OSDN Git Service

radeon/ac: move common llvm build functions to a separate file.
authorDave Airlie <airlied@redhat.com>
Fri, 3 Feb 2017 00:05:00 +0000 (10:05 +1000)
committerDave Airlie <airlied@redhat.com>
Mon, 6 Feb 2017 19:46:35 +0000 (05:46 +1000)
Suggested by Marek.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Acked-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Signed-off-by: Dave Airlie <airlied@redhat.com>
src/amd/Makefile.sources
src/amd/common/ac_llvm_build.c [new file with mode: 0644]
src/amd/common/ac_llvm_build.h [new file with mode: 0644]
src/amd/common/ac_llvm_util.c
src/amd/common/ac_llvm_util.h
src/amd/common/ac_nir_to_llvm.c
src/gallium/drivers/radeonsi/si_shader_internal.h

index d981453..7aaa90a 100644 (file)
@@ -29,6 +29,8 @@ ADDRLIB_FILES = \
 AMD_COMPILER_FILES = \
        common/ac_binary.c \
        common/ac_binary.h \
+       common/ac_llvm_build.c \
+       common/ac_llvm_build.h \
        common/ac_llvm_helper.cpp \
        common/ac_llvm_util.c \
        common/ac_llvm_util.h
diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
new file mode 100644 (file)
index 0000000..afcbf31
--- /dev/null
@@ -0,0 +1,752 @@
+/*
+ * Copyright 2014 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 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 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 COPYRIGHT HOLDERS, AUTHORS AND/OR ITS 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.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ */
+/* based on pieces from si_pipe.c and radeon_llvm_emit.c */
+#include "ac_llvm_build.h"
+
+#include <llvm-c/Core.h>
+
+#include "c11/threads.h"
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "ac_llvm_util.h"
+
+#include "util/bitscan.h"
+#include "util/macros.h"
+#include "sid.h"
+
+/* Initialize module-independent parts of the context.
+ *
+ * The caller is responsible for initializing ctx::module and ctx::builder.
+ */
+void
+ac_llvm_context_init(struct ac_llvm_context *ctx, LLVMContextRef context)
+{
+       LLVMValueRef args[1];
+
+       ctx->context = context;
+       ctx->module = NULL;
+       ctx->builder = NULL;
+
+       ctx->voidt = LLVMVoidTypeInContext(ctx->context);
+       ctx->i1 = LLVMInt1TypeInContext(ctx->context);
+       ctx->i8 = LLVMInt8TypeInContext(ctx->context);
+       ctx->i32 = LLVMIntTypeInContext(ctx->context, 32);
+       ctx->f32 = LLVMFloatTypeInContext(ctx->context);
+       ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
+       ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
+       ctx->v16i8 = LLVMVectorType(ctx->i8, 16);
+
+       ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context,
+                                                    "range", 5);
+
+       ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context,
+                                                              "invariant.load", 14);
+
+       ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->context, "fpmath", 6);
+
+       args[0] = LLVMConstReal(ctx->f32, 2.5);
+       ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->context, args, 1);
+
+       ctx->uniform_md_kind = LLVMGetMDKindIDInContext(ctx->context,
+                                                       "amdgpu.uniform", 14);
+
+       ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
+}
+
+LLVMValueRef
+ac_emit_llvm_intrinsic(struct ac_llvm_context *ctx, const char *name,
+                      LLVMTypeRef return_type, LLVMValueRef *params,
+                      unsigned param_count, unsigned attrib_mask)
+{
+       LLVMValueRef function;
+
+       function = LLVMGetNamedFunction(ctx->module, name);
+       if (!function) {
+               LLVMTypeRef param_types[32], function_type;
+               unsigned i;
+
+               assert(param_count <= 32);
+
+               for (i = 0; i < param_count; ++i) {
+                       assert(params[i]);
+                       param_types[i] = LLVMTypeOf(params[i]);
+               }
+               function_type =
+                   LLVMFunctionType(return_type, param_types, param_count, 0);
+               function = LLVMAddFunction(ctx->module, name, function_type);
+
+               LLVMSetFunctionCallConv(function, LLVMCCallConv);
+               LLVMSetLinkage(function, LLVMExternalLinkage);
+
+               attrib_mask |= AC_FUNC_ATTR_NOUNWIND;
+               while (attrib_mask) {
+                       enum ac_func_attr attr = 1u << u_bit_scan(&attrib_mask);
+                       ac_add_function_attr(function, -1, attr);
+               }
+       }
+       return LLVMBuildCall(ctx->builder, function, params, param_count, "");
+}
+
+LLVMValueRef
+ac_build_gather_values_extended(struct ac_llvm_context *ctx,
+                               LLVMValueRef *values,
+                               unsigned value_count,
+                               unsigned value_stride,
+                               bool load)
+{
+       LLVMBuilderRef builder = ctx->builder;
+       LLVMValueRef vec;
+       unsigned i;
+
+
+       if (value_count == 1) {
+               if (load)
+                       return LLVMBuildLoad(builder, values[0], "");
+               return values[0];
+       } else if (!value_count)
+               unreachable("value_count is 0");
+
+       for (i = 0; i < value_count; i++) {
+               LLVMValueRef value = values[i * value_stride];
+               if (load)
+                       value = LLVMBuildLoad(builder, value, "");
+
+               if (!i)
+                       vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
+               LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
+               vec = LLVMBuildInsertElement(builder, vec, value, index, "");
+       }
+       return vec;
+}
+
+LLVMValueRef
+ac_build_gather_values(struct ac_llvm_context *ctx,
+                      LLVMValueRef *values,
+                      unsigned value_count)
+{
+       return ac_build_gather_values_extended(ctx, values, value_count, 1, false);
+}
+
+LLVMValueRef
+ac_emit_fdiv(struct ac_llvm_context *ctx,
+            LLVMValueRef num,
+            LLVMValueRef den)
+{
+       LLVMValueRef ret = LLVMBuildFDiv(ctx->builder, num, den, "");
+
+       if (!LLVMIsConstant(ret))
+               LLVMSetMetadata(ret, ctx->fpmath_md_kind, ctx->fpmath_md_2p5_ulp);
+       return ret;
+}
+
+/* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
+ * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
+ * already multiplied by two. id is the cube face number.
+ */
+struct cube_selection_coords {
+       LLVMValueRef stc[2];
+       LLVMValueRef ma;
+       LLVMValueRef id;
+};
+
+static void
+build_cube_intrinsic(struct ac_llvm_context *ctx,
+                    LLVMValueRef in[3],
+                    struct cube_selection_coords *out)
+{
+       LLVMBuilderRef builder = ctx->builder;
+
+       if (HAVE_LLVM >= 0x0309) {
+               LLVMTypeRef f32 = ctx->f32;
+
+               out->stc[1] = ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.cubetc",
+                                       f32, in, 3, AC_FUNC_ATTR_READNONE);
+               out->stc[0] = ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.cubesc",
+                                       f32, in, 3, AC_FUNC_ATTR_READNONE);
+               out->ma = ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.cubema",
+                                       f32, in, 3, AC_FUNC_ATTR_READNONE);
+               out->id = ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.cubeid",
+                                       f32, in, 3, AC_FUNC_ATTR_READNONE);
+       } else {
+               LLVMValueRef c[4] = {
+                       in[0],
+                       in[1],
+                       in[2],
+                       LLVMGetUndef(LLVMTypeOf(in[0]))
+               };
+               LLVMValueRef vec = ac_build_gather_values(ctx, c, 4);
+
+               LLVMValueRef tmp =
+                       ac_emit_llvm_intrinsic(ctx, "llvm.AMDGPU.cube",
+                                         LLVMTypeOf(vec), &vec, 1,
+                                         AC_FUNC_ATTR_READNONE);
+
+               out->stc[1] = LLVMBuildExtractElement(builder, tmp,
+                               LLVMConstInt(ctx->i32, 0, 0), "");
+               out->stc[0] = LLVMBuildExtractElement(builder, tmp,
+                               LLVMConstInt(ctx->i32, 1, 0), "");
+               out->ma = LLVMBuildExtractElement(builder, tmp,
+                               LLVMConstInt(ctx->i32, 2, 0), "");
+               out->id = LLVMBuildExtractElement(builder, tmp,
+                               LLVMConstInt(ctx->i32, 3, 0), "");
+       }
+}
+
+/**
+ * Build a manual selection sequence for cube face sc/tc coordinates and
+ * major axis vector (multiplied by 2 for consistency) for the given
+ * vec3 \p coords, for the face implied by \p selcoords.
