[Mesa-dev] [PATCH 1/6] radeonsi: move LLVM ALU codegen into radeonsi
Edward O'Callaghan
funfunctor at folklore1984.net
Mon Oct 17 22:54:17 UTC 2016
This series is,
Acked-by: Edward O'Callaghan <funfunctor at folklore1984.net>
On 10/18/2016 12:44 AM, Marek Olšák wrote:
> From: Marek Olšák <marek.olsak at amd.com>
>
> ---
> src/gallium/drivers/radeon/radeon_llvm.h | 9 -
> .../drivers/radeon/radeon_setup_tgsi_llvm.c | 979 +------------------
> src/gallium/drivers/radeonsi/Makefile.sources | 2 +
> src/gallium/drivers/radeonsi/si_shader.c | 9 +-
> src/gallium/drivers/radeonsi/si_shader_internal.h | 37 +
> src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c | 1012 ++++++++++++++++++++
> 6 files changed, 1056 insertions(+), 992 deletions(-)
> create mode 100644 src/gallium/drivers/radeonsi/si_shader_internal.h
> create mode 100644 src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
>
> diff --git a/src/gallium/drivers/radeon/radeon_llvm.h b/src/gallium/drivers/radeon/radeon_llvm.h
> index 6010254..367e4c0 100644
> --- a/src/gallium/drivers/radeon/radeon_llvm.h
> +++ b/src/gallium/drivers/radeon/radeon_llvm.h
> @@ -103,45 +103,36 @@ struct radeon_llvm_context {
> LLVMTypeRef tgsi2llvmtype(struct lp_build_tgsi_context *bld_base,
> enum tgsi_opcode_type type);
>
> LLVMValueRef bitcast(struct lp_build_tgsi_context *bld_base,
> enum tgsi_opcode_type type, LLVMValueRef value);
>
> LLVMValueRef radeon_llvm_bound_index(struct radeon_llvm_context *ctx,
> LLVMValueRef index,
> unsigned num);
>
> -void radeon_llvm_emit_prepare_cube_coords(struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data,
> - LLVMValueRef *coords_arg,
> - LLVMValueRef *derivs_arg);
> -
> void radeon_llvm_context_init(struct radeon_llvm_context *ctx,
> const char *triple,
> const struct tgsi_shader_info *info,
> const struct tgsi_token *tokens);
>
> void radeon_llvm_create_func(struct radeon_llvm_context *ctx,
> LLVMTypeRef *return_types, unsigned num_return_elems,
> LLVMTypeRef *ParamTypes, unsigned ParamCount);
>
> void radeon_llvm_dispose(struct radeon_llvm_context *ctx);
>
> unsigned radeon_llvm_reg_index_soa(unsigned index, unsigned chan);
>
> void radeon_llvm_finalize_module(struct radeon_llvm_context *ctx,
> bool run_verifier);
>
> -void build_tgsi_intrinsic_nomem(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data);
> -
> LLVMValueRef radeon_llvm_emit_fetch_64bit(struct lp_build_tgsi_context *bld_base,
> enum tgsi_opcode_type type,
> LLVMValueRef ptr,
> LLVMValueRef ptr2);
>
> LLVMValueRef radeon_llvm_saturate(struct lp_build_tgsi_context *bld_base,
> LLVMValueRef value);
>
> LLVMValueRef radeon_llvm_emit_fetch(struct lp_build_tgsi_context *bld_base,
> const struct tgsi_full_src_register *reg,
> diff --git a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
> index c843541..2713cc6 100644
> --- a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
> +++ b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
> @@ -1008,871 +1008,34 @@ static void uif_emit(const struct lp_build_tgsi_action *action,
> struct gallivm_state *gallivm = bld_base->base.gallivm;
> LLVMValueRef cond;
>
> cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE,
> bitcast(bld_base, TGSI_TYPE_UNSIGNED, emit_data->args[0]),
> bld_base->int_bld.zero, "");
>
> if_cond_emit(action, bld_base, emit_data, cond);
> }
>
> -static void kill_if_fetch_args(struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - const struct tgsi_full_instruction *inst = emit_data->inst;
> - struct gallivm_state *gallivm = bld_base->base.gallivm;
> - LLVMBuilderRef builder = gallivm->builder;
> - unsigned i;
> - LLVMValueRef conds[TGSI_NUM_CHANNELS];
> -
> - for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
> - LLVMValueRef value = lp_build_emit_fetch(bld_base, inst, 0, i);
> - conds[i] = LLVMBuildFCmp(builder, LLVMRealOLT, value,
> - bld_base->base.zero, "");
> - }
> -
> - /* Or the conditions together */
> - for (i = TGSI_NUM_CHANNELS - 1; i > 0; i--) {
> - conds[i - 1] = LLVMBuildOr(builder, conds[i], conds[i - 1], "");
> - }
> -
> - emit_data->dst_type = LLVMVoidTypeInContext(gallivm->context);
> - emit_data->arg_count = 1;
> - emit_data->args[0] = LLVMBuildSelect(builder, conds[0],
> - lp_build_const_float(gallivm, -1.0f),
> - bld_base->base.zero, "");
> -}
> -
> -static void kil_emit(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - unsigned i;
> - for (i = 0; i < emit_data->arg_count; i++) {
> - emit_data->output[i] = lp_build_intrinsic_unary(
> - bld_base->base.gallivm->builder,
> - action->intr_name,
> - emit_data->dst_type, emit_data->args[i]);
> - }
> -}
> -
> -static LLVMValueRef build_cube_intrinsic(struct gallivm_state *gallivm,
> - LLVMValueRef in[3])
> -{
> - if (HAVE_LLVM >= 0x0309) {
> - LLVMTypeRef f32 = LLVMTypeOf(in[0]);
> - LLVMValueRef out[4];
> -
> - out[0] = lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.cubetc",
> - f32, in, 3, LLVMReadNoneAttribute);
> - out[1] = lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.cubesc",
> - f32, in, 3, LLVMReadNoneAttribute);
> - out[2] = lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.cubema",
> - f32, in, 3, LLVMReadNoneAttribute);
> - out[3] = lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.cubeid",
> - f32, in, 3, LLVMReadNoneAttribute);
> -
> - return lp_build_gather_values(gallivm, out, 4);
> - } else {
> - LLVMValueRef c[4] = {
> - in[0],
> - in[1],
> - in[2],
> - LLVMGetUndef(LLVMTypeOf(in[0]))
> - };
> - LLVMValueRef vec = lp_build_gather_values(gallivm, c, 4);
> -
> - return lp_build_intrinsic(gallivm->builder, "llvm.AMDGPU.cube",
> - LLVMTypeOf(vec), &vec, 1,
> - LLVMReadNoneAttribute);
> - }
> -}
> -
> -static void radeon_llvm_cube_to_2d_coords(struct lp_build_tgsi_context *bld_base,
> - LLVMValueRef *in, LLVMValueRef *out)
> -{
> - struct gallivm_state *gallivm = bld_base->base.gallivm;
> - LLVMBuilderRef builder = gallivm->builder;
> - LLVMTypeRef type = bld_base->base.elem_type;
> - LLVMValueRef coords[4];
> - LLVMValueRef mad_args[3];
> - LLVMValueRef v;
> - unsigned i;
> -
> - v = build_cube_intrinsic(gallivm, in);
> -
> - for (i = 0; i < 4; ++i)
> - coords[i] = LLVMBuildExtractElement(builder, v,
> - lp_build_const_int32(gallivm, i), "");
> -
> - coords[2] = lp_build_intrinsic(builder, "llvm.fabs.f32",
> - type, &coords[2], 1, LLVMReadNoneAttribute);
> - coords[2] = lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_RCP, coords[2]);
> -
> - mad_args[1] = coords[2];
> - mad_args[2] = LLVMConstReal(type, 1.5);
> -
> - mad_args[0] = coords[0];
> - coords[0] = lp_build_emit_llvm_ternary(bld_base, TGSI_OPCODE_MAD,
> - mad_args[0], mad_args[1], mad_args[2]);
> -
> - mad_args[0] = coords[1];
> - coords[1] = lp_build_emit_llvm_ternary(bld_base, TGSI_OPCODE_MAD,
> - mad_args[0], mad_args[1], mad_args[2]);
> -
> - /* apply xyz = yxw swizzle to cooords */
> - out[0] = coords[1];
> - out[1] = coords[0];
> - out[2] = coords[3];
> -}
> -
> -void radeon_llvm_emit_prepare_cube_coords(struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data,
> - LLVMValueRef *coords_arg,
> - LLVMValueRef *derivs_arg)
> -{
> -
> - unsigned target = emit_data->inst->Texture.Texture;
> - unsigned opcode = emit_data->inst->Instruction.Opcode;
> - struct gallivm_state *gallivm = bld_base->base.gallivm;
> - LLVMBuilderRef builder = gallivm->builder;
> - LLVMValueRef coords[4];
> - unsigned i;
> -
> - radeon_llvm_cube_to_2d_coords(bld_base, coords_arg, coords);
> -
> - if (opcode == TGSI_OPCODE_TXD && derivs_arg) {
> - LLVMValueRef derivs[4];
> - int axis;
> -
> - /* Convert cube derivatives to 2D derivatives. */
> - for (axis = 0; axis < 2; axis++) {
> - LLVMValueRef shifted_cube_coords[4], shifted_coords[4];
> -
> - /* Shift the cube coordinates by the derivatives to get
> - * the cube coordinates of the "neighboring pixel".
> - */
> - for (i = 0; i < 3; i++)
> - shifted_cube_coords[i] =
> - LLVMBuildFAdd(builder, coords_arg[i],
> - derivs_arg[axis*3+i], "");
> - shifted_cube_coords[3] = LLVMGetUndef(bld_base->base.elem_type);
> -
> - /* Project the shifted cube coordinates onto the face. */
> - radeon_llvm_cube_to_2d_coords(bld_base, shifted_cube_coords,
> - shifted_coords);
> -
> - /* Subtract both sets of 2D coordinates to get 2D derivatives.
> - * This won't work if the shifted coordinates ended up
> - * in a different face.
