fp->insn_len);
 
-       for(unsigned i = 0; i < fpc->label_relocs.size; i += sizeof(struct nvfx_label_relocation))
+       for(unsigned i = 0; i < fpc->label_relocs.size; i += sizeof(struct nvfx_relocation))
        {
-               struct nvfx_label_relocation* label_reloc = (struct nvfx_label_relocation*)((char*)fpc->label_relocs.data + i);
+               struct nvfx_relocation* label_reloc = (struct nvfx_relocation*)((char*)fpc->label_relocs.data + i);
                fp->insn[label_reloc->location] |= ((unsigned*)insns.data)[label_reloc->target];
        }
        util_dynarray_fini(&insns);

diff --git a/src/gallium/drivers/nvfx/nvfx_shader.h b/src/gallium/drivers/nvfx/nvfx_shader.h
index 52e684aec3be432ee46df38113e88f234822dd86..c711484ee2bf2499e9aab12f456ec69ea139edb7 100644 (file)
--- a/src/gallium/drivers/nvfx/nvfx_shader.h
+++ b/src/gallium/drivers/nvfx/nvfx_shader.h
@@ -509,7 +509,7 @@ nvfx_src_abs(struct nvfx_src src)
        return src;
 }
 
-struct nvfx_label_relocation {
+struct nvfx_relocation {
         unsigned location;
         unsigned target;
 };

diff --git a/src/gallium/drivers/nvfx/nvfx_state.h b/src/gallium/drivers/nvfx/nvfx_state.h
index 6d589af5f3555bcdf11c0849881668625da91ae0..1247abcfa219208564747e57239aaf035099c6e2 100644 (file)
--- a/src/gallium/drivers/nvfx/nvfx_state.h
+++ b/src/gallium/drivers/nvfx/nvfx_state.h
@@ -9,7 +9,6 @@
 
 struct nvfx_vertex_program_exec {
        uint32_t data[4];
-       int const_index;
 };
 
 struct nvfx_vertex_program_data {
@@ -46,6 +45,7 @@ struct nvfx_vertex_program {
        uint32_t clip_ctrl;
 
        struct util_dynarray branch_relocs;
+       struct util_dynarray const_relocs;
 };
 
 struct nvfx_fragment_program_data {

diff --git a/src/gallium/drivers/nvfx/nvfx_vertprog.c b/src/gallium/drivers/nvfx/nvfx_vertprog.c
index 38f37168a18885b5c00ccb62563ac905b29417ab..996680e32cc459bf52c44aad17e38d78fbd006c3 100644 (file)
--- a/src/gallium/drivers/nvfx/nvfx_vertprog.c
+++ b/src/gallium/drivers/nvfx/nvfx_vertprog.c
@@ -113,6 +113,7 @@ emit_src(struct nvfx_context* nvfx, struct nvfx_vpc *vpc, uint32_t *hw, int pos,
 {
        struct nvfx_vertex_program *vp = vpc->vp;
        uint32_t sr = 0;
+       struct nvfx_relocation reloc;
 
        switch (src.reg.type) {
        case NVFXSR_TEMP:
@@ -128,9 +129,9 @@ emit_src(struct nvfx_context* nvfx, struct nvfx_vpc *vpc, uint32_t *hw, int pos,
        case NVFXSR_CONST:
                sr |= (NVFX_VP(SRC_REG_TYPE_CONST) <<
                       NVFX_VP(SRC_REG_TYPE_SHIFT));
-               assert(vpc->vpi->const_index == -1 ||
-                      vpc->vpi->const_index == src.reg.index);
-               vpc->vpi->const_index = src.reg.index;
+               reloc.location = vp->nr_insns - 1;
+               reloc.target = src.reg.index;
+               util_dynarray_append(&vp->const_relocs, struct nvfx_relocation, reloc);
                break;
        case NVFXSR_NONE:
                sr |= (NVFX_VP(SRC_REG_TYPE_INPUT) <<
@@ -308,7 +309,6 @@ nvfx_vp_emit(struct nvfx_vpc *vpc, struct nvfx_insn insn)
        vp->insns = realloc(vp->insns, ++vp->nr_insns * sizeof(*vpc->vpi));
        vpc->vpi = &vp->insns[vp->nr_insns - 1];
        memset(vpc->vpi, 0, sizeof(*vpc->vpi));
-       vpc->vpi->const_index = -1;
 
        hw = vpc->vpi->data;
 
@@ -440,7 +440,7 @@ nvfx_vertprog_parse_instruction(struct nvfx_context* nvfx, struct nvfx_vpc *vpc,
        struct nvfx_reg dst;
        struct nvfx_src none = nvfx_src(nvfx_reg(NVFXSR_NONE, 0));
        struct nvfx_insn insn;
-       struct nvfx_label_relocation reloc;
+       struct nvfx_relocation reloc;
        struct nvfx_loop_entry loop;
        int mask;
        int ai = -1, ci = -1, ii = -1;
@@ -662,7 +662,7 @@ nvfx_vertprog_parse_instruction(struct nvfx_context* nvfx, struct nvfx_vpc *vpc,
 
                reloc.location = vpc->vp->nr_insns;
                reloc.target = finst->Label.Label + 1;
-               util_dynarray_append(&vpc->label_relocs, struct nvfx_label_relocation, reloc);
+               util_dynarray_append(&vpc->label_relocs, struct nvfx_relocation, reloc);
 
                insn = arith(SCA, BRA, none.reg, 0, none, none, none);
                insn.cc_test = NVFX_COND_EQ;
@@ -675,7 +675,7 @@ nvfx_vertprog_parse_instruction(struct nvfx_context* nvfx, struct nvfx_vpc *vpc,
        case TGSI_OPCODE_CAL:
                reloc.location = vpc->vp->nr_insns;
                reloc.target = finst->Label.Label;
-               util_dynarray_append(&vpc->label_relocs, struct nvfx_label_relocation, reloc);
+               util_dynarray_append(&vpc->label_relocs, struct nvfx_relocation, reloc);
 