+ *
+ * For the major axis, we always adjust the sign to be in the direction of
+ * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
+ * the selcoords major axis.
+ */
+static void build_cube_select(LLVMBuilderRef builder,
+                             const struct cube_selection_coords *selcoords,
+                             const LLVMValueRef *coords,
+                             LLVMValueRef *out_st,
+                             LLVMValueRef *out_ma)
+{
+       LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
+       LLVMValueRef is_ma_positive;
+       LLVMValueRef sgn_ma;
+       LLVMValueRef is_ma_z, is_not_ma_z;
+       LLVMValueRef is_ma_y;
+       LLVMValueRef is_ma_x;
+       LLVMValueRef sgn;
+       LLVMValueRef tmp;
+
+       is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
+               selcoords->ma, LLVMConstReal(f32, 0.0), "");
+       sgn_ma = LLVMBuildSelect(builder, is_ma_positive,
+               LLVMConstReal(f32, 1.0), LLVMConstReal(f32, -1.0), "");
+
+       is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
+       is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
+       is_ma_y = LLVMBuildAnd(builder, is_not_ma_z,
+               LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
+       is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
+
+       /* Select sc */
+       tmp = LLVMBuildSelect(builder, is_ma_z, coords[2], coords[0], "");
+       sgn = LLVMBuildSelect(builder, is_ma_y, LLVMConstReal(f32, 1.0),
+               LLVMBuildSelect(builder, is_ma_x, sgn_ma,
+                       LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
+       out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
+
+       /* Select tc */
+       tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
+       sgn = LLVMBuildSelect(builder, is_ma_y, LLVMBuildFNeg(builder, sgn_ma, ""),
+               LLVMConstReal(f32, -1.0), "");
+       out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
+
+       /* Select ma */
+       tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
+               LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
+       sgn = LLVMBuildSelect(builder, is_ma_positive,
+               LLVMConstReal(f32, 2.0), LLVMConstReal(f32, -2.0), "");
+       *out_ma = LLVMBuildFMul(builder, tmp, sgn, "");
+}
+
+void
+ac_prepare_cube_coords(struct ac_llvm_context *ctx,
+                      bool is_deriv, bool is_array,
+                      LLVMValueRef *coords_arg,
+                      LLVMValueRef *derivs_arg)
+{
+
+       LLVMBuilderRef builder = ctx->builder;
+       struct cube_selection_coords selcoords;
+       LLVMValueRef coords[3];
+       LLVMValueRef invma;
+
+       build_cube_intrinsic(ctx, coords_arg, &selcoords);
+
+       invma = ac_emit_llvm_intrinsic(ctx, "llvm.fabs.f32",
+                       ctx->f32, &selcoords.ma, 1, AC_FUNC_ATTR_READNONE);
+       invma = ac_emit_fdiv(ctx, LLVMConstReal(ctx->f32, 1.0), invma);
+
+       for (int i = 0; i < 2; ++i)
+               coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
+
+       coords[2] = selcoords.id;
+
+       if (is_deriv && derivs_arg) {
+               LLVMValueRef derivs[4];
+               int axis;
+
+               /* Convert cube derivatives to 2D derivatives. */
+               for (axis = 0; axis < 2; axis++) {
+                       LLVMValueRef deriv_st[2];
+                       LLVMValueRef deriv_ma;
+
+                       /* Transform the derivative alongside the texture
+                        * coordinate. Mathematically, the correct formula is
+                        * as follows. Assume we're projecting onto the +Z face
+                        * and denote by dx/dh the derivative of the (original)
+                        * X texture coordinate with respect to horizontal
+                        * window coordinates. The projection onto the +Z face
+                        * plane is:
+                        *
+                        *   f(x,z) = x/z
+                        *
+                        * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
+                        *            = 1/z * dx/dh - x/z * 1/z * dz/dh.
+                        *
+                        * This motivatives the implementation below.
+                        *
+                        * Whether this actually gives the expected results for
+                        * apps that might feed in derivatives obtained via
+                        * finite differences is anyone's guess. The OpenGL spec
+                        * seems awfully quiet about how textureGrad for cube
+                        * maps should be handled.
+                        */
+                       build_cube_select(builder, &selcoords, &derivs_arg[axis * 3],
+                                         deriv_st, &deriv_ma);
+
+                       deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
+
+                       for (int i = 0; i < 2; ++i)
+                               derivs[axis * 2 + i] =
+                                       LLVMBuildFSub(builder,
+                                               LLVMBuildFMul(builder, deriv_st[i], invma, ""),
+                                               LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
+               }
+
+               memcpy(derivs_arg, derivs, sizeof(derivs));
+       }
+
+       /* Shift the texture coordinate. This must be applied after the
+        * derivative calculation.
+        */
+       for (int i = 0; i < 2; ++i)
+               coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
+
+       if (is_array) {
+               /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
+               /* coords_arg.w component - array_index for cube arrays */
+               LLVMValueRef tmp = LLVMBuildFMul(ctx->builder, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), "");
+               coords[2] = LLVMBuildFAdd(ctx->builder, tmp, coords[2], "");
+       }
+
+       memcpy(coords_arg, coords, sizeof(coords));
+}
+
+
+LLVMValueRef
+ac_build_fs_interp(struct ac_llvm_context *ctx,
+                  LLVMValueRef llvm_chan,
+                  LLVMValueRef attr_number,
+                  LLVMValueRef params,
+                  LLVMValueRef i,
+                  LLVMValueRef j)
+{
+       LLVMValueRef args[5];
+       LLVMValueRef p1;
+       
+       if (HAVE_LLVM < 0x0400) {
+               LLVMValueRef ij[2];
+               ij[0] = LLVMBuildBitCast(ctx->builder, i, ctx->i32, "");
+               ij[1] = LLVMBuildBitCast(ctx->builder, j, ctx->i32, "");
+
+               args[0] = llvm_chan;
+               args[1] = attr_number;
+               args[2] = params;
+               args[3] = ac_build_gather_values(ctx, ij, 2);
+               return ac_emit_llvm_intrinsic(ctx, "llvm.SI.fs.interp",
+                                             ctx->f32, args, 4,
+                                             AC_FUNC_ATTR_READNONE);
+       }
+
+       args[0] = i;
+       args[1] = llvm_chan;
+       args[2] = attr_number;
+       args[3] = params;
+
+       p1 = ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.interp.p1",
+                                   ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
+
+       args[0] = p1;
+       args[1] = j;
+       args[2] = llvm_chan;
+       args[3] = attr_number;
+       args[4] = params;
+
+       return ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.interp.p2",
+                                     ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
+}
+
+LLVMValueRef
+ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
+                      LLVMValueRef parameter,
+                      LLVMValueRef llvm_chan,
+                      LLVMValueRef attr_number,
+                      LLVMValueRef params)
+{
+       LLVMValueRef args[4];
+       if (HAVE_LLVM < 0x0400) {
+               args[0] = llvm_chan;
+               args[1] = attr_number;
+               args[2] = params;
+
+               return ac_emit_llvm_intrinsic(ctx,
+                                             "llvm.SI.fs.constant",
+                                             ctx->f32, args, 3,
+                                             AC_FUNC_ATTR_READNONE);
+       }
+
+       args[0] = parameter;
+       args[1] = llvm_chan;
+       args[2] = attr_number;
+       args[3] = params;
+
+       return ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.interp.mov",
+                                     ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
+}
+
+LLVMValueRef
+ac_build_gep0(struct ac_llvm_context *ctx,
+             LLVMValueRef base_ptr,
+             LLVMValueRef index)
+{
+       LLVMValueRef indices[2] = {
+               LLVMConstInt(ctx->i32, 0, 0),
+               index,
+       };
+       return LLVMBuildGEP(ctx->builder, base_ptr,
+                           indices, 2, "");
+}
+
+void
+ac_build_indexed_store(struct ac_llvm_context *ctx,
+                      LLVMValueRef base_ptr, LLVMValueRef index,
+                      LLVMValueRef value)
+{
+       LLVMBuildStore(ctx->builder, value,
+                      ac_build_gep0(ctx, base_ptr, index));
+}
+
+/**
+ * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
+ * It's equivalent to doing a load from &base_ptr[index].
+ *
+ * \param base_ptr  Where the array starts.
+ * \param index     The element index into the array.