> - */
> - for (i = 0; i < 2; i++)
> - derivs[axis * 2 + i] =
> - LLVMBuildFSub(builder, shifted_coords[i],
> - coords[i], "");
> - }
> -
> - memcpy(derivs_arg, derivs, sizeof(derivs));
> - }
> -
> - if (target == TGSI_TEXTURE_CUBE_ARRAY ||
> - target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
> - /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
> - /* coords_arg.w component - array_index for cube arrays */
> - coords[2] = lp_build_emit_llvm_ternary(bld_base, TGSI_OPCODE_MAD,
> - coords_arg[3], lp_build_const_float(gallivm, 8.0), coords[2]);
> - }
> -
> - /* Preserve compare/lod/bias. Put it in coords.w. */
> - if (opcode == TGSI_OPCODE_TEX2 ||
> - opcode == TGSI_OPCODE_TXB2 ||
> - opcode == TGSI_OPCODE_TXL2) {
> - coords[3] = coords_arg[4];
> - } else if (opcode == TGSI_OPCODE_TXB ||
> - opcode == TGSI_OPCODE_TXL ||
> - target == TGSI_TEXTURE_SHADOWCUBE) {
> - coords[3] = coords_arg[3];
> - }
> -
> - memcpy(coords_arg, coords, sizeof(coords));
> -}
> -
> -static void emit_icmp(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - unsigned pred;
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - LLVMContextRef context = bld_base->base.gallivm->context;
> -
> - switch (emit_data->inst->Instruction.Opcode) {
> - case TGSI_OPCODE_USEQ:
> - case TGSI_OPCODE_U64SEQ: pred = LLVMIntEQ; break;
> - case TGSI_OPCODE_USNE:
> - case TGSI_OPCODE_U64SNE: pred = LLVMIntNE; break;
> - case TGSI_OPCODE_USGE:
> - case TGSI_OPCODE_U64SGE: pred = LLVMIntUGE; break;
> - case TGSI_OPCODE_USLT:
> - case TGSI_OPCODE_U64SLT: pred = LLVMIntULT; break;
> - case TGSI_OPCODE_ISGE:
> - case TGSI_OPCODE_I64SGE: pred = LLVMIntSGE; break;
> - case TGSI_OPCODE_ISLT:
> - case TGSI_OPCODE_I64SLT: pred = LLVMIntSLT; break;
> - default:
> - assert(!"unknown instruction");
> - pred = 0;
> - break;
> - }
> -
> - LLVMValueRef v = LLVMBuildICmp(builder, pred,
> - emit_data->args[0], emit_data->args[1],"");
> -
> - v = LLVMBuildSExtOrBitCast(builder, v,
> - LLVMInt32TypeInContext(context), "");
> -
> - emit_data->output[emit_data->chan] = v;
> -}
> -
> -static void emit_ucmp(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> -
> - LLVMValueRef arg0 = LLVMBuildBitCast(builder, emit_data->args[0],
> - bld_base->uint_bld.elem_type, "");
> -
> - LLVMValueRef v = LLVMBuildICmp(builder, LLVMIntNE, arg0,
> - bld_base->uint_bld.zero, "");
> -
> - emit_data->output[emit_data->chan] =
> - LLVMBuildSelect(builder, v, emit_data->args[1], emit_data->args[2], "");
> -}
> -
> -static void emit_cmp(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - LLVMValueRef cond, *args = emit_data->args;
> -
> - cond = LLVMBuildFCmp(builder, LLVMRealOLT, args[0],
> - bld_base->base.zero, "");
> -
> - emit_data->output[emit_data->chan] =
> - LLVMBuildSelect(builder, cond, args[1], args[2], "");
> -}
> -
> -static void emit_set_cond(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - LLVMRealPredicate pred;
> - LLVMValueRef cond;
> -
> - /* Use ordered for everything but NE (which is usual for
> - * float comparisons)
> - */
> - switch (emit_data->inst->Instruction.Opcode) {
> - case TGSI_OPCODE_SGE: pred = LLVMRealOGE; break;
> - case TGSI_OPCODE_SEQ: pred = LLVMRealOEQ; break;
> - case TGSI_OPCODE_SLE: pred = LLVMRealOLE; break;
> - case TGSI_OPCODE_SLT: pred = LLVMRealOLT; break;
> - case TGSI_OPCODE_SNE: pred = LLVMRealUNE; break;
> - case TGSI_OPCODE_SGT: pred = LLVMRealOGT; break;
> - default: assert(!"unknown instruction"); pred = 0; break;
> - }
> -
> - cond = LLVMBuildFCmp(builder,
> - pred, emit_data->args[0], emit_data->args[1], "");
> -
> - emit_data->output[emit_data->chan] = LLVMBuildSelect(builder,
> - cond, bld_base->base.one, bld_base->base.zero, "");
> -}
> -
> -static void emit_fcmp(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - LLVMContextRef context = bld_base->base.gallivm->context;
> - LLVMRealPredicate pred;
> -
> - /* Use ordered for everything but NE (which is usual for
> - * float comparisons)
> - */
> - switch (emit_data->inst->Instruction.Opcode) {
> - case TGSI_OPCODE_FSEQ: pred = LLVMRealOEQ; break;
> - case TGSI_OPCODE_FSGE: pred = LLVMRealOGE; break;
> - case TGSI_OPCODE_FSLT: pred = LLVMRealOLT; break;
> - case TGSI_OPCODE_FSNE: pred = LLVMRealUNE; break;
> - default: assert(!"unknown instruction"); pred = 0; break;
> - }
> -
> - LLVMValueRef v = LLVMBuildFCmp(builder, pred,
> - emit_data->args[0], emit_data->args[1],"");
> -
> - v = LLVMBuildSExtOrBitCast(builder, v,
> - LLVMInt32TypeInContext(context), "");
> -
> - emit_data->output[emit_data->chan] = v;
> -}
> -
> -static void emit_dcmp(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - LLVMContextRef context = bld_base->base.gallivm->context;
> - LLVMRealPredicate pred;
> -
> - /* Use ordered for everything but NE (which is usual for
> - * float comparisons)
> - */
> - switch (emit_data->inst->Instruction.Opcode) {
> - case TGSI_OPCODE_DSEQ: pred = LLVMRealOEQ; break;
> - case TGSI_OPCODE_DSGE: pred = LLVMRealOGE; break;
> - case TGSI_OPCODE_DSLT: pred = LLVMRealOLT; break;
> - case TGSI_OPCODE_DSNE: pred = LLVMRealUNE; break;
> - default: assert(!"unknown instruction"); pred = 0; break;
> - }
> -
> - LLVMValueRef v = LLVMBuildFCmp(builder, pred,
> - emit_data->args[0], emit_data->args[1],"");
> -
> - v = LLVMBuildSExtOrBitCast(builder, v,
> - LLVMInt32TypeInContext(context), "");
> -
> - emit_data->output[emit_data->chan] = v;
> -}
> -
> -static void emit_not(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - LLVMValueRef v = bitcast(bld_base, TGSI_TYPE_UNSIGNED,
> - emit_data->args[0]);
> - emit_data->output[emit_data->chan] = LLVMBuildNot(builder, v, "");
> -}
> -
> -static void emit_arl(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - LLVMValueRef floor_index = lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_FLR, emit_data->args[0]);
> - emit_data->output[emit_data->chan] = LLVMBuildFPToSI(builder,
> - floor_index, bld_base->base.int_elem_type , "");
> -}
> -
> -static void emit_and(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - emit_data->output[emit_data->chan] = LLVMBuildAnd(builder,
> - emit_data->args[0], emit_data->args[1], "");
> -}
> -
> -static void emit_or(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - emit_data->output[emit_data->chan] = LLVMBuildOr(builder,
> - emit_data->args[0], emit_data->args[1], "");
> -}
> -
> -static void emit_uadd(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - emit_data->output[emit_data->chan] = LLVMBuildAdd(builder,
> - emit_data->args[0], emit_data->args[1], "");
> -}
> -
> -static void emit_udiv(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - emit_data->output[emit_data->chan] = LLVMBuildUDiv(builder,
> - emit_data->args[0], emit_data->args[1], "");
> -}
> -
> -static void emit_idiv(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - emit_data->output[emit_data->chan] = LLVMBuildSDiv(builder,
> - emit_data->args[0], emit_data->args[1], "");
> -}
> -
> -static void emit_mod(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - emit_data->output[emit_data->chan] = LLVMBuildSRem(builder,
> - emit_data->args[0], emit_data->args[1], "");
> -}
> -
> -static void emit_umod(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - emit_data->output[emit_data->chan] = LLVMBuildURem(builder,
> - emit_data->args[0], emit_data->args[1], "");
> -}
> -
> -static void emit_shl(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - emit_data->output[emit_data->chan] = LLVMBuildShl(builder,
> - emit_data->args[0], emit_data->args[1], "");
> -}
> -
> -static void emit_ushr(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - emit_data->output[emit_data->chan] = LLVMBuildLShr(builder,
> - emit_data->args[0], emit_data->args[1], "");
> -}
> -static void emit_ishr(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - emit_data->output[emit_data->chan] = LLVMBuildAShr(builder,
> - emit_data->args[0], emit_data->args[1], "");
> -}
> -
> -static void emit_xor(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - emit_data->output[emit_data->chan] = LLVMBuildXor(builder,
> - emit_data->args[0], emit_data->args[1], "");
> -}
> -
> -static void emit_ssg(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> -
> - LLVMValueRef cmp, val;
> -
> - if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_I64SSG) {
> - cmp = LLVMBuildICmp(builder, LLVMIntSGT, emit_data->args[0], bld_base->int64_bld.zero, "");
> - val = LLVMBuildSelect(builder, cmp, bld_base->int64_bld.one, emit_data->args[0], "");
> - cmp = LLVMBuildICmp(builder, LLVMIntSGE, val, bld_base->int64_bld.zero, "");
> - val = LLVMBuildSelect(builder, cmp, val, LLVMConstInt(bld_base->int64_bld.elem_type, -1, true), "");
> - } else if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_ISSG) {
> - cmp = LLVMBuildICmp(builder, LLVMIntSGT, emit_data->args[0], bld_base->int_bld.zero, "");
> - val = LLVMBuildSelect(builder, cmp, bld_base->int_bld.one, emit_data->args[0], "");
> - cmp = LLVMBuildICmp(builder, LLVMIntSGE, val, bld_base->int_bld.zero, "");
> - val = LLVMBuildSelect(builder, cmp, val, LLVMConstInt(bld_base->int_bld.elem_type, -1, true), "");
> - } else { // float SSG
> - cmp = LLVMBuildFCmp(builder, LLVMRealOGT, emit_data->args[0], bld_base->base.zero, "");
> - val = LLVMBuildSelect(builder, cmp, bld_base->base.one, emit_data->args[0], "");
> - cmp = LLVMBuildFCmp(builder, LLVMRealOGE, val, bld_base->base.zero, "");
> - val = LLVMBuildSelect(builder, cmp, val, LLVMConstReal(bld_base->base.elem_type, -1), "");
> - }
> -
> - emit_data->output[emit_data->chan] = val;
> -}
> -
> -static void emit_ineg(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - emit_data->output[emit_data->chan] = LLVMBuildNeg(builder,
> - emit_data->args[0], "");
> -}
> -
> -static void emit_dneg(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - emit_data->output[emit_data->chan] = LLVMBuildFNeg(builder,
> - emit_data->args[0], "");
> -}
> -
> -static void emit_frac(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - char *intr;
> -
> - if (emit_data->info->opcode == TGSI_OPCODE_FRC)
> - intr = "llvm.floor.f32";
> - else if (emit_data->info->opcode == TGSI_OPCODE_DFRAC)
> - intr = "llvm.floor.f64";
> - else {
> - assert(0);
> - return;
> - }
> -
> - LLVMValueRef floor = lp_build_intrinsic(builder, intr, emit_data->dst_type,
> - &emit_data->args[0], 1,
> - LLVMReadNoneAttribute);
> - emit_data->output[emit_data->chan] = LLVMBuildFSub(builder,
> - emit_data->args[0], floor, "");
> -}
> -
> -static void emit_f2i(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - emit_data->output[emit_data->chan] = LLVMBuildFPToSI(builder,
> - emit_data->args[0], bld_base->int_bld.elem_type, "");
> -}
> -
> -static void emit_f2u(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - emit_data->output[emit_data->chan] = LLVMBuildFPToUI(builder,
> - emit_data->args[0], bld_base->uint_bld.elem_type, "");
> -}
> -
> -static void emit_i2f(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - emit_data->output[emit_data->chan] = LLVMBuildSIToFP(builder,
> - emit_data->args[0], bld_base->base.elem_type, "");
> -}
> -
> -static void emit_u2f(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - emit_data->output[emit_data->chan] = LLVMBuildUIToFP(builder,
> - emit_data->args[0], bld_base->base.elem_type, "");
> -}
> -
> static void emit_immediate(struct lp_build_tgsi_context *bld_base,
> const struct tgsi_full_immediate *imm)
> {
> unsigned i;
> struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
>
> for (i = 0; i < 4; ++i) {
> ctx->soa.immediates[ctx->soa.num_immediates][i] =
> LLVMConstInt(bld_base->uint_bld.elem_type, imm->u[i].Uint, false );
> }
>
> ctx->soa.num_immediates++;
> }
>
> -void
> -build_tgsi_intrinsic_nomem(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - struct lp_build_context *base = &bld_base->base;
> - emit_data->output[emit_data->chan] =
> - lp_build_intrinsic(base->gallivm->builder, action->intr_name,
> - emit_data->dst_type, emit_data->args,
> - emit_data->arg_count, LLVMReadNoneAttribute);
> -}
> -
> -static void emit_bfi(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - struct gallivm_state *gallivm = bld_base->base.gallivm;
> - LLVMBuilderRef builder = gallivm->builder;
> - LLVMValueRef bfi_args[3];
> -
> - // Calculate the bitmask: (((1 << src3) - 1) << src2
> - bfi_args[0] = LLVMBuildShl(builder,
> - LLVMBuildSub(builder,
> - LLVMBuildShl(builder,
> - bld_base->int_bld.one,
> - emit_data->args[3], ""),
> - bld_base->int_bld.one, ""),
> - emit_data->args[2], "");
> -
> - bfi_args[1] = LLVMBuildShl(builder, emit_data->args[1],
> - emit_data->args[2], "");
> -
> - bfi_args[2] = emit_data->args[0];
> -
> - /* Calculate:
> - * (arg0 & arg1) | (~arg0 & arg2) = arg2 ^ (arg0 & (arg1 ^ arg2)
> - * Use the right-hand side, which the LLVM backend can convert to V_BFI.
