diff options
author | Brian Paul <brian.paul@tungstengraphics.com> | 2008-10-29 12:14:11 -0600 |
---|---|---|
committer | Brian Paul <brian.paul@tungstengraphics.com> | 2008-10-29 12:14:11 -0600 |
commit | 1f7a323a138e6cc43b1192022b071c606a5ee6f4 (patch) | |
tree | f3d0b3fd54b67f498d94b926c39b5494ac3b4b62 /src/gallium/drivers/cell/ppu | |
parent | 8b3af5c5d6fe100707da0d9dcc42500921792638 (diff) |
cell: add scalar param to emit_function_call() to indicate scalar function calls
Scalar calls only use the X component of the src regs and smear the
result across the dest register's X/Y/Z/W.
Diffstat (limited to 'src/gallium/drivers/cell/ppu')
-rw-r--r-- | src/gallium/drivers/cell/ppu/cell_gen_fp.c | 103 |
1 files changed, 69 insertions, 34 deletions
diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fp.c b/src/gallium/drivers/cell/ppu/cell_gen_fp.c index d4d644d6e8..5c41b264ac 100644 --- a/src/gallium/drivers/cell/ppu/cell_gen_fp.c +++ b/src/gallium/drivers/cell/ppu/cell_gen_fp.c @@ -1303,60 +1303,91 @@ lookup_function(struct cell_context *cell, const char *funcname) /** * Emit code to call a SPU function. * Used to implement instructions like SIN/COS/POW/TEX/etc. + * If scalar, only the X components of the src regs are used, and the + * result is replicated across the dest register's XYZW components. */ static boolean emit_function_call(struct codegen *gen, const struct tgsi_full_instruction *inst, - char *funcname, uint num_args) + char *funcname, uint num_args, boolean scalar) { const uint addr = lookup_function(gen->cell, funcname); char comment[100]; - int ch; + int s_regs[3]; + int func_called = FALSE; + uint a, ch; + int retval_reg = -1; assert(num_args <= 3); snprintf(comment, sizeof(comment), "CALL %s:", funcname); spe_comment(gen->f, -4, comment); + if (scalar) { + for (a = 0; a < num_args; a++) { + s_regs[a] = get_src_reg(gen, CHAN_X, &inst->FullSrcRegisters[a]); + } + /* we'll call the function, put the return value in this register, + * then replicate it across all write-enabled components in d_reg. + */ + retval_reg = spe_allocate_available_register(gen->f); + } + for (ch = 0; ch < 4; ch++) { if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { - int s_regs[3], d_reg; + int d_reg; ubyte usedRegs[SPE_NUM_REGS]; - uint a, i, numUsed; + uint i, numUsed; - for (a = 0; a < num_args; a++) { - s_regs[a] = get_src_reg(gen, ch, &inst->FullSrcRegisters[a]); + if (!scalar) { + for (a = 0; a < num_args; a++) { + s_regs[a] = get_src_reg(gen, ch, &inst->FullSrcRegisters[a]); + } } - d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); - numUsed = spe_get_registers_used(gen->f, usedRegs); - assert(numUsed < gen->frame_size / 16 - 2); + d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); - /* save registers to stack */ - for (i = 0; i < numUsed; i++) { - uint reg = usedRegs[i]; - int offset = 2 + i; - spe_stqd(gen->f, reg, SPE_REG_SP, 16 * offset); - } + if (!scalar || !func_called) { + /* for a scalar function, we'll really only call the function once */ - /* setup function arguments */ - for (a = 0; a < num_args; a++) { - spe_move(gen->f, 3 + a, s_regs[a]); - } + numUsed = spe_get_registers_used(gen->f, usedRegs); + assert(numUsed < gen->frame_size / 16 - 2); - /* branch to function, save return addr */ - spe_brasl(gen->f, SPE_REG_RA, addr); + /* save registers to stack */ + for (i = 0; i < numUsed; i++) { + uint reg = usedRegs[i]; + int offset = 2 + i; + spe_stqd(gen->f, reg, SPE_REG_SP, 16 * offset); + } - /* save function's return value */ - spe_move(gen->f, d_reg, 3); + /* setup function arguments */ + for (a = 0; a < num_args; a++) { + spe_move(gen->f, 3 + a, s_regs[a]); + } - /* restore registers from stack */ - for (i = 0; i < numUsed; i++) { - uint reg = usedRegs[i]; - if (reg != d_reg) { - int offset = 2 + i; - spe_lqd(gen->f, reg, SPE_REG_SP, 16 * offset); + /* branch to function, save return addr */ + spe_brasl(gen->f, SPE_REG_RA, addr); + + /* save function's return value */ + if (scalar) + spe_move(gen->f, retval_reg, 3); + else + spe_move(gen->f, d_reg, 3); + + /* restore registers from stack */ + for (i = 0; i < numUsed; i++) { + uint reg = usedRegs[i]; + if (reg != d_reg && reg != retval_reg) { + int offset = 2 + i; + spe_lqd(gen->f, reg, SPE_REG_SP, 16 * offset); + } } + + func_called = TRUE; + } + + if (scalar) { + spe_move(gen->f, d_reg, retval_reg); } store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); @@ -1364,6 +1395,10 @@ emit_function_call(struct codegen *gen, } } + if (scalar) { + spe_release_register(gen->f, retval_reg); + } + return true; } @@ -1770,15 +1805,15 @@ emit_instruction(struct codegen *gen, return emit_END(gen); case TGSI_OPCODE_COS: - return emit_function_call(gen, inst, "spu_cos", 1); + return emit_function_call(gen, inst, "spu_cos", 1, TRUE); case TGSI_OPCODE_SIN: - return emit_function_call(gen, inst, "spu_sin", 1); + return emit_function_call(gen, inst, "spu_sin", 1, TRUE); case TGSI_OPCODE_POW: - return emit_function_call(gen, inst, "spu_pow", 2); + return emit_function_call(gen, inst, "spu_pow", 2, TRUE); case TGSI_OPCODE_EXPBASE2: - return emit_function_call(gen, inst, "spu_exp2", 1); + return emit_function_call(gen, inst, "spu_exp2", 1, TRUE); case TGSI_OPCODE_LOGBASE2: - return emit_function_call(gen, inst, "spu_log2", 1); + return emit_function_call(gen, inst, "spu_log2", 1, TRUE); case TGSI_OPCODE_TEX: /* fall-through for now */ case TGSI_OPCODE_TXD: |