+ * \param uniform   Whether the base_ptr and index can be assumed to be
+ *                  dynamically uniform
+ */
+LLVMValueRef
+ac_build_indexed_load(struct ac_llvm_context *ctx,
+                     LLVMValueRef base_ptr, LLVMValueRef index,
+                     bool uniform)
+{
+       LLVMValueRef pointer;
+
+       pointer = ac_build_gep0(ctx, base_ptr, index);
+       if (uniform)
+               LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
+       return LLVMBuildLoad(ctx->builder, pointer, "");
+}
+
+/**
+ * Do a load from &base_ptr[index], but also add a flag that it's loading
+ * a constant from a dynamically uniform index.
+ */
+LLVMValueRef
+ac_build_indexed_load_const(struct ac_llvm_context *ctx,
+                           LLVMValueRef base_ptr, LLVMValueRef index)
+{
+       LLVMValueRef result = ac_build_indexed_load(ctx, base_ptr, index, true);
+       LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
+       return result;
+}
+
+/* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
+ * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
+ * or v4i32 (num_channels=3,4).
+ */
+void
+ac_build_tbuffer_store(struct ac_llvm_context *ctx,
+                      LLVMValueRef rsrc,
+                      LLVMValueRef vdata,
+                      unsigned num_channels,
+                      LLVMValueRef vaddr,
+                      LLVMValueRef soffset,
+                      unsigned inst_offset,
+                      unsigned dfmt,
+                      unsigned nfmt,
+                      unsigned offen,
+                      unsigned idxen,
+                      unsigned glc,
+                      unsigned slc,
+                      unsigned tfe)
+{
+       LLVMValueRef args[] = {
+               rsrc,
+               vdata,
+               LLVMConstInt(ctx->i32, num_channels, 0),
+               vaddr,
+               soffset,
+               LLVMConstInt(ctx->i32, inst_offset, 0),
+               LLVMConstInt(ctx->i32, dfmt, 0),
+               LLVMConstInt(ctx->i32, nfmt, 0),
+               LLVMConstInt(ctx->i32, offen, 0),
+               LLVMConstInt(ctx->i32, idxen, 0),
+               LLVMConstInt(ctx->i32, glc, 0),
+               LLVMConstInt(ctx->i32, slc, 0),
+               LLVMConstInt(ctx->i32, tfe, 0)
+       };
+
+       /* The instruction offset field has 12 bits */
+       assert(offen || inst_offset < (1 << 12));
+
+       /* The intrinsic is overloaded, we need to add a type suffix for overloading to work. */
+       unsigned func = CLAMP(num_channels, 1, 3) - 1;
+       const char *types[] = {"i32", "v2i32", "v4i32"};
+       char name[256];
+       snprintf(name, sizeof(name), "llvm.SI.tbuffer.store.%s", types[func]);
+
+       ac_emit_llvm_intrinsic(ctx, name, ctx->voidt,
+                              args, ARRAY_SIZE(args), 0);
+}
+
+void
+ac_build_tbuffer_store_dwords(struct ac_llvm_context *ctx,
+                             LLVMValueRef rsrc,
+                             LLVMValueRef vdata,
+                             unsigned num_channels,
+                             LLVMValueRef vaddr,
+                             LLVMValueRef soffset,
+                             unsigned inst_offset)
+{
+       static unsigned dfmt[] = {
+               V_008F0C_BUF_DATA_FORMAT_32,
+               V_008F0C_BUF_DATA_FORMAT_32_32,
+               V_008F0C_BUF_DATA_FORMAT_32_32_32,
+               V_008F0C_BUF_DATA_FORMAT_32_32_32_32
+       };
+       assert(num_channels >= 1 && num_channels <= 4);
+
+       ac_build_tbuffer_store(ctx, rsrc, vdata, num_channels, vaddr, soffset,
+                              inst_offset, dfmt[num_channels - 1],
+                              V_008F0C_BUF_NUM_FORMAT_UINT, 1, 0, 1, 1, 0);
+}
+
+LLVMValueRef
+ac_build_buffer_load(struct ac_llvm_context *ctx,
+                    LLVMValueRef rsrc,
+                    int num_channels,
+                    LLVMValueRef vindex,
+                    LLVMValueRef voffset,
+                    LLVMValueRef soffset,
+                    unsigned inst_offset,
+                    unsigned glc,
+                    unsigned slc)
+{
+       unsigned func = CLAMP(num_channels, 1, 3) - 1;
+
+       if (HAVE_LLVM >= 0x309) {
+               LLVMValueRef args[] = {
+                       LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
+                       vindex ? vindex : LLVMConstInt(ctx->i32, 0, 0),
+                       LLVMConstInt(ctx->i32, inst_offset, 0),
+                       LLVMConstInt(ctx->i1, glc, 0),
+                       LLVMConstInt(ctx->i1, slc, 0)
+               };
+
+               LLVMTypeRef types[] = {ctx->f32, LLVMVectorType(ctx->f32, 2),
+                                      ctx->v4f32};
+               const char *type_names[] = {"f32", "v2f32", "v4f32"};
+               char name[256];
+
+               if (voffset) {
+                       args[2] = LLVMBuildAdd(ctx->builder, args[2], voffset,
+                                              "");
+               }
+
+               if (soffset) {
+                       args[2] = LLVMBuildAdd(ctx->builder, args[2], soffset,
+                                              "");
+               }
+
+               snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.%s",
+                        type_names[func]);
+
+               return ac_emit_llvm_intrinsic(ctx, name, types[func], args,
+                                             ARRAY_SIZE(args), AC_FUNC_ATTR_READONLY);
+       } else {
+               LLVMValueRef args[] = {
+                       LLVMBuildBitCast(ctx->builder, rsrc, ctx->v16i8, ""),
+                       voffset ? voffset : vindex,
+                       soffset,
+                       LLVMConstInt(ctx->i32, inst_offset, 0),
+                       LLVMConstInt(ctx->i32, voffset ? 1 : 0, 0), // offen
+                       LLVMConstInt(ctx->i32, vindex ? 1 : 0, 0), //idxen
+                       LLVMConstInt(ctx->i32, glc, 0),
+                       LLVMConstInt(ctx->i32, slc, 0),
+                       LLVMConstInt(ctx->i32, 0, 0), // TFE
+               };
+
+               LLVMTypeRef types[] = {ctx->i32, LLVMVectorType(ctx->i32, 2),
+                                      ctx->v4i32};
+               const char *type_names[] = {"i32", "v2i32", "v4i32"};
+               const char *arg_type = "i32";
+               char name[256];
+
+               if (voffset && vindex) {
+                       LLVMValueRef vaddr[] = {vindex, voffset};
+
+                       arg_type = "v2i32";
+                       args[1] = ac_build_gather_values(ctx, vaddr, 2);
+               }
+
+               snprintf(name, sizeof(name), "llvm.SI.buffer.load.dword.%s.%s",
+                        type_names[func], arg_type);
+
+               return ac_emit_llvm_intrinsic(ctx, name, types[func], args,
+                                              ARRAY_SIZE(args), AC_FUNC_ATTR_READONLY);
+       }
+}
+
+/**
+ * Set range metadata on an instruction.  This can only be used on load and
+ * call instructions.  If you know an instruction can only produce the values
+ * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
+ * \p lo is the minimum value inclusive.
+ * \p hi is the maximum value exclusive.
+ */
+static void set_range_metadata(struct ac_llvm_context *ctx,
+                              LLVMValueRef value, unsigned lo, unsigned hi)
+{
+       LLVMValueRef range_md, md_args[2];
+       LLVMTypeRef type = LLVMTypeOf(value);
+       LLVMContextRef context = LLVMGetTypeContext(type);
+
+       md_args[0] = LLVMConstInt(type, lo, false);
+       md_args[1] = LLVMConstInt(type, hi, false);
+       range_md = LLVMMDNodeInContext(context, md_args, 2);
+       LLVMSetMetadata(value, ctx->range_md_kind, range_md);
+}
+
+LLVMValueRef
+ac_get_thread_id(struct ac_llvm_context *ctx)
+{
+       LLVMValueRef tid;
+
+       if (HAVE_LLVM < 0x0308) {
+               tid = ac_emit_llvm_intrinsic(ctx, "llvm.SI.tid",
+                                            ctx->i32,
+                                            NULL, 0, AC_FUNC_ATTR_READNONE);
+       } else {
+               LLVMValueRef tid_args[2];
+               tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
+               tid_args[1] = LLVMConstInt(ctx->i32, 0, false);
+               tid_args[1] = ac_emit_llvm_intrinsic(ctx,
+                                                    "llvm.amdgcn.mbcnt.lo", ctx->i32,
+                                                    tid_args, 2, AC_FUNC_ATTR_READNONE);
+
+               tid = ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
+                                            ctx->i32, tid_args,
+                                            2, AC_FUNC_ATTR_READNONE);
+       }
+       set_range_metadata(ctx, tid, 0, 64);
+       return tid;
+}
+
+/*
+ * SI implements derivatives using the local data store (LDS)
+ * All writes to the LDS happen in all executing threads at
+ * the same time. TID is the Thread ID for the current
+ * thread and is a value between 0 and 63, representing
+ * the thread's position in the wavefront.