> - */
> - emit_data->output[emit_data->chan] =
> - LLVMBuildXor(builder, bfi_args[2],
> - LLVMBuildAnd(builder, bfi_args[0],
> - LLVMBuildXor(builder, bfi_args[1], bfi_args[2],
> - ""), ""), "");
> -}
> -
> -/* this is ffs in C */
> -static void emit_lsb(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - struct gallivm_state *gallivm = bld_base->base.gallivm;
> - LLVMValueRef args[2] = {
> - emit_data->args[0],
> -
> - /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
> - * add special code to check for x=0. The reason is that
> - * the LLVM behavior for x=0 is different from what we
> - * need here.
> - *
> - * The hardware already implements the correct behavior.
> - */
> - LLVMConstInt(LLVMInt1TypeInContext(gallivm->context), 1, 0)
> - };
> -
> - emit_data->output[emit_data->chan] =
> - lp_build_intrinsic(gallivm->builder, "llvm.cttz.i32",
> - emit_data->dst_type, args, ARRAY_SIZE(args),
> - LLVMReadNoneAttribute);
> -}
> -
> -/* Find the last bit set. */
> -static void emit_umsb(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - struct gallivm_state *gallivm = bld_base->base.gallivm;
> - LLVMBuilderRef builder = gallivm->builder;
> - LLVMValueRef args[2] = {
> - emit_data->args[0],
> - /* Don't generate code for handling zero: */
> - LLVMConstInt(LLVMInt1TypeInContext(gallivm->context), 1, 0)
> - };
> -
> - LLVMValueRef msb =
> - lp_build_intrinsic(builder, "llvm.ctlz.i32",
> - emit_data->dst_type, args, ARRAY_SIZE(args),
> - LLVMReadNoneAttribute);
> -
> - /* The HW returns the last bit index from MSB, but TGSI wants
> - * the index from LSB. Invert it by doing "31 - msb". */
> - msb = LLVMBuildSub(builder, lp_build_const_int32(gallivm, 31),
> - msb, "");
> -
> - /* Check for zero: */
> - emit_data->output[emit_data->chan] =
> - LLVMBuildSelect(builder,
> - LLVMBuildICmp(builder, LLVMIntEQ, args[0],
> - bld_base->uint_bld.zero, ""),
> - lp_build_const_int32(gallivm, -1), msb, "");
> -}
> -
> -/* Find the last bit opposite of the sign bit. */
> -static void emit_imsb(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - struct gallivm_state *gallivm = bld_base->base.gallivm;
> - LLVMBuilderRef builder = gallivm->builder;
> - LLVMValueRef arg = emit_data->args[0];
> -
> - LLVMValueRef msb =
> - lp_build_intrinsic(builder, "llvm.AMDGPU.flbit.i32",
> - emit_data->dst_type, &arg, 1,
> - LLVMReadNoneAttribute);
> -
> - /* The HW returns the last bit index from MSB, but TGSI wants
> - * the index from LSB. Invert it by doing "31 - msb". */
> - msb = LLVMBuildSub(builder, lp_build_const_int32(gallivm, 31),
> - msb, "");
> -
> - /* If arg == 0 || arg == -1 (0xffffffff), return -1. */
> - LLVMValueRef all_ones = lp_build_const_int32(gallivm, -1);
> -
> - LLVMValueRef cond =
> - LLVMBuildOr(builder,
> - LLVMBuildICmp(builder, LLVMIntEQ, arg,
> - bld_base->uint_bld.zero, ""),
> - LLVMBuildICmp(builder, LLVMIntEQ, arg,
> - all_ones, ""), "");
> -
> - emit_data->output[emit_data->chan] =
> - LLVMBuildSelect(builder, cond, all_ones, msb, "");
> -}
> -
> -static void emit_iabs(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> -
> - emit_data->output[emit_data->chan] =
> - lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_IMAX,
> - emit_data->args[0],
> - LLVMBuildNeg(builder,
> - emit_data->args[0], ""));
> -}
> -
> -static void emit_minmax_int(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - LLVMIntPredicate op;
> -
> - switch (emit_data->info->opcode) {
> - default:
> - assert(0);
> - case TGSI_OPCODE_IMAX:
> - case TGSI_OPCODE_I64MAX:
> - op = LLVMIntSGT;
> - break;
> - case TGSI_OPCODE_IMIN:
> - case TGSI_OPCODE_I64MIN:
> - op = LLVMIntSLT;
> - break;
> - case TGSI_OPCODE_UMAX:
> - case TGSI_OPCODE_U64MAX:
> - op = LLVMIntUGT;
> - break;
> - case TGSI_OPCODE_UMIN:
> - case TGSI_OPCODE_U64MIN:
> - op = LLVMIntULT;
> - break;
> - }
> -
> - emit_data->output[emit_data->chan] =
> - LLVMBuildSelect(builder,
> - LLVMBuildICmp(builder, op, emit_data->args[0],
> - emit_data->args[1], ""),
> - emit_data->args[0],
> - emit_data->args[1], "");
> -}
> -
> -static void pk2h_fetch_args(struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst,
> - 0, TGSI_CHAN_X);
> - emit_data->args[1] = lp_build_emit_fetch(bld_base, emit_data->inst,
> - 0, TGSI_CHAN_Y);
> -}
> -
> -static void emit_pk2h(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - LLVMContextRef context = bld_base->base.gallivm->context;
> - struct lp_build_context *uint_bld = &bld_base->uint_bld;
> - LLVMTypeRef fp16, i16;
> - LLVMValueRef const16, comp[2];
> - unsigned i;
> -
> - fp16 = LLVMHalfTypeInContext(context);
> - i16 = LLVMInt16TypeInContext(context);
> - const16 = lp_build_const_int32(uint_bld->gallivm, 16);
> -
> - for (i = 0; i < 2; i++) {
> - comp[i] = LLVMBuildFPTrunc(builder, emit_data->args[i], fp16, "");
> - comp[i] = LLVMBuildBitCast(builder, comp[i], i16, "");
> - comp[i] = LLVMBuildZExt(builder, comp[i], uint_bld->elem_type, "");
> - }
> -
> - comp[1] = LLVMBuildShl(builder, comp[1], const16, "");
> - comp[0] = LLVMBuildOr(builder, comp[0], comp[1], "");
> -
> - emit_data->output[emit_data->chan] = comp[0];
> -}
> -
> -static void up2h_fetch_args(struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst,
> - 0, TGSI_CHAN_X);
> -}
> -
> -static void emit_up2h(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> - LLVMContextRef context = bld_base->base.gallivm->context;
> - struct lp_build_context *uint_bld = &bld_base->uint_bld;
> - LLVMTypeRef fp16, i16;
> - LLVMValueRef const16, input, val;
> - unsigned i;
> -
> - fp16 = LLVMHalfTypeInContext(context);
> - i16 = LLVMInt16TypeInContext(context);
> - const16 = lp_build_const_int32(uint_bld->gallivm, 16);
> - input = emit_data->args[0];
> -
> - for (i = 0; i < 2; i++) {
> - val = i == 1 ? LLVMBuildLShr(builder, input, const16, "") : input;
> - val = LLVMBuildTrunc(builder, val, i16, "");
> - val = LLVMBuildBitCast(builder, val, fp16, "");
> - emit_data->output[i] =
> - LLVMBuildFPExt(builder, val, bld_base->base.elem_type, "");
> - }
> -}
> -
> -static void emit_fdiv(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
> -
> - emit_data->output[emit_data->chan] =
> - LLVMBuildFDiv(bld_base->base.gallivm->builder,
> - emit_data->args[0], emit_data->args[1], "");
> -
> - /* Use v_rcp_f32 instead of precise division. */
> - if (HAVE_LLVM >= 0x0309 &&
> - !LLVMIsConstant(emit_data->output[emit_data->chan]))
> - LLVMSetMetadata(emit_data->output[emit_data->chan],
> - ctx->fpmath_md_kind, ctx->fpmath_md_2p5_ulp);
> -}
> -
> -/* 1/sqrt is translated to rsq for f32 if fp32 denormals are not enabled in
> - * the target machine. f64 needs global unsafe math flags to get rsq. */
> -static void emit_rsq(const struct lp_build_tgsi_action *action,
> - struct lp_build_tgsi_context *bld_base,
> - struct lp_build_emit_data *emit_data)
> -{
> - LLVMValueRef sqrt =
> - lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_SQRT,
> - emit_data->args[0]);
> -
> - emit_data->output[emit_data->chan] =
> - lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_DIV,
> - bld_base->base.one, sqrt);
> -}
> -
> void radeon_llvm_context_init(struct radeon_llvm_context *ctx, const char *triple,
> const struct tgsi_shader_info *info,
> const struct tgsi_token *tokens)
> {
> struct lp_type type;
>
> /* Initialize the gallivm object:
> * We are only using the module, context, and builder fields of this struct.
> * This should be enough for us to be able to pass our gallivm struct to the
> * helper functions in the gallivm module.