                if(finst->Instruction.Opcode == TGSI_OPCODE_CAL)
                        insn = arith(SCA, CAL, none.reg, 0, none, none, none);
@@ -707,7 +707,7 @@ nvfx_vertprog_parse_instruction(struct nvfx_context* nvfx, struct nvfx_vpc *vpc,
 
                reloc.location = vpc->vp->nr_insns;
                reloc.target = loop.cont_target;
-               util_dynarray_append(&vpc->label_relocs, struct nvfx_label_relocation, reloc);
+               util_dynarray_append(&vpc->label_relocs, struct nvfx_relocation, reloc);
 
                nvfx_vp_emit(vpc, arith(SCA, BRA, none.reg, 0, none, none, none));
                break;
@@ -717,7 +717,7 @@ nvfx_vertprog_parse_instruction(struct nvfx_context* nvfx, struct nvfx_vpc *vpc,
 
                reloc.location = vpc->vp->nr_insns;
                reloc.target = loop.cont_target;
-               util_dynarray_append(&vpc->label_relocs, struct nvfx_label_relocation, reloc);
+               util_dynarray_append(&vpc->label_relocs, struct nvfx_relocation, reloc);
 
                nvfx_vp_emit(vpc, arith(SCA, BRA, none.reg, 0, none, none, none));
                break;
@@ -727,7 +727,7 @@ nvfx_vertprog_parse_instruction(struct nvfx_context* nvfx, struct nvfx_vpc *vpc,
 
                reloc.location = vpc->vp->nr_insns;
                reloc.target = loop.brk_target;
-               util_dynarray_append(&vpc->label_relocs, struct nvfx_label_relocation, reloc);
+               util_dynarray_append(&vpc->label_relocs, struct nvfx_relocation, reloc);
 
                nvfx_vp_emit(vpc, arith(SCA, BRA, none.reg, 0, none, none, none));
                break;
@@ -979,17 +979,17 @@ nvfx_vertprog_translate(struct nvfx_context *nvfx,
 
        util_dynarray_append(&insns, unsigned, vp->nr_insns);
 
-       for(unsigned i = 0; i < vpc->label_relocs.size; i += sizeof(struct nvfx_label_relocation))
+       for(unsigned i = 0; i < vpc->label_relocs.size; i += sizeof(struct nvfx_relocation))
        {
-               struct nvfx_label_relocation* label_reloc = (struct nvfx_label_relocation*)((char*)vpc->label_relocs.data + i);
-               struct nvfx_label_relocation hw_reloc;
+               struct nvfx_relocation* label_reloc = (struct nvfx_relocation*)((char*)vpc->label_relocs.data + i);
+               struct nvfx_relocation hw_reloc;
 
                hw_reloc.location = label_reloc->location;
                hw_reloc.target = ((unsigned*)insns.data)[label_reloc->target];
 
                //debug_printf("hw %u -> tgsi %u = hw %u\n", hw_reloc.location, label_reloc->target, hw_reloc.target);
 
-               util_dynarray_append(&vp->branch_relocs, struct nvfx_label_relocation, hw_reloc);
+               util_dynarray_append(&vp->branch_relocs, struct nvfx_relocation, hw_reloc);
        }
        util_dynarray_fini(&insns);
        util_dynarray_trim(&vp->branch_relocs);
@@ -1155,9 +1155,9 @@ nvfx_vertprog_validate(struct nvfx_context *nvfx)
         */
        if (vp->exec_start != vp->exec->start) {
                //printf("vp_relocs %u -> %u\n", vp->exec_start, vp->exec->start);
-               for(unsigned i = 0; i < vp->branch_relocs.size; i += sizeof(struct nvfx_label_relocation))
+               for(unsigned i = 0; i < vp->branch_relocs.size; i += sizeof(struct nvfx_relocation))
                {
-                       struct nvfx_label_relocation* reloc = (struct nvfx_label_relocation*)((char*)vp->branch_relocs.data + i);
+                       struct nvfx_relocation* reloc = (struct nvfx_relocation*)((char*)vp->branch_relocs.data + i);
                        uint32_t* hw = vp->insns[reloc->location].data;
                        unsigned target = vp->exec->start + reloc->target;
 
@@ -1182,16 +1182,15 @@ nvfx_vertprog_validate(struct nvfx_context *nvfx)
        }
 
        if (vp->nr_consts && vp->data_start != vp->data->start) {
-               for (i = 0; i < vp->nr_insns; i++) {
-                       struct nvfx_vertex_program_exec *vpi = &vp->insns[i];
+               for(unsigned i = 0; i < vp->const_relocs.size; i += sizeof(struct nvfx_relocation))
+               {
+                       struct nvfx_relocation* reloc = (struct nvfx_relocation*)((char*)vp->const_relocs.data + i);
+                       struct nvfx_vertex_program_exec *vpi = &vp->insns[reloc->location];
 
-                       if (vpi->const_index >= 0) {
-                               vpi->data[1] &= ~NVFX_VP(INST_CONST_SRC_MASK);
-                               vpi->data[1] |=
-                                       (vpi->const_index + vp->data->start) <<
+                       vpi->data[1] &= ~NVFX_VP(INST_CONST_SRC_MASK);
+                       vpi->data[1] |=
+                                       (reloc->target + vp->data->start) <<
                                        NVFX_VP(INST_CONST_SRC_SHIFT);
-
-                       }
                }
 
                vp->data_start = vp->data->start;