+ *
+ * For the pixel shader threads are grouped into quads of four pixels.
+ * The TIDs of the pixels of a quad are:
+ *
+ *  +------+------+
+ *  |4n + 0|4n + 1|
+ *  +------+------+
+ *  |4n + 2|4n + 3|
+ *  +------+------+
+ *
+ * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
+ * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
+ * the current pixel's column, and masking with 0xfffffffe yields the TID
+ * of the left pixel of the current pixel's row.
+ *
+ * Adding 1 yields the TID of the pixel to the right of the left pixel, and
+ * adding 2 yields the TID of the pixel below the top pixel.
+ */
+LLVMValueRef
+ac_emit_ddxy(struct ac_llvm_context *ctx,
+            bool has_ds_bpermute,
+            uint32_t mask,
+            int idx,
+            LLVMValueRef lds,
+            LLVMValueRef val)
+{
+       LLVMValueRef thread_id, tl, trbl, tl_tid, trbl_tid, args[2];
+       LLVMValueRef result;
+
+       thread_id = ac_get_thread_id(ctx);
+
+       tl_tid = LLVMBuildAnd(ctx->builder, thread_id,
+                             LLVMConstInt(ctx->i32, mask, false), "");
+
+       trbl_tid = LLVMBuildAdd(ctx->builder, tl_tid,
+                               LLVMConstInt(ctx->i32, idx, false), "");
+
+       if (has_ds_bpermute) {
+               args[0] = LLVMBuildMul(ctx->builder, tl_tid,
+                                      LLVMConstInt(ctx->i32, 4, false), "");
+               args[1] = val;
+               tl = ac_emit_llvm_intrinsic(ctx,
+                                           "llvm.amdgcn.ds.bpermute", ctx->i32,
+                                           args, 2, AC_FUNC_ATTR_READNONE);
+
+               args[0] = LLVMBuildMul(ctx->builder, trbl_tid,
+                                      LLVMConstInt(ctx->i32, 4, false), "");
+               trbl = ac_emit_llvm_intrinsic(ctx,
+                                             "llvm.amdgcn.ds.bpermute", ctx->i32,
+                                             args, 2, AC_FUNC_ATTR_READNONE);
+       } else {
+               LLVMValueRef store_ptr, load_ptr0, load_ptr1;
+
+               store_ptr = ac_build_gep0(ctx, lds, thread_id);
+               load_ptr0 = ac_build_gep0(ctx, lds, tl_tid);
+               load_ptr1 = ac_build_gep0(ctx, lds, trbl_tid);
+
+               LLVMBuildStore(ctx->builder, val, store_ptr);
+               tl = LLVMBuildLoad(ctx->builder, load_ptr0, "");
+               trbl = LLVMBuildLoad(ctx->builder, load_ptr1, "");
+       }
+
+       tl = LLVMBuildBitCast(ctx->builder, tl, ctx->f32, "");
+       trbl = LLVMBuildBitCast(ctx->builder, trbl, ctx->f32, "");
+       result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
+       return result;
+}
diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
new file mode 100644 (file)
index 0000000..37a5bea
--- /dev/null
@@ -0,0 +1,177 @@
+/*
+ * Copyright 2016 Bas Nieuwenhuizen
+ *
+ * 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 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 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 COPYRIGHT HOLDERS, AUTHORS AND/OR ITS 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.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ */
+#ifndef AC_LLVM_BUILD_H
+#define AC_LLVM_BUILD_H
+
+#include <stdbool.h>
+#include <llvm-c/TargetMachine.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct ac_llvm_context {
+       LLVMContextRef context;
+       LLVMModuleRef module;
+       LLVMBuilderRef builder;
+
+       LLVMTypeRef voidt;
+       LLVMTypeRef i1;
+       LLVMTypeRef i8;
+       LLVMTypeRef i32;
+       LLVMTypeRef f32;
+       LLVMTypeRef v4i32;
+       LLVMTypeRef v4f32;
+       LLVMTypeRef v16i8;
+
+       unsigned range_md_kind;
+       unsigned invariant_load_md_kind;
+       unsigned uniform_md_kind;
+       unsigned fpmath_md_kind;
+       LLVMValueRef fpmath_md_2p5_ulp;
+       LLVMValueRef empty_md;
+};
+
+void
+ac_llvm_context_init(struct ac_llvm_context *ctx, LLVMContextRef context);
+
+LLVMValueRef
+ac_emit_llvm_intrinsic(struct ac_llvm_context *ctx, const char *name,
+                      LLVMTypeRef return_type, LLVMValueRef *params,
+                      unsigned param_count, unsigned attrib_mask);
+
+LLVMValueRef
+ac_build_gather_values_extended(struct ac_llvm_context *ctx,
+                               LLVMValueRef *values,
+                               unsigned value_count,
+                               unsigned value_stride,
+                               bool load);
+LLVMValueRef
+ac_build_gather_values(struct ac_llvm_context *ctx,
+                      LLVMValueRef *values,
+                      unsigned value_count);
+
+LLVMValueRef
+ac_emit_fdiv(struct ac_llvm_context *ctx,
+            LLVMValueRef num,
+            LLVMValueRef den);
+
+void
+ac_prepare_cube_coords(struct ac_llvm_context *ctx,
+                      bool is_deriv, bool is_array,
+                      LLVMValueRef *coords_arg,
+                      LLVMValueRef *derivs_arg);
+
+
+LLVMValueRef
+ac_build_fs_interp(struct ac_llvm_context *ctx,
+                  LLVMValueRef llvm_chan,
+                  LLVMValueRef attr_number,
+                  LLVMValueRef params,
+                  LLVMValueRef i,
+                  LLVMValueRef j);
+
+LLVMValueRef
+ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
+                      LLVMValueRef parameter,
+                      LLVMValueRef llvm_chan,
+                      LLVMValueRef attr_number,
+                      LLVMValueRef params);
+
+LLVMValueRef
+ac_build_gep0(struct ac_llvm_context *ctx,
+             LLVMValueRef base_ptr,
+             LLVMValueRef index);
+
+void
+ac_build_indexed_store(struct ac_llvm_context *ctx,
+                      LLVMValueRef base_ptr, LLVMValueRef index,
+                      LLVMValueRef value);
+
+LLVMValueRef
+ac_build_indexed_load(struct ac_llvm_context *ctx,
+                     LLVMValueRef base_ptr, LLVMValueRef index,
+                     bool uniform);
+
+LLVMValueRef
+ac_build_indexed_load_const(struct ac_llvm_context *ctx,
+                           LLVMValueRef base_ptr, LLVMValueRef index);
+
+void
+ac_build_tbuffer_store_dwords(struct ac_llvm_context *ctx,
+                             LLVMValueRef rsrc,
+                             LLVMValueRef vdata,
+                             unsigned num_channels,
+                             LLVMValueRef vaddr,
+                             LLVMValueRef soffset,
+                             unsigned inst_offset);
+
+void
+ac_build_tbuffer_store(struct ac_llvm_context *ctx,
+                      LLVMValueRef rsrc,
+                      LLVMValueRef vdata,
+                      unsigned num_channels,
+                      LLVMValueRef vaddr,
+                      LLVMValueRef soffset,
+                      unsigned inst_offset,
+                      unsigned dfmt,
+                      unsigned nfmt,
+                      unsigned offen,
+                      unsigned idxen,
+                      unsigned glc,
+                      unsigned slc,
+                      unsigned tfe);
+
+LLVMValueRef
+ac_build_buffer_load(struct ac_llvm_context *ctx,
+                    LLVMValueRef rsrc,
+                    int num_channels,
+                    LLVMValueRef vindex,
+                    LLVMValueRef voffset,
+                    LLVMValueRef soffset,
+                    unsigned inst_offset,
+                    unsigned glc,
+                    unsigned slc);
+
+LLVMValueRef
+ac_get_thread_id(struct ac_llvm_context *ctx);
+
+#define AC_TID_MASK_TOP_LEFT 0xfffffffc
+#define AC_TID_MASK_TOP      0xfffffffd
+#define AC_TID_MASK_LEFT     0xfffffffe
+
+LLVMValueRef
+ac_emit_ddxy(struct ac_llvm_context *ctx,
+            bool has_ds_bpermute,
+            uint32_t mask,
+            int idx,
+            LLVMValueRef lds,
+            LLVMValueRef val);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
index 156d874..be127c5 100644 (file)
 
 #include <assert.h>
 #include <stdio.h>
-
-#include "util/bitscan.h"
-#include "util/macros.h"
-
-#include "sid.h"
+#include <string.h>
 
 static void ac_init_llvm_target()
 {
@@ -146,44 +142,6 @@ LLVMTargetMachineRef ac_create_target_machine(enum radeon_family family, bool su
        return tm;
 }
 
-/* Initialize module-independent parts of the context.