> @@ -1944,166 +1107,28 @@ void radeon_llvm_context_init(struct radeon_llvm_context *ctx, const char *tripl
> /* metadata allowing 2.5 ULP */
> ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->gallivm.context,
> "fpmath", 6);
> LLVMValueRef arg = lp_build_const_float(&ctx->gallivm, 2.5);
> ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->gallivm.context,
> &arg, 1);
>
> /* Allocate outputs */
> ctx->soa.outputs = ctx->outputs;
>
> - lp_set_default_actions(bld_base);
> -
> - bld_base->op_actions[TGSI_OPCODE_ABS].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_ABS].intr_name = "llvm.fabs.f32";
> - bld_base->op_actions[TGSI_OPCODE_AND].emit = emit_and;
> - bld_base->op_actions[TGSI_OPCODE_ARL].emit = emit_arl;
> - bld_base->op_actions[TGSI_OPCODE_BFI].emit = emit_bfi;
> bld_base->op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
> - bld_base->op_actions[TGSI_OPCODE_BREV].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_BREV].intr_name =
> - HAVE_LLVM >= 0x0308 ? "llvm.bitreverse.i32" : "llvm.AMDGPU.brev";
> bld_base->op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
> - bld_base->op_actions[TGSI_OPCODE_CEIL].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_CEIL].intr_name = "llvm.ceil.f32";
> - bld_base->op_actions[TGSI_OPCODE_CLAMP].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_CLAMP].intr_name =
> - HAVE_LLVM >= 0x0308 ? "llvm.AMDGPU.clamp." : "llvm.AMDIL.clamp.";
> - bld_base->op_actions[TGSI_OPCODE_CMP].emit = emit_cmp;
> bld_base->op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
> - bld_base->op_actions[TGSI_OPCODE_COS].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_COS].intr_name = "llvm.cos.f32";
> - bld_base->op_actions[TGSI_OPCODE_DABS].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_DABS].intr_name = "llvm.fabs.f64";
> - bld_base->op_actions[TGSI_OPCODE_DFMA].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_DFMA].intr_name = "llvm.fma.f64";
> - bld_base->op_actions[TGSI_OPCODE_DFRAC].emit = emit_frac;
> - bld_base->op_actions[TGSI_OPCODE_DIV].emit = emit_fdiv;
> - bld_base->op_actions[TGSI_OPCODE_DNEG].emit = emit_dneg;
> - bld_base->op_actions[TGSI_OPCODE_DSEQ].emit = emit_dcmp;
> - bld_base->op_actions[TGSI_OPCODE_DSGE].emit = emit_dcmp;
> - bld_base->op_actions[TGSI_OPCODE_DSLT].emit = emit_dcmp;
> - bld_base->op_actions[TGSI_OPCODE_DSNE].emit = emit_dcmp;
> - bld_base->op_actions[TGSI_OPCODE_DRSQ].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_DRSQ].intr_name =
> - HAVE_LLVM >= 0x0309 ? "llvm.amdgcn.rsq.f64" : "llvm.AMDGPU.rsq.f64";
> - bld_base->op_actions[TGSI_OPCODE_DSQRT].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_DSQRT].intr_name = "llvm.sqrt.f64";
> + bld_base->op_actions[TGSI_OPCODE_IF].emit = if_emit;
> + bld_base->op_actions[TGSI_OPCODE_UIF].emit = uif_emit;
> bld_base->op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
> bld_base->op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
> bld_base->op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
> - bld_base->op_actions[TGSI_OPCODE_EX2].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_EX2].intr_name =
> - HAVE_LLVM >= 0x0308 ? "llvm.exp2.f32" : "llvm.AMDIL.exp.";
> - bld_base->op_actions[TGSI_OPCODE_FLR].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_FLR].intr_name = "llvm.floor.f32";
> - bld_base->op_actions[TGSI_OPCODE_FMA].emit =
> - bld_base->op_actions[TGSI_OPCODE_MAD].emit;
> - bld_base->op_actions[TGSI_OPCODE_FRC].emit = emit_frac;
> - bld_base->op_actions[TGSI_OPCODE_F2I].emit = emit_f2i;
> - bld_base->op_actions[TGSI_OPCODE_F2U].emit = emit_f2u;
> - bld_base->op_actions[TGSI_OPCODE_FSEQ].emit = emit_fcmp;
> - bld_base->op_actions[TGSI_OPCODE_FSGE].emit = emit_fcmp;
> - bld_base->op_actions[TGSI_OPCODE_FSLT].emit = emit_fcmp;
> - bld_base->op_actions[TGSI_OPCODE_FSNE].emit = emit_fcmp;
> - bld_base->op_actions[TGSI_OPCODE_IABS].emit = emit_iabs;
> - bld_base->op_actions[TGSI_OPCODE_IBFE].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_IBFE].intr_name = "llvm.AMDGPU.bfe.i32";
> - bld_base->op_actions[TGSI_OPCODE_IDIV].emit = emit_idiv;
> - bld_base->op_actions[TGSI_OPCODE_IF].emit = if_emit;
> - bld_base->op_actions[TGSI_OPCODE_UIF].emit = uif_emit;
> - bld_base->op_actions[TGSI_OPCODE_IMAX].emit = emit_minmax_int;
> - bld_base->op_actions[TGSI_OPCODE_IMIN].emit = emit_minmax_int;
> - bld_base->op_actions[TGSI_OPCODE_IMSB].emit = emit_imsb;
> - bld_base->op_actions[TGSI_OPCODE_INEG].emit = emit_ineg;
> - bld_base->op_actions[TGSI_OPCODE_ISHR].emit = emit_ishr;
> - bld_base->op_actions[TGSI_OPCODE_ISGE].emit = emit_icmp;
> - bld_base->op_actions[TGSI_OPCODE_ISLT].emit = emit_icmp;
> - bld_base->op_actions[TGSI_OPCODE_ISSG].emit = emit_ssg;
> - bld_base->op_actions[TGSI_OPCODE_I2F].emit = emit_i2f;
> - bld_base->op_actions[TGSI_OPCODE_KILL_IF].fetch_args = kill_if_fetch_args;
> - bld_base->op_actions[TGSI_OPCODE_KILL_IF].emit = kil_emit;
> - bld_base->op_actions[TGSI_OPCODE_KILL_IF].intr_name = "llvm.AMDGPU.kill";
> - bld_base->op_actions[TGSI_OPCODE_KILL].emit = lp_build_tgsi_intrinsic;
> - bld_base->op_actions[TGSI_OPCODE_KILL].intr_name = "llvm.AMDGPU.kilp";
> - bld_base->op_actions[TGSI_OPCODE_LSB].emit = emit_lsb;
> - bld_base->op_actions[TGSI_OPCODE_LG2].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_LG2].intr_name = "llvm.log2.f32";
> - bld_base->op_actions[TGSI_OPCODE_MOD].emit = emit_mod;
> - bld_base->op_actions[TGSI_OPCODE_UMSB].emit = emit_umsb;
> - bld_base->op_actions[TGSI_OPCODE_NOT].emit = emit_not;
> - bld_base->op_actions[TGSI_OPCODE_OR].emit = emit_or;
> - bld_base->op_actions[TGSI_OPCODE_PK2H].fetch_args = pk2h_fetch_args;
> - bld_base->op_actions[TGSI_OPCODE_PK2H].emit = emit_pk2h;
> - bld_base->op_actions[TGSI_OPCODE_POPC].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_POPC].intr_name = "llvm.ctpop.i32";
> - bld_base->op_actions[TGSI_OPCODE_POW].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_POW].intr_name = "llvm.pow.f32";
> - bld_base->op_actions[TGSI_OPCODE_ROUND].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_ROUND].intr_name = "llvm.rint.f32";
> - bld_base->op_actions[TGSI_OPCODE_RSQ].emit = emit_rsq;
> - bld_base->op_actions[TGSI_OPCODE_SGE].emit = emit_set_cond;
> - bld_base->op_actions[TGSI_OPCODE_SEQ].emit = emit_set_cond;
> - bld_base->op_actions[TGSI_OPCODE_SHL].emit = emit_shl;
> - bld_base->op_actions[TGSI_OPCODE_SLE].emit = emit_set_cond;
> - bld_base->op_actions[TGSI_OPCODE_SLT].emit = emit_set_cond;
> - bld_base->op_actions[TGSI_OPCODE_SNE].emit = emit_set_cond;
> - bld_base->op_actions[TGSI_OPCODE_SGT].emit = emit_set_cond;
> - bld_base->op_actions[TGSI_OPCODE_SIN].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_SIN].intr_name = "llvm.sin.f32";
> - bld_base->op_actions[TGSI_OPCODE_SQRT].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_SQRT].intr_name = "llvm.sqrt.f32";
> - bld_base->op_actions[TGSI_OPCODE_SSG].emit = emit_ssg;
> - bld_base->op_actions[TGSI_OPCODE_TRUNC].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_TRUNC].intr_name = "llvm.trunc.f32";
> - bld_base->op_actions[TGSI_OPCODE_UADD].emit = emit_uadd;
> - bld_base->op_actions[TGSI_OPCODE_UBFE].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_UBFE].intr_name = "llvm.AMDGPU.bfe.u32";
> - bld_base->op_actions[TGSI_OPCODE_UDIV].emit = emit_udiv;
> - bld_base->op_actions[TGSI_OPCODE_UMAX].emit = emit_minmax_int;
> - bld_base->op_actions[TGSI_OPCODE_UMIN].emit = emit_minmax_int;
> - bld_base->op_actions[TGSI_OPCODE_UMOD].emit = emit_umod;
> - bld_base->op_actions[TGSI_OPCODE_USEQ].emit = emit_icmp;
> - bld_base->op_actions[TGSI_OPCODE_USGE].emit = emit_icmp;
> - bld_base->op_actions[TGSI_OPCODE_USHR].emit = emit_ushr;
> - bld_base->op_actions[TGSI_OPCODE_USLT].emit = emit_icmp;
> - bld_base->op_actions[TGSI_OPCODE_USNE].emit = emit_icmp;
> - bld_base->op_actions[TGSI_OPCODE_U2F].emit = emit_u2f;
> - bld_base->op_actions[TGSI_OPCODE_XOR].emit = emit_xor;
> - bld_base->op_actions[TGSI_OPCODE_UCMP].emit = emit_ucmp;
> - bld_base->op_actions[TGSI_OPCODE_UP2H].fetch_args = up2h_fetch_args;
> - bld_base->op_actions[TGSI_OPCODE_UP2H].emit = emit_up2h;
> -
> - bld_base->op_actions[TGSI_OPCODE_I64MAX].emit = emit_minmax_int;
> - bld_base->op_actions[TGSI_OPCODE_I64MIN].emit = emit_minmax_int;
> - bld_base->op_actions[TGSI_OPCODE_U64MAX].emit = emit_minmax_int;
> - bld_base->op_actions[TGSI_OPCODE_U64MIN].emit = emit_minmax_int;
> - bld_base->op_actions[TGSI_OPCODE_I64ABS].emit = emit_iabs;
> - bld_base->op_actions[TGSI_OPCODE_I64SSG].emit = emit_ssg;
> - bld_base->op_actions[TGSI_OPCODE_I64NEG].emit = emit_ineg;
> -
> - bld_base->op_actions[TGSI_OPCODE_U64SEQ].emit = emit_icmp;
> - bld_base->op_actions[TGSI_OPCODE_U64SNE].emit = emit_icmp;
> - bld_base->op_actions[TGSI_OPCODE_U64SGE].emit = emit_icmp;
> - bld_base->op_actions[TGSI_OPCODE_U64SLT].emit = emit_icmp;
> - bld_base->op_actions[TGSI_OPCODE_I64SGE].emit = emit_icmp;
> - bld_base->op_actions[TGSI_OPCODE_I64SLT].emit = emit_icmp;
> -
> - bld_base->op_actions[TGSI_OPCODE_U64ADD].emit = emit_uadd;
> - bld_base->op_actions[TGSI_OPCODE_U64SHL].emit = emit_shl;
> - bld_base->op_actions[TGSI_OPCODE_U64SHR].emit = emit_ushr;
> - bld_base->op_actions[TGSI_OPCODE_I64SHR].emit = emit_ishr;
> -
> - bld_base->op_actions[TGSI_OPCODE_U64MOD].emit = emit_umod;
> - bld_base->op_actions[TGSI_OPCODE_I64MOD].emit = emit_mod;
> - bld_base->op_actions[TGSI_OPCODE_U64DIV].emit = emit_udiv;
> - bld_base->op_actions[TGSI_OPCODE_I64DIV].emit = emit_idiv;
> }
>
> void radeon_llvm_create_func(struct radeon_llvm_context *ctx,
> LLVMTypeRef *return_types, unsigned num_return_elems,
> LLVMTypeRef *ParamTypes, unsigned ParamCount)
> {
> LLVMTypeRef main_fn_type, ret_type;
> LLVMBasicBlockRef main_fn_body;
>
> if (num_return_elems)
> diff --git a/src/gallium/drivers/radeonsi/Makefile.sources b/src/gallium/drivers/radeonsi/Makefile.sources
> index 96b8839..a16f19c 100644
> --- a/src/gallium/drivers/radeonsi/Makefile.sources