- *
- * The caller is responsible for initializing ctx::module and ctx::builder.
- */
-void
-ac_llvm_context_init(struct ac_llvm_context *ctx, LLVMContextRef context)
-{
-       LLVMValueRef args[1];
-
-       ctx->context = context;
-       ctx->module = NULL;
-       ctx->builder = NULL;
-
-       ctx->voidt = LLVMVoidTypeInContext(ctx->context);
-       ctx->i1 = LLVMInt1TypeInContext(ctx->context);
-       ctx->i8 = LLVMInt8TypeInContext(ctx->context);
-       ctx->i32 = LLVMIntTypeInContext(ctx->context, 32);
-       ctx->f32 = LLVMFloatTypeInContext(ctx->context);
-       ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
-       ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
-       ctx->v16i8 = LLVMVectorType(ctx->i8, 16);
-
-       ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context,
-                                                    "range", 5);
-
-       ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context,
-                                                              "invariant.load", 14);
-
-       ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->context, "fpmath", 6);
-
-       args[0] = LLVMConstReal(ctx->f32, 2.5);
-       ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->context, args, 1);
-
-       ctx->uniform_md_kind = LLVMGetMDKindIDInContext(ctx->context,
-                                                       "amdgpu.uniform", 14);
-
-       ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
-}
 
 #if HAVE_LLVM < 0x0400
 static LLVMAttribute ac_attr_to_llvm_attr(enum ac_func_attr attr)
@@ -245,285 +203,6 @@ ac_add_function_attr(LLVMValueRef function,
 #endif
 }
 
-LLVMValueRef
-ac_emit_llvm_intrinsic(struct ac_llvm_context *ctx, const char *name,
-                      LLVMTypeRef return_type, LLVMValueRef *params,
-                      unsigned param_count, unsigned attrib_mask)
-{
-       LLVMValueRef function;
-
-       function = LLVMGetNamedFunction(ctx->module, name);
-       if (!function) {
-               LLVMTypeRef param_types[32], function_type;
-               unsigned i;
-
-               assert(param_count <= 32);
-
-               for (i = 0; i < param_count; ++i) {
-                       assert(params[i]);
-                       param_types[i] = LLVMTypeOf(params[i]);
-               }
-               function_type =
-                   LLVMFunctionType(return_type, param_types, param_count, 0);
-               function = LLVMAddFunction(ctx->module, name, function_type);
-
-               LLVMSetFunctionCallConv(function, LLVMCCallConv);
-               LLVMSetLinkage(function, LLVMExternalLinkage);
-
-               attrib_mask |= AC_FUNC_ATTR_NOUNWIND;
-               while (attrib_mask) {
-                       enum ac_func_attr attr = 1u << u_bit_scan(&attrib_mask);
-                       ac_add_function_attr(function, -1, attr);
-               }
-       }
-       return LLVMBuildCall(ctx->builder, function, params, param_count, "");
-}
-
-LLVMValueRef
-ac_build_gather_values_extended(struct ac_llvm_context *ctx,
-                               LLVMValueRef *values,
-                               unsigned value_count,
-                               unsigned value_stride,
-                               bool load)
-{
-       LLVMBuilderRef builder = ctx->builder;
-       LLVMValueRef vec;
-       unsigned i;
-
-
-       if (value_count == 1) {
-               if (load)
-                       return LLVMBuildLoad(builder, values[0], "");
-               return values[0];
-       } else if (!value_count)
-               unreachable("value_count is 0");
-
-       for (i = 0; i < value_count; i++) {
-               LLVMValueRef value = values[i * value_stride];
-               if (load)
-                       value = LLVMBuildLoad(builder, value, "");
-
-               if (!i)
-                       vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
-               LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
-               vec = LLVMBuildInsertElement(builder, vec, value, index, "");
-       }
-       return vec;
-}
-
-LLVMValueRef
-ac_build_gather_values(struct ac_llvm_context *ctx,
-                      LLVMValueRef *values,
-                      unsigned value_count)
-{
-       return ac_build_gather_values_extended(ctx, values, value_count, 1, false);
-}
-
-LLVMValueRef
-ac_emit_fdiv(struct ac_llvm_context *ctx,
-            LLVMValueRef num,
-            LLVMValueRef den)
-{
-       LLVMValueRef ret = LLVMBuildFDiv(ctx->builder, num, den, "");
-
-       if (!LLVMIsConstant(ret))
-               LLVMSetMetadata(ret, ctx->fpmath_md_kind, ctx->fpmath_md_2p5_ulp);
-       return ret;
-}
-
-/* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
- * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
- * already multiplied by two. id is the cube face number.
- */
-struct cube_selection_coords {
-       LLVMValueRef stc[2];
-       LLVMValueRef ma;
-       LLVMValueRef id;
-};
-
-static void
-build_cube_intrinsic(struct ac_llvm_context *ctx,
-                    LLVMValueRef in[3],
-                    struct cube_selection_coords *out)
-{
-       LLVMBuilderRef builder = ctx->builder;
-
-       if (HAVE_LLVM >= 0x0309) {
-               LLVMTypeRef f32 = ctx->f32;
-
-               out->stc[1] = ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.cubetc",
-                                       f32, in, 3, AC_FUNC_ATTR_READNONE);
-               out->stc[0] = ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.cubesc",
-                                       f32, in, 3, AC_FUNC_ATTR_READNONE);
-               out->ma = ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.cubema",
-                                       f32, in, 3, AC_FUNC_ATTR_READNONE);
-               out->id = ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.cubeid",
-                                       f32, in, 3, AC_FUNC_ATTR_READNONE);
-       } else {
-               LLVMValueRef c[4] = {
-                       in[0],
-                       in[1],
-                       in[2],
-                       LLVMGetUndef(LLVMTypeOf(in[0]))
-               };
-               LLVMValueRef vec = ac_build_gather_values(ctx, c, 4);
-
-               LLVMValueRef tmp =
-                       ac_emit_llvm_intrinsic(ctx, "llvm.AMDGPU.cube",
-                                         LLVMTypeOf(vec), &vec, 1,
-                                         AC_FUNC_ATTR_READNONE);
-
-               out->stc[1] = LLVMBuildExtractElement(builder, tmp,
-                               LLVMConstInt(ctx->i32, 0, 0), "");
-               out->stc[0] = LLVMBuildExtractElement(builder, tmp,
-                               LLVMConstInt(ctx->i32, 1, 0), "");
-               out->ma = LLVMBuildExtractElement(builder, tmp,
-                               LLVMConstInt(ctx->i32, 2, 0), "");
-               out->id = LLVMBuildExtractElement(builder, tmp,
-                               LLVMConstInt(ctx->i32, 3, 0), "");
-       }
-}
-
-/**
- * Build a manual selection sequence for cube face sc/tc coordinates and
- * major axis vector (multiplied by 2 for consistency) for the given
- * vec3 \p coords, for the face implied by \p selcoords.
- *
- * For the major axis, we always adjust the sign to be in the direction of
- * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
- * the selcoords major axis.
- */
-static void build_cube_select(LLVMBuilderRef builder,
-                             const struct cube_selection_coords *selcoords,
-                             const LLVMValueRef *coords,
-                             LLVMValueRef *out_st,
-                             LLVMValueRef *out_ma)
-{
-       LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
-       LLVMValueRef is_ma_positive;
-       LLVMValueRef sgn_ma;
-       LLVMValueRef is_ma_z, is_not_ma_z;
-       LLVMValueRef is_ma_y;
-       LLVMValueRef is_ma_x;
-       LLVMValueRef sgn;
-       LLVMValueRef tmp;
-
-       is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
-               selcoords->ma, LLVMConstReal(f32, 0.0), "");
-       sgn_ma = LLVMBuildSelect(builder, is_ma_positive,
-               LLVMConstReal(f32, 1.0), LLVMConstReal(f32, -1.0), "");
-
-       is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
-       is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
-       is_ma_y = LLVMBuildAnd(builder, is_not_ma_z,
-               LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
-       is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
-
-       /* Select sc */
-       tmp = LLVMBuildSelect(builder, is_ma_z, coords[2], coords[0], "");
-       sgn = LLVMBuildSelect(builder, is_ma_y, LLVMConstReal(f32, 1.0),
-               LLVMBuildSelect(builder, is_ma_x, sgn_ma,
-                       LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
-       out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
-
-       /* Select tc */
-       tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
-       sgn = LLVMBuildSelect(builder, is_ma_y, LLVMBuildFNeg(builder, sgn_ma, ""),
-               LLVMConstReal(f32, -1.0), "");
-       out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
-
-       /* Select ma */
-       tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
-               LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
-       sgn = LLVMBuildSelect(builder, is_ma_positive,
-               LLVMConstReal(f32, 2.0), LLVMConstReal(f32, -2.0), "");
-       *out_ma = LLVMBuildFMul(builder, tmp, sgn, "");
-}
-
-void
-ac_prepare_cube_coords(struct ac_llvm_context *ctx,
-                      bool is_deriv, bool is_array,
-                      LLVMValueRef *coords_arg,
-                      LLVMValueRef *derivs_arg)
-{
-
-       LLVMBuilderRef builder = ctx->builder;
-       struct cube_selection_coords selcoords;
-       LLVMValueRef coords[3];
-       LLVMValueRef invma;
-
-       build_cube_intrinsic(ctx, coords_arg, &selcoords);
-
-       invma = ac_emit_llvm_intrinsic(ctx, "llvm.fabs.f32",
-                       ctx->f32, &selcoords.ma, 1, AC_FUNC_ATTR_READNONE);
-       invma = ac_emit_fdiv(ctx, LLVMConstReal(ctx->f32, 1.0), invma);
-
-       for (int i = 0; i < 2; ++i)
-               coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
-
-       coords[2] = selcoords.id;
-
-       if (is_deriv && derivs_arg) {
-               LLVMValueRef derivs[4];
-               int axis;
-
-               /* Convert cube derivatives to 2D derivatives. */
-               for (axis = 0; axis < 2; axis++) {
-                       LLVMValueRef deriv_st[2];
-                       LLVMValueRef deriv_ma;
-
-                       /* Transform the derivative alongside the texture
-                        * coordinate. Mathematically, the correct formula is
-                        * as follows. Assume we're projecting onto the +Z face
-                        * and denote by dx/dh the derivative of the (original)
-                        * X texture coordinate with respect to horizontal
-                        * window coordinates. The projection onto the +Z face
-                        * plane is:
-                        *
-                        *   f(x,z) = x/z
-                        *
-                        * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
-                        *            = 1/z * dx/dh - x/z * 1/z * dz/dh.
-                        *
-                        * This motivatives the implementation below.
-                        *
-                        * Whether this actually gives the expected results for
-                        * apps that might feed in derivatives obtained via
-                        * finite differences is anyone's guess. The OpenGL spec
-                        * seems awfully quiet about how textureGrad for cube
-                        * maps should be handled.
-                        */
-                       build_cube_select(builder, &selcoords, &derivs_arg[axis * 3],
-                                         deriv_st, &deriv_ma);
-
-                       deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
-
-                       for (int i = 0; i < 2; ++i)
-                               derivs[axis * 2 + i] =
-                                       LLVMBuildFSub(builder,
-                                               LLVMBuildFMul(builder, deriv_st[i], invma, ""),
-                                               LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
-               }
-
-               memcpy(derivs_arg, derivs, sizeof(derivs));
-       }
-
-       /* Shift the texture coordinate. This must be applied after the
-        * derivative calculation.
-        */
-       for (int i = 0; i < 2; ++i)
-               coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
-
-       if (is_array) {
-               /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
-               /* coords_arg.w component - array_index for cube arrays */
-               LLVMValueRef tmp = LLVMBuildFMul(ctx->builder, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), "");
-               coords[2] = LLVMBuildFAdd(ctx->builder, tmp, coords[2], "");
-       }
-
-       memcpy(coords_arg, coords, sizeof(coords));
-}
-
 void
 ac_dump_module(LLVMModuleRef module)
 {
@@ -531,397 +210,3 @@ ac_dump_module(LLVMModuleRef module)
        fprintf(stderr, "%s", str);
        LLVMDisposeMessage(str);
 }
-
-LLVMValueRef
-ac_build_fs_interp(struct ac_llvm_context *ctx,
-                  LLVMValueRef llvm_chan,
-                  LLVMValueRef attr_number,
-                  LLVMValueRef params,
-                  LLVMValueRef i,
-                  LLVMValueRef j)
-{
-       LLVMValueRef args[5];
-       LLVMValueRef p1;
-       
-       if (HAVE_LLVM < 0x0400) {
-               LLVMValueRef ij[2];
-               ij[0] = LLVMBuildBitCast(ctx->builder, i, ctx->i32, "");
-               ij[1] = LLVMBuildBitCast(ctx->builder, j, ctx->i32, "");
-
-               args[0] = llvm_chan;
-               args[1] = attr_number;
-               args[2] = params;
-               args[3] = ac_build_gather_values(ctx, ij, 2);
-               return ac_emit_llvm_intrinsic(ctx, "llvm.SI.fs.interp",
-                                             ctx->f32, args, 4,
-                                             AC_FUNC_ATTR_READNONE);
-       }
-
-       args[0] = i;
-       args[1] = llvm_chan;
-       args[2] = attr_number;
-       args[3] = params;
-
-       p1 = ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.interp.p1",
-                                   ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
-
-       args[0] = p1;
-       args[1] = j;
-       args[2] = llvm_chan;
-       args[3] = attr_number;
-       args[4] = params;
-
-       return ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.interp.p2",
-                                     ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
-}
-
-LLVMValueRef
-ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
-                      LLVMValueRef parameter,
-                      LLVMValueRef llvm_chan,
-                      LLVMValueRef attr_number,
-                      LLVMValueRef params)
-{
-       LLVMValueRef args[4];
-       if (HAVE_LLVM < 0x0400) {
-               args[0] = llvm_chan;
-               args[1] = attr_number;
-               args[2] = params;
-
-               return ac_emit_llvm_intrinsic(ctx,
-                                             "llvm.SI.fs.constant",
-                                             ctx->f32, args, 3,
-                                             AC_FUNC_ATTR_READNONE);
-       }
-
-       args[0] = parameter;
-       args[1] = llvm_chan;
-       args[2] = attr_number;
-       args[3] = params;
-
-       return ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.interp.mov",
-                                     ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
-}
-
-LLVMValueRef
-ac_build_gep0(struct ac_llvm_context *ctx,
-             LLVMValueRef base_ptr,
-             LLVMValueRef index)
-{
-       LLVMValueRef indices[2] = {
-               LLVMConstInt(ctx->i32, 0, 0),
-               index,
-       };
-       return LLVMBuildGEP(ctx->builder, base_ptr,
-                           indices, 2, "");
-}
-
-void
-ac_build_indexed_store(struct ac_llvm_context *ctx,
-                      LLVMValueRef base_ptr, LLVMValueRef index,
-                      LLVMValueRef value)
-{
-       LLVMBuildStore(ctx->builder, value,
-                      ac_build_gep0(ctx, base_ptr, index));
-}
-
-/**
- * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
- * It's equivalent to doing a load from &base_ptr[index].
- *
- * \param base_ptr  Where the array starts.
- * \param index     The element index into the array.
- * \param uniform   Whether the base_ptr and index can be assumed to be
- *                  dynamically uniform
- */
-LLVMValueRef
-ac_build_indexed_load(struct ac_llvm_context *ctx,
-                     LLVMValueRef base_ptr, LLVMValueRef index,
-                     bool uniform)
-{
-       LLVMValueRef pointer;
-
-       pointer = ac_build_gep0(ctx, base_ptr, index);
-       if (uniform)
-               LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
-       return LLVMBuildLoad(ctx->builder, pointer, "");
-}
-
-/**
- * Do a load from &base_ptr[index], but also add a flag that it's loading
- * a constant from a dynamically uniform index.