> +++ b/src/gallium/drivers/radeonsi/Makefile.sources
> @@ -8,18 +8,20 @@ C_SOURCES := \
> si_dma.c \
> si_hw_context.c \
> si_pipe.c \
> si_pipe.h \
> si_pm4.c \
> si_pm4.h \
> si_perfcounter.c \
> si_public.h \
> si_shader.c \
> si_shader.h \
> + si_shader_internal.h \
> + si_shader_tgsi_alu.c \
> si_state.c \
> si_state_draw.c \
> si_state_shaders.c \
> si_state.h \
> si_uvd.c
>
> GENERATED_SOURCES := \
> sid_tables.h
> diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
> index e6edd90..e3c27b1 100644
> --- a/src/gallium/drivers/radeonsi/si_shader.c
> +++ b/src/gallium/drivers/radeonsi/si_shader.c
> @@ -35,20 +35,21 @@
> #include "gallivm/lp_bld_misc.h"
> #include "radeon/radeon_llvm.h"
> #include "radeon/radeon_elf_util.h"
> #include "radeon/radeon_llvm_emit.h"
> #include "util/u_memory.h"
> #include "util/u_string.h"
> #include "tgsi/tgsi_build.h"
> #include "tgsi/tgsi_util.h"
> #include "tgsi/tgsi_dump.h"
>
> +#include "si_shader_internal.h"
> #include "si_pipe.h"
> #include "sid.h"
>
>
> static const char *scratch_rsrc_dword0_symbol =
> "SCRATCH_RSRC_DWORD0";
>
> static const char *scratch_rsrc_dword1_symbol =
> "SCRATCH_RSRC_DWORD1";
>
> @@ -4669,21 +4670,21 @@ static void tex_fetch_args(
> for (param = 0; param < 2; param++)
> for (chan = 0; chan < num_src_deriv_channels; chan++)
> derivs[param * num_src_deriv_channels + chan] =
> lp_build_emit_fetch(bld_base, inst, param+1, chan);
> }
>
> if (target == TGSI_TEXTURE_CUBE ||
> target == TGSI_TEXTURE_CUBE_ARRAY ||
> target == TGSI_TEXTURE_SHADOWCUBE ||
> target == TGSI_TEXTURE_SHADOWCUBE_ARRAY)
> - radeon_llvm_emit_prepare_cube_coords(bld_base, emit_data, coords, derivs);
> + si_prepare_cube_coords(bld_base, emit_data, coords, derivs);
>
> if (opcode == TGSI_OPCODE_TXD)
> for (int i = 0; i < num_deriv_channels * 2; i++)
> address[count++] = derivs[i];
>
> /* Pack texture coordinates */
> address[count++] = coords[0];
> if (num_coords > 1)
> address[count++] = coords[1];
> if (num_coords > 2)
> @@ -6520,20 +6521,21 @@ static void si_init_shader_ctx(struct si_shader_context *ctx,
> LLVMTargetMachineRef tm)
> {
> struct lp_build_tgsi_context *bld_base;
> struct lp_build_tgsi_action tmpl = {};
>
> memset(ctx, 0, sizeof(*ctx));
> radeon_llvm_context_init(
> &ctx->radeon_bld, "amdgcn--",
> (shader && shader->selector) ? &shader->selector->info : NULL,
> (shader && shader->selector) ? shader->selector->tokens : NULL);
> + si_shader_context_init_alu(&ctx->radeon_bld.soa.bld_base);
> ctx->tm = tm;
> ctx->screen = sscreen;
> if (shader && shader->selector)
> ctx->type = shader->selector->info.processor;
> else
> ctx->type = -1;
> ctx->shader = shader;
>
> ctx->voidt = LLVMVoidTypeInContext(ctx->radeon_bld.gallivm.context);
> ctx->i1 = LLVMInt1TypeInContext(ctx->radeon_bld.gallivm.context);
> @@ -6603,25 +6605,20 @@ static void si_init_shader_ctx(struct si_shader_context *ctx,
> bld_base->op_actions[TGSI_OPCODE_MEMBAR].emit = membar_emit;
>
> bld_base->op_actions[TGSI_OPCODE_DDX].emit = si_llvm_emit_ddxy;
> bld_base->op_actions[TGSI_OPCODE_DDY].emit = si_llvm_emit_ddxy;
> bld_base->op_actions[TGSI_OPCODE_DDX_FINE].emit = si_llvm_emit_ddxy;
> bld_base->op_actions[TGSI_OPCODE_DDY_FINE].emit = si_llvm_emit_ddxy;
>
> bld_base->op_actions[TGSI_OPCODE_EMIT].emit = si_llvm_emit_vertex;
> bld_base->op_actions[TGSI_OPCODE_ENDPRIM].emit = si_llvm_emit_primitive;
> bld_base->op_actions[TGSI_OPCODE_BARRIER].emit = si_llvm_emit_barrier;
> -
> - bld_base->op_actions[TGSI_OPCODE_MAX].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_MAX].intr_name = "llvm.maxnum.f32";
> - bld_base->op_actions[TGSI_OPCODE_MIN].emit = build_tgsi_intrinsic_nomem;
> - bld_base->op_actions[TGSI_OPCODE_MIN].intr_name = "llvm.minnum.f32";
> }
>
> int si_compile_tgsi_shader(struct si_screen *sscreen,
> LLVMTargetMachineRef tm,
> struct si_shader *shader,
> bool is_monolithic,
> struct pipe_debug_callback *debug)
> {
> struct si_shader_selector *sel = shader->selector;
> struct si_shader_context ctx;
> diff --git a/src/gallium/drivers/radeonsi/si_shader_internal.h b/src/gallium/drivers/radeonsi/si_shader_internal.h
> new file mode 100644
> index 0000000..44dd5fd
> --- /dev/null
> +++ b/src/gallium/drivers/radeonsi/si_shader_internal.h
> @@ -0,0 +1,37 @@
> +/*
> + * Copyright 2016 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * on the rights to use, copy, modify, merge, publish, distribute, sub
> + * license, and/or sell copies of the Software, and to permit persons to whom
> + * the Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice (including the next
> + * paragraph) shall be included in all copies or substantial portions of the
> + * Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
> + * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
> + * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
> + * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
> + * USE OR OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#ifndef SI_SHADER_PRIVATE_H
> +#define SI_SHADER_PRIVATE_H
> +
> +#include "si_shader.h"
> +#include "gallivm/lp_bld_init.h"
> +#include "gallivm/lp_bld_tgsi.h"
> +
> +void si_shader_context_init_alu(struct lp_build_tgsi_context *bld_base);
> +void si_prepare_cube_coords(struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data,
> + LLVMValueRef *coords_arg,
> + LLVMValueRef *derivs_arg);
> +
> +#endif
> diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c b/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
> new file mode 100644
> index 0000000..192a2cb
> --- /dev/null
> +++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
> @@ -0,0 +1,1012 @@
> +/*
> + * Copyright 2016 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * on the rights to use, copy, modify, merge, publish, distribute, sub
> + * license, and/or sell copies of the Software, and to permit persons to whom
> + * the Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice (including the next
> + * paragraph) shall be included in all copies or substantial portions of the
> + * Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
> + * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
> + * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
> + * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
> + * USE OR OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#include "si_shader_internal.h"
> +#include "gallivm/lp_bld_const.h"
> +#include "gallivm/lp_bld_intr.h"
> +#include "gallivm/lp_bld_gather.h"
> +#include "tgsi/tgsi_parse.h"
> +#include "radeon/radeon_llvm.h"
> +
> +static void kill_if_fetch_args(struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + const struct tgsi_full_instruction *inst = emit_data->inst;
> + struct gallivm_state *gallivm = bld_base->base.gallivm;
> + LLVMBuilderRef builder = gallivm->builder;
> + unsigned i;
> + LLVMValueRef conds[TGSI_NUM_CHANNELS];
> +
> + for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
> + LLVMValueRef value = lp_build_emit_fetch(bld_base, inst, 0, i);
> + conds[i] = LLVMBuildFCmp(builder, LLVMRealOLT, value,
> + bld_base->base.zero, "");
> + }
> +
> + /* Or the conditions together */
> + for (i = TGSI_NUM_CHANNELS - 1; i > 0; i--) {
> + conds[i - 1] = LLVMBuildOr(builder, conds[i], conds[i - 1], "");
> + }
> +
> + emit_data->dst_type = LLVMVoidTypeInContext(gallivm->context);
> + emit_data->arg_count = 1;
> + emit_data->args[0] = LLVMBuildSelect(builder, conds[0],
> + lp_build_const_float(gallivm, -1.0f),
> + bld_base->base.zero, "");
> +}
> +
> +static void kil_emit(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + unsigned i;
> + for (i = 0; i < emit_data->arg_count; i++) {
> + emit_data->output[i] = lp_build_intrinsic_unary(
> + bld_base->base.gallivm->builder,
> + action->intr_name,
> + emit_data->dst_type, emit_data->args[i]);
> + }
> +}
> +
> +static void emit_icmp(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + unsigned pred;
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + LLVMContextRef context = bld_base->base.gallivm->context;
> +
> + switch (emit_data->inst->Instruction.Opcode) {
> + case TGSI_OPCODE_USEQ:
> + case TGSI_OPCODE_U64SEQ: pred = LLVMIntEQ; break;
> + case TGSI_OPCODE_USNE:
> + case TGSI_OPCODE_U64SNE: pred = LLVMIntNE; break;
> + case TGSI_OPCODE_USGE:
> + case TGSI_OPCODE_U64SGE: pred = LLVMIntUGE; break;
> + case TGSI_OPCODE_USLT:
> + case TGSI_OPCODE_U64SLT: pred = LLVMIntULT; break;
> + case TGSI_OPCODE_ISGE:
> + case TGSI_OPCODE_I64SGE: pred = LLVMIntSGE; break;
> + case TGSI_OPCODE_ISLT:
> + case TGSI_OPCODE_I64SLT: pred = LLVMIntSLT; break;
> + default:
> + assert(!"unknown instruction");
> + pred = 0;
> + break;
> + }
> +
> + LLVMValueRef v = LLVMBuildICmp(builder, pred,
> + emit_data->args[0], emit_data->args[1],"");
> +
> + v = LLVMBuildSExtOrBitCast(builder, v,
> + LLVMInt32TypeInContext(context), "");
> +
> + emit_data->output[emit_data->chan] = v;
> +}
> +
> +static void emit_ucmp(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> +
> + LLVMValueRef arg0 = LLVMBuildBitCast(builder, emit_data->args[0],
> + bld_base->uint_bld.elem_type, "");
> +
> + LLVMValueRef v = LLVMBuildICmp(builder, LLVMIntNE, arg0,
> + bld_base->uint_bld.zero, "");
> +
> + emit_data->output[emit_data->chan] =
> + LLVMBuildSelect(builder, v, emit_data->args[1], emit_data->args[2], "");
> +}
> +
> +static void emit_cmp(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + LLVMValueRef cond, *args = emit_data->args;
> +
> + cond = LLVMBuildFCmp(builder, LLVMRealOLT, args[0],
> + bld_base->base.zero, "");
> +
> + emit_data->output[emit_data->chan] =
> + LLVMBuildSelect(builder, cond, args[1], args[2], "");
> +}
> +
> +static void emit_set_cond(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + LLVMRealPredicate pred;
> + LLVMValueRef cond;
> +
> + /* Use ordered for everything but NE (which is usual for
> + * float comparisons)
> + */
> + switch (emit_data->inst->Instruction.Opcode) {
> + case TGSI_OPCODE_SGE: pred = LLVMRealOGE; break;