- */
-LLVMValueRef
-ac_build_indexed_load_const(struct ac_llvm_context *ctx,
-                           LLVMValueRef base_ptr, LLVMValueRef index)
-{
-       LLVMValueRef result = ac_build_indexed_load(ctx, base_ptr, index, true);
-       LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
-       return result;
-}
-
-/* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
- * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
- * or v4i32 (num_channels=3,4).
- */
-void
-ac_build_tbuffer_store(struct ac_llvm_context *ctx,
-                      LLVMValueRef rsrc,
-                      LLVMValueRef vdata,
-                      unsigned num_channels,
-                      LLVMValueRef vaddr,
-                      LLVMValueRef soffset,
-                      unsigned inst_offset,
-                      unsigned dfmt,
-                      unsigned nfmt,
-                      unsigned offen,
-                      unsigned idxen,
-                      unsigned glc,
-                      unsigned slc,
-                      unsigned tfe)
-{
-       LLVMValueRef args[] = {
-               rsrc,
-               vdata,
-               LLVMConstInt(ctx->i32, num_channels, 0),
-               vaddr,
-               soffset,
-               LLVMConstInt(ctx->i32, inst_offset, 0),
-               LLVMConstInt(ctx->i32, dfmt, 0),
-               LLVMConstInt(ctx->i32, nfmt, 0),
-               LLVMConstInt(ctx->i32, offen, 0),
-               LLVMConstInt(ctx->i32, idxen, 0),
-               LLVMConstInt(ctx->i32, glc, 0),
-               LLVMConstInt(ctx->i32, slc, 0),
-               LLVMConstInt(ctx->i32, tfe, 0)
-       };
-
-       /* The instruction offset field has 12 bits */
-       assert(offen || inst_offset < (1 << 12));
-
-       /* The intrinsic is overloaded, we need to add a type suffix for overloading to work. */
-       unsigned func = CLAMP(num_channels, 1, 3) - 1;
-       const char *types[] = {"i32", "v2i32", "v4i32"};
-       char name[256];
-       snprintf(name, sizeof(name), "llvm.SI.tbuffer.store.%s", types[func]);
-
-       ac_emit_llvm_intrinsic(ctx, name, ctx->voidt,
-                              args, ARRAY_SIZE(args), 0);
-}
-
-void
-ac_build_tbuffer_store_dwords(struct ac_llvm_context *ctx,
-                             LLVMValueRef rsrc,
-                             LLVMValueRef vdata,
-                             unsigned num_channels,
-                             LLVMValueRef vaddr,
-                             LLVMValueRef soffset,
-                             unsigned inst_offset)
-{
-       static unsigned dfmt[] = {
-               V_008F0C_BUF_DATA_FORMAT_32,
-               V_008F0C_BUF_DATA_FORMAT_32_32,
-               V_008F0C_BUF_DATA_FORMAT_32_32_32,
-               V_008F0C_BUF_DATA_FORMAT_32_32_32_32
-       };
-       assert(num_channels >= 1 && num_channels <= 4);
-
-       ac_build_tbuffer_store(ctx, rsrc, vdata, num_channels, vaddr, soffset,
-                              inst_offset, dfmt[num_channels - 1],
-                              V_008F0C_BUF_NUM_FORMAT_UINT, 1, 0, 1, 1, 0);
-}
-
-LLVMValueRef
-ac_build_buffer_load(struct ac_llvm_context *ctx,
-                    LLVMValueRef rsrc,
-                    int num_channels,
-                    LLVMValueRef vindex,
-                    LLVMValueRef voffset,
-                    LLVMValueRef soffset,
-                    unsigned inst_offset,
-                    unsigned glc,
-                    unsigned slc)
-{
-       unsigned func = CLAMP(num_channels, 1, 3) - 1;
-
-       if (HAVE_LLVM >= 0x309) {
-               LLVMValueRef args[] = {
-                       LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
-                       vindex ? vindex : LLVMConstInt(ctx->i32, 0, 0),
-                       LLVMConstInt(ctx->i32, inst_offset, 0),
-                       LLVMConstInt(ctx->i1, glc, 0),
-                       LLVMConstInt(ctx->i1, slc, 0)
-               };
-
-               LLVMTypeRef types[] = {ctx->f32, LLVMVectorType(ctx->f32, 2),
-                                      ctx->v4f32};
-               const char *type_names[] = {"f32", "v2f32", "v4f32"};
-               char name[256];
-
-               if (voffset) {
-                       args[2] = LLVMBuildAdd(ctx->builder, args[2], voffset,
-                                              "");
-               }
-
-               if (soffset) {
-                       args[2] = LLVMBuildAdd(ctx->builder, args[2], soffset,
-                                              "");
-               }
-
-               snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.%s",
-                        type_names[func]);
-
-               return ac_emit_llvm_intrinsic(ctx, name, types[func], args,
-                                             ARRAY_SIZE(args), AC_FUNC_ATTR_READONLY);
-       } else {
-               LLVMValueRef args[] = {
-                       LLVMBuildBitCast(ctx->builder, rsrc, ctx->v16i8, ""),
-                       voffset ? voffset : vindex,
-                       soffset,
-                       LLVMConstInt(ctx->i32, inst_offset, 0),
-                       LLVMConstInt(ctx->i32, voffset ? 1 : 0, 0), // offen
-                       LLVMConstInt(ctx->i32, vindex ? 1 : 0, 0), //idxen
-                       LLVMConstInt(ctx->i32, glc, 0),
-                       LLVMConstInt(ctx->i32, slc, 0),
-                       LLVMConstInt(ctx->i32, 0, 0), // TFE
-               };
-
-               LLVMTypeRef types[] = {ctx->i32, LLVMVectorType(ctx->i32, 2),
-                                      ctx->v4i32};
-               const char *type_names[] = {"i32", "v2i32", "v4i32"};
-               const char *arg_type = "i32";
-               char name[256];
-
-               if (voffset && vindex) {
-                       LLVMValueRef vaddr[] = {vindex, voffset};
-
-                       arg_type = "v2i32";
-                       args[1] = ac_build_gather_values(ctx, vaddr, 2);
-               }
-
-               snprintf(name, sizeof(name), "llvm.SI.buffer.load.dword.%s.%s",
-                        type_names[func], arg_type);
-
-               return ac_emit_llvm_intrinsic(ctx, name, types[func], args,
-                                              ARRAY_SIZE(args), AC_FUNC_ATTR_READONLY);
-       }
-}
-
-/**
- * Set range metadata on an instruction.  This can only be used on load and
- * call instructions.  If you know an instruction can only produce the values
- * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
- * \p lo is the minimum value inclusive.
- * \p hi is the maximum value exclusive.
- */
-static void set_range_metadata(struct ac_llvm_context *ctx,
-                              LLVMValueRef value, unsigned lo, unsigned hi)
-{
-       LLVMValueRef range_md, md_args[2];
-       LLVMTypeRef type = LLVMTypeOf(value);
-       LLVMContextRef context = LLVMGetTypeContext(type);
-
-       md_args[0] = LLVMConstInt(type, lo, false);
-       md_args[1] = LLVMConstInt(type, hi, false);
-       range_md = LLVMMDNodeInContext(context, md_args, 2);
-       LLVMSetMetadata(value, ctx->range_md_kind, range_md);
-}
-
-LLVMValueRef
-ac_get_thread_id(struct ac_llvm_context *ctx)
-{
-       LLVMValueRef tid;
-
-       if (HAVE_LLVM < 0x0308) {
-               tid = ac_emit_llvm_intrinsic(ctx, "llvm.SI.tid",
-                                            ctx->i32,
-                                            NULL, 0, AC_FUNC_ATTR_READNONE);
-       } else {
-               LLVMValueRef tid_args[2];
-               tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
-               tid_args[1] = LLVMConstInt(ctx->i32, 0, false);
-               tid_args[1] = ac_emit_llvm_intrinsic(ctx,
-                                                    "llvm.amdgcn.mbcnt.lo", ctx->i32,
-                                                    tid_args, 2, AC_FUNC_ATTR_READNONE);
-
-               tid = ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
-                                            ctx->i32, tid_args,
-                                            2, AC_FUNC_ATTR_READNONE);
-       }
-       set_range_metadata(ctx, tid, 0, 64);
-       return tid;
-}
-
-/*
- * SI implements derivatives using the local data store (LDS)
- * All writes to the LDS happen in all executing threads at
- * the same time. TID is the Thread ID for the current
- * thread and is a value between 0 and 63, representing
- * the thread's position in the wavefront.