> + case TGSI_OPCODE_SEQ: pred = LLVMRealOEQ; break;
> + case TGSI_OPCODE_SLE: pred = LLVMRealOLE; break;
> + case TGSI_OPCODE_SLT: pred = LLVMRealOLT; break;
> + case TGSI_OPCODE_SNE: pred = LLVMRealUNE; break;
> + case TGSI_OPCODE_SGT: pred = LLVMRealOGT; break;
> + default: assert(!"unknown instruction"); pred = 0; break;
> + }
> +
> + cond = LLVMBuildFCmp(builder,
> + pred, emit_data->args[0], emit_data->args[1], "");
> +
> + emit_data->output[emit_data->chan] = LLVMBuildSelect(builder,
> + cond, bld_base->base.one, bld_base->base.zero, "");
> +}
> +
> +static void emit_fcmp(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + LLVMContextRef context = bld_base->base.gallivm->context;
> + LLVMRealPredicate pred;
> +
> + /* Use ordered for everything but NE (which is usual for
> + * float comparisons)
> + */
> + switch (emit_data->inst->Instruction.Opcode) {
> + case TGSI_OPCODE_FSEQ: pred = LLVMRealOEQ; break;
> + case TGSI_OPCODE_FSGE: pred = LLVMRealOGE; break;
> + case TGSI_OPCODE_FSLT: pred = LLVMRealOLT; break;
> + case TGSI_OPCODE_FSNE: pred = LLVMRealUNE; break;
> + default: assert(!"unknown instruction"); pred = 0; break;
> + }
> +
> + LLVMValueRef v = LLVMBuildFCmp(builder, pred,
> + emit_data->args[0], emit_data->args[1],"");
> +
> + v = LLVMBuildSExtOrBitCast(builder, v,
> + LLVMInt32TypeInContext(context), "");
> +
> + emit_data->output[emit_data->chan] = v;
> +}
> +
> +static void emit_dcmp(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + LLVMContextRef context = bld_base->base.gallivm->context;
> + LLVMRealPredicate pred;
> +
> + /* Use ordered for everything but NE (which is usual for
> + * float comparisons)
> + */
> + switch (emit_data->inst->Instruction.Opcode) {
> + case TGSI_OPCODE_DSEQ: pred = LLVMRealOEQ; break;
> + case TGSI_OPCODE_DSGE: pred = LLVMRealOGE; break;
> + case TGSI_OPCODE_DSLT: pred = LLVMRealOLT; break;
> + case TGSI_OPCODE_DSNE: pred = LLVMRealUNE; break;
> + default: assert(!"unknown instruction"); pred = 0; break;
> + }
> +
> + LLVMValueRef v = LLVMBuildFCmp(builder, pred,
> + emit_data->args[0], emit_data->args[1],"");
> +
> + v = LLVMBuildSExtOrBitCast(builder, v,
> + LLVMInt32TypeInContext(context), "");
> +
> + emit_data->output[emit_data->chan] = v;
> +}
> +
> +static void emit_not(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + LLVMValueRef v = bitcast(bld_base, TGSI_TYPE_UNSIGNED,
> + emit_data->args[0]);
> + emit_data->output[emit_data->chan] = LLVMBuildNot(builder, v, "");
> +}
> +
> +static void emit_arl(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + LLVMValueRef floor_index = lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_FLR, emit_data->args[0]);
> + emit_data->output[emit_data->chan] = LLVMBuildFPToSI(builder,
> + floor_index, bld_base->base.int_elem_type , "");
> +}
> +
> +static void emit_and(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + emit_data->output[emit_data->chan] = LLVMBuildAnd(builder,
> + emit_data->args[0], emit_data->args[1], "");
> +}
> +
> +static void emit_or(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + emit_data->output[emit_data->chan] = LLVMBuildOr(builder,
> + emit_data->args[0], emit_data->args[1], "");
> +}
> +
> +static void emit_uadd(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + emit_data->output[emit_data->chan] = LLVMBuildAdd(builder,
> + emit_data->args[0], emit_data->args[1], "");
> +}
> +
> +static void emit_udiv(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + emit_data->output[emit_data->chan] = LLVMBuildUDiv(builder,
> + emit_data->args[0], emit_data->args[1], "");
> +}
> +
> +static void emit_idiv(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + emit_data->output[emit_data->chan] = LLVMBuildSDiv(builder,
> + emit_data->args[0], emit_data->args[1], "");
> +}
> +
> +static void emit_mod(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + emit_data->output[emit_data->chan] = LLVMBuildSRem(builder,
> + emit_data->args[0], emit_data->args[1], "");
> +}
> +
> +static void emit_umod(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + emit_data->output[emit_data->chan] = LLVMBuildURem(builder,
> + emit_data->args[0], emit_data->args[1], "");
> +}
> +
> +static void emit_shl(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + emit_data->output[emit_data->chan] = LLVMBuildShl(builder,
> + emit_data->args[0], emit_data->args[1], "");
> +}
> +
> +static void emit_ushr(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + emit_data->output[emit_data->chan] = LLVMBuildLShr(builder,
> + emit_data->args[0], emit_data->args[1], "");
> +}
> +static void emit_ishr(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + emit_data->output[emit_data->chan] = LLVMBuildAShr(builder,
> + emit_data->args[0], emit_data->args[1], "");
> +}
> +
> +static void emit_xor(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + emit_data->output[emit_data->chan] = LLVMBuildXor(builder,
> + emit_data->args[0], emit_data->args[1], "");
> +}
> +
> +static void emit_ssg(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> +
> + LLVMValueRef cmp, val;
> +
> + if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_I64SSG) {
> + cmp = LLVMBuildICmp(builder, LLVMIntSGT, emit_data->args[0], bld_base->int64_bld.zero, "");
> + val = LLVMBuildSelect(builder, cmp, bld_base->int64_bld.one, emit_data->args[0], "");
> + cmp = LLVMBuildICmp(builder, LLVMIntSGE, val, bld_base->int64_bld.zero, "");
> + val = LLVMBuildSelect(builder, cmp, val, LLVMConstInt(bld_base->int64_bld.elem_type, -1, true), "");
> + } else if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_ISSG) {
> + cmp = LLVMBuildICmp(builder, LLVMIntSGT, emit_data->args[0], bld_base->int_bld.zero, "");
> + val = LLVMBuildSelect(builder, cmp, bld_base->int_bld.one, emit_data->args[0], "");
> + cmp = LLVMBuildICmp(builder, LLVMIntSGE, val, bld_base->int_bld.zero, "");
> + val = LLVMBuildSelect(builder, cmp, val, LLVMConstInt(bld_base->int_bld.elem_type, -1, true), "");
> + } else { // float SSG
> + cmp = LLVMBuildFCmp(builder, LLVMRealOGT, emit_data->args[0], bld_base->base.zero, "");
> + val = LLVMBuildSelect(builder, cmp, bld_base->base.one, emit_data->args[0], "");
> + cmp = LLVMBuildFCmp(builder, LLVMRealOGE, val, bld_base->base.zero, "");
> + val = LLVMBuildSelect(builder, cmp, val, LLVMConstReal(bld_base->base.elem_type, -1), "");
> + }
> +
> + emit_data->output[emit_data->chan] = val;
> +}
> +
> +static void emit_ineg(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + emit_data->output[emit_data->chan] = LLVMBuildNeg(builder,
> + emit_data->args[0], "");
> +}
> +
> +static void emit_dneg(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + emit_data->output[emit_data->chan] = LLVMBuildFNeg(builder,
> + emit_data->args[0], "");
> +}
> +
> +static void emit_frac(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + char *intr;
> +
> + if (emit_data->info->opcode == TGSI_OPCODE_FRC)
> + intr = "llvm.floor.f32";
> + else if (emit_data->info->opcode == TGSI_OPCODE_DFRAC)
> + intr = "llvm.floor.f64";
> + else {
> + assert(0);
> + return;
> + }
> +
> + LLVMValueRef floor = lp_build_intrinsic(builder, intr, emit_data->dst_type,
> + &emit_data->args[0], 1,
> + LLVMReadNoneAttribute);
> + emit_data->output[emit_data->chan] = LLVMBuildFSub(builder,
> + emit_data->args[0], floor, "");
> +}
> +
> +static void emit_f2i(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + emit_data->output[emit_data->chan] = LLVMBuildFPToSI(builder,
> + emit_data->args[0], bld_base->int_bld.elem_type, "");
> +}
> +
> +static void emit_f2u(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + emit_data->output[emit_data->chan] = LLVMBuildFPToUI(builder,
> + emit_data->args[0], bld_base->uint_bld.elem_type, "");
> +}
> +
> +static void emit_i2f(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + emit_data->output[emit_data->chan] = LLVMBuildSIToFP(builder,
> + emit_data->args[0], bld_base->base.elem_type, "");
> +}
> +
> +static void emit_u2f(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + emit_data->output[emit_data->chan] = LLVMBuildUIToFP(builder,
> + emit_data->args[0], bld_base->base.elem_type, "");
> +}
> +
> +static void
> +build_tgsi_intrinsic_nomem(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + struct lp_build_context *base = &bld_base->base;
> + emit_data->output[emit_data->chan] =
> + lp_build_intrinsic(base->gallivm->builder, action->intr_name,
> + emit_data->dst_type, emit_data->args,
> + emit_data->arg_count, LLVMReadNoneAttribute);
> +}
> +
> +static void emit_bfi(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + struct gallivm_state *gallivm = bld_base->base.gallivm;
> + LLVMBuilderRef builder = gallivm->builder;
> + LLVMValueRef bfi_args[3];
> +
> + // Calculate the bitmask: (((1 << src3) - 1) << src2
> + bfi_args[0] = LLVMBuildShl(builder,
> + LLVMBuildSub(builder,
> + LLVMBuildShl(builder,
> + bld_base->int_bld.one,
> + emit_data->args[3], ""),
> + bld_base->int_bld.one, ""),
> + emit_data->args[2], "");
> +
> + bfi_args[1] = LLVMBuildShl(builder, emit_data->args[1],
> + emit_data->args[2], "");
> +
> + bfi_args[2] = emit_data->args[0];
> +
> + /* Calculate:
> + * (arg0 & arg1) | (~arg0 & arg2) = arg2 ^ (arg0 & (arg1 ^ arg2)
> + * Use the right-hand side, which the LLVM backend can convert to V_BFI.
> + */
> + emit_data->output[emit_data->chan] =
> + LLVMBuildXor(builder, bfi_args[2],
> + LLVMBuildAnd(builder, bfi_args[0],
> + LLVMBuildXor(builder, bfi_args[1], bfi_args[2],
> + ""), ""), "");
> +}
> +
> +/* this is ffs in C */
> +static void emit_lsb(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + struct gallivm_state *gallivm = bld_base->base.gallivm;
> + LLVMValueRef args[2] = {
> + emit_data->args[0],
> +
> + /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
> + * add special code to check for x=0. The reason is that
> + * the LLVM behavior for x=0 is different from what we
> + * need here.
> + *
> + * The hardware already implements the correct behavior.
> + */
> + LLVMConstInt(LLVMInt1TypeInContext(gallivm->context), 1, 0)
> + };
> +
> + emit_data->output[emit_data->chan] =
> + lp_build_intrinsic(gallivm->builder, "llvm.cttz.i32",
> + emit_data->dst_type, args, ARRAY_SIZE(args),
> + LLVMReadNoneAttribute);
> +}
> +
> +/* Find the last bit set. */
> +static void emit_umsb(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + struct gallivm_state *gallivm = bld_base->base.gallivm;
> + LLVMBuilderRef builder = gallivm->builder;
> + LLVMValueRef args[2] = {
> + emit_data->args[0],
> + /* Don't generate code for handling zero: */
> + LLVMConstInt(LLVMInt1TypeInContext(gallivm->context), 1, 0)
> + };
> +
> + LLVMValueRef msb =
> + lp_build_intrinsic(builder, "llvm.ctlz.i32",
> + emit_data->dst_type, args, ARRAY_SIZE(args),
> + LLVMReadNoneAttribute);
> +
> + /* The HW returns the last bit index from MSB, but TGSI wants
> + * the index from LSB. Invert it by doing "31 - msb". */
> + msb = LLVMBuildSub(builder, lp_build_const_int32(gallivm, 31),
> + msb, "");
> +
> + /* Check for zero: */
> + emit_data->output[emit_data->chan] =
> + LLVMBuildSelect(builder,
> + LLVMBuildICmp(builder, LLVMIntEQ, args[0],
> + bld_base->uint_bld.zero, ""),
> + lp_build_const_int32(gallivm, -1), msb, "");
> +}
> +
> +/* Find the last bit opposite of the sign bit. */
> +static void emit_imsb(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + struct gallivm_state *gallivm = bld_base->base.gallivm;
> + LLVMBuilderRef builder = gallivm->builder;
> + LLVMValueRef arg = emit_data->args[0];
> +
> + LLVMValueRef msb =
> + lp_build_intrinsic(builder, "llvm.AMDGPU.flbit.i32",
> + emit_data->dst_type, &arg, 1,
> + LLVMReadNoneAttribute);
> +
> + /* The HW returns the last bit index from MSB, but TGSI wants
> + * the index from LSB. Invert it by doing "31 - msb". */
> + msb = LLVMBuildSub(builder, lp_build_const_int32(gallivm, 31),
> + msb, "");
> +
> + /* If arg == 0 || arg == -1 (0xffffffff), return -1. */
> + LLVMValueRef all_ones = lp_build_const_int32(gallivm, -1);
> +
> + LLVMValueRef cond =
> + LLVMBuildOr(builder,
> + LLVMBuildICmp(builder, LLVMIntEQ, arg,
> + bld_base->uint_bld.zero, ""),
> + LLVMBuildICmp(builder, LLVMIntEQ, arg,
> + all_ones, ""), "");
> +
> + emit_data->output[emit_data->chan] =
> + LLVMBuildSelect(builder, cond, all_ones, msb, "");
> +}
> +
> +static void emit_iabs(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> +
> + emit_data->output[emit_data->chan] =
> + lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_IMAX,
> + emit_data->args[0],
> + LLVMBuildNeg(builder,
> + emit_data->args[0], ""));
> +}
> +
> +static void emit_minmax_int(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + LLVMIntPredicate op;
> +
> + switch (emit_data->info->opcode) {
> + default:
> + assert(0);
> + case TGSI_OPCODE_IMAX:
> + case TGSI_OPCODE_I64MAX:
> + op = LLVMIntSGT;
> + break;
> + case TGSI_OPCODE_IMIN:
> + case TGSI_OPCODE_I64MIN:
> + op = LLVMIntSLT;
> + break;
> + case TGSI_OPCODE_UMAX:
> + case TGSI_OPCODE_U64MAX:
> + op = LLVMIntUGT;
> + break;
> + case TGSI_OPCODE_UMIN:
> + case TGSI_OPCODE_U64MIN:
> + op = LLVMIntULT;
> + break;
> + }
> +
> + emit_data->output[emit_data->chan] =
> + LLVMBuildSelect(builder,
> + LLVMBuildICmp(builder, op, emit_data->args[0],
> + emit_data->args[1], ""),
> + emit_data->args[0],
> + emit_data->args[1], "");
> +}
> +
> +static void pk2h_fetch_args(struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst,
> + 0, TGSI_CHAN_X);
> + emit_data->args[1] = lp_build_emit_fetch(bld_base, emit_data->inst,
> + 0, TGSI_CHAN_Y);
> +}
> +
> +static void emit_pk2h(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + LLVMContextRef context = bld_base->base.gallivm->context;
> + struct lp_build_context *uint_bld = &bld_base->uint_bld;
> + LLVMTypeRef fp16, i16;
> + LLVMValueRef const16, comp[2];
> + unsigned i;
> +
> + fp16 = LLVMHalfTypeInContext(context);
> + i16 = LLVMInt16TypeInContext(context);
> + const16 = lp_build_const_int32(uint_bld->gallivm, 16);
> +
> + for (i = 0; i < 2; i++) {
> + comp[i] = LLVMBuildFPTrunc(builder, emit_data->args[i], fp16, "");
> + comp[i] = LLVMBuildBitCast(builder, comp[i], i16, "");
> + comp[i] = LLVMBuildZExt(builder, comp[i], uint_bld->elem_type, "");
> + }
> +
> + comp[1] = LLVMBuildShl(builder, comp[1], const16, "");
> + comp[0] = LLVMBuildOr(builder, comp[0], comp[1], "");
> +
> + emit_data->output[emit_data->chan] = comp[0];
> +}
> +
> +static void up2h_fetch_args(struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst,
> + 0, TGSI_CHAN_X);
> +}
> +
> +static void emit_up2h(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMBuilderRef builder = bld_base->base.gallivm->builder;
> + LLVMContextRef context = bld_base->base.gallivm->context;
> + struct lp_build_context *uint_bld = &bld_base->uint_bld;
> + LLVMTypeRef fp16, i16;
> + LLVMValueRef const16, input, val;
> + unsigned i;
> +
> + fp16 = LLVMHalfTypeInContext(context);
> + i16 = LLVMInt16TypeInContext(context);
> + const16 = lp_build_const_int32(uint_bld->gallivm, 16);
> + input = emit_data->args[0];
> +
> + for (i = 0; i < 2; i++) {
> + val = i == 1 ? LLVMBuildLShr(builder, input, const16, "") : input;
> + val = LLVMBuildTrunc(builder, val, i16, "");
> + val = LLVMBuildBitCast(builder, val, fp16, "");
> + emit_data->output[i] =
> + LLVMBuildFPExt(builder, val, bld_base->base.elem_type, "");
> + }
> +}
> +
> +static void emit_fdiv(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
> +
> + emit_data->output[emit_data->chan] =
> + LLVMBuildFDiv(bld_base->base.gallivm->builder,
> + emit_data->args[0], emit_data->args[1], "");
> +
> + /* Use v_rcp_f32 instead of precise division. */
> + if (HAVE_LLVM >= 0x0309 &&
> + !LLVMIsConstant(emit_data->output[emit_data->chan]))
> + LLVMSetMetadata(emit_data->output[emit_data->chan],
> + ctx->fpmath_md_kind, ctx->fpmath_md_2p5_ulp);
> +}
> +
> +/* 1/sqrt is translated to rsq for f32 if fp32 denormals are not enabled in
> + * the target machine. f64 needs global unsafe math flags to get rsq. */
> +static void emit_rsq(const struct lp_build_tgsi_action *action,
> + struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data)
> +{
> + LLVMValueRef sqrt =
> + lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_SQRT,
> + emit_data->args[0]);
> +
> + emit_data->output[emit_data->chan] =
> + lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_DIV,
> + bld_base->base.one, sqrt);
> +}
> +
> +void si_shader_context_init_alu(struct lp_build_tgsi_context *bld_base)
> +{
> + lp_set_default_actions(bld_base);
> +
> + bld_base->op_actions[TGSI_OPCODE_ABS].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_ABS].intr_name = "llvm.fabs.f32";
> + bld_base->op_actions[TGSI_OPCODE_AND].emit = emit_and;
> + bld_base->op_actions[TGSI_OPCODE_ARL].emit = emit_arl;
> + bld_base->op_actions[TGSI_OPCODE_BFI].emit = emit_bfi;
> + bld_base->op_actions[TGSI_OPCODE_BREV].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_BREV].intr_name =
> + HAVE_LLVM >= 0x0308 ? "llvm.bitreverse.i32" : "llvm.AMDGPU.brev";
> + bld_base->op_actions[TGSI_OPCODE_CEIL].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_CEIL].intr_name = "llvm.ceil.f32";
> + bld_base->op_actions[TGSI_OPCODE_CLAMP].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_CLAMP].intr_name =
> + HAVE_LLVM >= 0x0308 ? "llvm.AMDGPU.clamp." : "llvm.AMDIL.clamp.";
> + bld_base->op_actions[TGSI_OPCODE_CMP].emit = emit_cmp;
> + bld_base->op_actions[TGSI_OPCODE_COS].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_COS].intr_name = "llvm.cos.f32";
> + bld_base->op_actions[TGSI_OPCODE_DABS].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_DABS].intr_name = "llvm.fabs.f64";
> + bld_base->op_actions[TGSI_OPCODE_DFMA].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_DFMA].intr_name = "llvm.fma.f64";
> + bld_base->op_actions[TGSI_OPCODE_DFRAC].emit = emit_frac;
> + bld_base->op_actions[TGSI_OPCODE_DIV].emit = emit_fdiv;
> + bld_base->op_actions[TGSI_OPCODE_DNEG].emit = emit_dneg;
> + bld_base->op_actions[TGSI_OPCODE_DSEQ].emit = emit_dcmp;
> + bld_base->op_actions[TGSI_OPCODE_DSGE].emit = emit_dcmp;
> + bld_base->op_actions[TGSI_OPCODE_DSLT].emit = emit_dcmp;
> + bld_base->op_actions[TGSI_OPCODE_DSNE].emit = emit_dcmp;
> + bld_base->op_actions[TGSI_OPCODE_DRSQ].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_DRSQ].intr_name =
> + HAVE_LLVM >= 0x0309 ? "llvm.amdgcn.rsq.f64" : "llvm.AMDGPU.rsq.f64";
> + bld_base->op_actions[TGSI_OPCODE_DSQRT].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_DSQRT].intr_name = "llvm.sqrt.f64";
> + bld_base->op_actions[TGSI_OPCODE_EX2].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_EX2].intr_name =
> + HAVE_LLVM >= 0x0308 ? "llvm.exp2.f32" : "llvm.AMDIL.exp.";
> + bld_base->op_actions[TGSI_OPCODE_FLR].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_FLR].intr_name = "llvm.floor.f32";
> + bld_base->op_actions[TGSI_OPCODE_FMA].emit =
> + bld_base->op_actions[TGSI_OPCODE_MAD].emit;
> + bld_base->op_actions[TGSI_OPCODE_FRC].emit = emit_frac;
> + bld_base->op_actions[TGSI_OPCODE_F2I].emit = emit_f2i;
> + bld_base->op_actions[TGSI_OPCODE_F2U].emit = emit_f2u;
> + bld_base->op_actions[TGSI_OPCODE_FSEQ].emit = emit_fcmp;
> + bld_base->op_actions[TGSI_OPCODE_FSGE].emit = emit_fcmp;
> + bld_base->op_actions[TGSI_OPCODE_FSLT].emit = emit_fcmp;
> + bld_base->op_actions[TGSI_OPCODE_FSNE].emit = emit_fcmp;
> + bld_base->op_actions[TGSI_OPCODE_IABS].emit = emit_iabs;
> + bld_base->op_actions[TGSI_OPCODE_IBFE].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_IBFE].intr_name = "llvm.AMDGPU.bfe.i32";
> + bld_base->op_actions[TGSI_OPCODE_IDIV].emit = emit_idiv;
> + bld_base->op_actions[TGSI_OPCODE_IMAX].emit = emit_minmax_int;
> + bld_base->op_actions[TGSI_OPCODE_IMIN].emit = emit_minmax_int;
> + bld_base->op_actions[TGSI_OPCODE_IMSB].emit = emit_imsb;
> + bld_base->op_actions[TGSI_OPCODE_INEG].emit = emit_ineg;
> + bld_base->op_actions[TGSI_OPCODE_ISHR].emit = emit_ishr;
> + bld_base->op_actions[TGSI_OPCODE_ISGE].emit = emit_icmp;
> + bld_base->op_actions[TGSI_OPCODE_ISLT].emit = emit_icmp;
> + bld_base->op_actions[TGSI_OPCODE_ISSG].emit = emit_ssg;
> + bld_base->op_actions[TGSI_OPCODE_I2F].emit = emit_i2f;
> + bld_base->op_actions[TGSI_OPCODE_KILL_IF].fetch_args = kill_if_fetch_args;
> + bld_base->op_actions[TGSI_OPCODE_KILL_IF].emit = kil_emit;
> + bld_base->op_actions[TGSI_OPCODE_KILL_IF].intr_name = "llvm.AMDGPU.kill";
> + bld_base->op_actions[TGSI_OPCODE_KILL].emit = lp_build_tgsi_intrinsic;
> + bld_base->op_actions[TGSI_OPCODE_KILL].intr_name = "llvm.AMDGPU.kilp";
> + bld_base->op_actions[TGSI_OPCODE_LSB].emit = emit_lsb;
> + bld_base->op_actions[TGSI_OPCODE_LG2].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_LG2].intr_name = "llvm.log2.f32";
> + bld_base->op_actions[TGSI_OPCODE_MAX].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_MAX].intr_name = "llvm.maxnum.f32";
> + bld_base->op_actions[TGSI_OPCODE_MIN].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_MIN].intr_name = "llvm.minnum.f32";
> + bld_base->op_actions[TGSI_OPCODE_MOD].emit = emit_mod;
> + bld_base->op_actions[TGSI_OPCODE_UMSB].emit = emit_umsb;
> + bld_base->op_actions[TGSI_OPCODE_NOT].emit = emit_not;
> + bld_base->op_actions[TGSI_OPCODE_OR].emit = emit_or;
> + bld_base->op_actions[TGSI_OPCODE_PK2H].fetch_args = pk2h_fetch_args;
> + bld_base->op_actions[TGSI_OPCODE_PK2H].emit = emit_pk2h;
> + bld_base->op_actions[TGSI_OPCODE_POPC].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_POPC].intr_name = "llvm.ctpop.i32";
> + bld_base->op_actions[TGSI_OPCODE_POW].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_POW].intr_name = "llvm.pow.f32";
> + bld_base->op_actions[TGSI_OPCODE_ROUND].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_ROUND].intr_name = "llvm.rint.f32";
> + bld_base->op_actions[TGSI_OPCODE_RSQ].emit = emit_rsq;
> + bld_base->op_actions[TGSI_OPCODE_SGE].emit = emit_set_cond;
> + bld_base->op_actions[TGSI_OPCODE_SEQ].emit = emit_set_cond;
> + bld_base->op_actions[TGSI_OPCODE_SHL].emit = emit_shl;
> + bld_base->op_actions[TGSI_OPCODE_SLE].emit = emit_set_cond;
> + bld_base->op_actions[TGSI_OPCODE_SLT].emit = emit_set_cond;
> + bld_base->op_actions[TGSI_OPCODE_SNE].emit = emit_set_cond;
> + bld_base->op_actions[TGSI_OPCODE_SGT].emit = emit_set_cond;
> + bld_base->op_actions[TGSI_OPCODE_SIN].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_SIN].intr_name = "llvm.sin.f32";
> + bld_base->op_actions[TGSI_OPCODE_SQRT].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_SQRT].intr_name = "llvm.sqrt.f32";
> + bld_base->op_actions[TGSI_OPCODE_SSG].emit = emit_ssg;
> + bld_base->op_actions[TGSI_OPCODE_TRUNC].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_TRUNC].intr_name = "llvm.trunc.f32";
> + bld_base->op_actions[TGSI_OPCODE_UADD].emit = emit_uadd;
> + bld_base->op_actions[TGSI_OPCODE_UBFE].emit = build_tgsi_intrinsic_nomem;
> + bld_base->op_actions[TGSI_OPCODE_UBFE].intr_name = "llvm.AMDGPU.bfe.u32";
> + bld_base->op_actions[TGSI_OPCODE_UDIV].emit = emit_udiv;
> + bld_base->op_actions[TGSI_OPCODE_UMAX].emit = emit_minmax_int;
> + bld_base->op_actions[TGSI_OPCODE_UMIN].emit = emit_minmax_int;
> + bld_base->op_actions[TGSI_OPCODE_UMOD].emit = emit_umod;
> + bld_base->op_actions[TGSI_OPCODE_USEQ].emit = emit_icmp;
> + bld_base->op_actions[TGSI_OPCODE_USGE].emit = emit_icmp;
> + bld_base->op_actions[TGSI_OPCODE_USHR].emit = emit_ushr;
> + bld_base->op_actions[TGSI_OPCODE_USLT].emit = emit_icmp;
> + bld_base->op_actions[TGSI_OPCODE_USNE].emit = emit_icmp;
> + bld_base->op_actions[TGSI_OPCODE_U2F].emit = emit_u2f;
> + bld_base->op_actions[TGSI_OPCODE_XOR].emit = emit_xor;
> + bld_base->op_actions[TGSI_OPCODE_UCMP].emit = emit_ucmp;
> + bld_base->op_actions[TGSI_OPCODE_UP2H].fetch_args = up2h_fetch_args;
> + bld_base->op_actions[TGSI_OPCODE_UP2H].emit = emit_up2h;
> +
> + bld_base->op_actions[TGSI_OPCODE_I64MAX].emit = emit_minmax_int;
> + bld_base->op_actions[TGSI_OPCODE_I64MIN].emit = emit_minmax_int;
> + bld_base->op_actions[TGSI_OPCODE_U64MAX].emit = emit_minmax_int;
> + bld_base->op_actions[TGSI_OPCODE_U64MIN].emit = emit_minmax_int;
> + bld_base->op_actions[TGSI_OPCODE_I64ABS].emit = emit_iabs;
> + bld_base->op_actions[TGSI_OPCODE_I64SSG].emit = emit_ssg;
> + bld_base->op_actions[TGSI_OPCODE_I64NEG].emit = emit_ineg;
> +
> + bld_base->op_actions[TGSI_OPCODE_U64SEQ].emit = emit_icmp;
> + bld_base->op_actions[TGSI_OPCODE_U64SNE].emit = emit_icmp;
> + bld_base->op_actions[TGSI_OPCODE_U64SGE].emit = emit_icmp;
> + bld_base->op_actions[TGSI_OPCODE_U64SLT].emit = emit_icmp;
> + bld_base->op_actions[TGSI_OPCODE_I64SGE].emit = emit_icmp;
> + bld_base->op_actions[TGSI_OPCODE_I64SLT].emit = emit_icmp;
> +
> + bld_base->op_actions[TGSI_OPCODE_U64ADD].emit = emit_uadd;
> + bld_base->op_actions[TGSI_OPCODE_U64SHL].emit = emit_shl;
> + bld_base->op_actions[TGSI_OPCODE_U64SHR].emit = emit_ushr;
> + bld_base->op_actions[TGSI_OPCODE_I64SHR].emit = emit_ishr;
> +
> + bld_base->op_actions[TGSI_OPCODE_U64MOD].emit = emit_umod;
> + bld_base->op_actions[TGSI_OPCODE_I64MOD].emit = emit_mod;
> + bld_base->op_actions[TGSI_OPCODE_U64DIV].emit = emit_udiv;
> + bld_base->op_actions[TGSI_OPCODE_I64DIV].emit = emit_idiv;
> +}
> +
> +static LLVMValueRef build_cube_intrinsic(struct gallivm_state *gallivm,
> + LLVMValueRef in[3])
> +{
> + if (HAVE_LLVM >= 0x0309) {
> + LLVMTypeRef f32 = LLVMTypeOf(in[0]);
> + LLVMValueRef out[4];
> +
> + out[0] = lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.cubetc",
> + f32, in, 3, LLVMReadNoneAttribute);
> + out[1] = lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.cubesc",
> + f32, in, 3, LLVMReadNoneAttribute);
> + out[2] = lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.cubema",
> + f32, in, 3, LLVMReadNoneAttribute);
> + out[3] = lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.cubeid",
> + f32, in, 3, LLVMReadNoneAttribute);
> +
> + return lp_build_gather_values(gallivm, out, 4);
> + } else {
> + LLVMValueRef c[4] = {
> + in[0],
> + in[1],
> + in[2],
> + LLVMGetUndef(LLVMTypeOf(in[0]))
> + };
> + LLVMValueRef vec = lp_build_gather_values(gallivm, c, 4);
> +
> + return lp_build_intrinsic(gallivm->builder, "llvm.AMDGPU.cube",
> + LLVMTypeOf(vec), &vec, 1,
> + LLVMReadNoneAttribute);
> + }
> +}
> +
> +static void si_llvm_cube_to_2d_coords(struct lp_build_tgsi_context *bld_base,
> + LLVMValueRef *in, LLVMValueRef *out)
> +{
> + struct gallivm_state *gallivm = bld_base->base.gallivm;
> + LLVMBuilderRef builder = gallivm->builder;
> + LLVMTypeRef type = bld_base->base.elem_type;
> + LLVMValueRef coords[4];
> + LLVMValueRef mad_args[3];
> + LLVMValueRef v;
> + unsigned i;
> +
> + v = build_cube_intrinsic(gallivm, in);
> +
> + for (i = 0; i < 4; ++i)
> + coords[i] = LLVMBuildExtractElement(builder, v,
> + lp_build_const_int32(gallivm, i), "");
> +
> + coords[2] = lp_build_intrinsic(builder, "llvm.fabs.f32",
> + type, &coords[2], 1, LLVMReadNoneAttribute);
> + coords[2] = lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_RCP, coords[2]);
> +
> + mad_args[1] = coords[2];
> + mad_args[2] = LLVMConstReal(type, 1.5);
> +
> + mad_args[0] = coords[0];
> + coords[0] = lp_build_emit_llvm_ternary(bld_base, TGSI_OPCODE_MAD,
> + mad_args[0], mad_args[1], mad_args[2]);
> +
> + mad_args[0] = coords[1];
> + coords[1] = lp_build_emit_llvm_ternary(bld_base, TGSI_OPCODE_MAD,
> + mad_args[0], mad_args[1], mad_args[2]);
> +
> + /* apply xyz = yxw swizzle to cooords */
> + out[0] = coords[1];
> + out[1] = coords[0];
> + out[2] = coords[3];
> +}
> +
> +void si_prepare_cube_coords(struct lp_build_tgsi_context *bld_base,
> + struct lp_build_emit_data *emit_data,
> + LLVMValueRef *coords_arg,
> + LLVMValueRef *derivs_arg)
> +{
> +
> + unsigned target = emit_data->inst->Texture.Texture;
> + unsigned opcode = emit_data->inst->Instruction.Opcode;
> + struct gallivm_state *gallivm = bld_base->base.gallivm;
> + LLVMBuilderRef builder = gallivm->builder;
> + LLVMValueRef coords[4];
> + unsigned i;
> +
> + si_llvm_cube_to_2d_coords(bld_base, coords_arg, coords);
> +
> + if (opcode == TGSI_OPCODE_TXD && derivs_arg) {
> + LLVMValueRef derivs[4];
> + int axis;
> +
> + /* Convert cube derivatives to 2D derivatives. */
> + for (axis = 0; axis < 2; axis++) {
> + LLVMValueRef shifted_cube_coords[4], shifted_coords[4];
> +
> + /* Shift the cube coordinates by the derivatives to get
> + * the cube coordinates of the "neighboring pixel".
> + */
> + for (i = 0; i < 3; i++)
> + shifted_cube_coords[i] =
> + LLVMBuildFAdd(builder, coords_arg[i],
> + derivs_arg[axis*3+i], "");
> + shifted_cube_coords[3] = LLVMGetUndef(bld_base->base.elem_type);
> +
> + /* Project the shifted cube coordinates onto the face. */
> + si_llvm_cube_to_2d_coords(bld_base, shifted_cube_coords,
> + shifted_coords);
> +
> + /* Subtract both sets of 2D coordinates to get 2D derivatives.
> + * This won't work if the shifted coordinates ended up
> + * in a different face.
> + */
> + for (i = 0; i < 2; i++)
> + derivs[axis * 2 + i] =
> + LLVMBuildFSub(builder, shifted_coords[i],
> + coords[i], "");
> + }
> +
> + memcpy(derivs_arg, derivs, sizeof(derivs));
> + }
> +
> + if (target == TGSI_TEXTURE_CUBE_ARRAY ||
> + target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
> + /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
> + /* coords_arg.w component - array_index for cube arrays */
> + coords[2] = lp_build_emit_llvm_ternary(bld_base, TGSI_OPCODE_MAD,
> + coords_arg[3], lp_build_const_float(gallivm, 8.0), coords[2]);
> + }
> +
> + /* Preserve compare/lod/bias. Put it in coords.w. */
> + if (opcode == TGSI_OPCODE_TEX2 ||
> + opcode == TGSI_OPCODE_TXB2 ||
> + opcode == TGSI_OPCODE_TXL2) {
> + coords[3] = coords_arg[4];
> + } else if (opcode == TGSI_OPCODE_TXB ||
> + opcode == TGSI_OPCODE_TXL ||
> + target == TGSI_TEXTURE_SHADOWCUBE) {
> + coords[3] = coords_arg[3];
> + }
> +
> + memcpy(coords_arg, coords, sizeof(coords));
> +}
>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: OpenPGP digital signature
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20161018/62d628ec/attachment-0001.sig>
More information about the mesa-dev
mailing list