- *
- * For the pixel shader threads are grouped into quads of four pixels.
- * The TIDs of the pixels of a quad are:
- *
- *  +------+------+
- *  |4n + 0|4n + 1|
- *  +------+------+
- *  |4n + 2|4n + 3|
- *  +------+------+
- *
- * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
- * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
- * the current pixel's column, and masking with 0xfffffffe yields the TID
- * of the left pixel of the current pixel's row.
- *
- * Adding 1 yields the TID of the pixel to the right of the left pixel, and
- * adding 2 yields the TID of the pixel below the top pixel.
- */
-LLVMValueRef
-ac_emit_ddxy(struct ac_llvm_context *ctx,
-            bool has_ds_bpermute,
-            uint32_t mask,
-            int idx,
-            LLVMValueRef lds,
-            LLVMValueRef val)
-{
-       LLVMValueRef thread_id, tl, trbl, tl_tid, trbl_tid, args[2];
-       LLVMValueRef result;
-
-       thread_id = ac_get_thread_id(ctx);
-
-       tl_tid = LLVMBuildAnd(ctx->builder, thread_id,
-                             LLVMConstInt(ctx->i32, mask, false), "");
-
-       trbl_tid = LLVMBuildAdd(ctx->builder, tl_tid,
-                               LLVMConstInt(ctx->i32, idx, false), "");
-
-       if (has_ds_bpermute) {
-               args[0] = LLVMBuildMul(ctx->builder, tl_tid,
-                                      LLVMConstInt(ctx->i32, 4, false), "");
-               args[1] = val;
-               tl = ac_emit_llvm_intrinsic(ctx,
-                                           "llvm.amdgcn.ds.bpermute", ctx->i32,
-                                           args, 2, AC_FUNC_ATTR_READNONE);
-
-               args[0] = LLVMBuildMul(ctx->builder, trbl_tid,
-                                      LLVMConstInt(ctx->i32, 4, false), "");
-               trbl = ac_emit_llvm_intrinsic(ctx,
-                                             "llvm.amdgcn.ds.bpermute", ctx->i32,
-                                             args, 2, AC_FUNC_ATTR_READNONE);
-       } else {
-               LLVMValueRef store_ptr, load_ptr0, load_ptr1;
-
-               store_ptr = ac_build_gep0(ctx, lds, thread_id);
-               load_ptr0 = ac_build_gep0(ctx, lds, tl_tid);
-               load_ptr1 = ac_build_gep0(ctx, lds, trbl_tid);
-
-               LLVMBuildStore(ctx->builder, val, store_ptr);
-               tl = LLVMBuildLoad(ctx->builder, load_ptr0, "");
-               trbl = LLVMBuildLoad(ctx->builder, load_ptr1, "");
-       }
-
-       tl = LLVMBuildBitCast(ctx->builder, tl, ctx->f32, "");
-       trbl = LLVMBuildBitCast(ctx->builder, trbl, ctx->f32, "");
-       result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
-       return result;
-}
index a457714..1f37a12 100644 (file)
@@ -44,154 +44,19 @@ enum ac_func_attr {
        AC_FUNC_ATTR_LAST         = (1 << 7)
 };
 
-struct ac_llvm_context {
-       LLVMContextRef context;
-       LLVMModuleRef module;
-       LLVMBuilderRef builder;
-
-       LLVMTypeRef voidt;
-       LLVMTypeRef i1;
-       LLVMTypeRef i8;
-       LLVMTypeRef i32;
-       LLVMTypeRef f32;
-       LLVMTypeRef v4i32;
-       LLVMTypeRef v4f32;
-       LLVMTypeRef v16i8;
-
-       unsigned range_md_kind;
-       unsigned invariant_load_md_kind;
-       unsigned uniform_md_kind;
-       unsigned fpmath_md_kind;
-       LLVMValueRef fpmath_md_2p5_ulp;
-       LLVMValueRef empty_md;
-};
-
 LLVMTargetMachineRef ac_create_target_machine(enum radeon_family family, bool supports_spill);
 
 void ac_add_attr_dereferenceable(LLVMValueRef val, uint64_t bytes);
 bool ac_is_sgpr_param(LLVMValueRef param);
 
 void
-ac_llvm_context_init(struct ac_llvm_context *ctx, LLVMContextRef context);
-
-void
 ac_add_function_attr(LLVMValueRef function,
                      int attr_idx,
                      enum ac_func_attr attr);
-LLVMValueRef
-ac_emit_llvm_intrinsic(struct ac_llvm_context *ctx, const char *name,
-                      LLVMTypeRef return_type, LLVMValueRef *params,
-                      unsigned param_count, unsigned attrib_mask);
-
-LLVMValueRef
-ac_build_gather_values_extended(struct ac_llvm_context *ctx,
-                               LLVMValueRef *values,
-                               unsigned value_count,
-                               unsigned value_stride,
-                               bool load);
-LLVMValueRef
-ac_build_gather_values(struct ac_llvm_context *ctx,
-                      LLVMValueRef *values,
-                      unsigned value_count);
-
-LLVMValueRef
-ac_emit_fdiv(struct ac_llvm_context *ctx,
-            LLVMValueRef num,
-            LLVMValueRef den);
-
-void
-ac_prepare_cube_coords(struct ac_llvm_context *ctx,
-                      bool is_deriv, bool is_array,
-                      LLVMValueRef *coords_arg,
-                      LLVMValueRef *derivs_arg);
 
 void
 ac_dump_module(LLVMModuleRef module);
 
-LLVMValueRef
-ac_build_fs_interp(struct ac_llvm_context *ctx,
-                  LLVMValueRef llvm_chan,
-                  LLVMValueRef attr_number,
-                  LLVMValueRef params,
-                  LLVMValueRef i,
-                  LLVMValueRef j);
-
-LLVMValueRef
-ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
-                      LLVMValueRef parameter,
-                      LLVMValueRef llvm_chan,
-                      LLVMValueRef attr_number,
-                      LLVMValueRef params);
-
-LLVMValueRef
-ac_build_gep0(struct ac_llvm_context *ctx,
-             LLVMValueRef base_ptr,
-             LLVMValueRef index);
-
-void
-ac_build_indexed_store(struct ac_llvm_context *ctx,
-                      LLVMValueRef base_ptr, LLVMValueRef index,
-                      LLVMValueRef value);
-
-LLVMValueRef
-ac_build_indexed_load(struct ac_llvm_context *ctx,
-                     LLVMValueRef base_ptr, LLVMValueRef index,
-                     bool uniform);
-
-LLVMValueRef
-ac_build_indexed_load_const(struct ac_llvm_context *ctx,
-                           LLVMValueRef base_ptr, LLVMValueRef index);
-
-void
-ac_build_tbuffer_store_dwords(struct ac_llvm_context *ctx,
-                             LLVMValueRef rsrc,
-                             LLVMValueRef vdata,
-                             unsigned num_channels,
-                             LLVMValueRef vaddr,
-                             LLVMValueRef soffset,
-                             unsigned inst_offset);
-
-void
-ac_build_tbuffer_store(struct ac_llvm_context *ctx,
-                      LLVMValueRef rsrc,
-                      LLVMValueRef vdata,
-                      unsigned num_channels,
-                      LLVMValueRef vaddr,
-                      LLVMValueRef soffset,
-                      unsigned inst_offset,
-                      unsigned dfmt,
-                      unsigned nfmt,
-                      unsigned offen,
-                      unsigned idxen,
-                      unsigned glc,
-                      unsigned slc,
-                      unsigned tfe);
-
-LLVMValueRef
-ac_build_buffer_load(struct ac_llvm_context *ctx,
-                    LLVMValueRef rsrc,
-                    int num_channels,
-                    LLVMValueRef vindex,
-                    LLVMValueRef voffset,
-                    LLVMValueRef soffset,
-                    unsigned inst_offset,
-                    unsigned glc,
-                    unsigned slc);
-
-LLVMValueRef
-ac_get_thread_id(struct ac_llvm_context *ctx);
-
-#define AC_TID_MASK_TOP_LEFT 0xfffffffc
-#define AC_TID_MASK_TOP      0xfffffffd
-#define AC_TID_MASK_LEFT     0xfffffffe
-
-LLVMValueRef
-ac_emit_ddxy(struct ac_llvm_context *ctx,
-            bool has_ds_bpermute,
-            uint32_t mask,
-            int idx,
-            LLVMValueRef lds,
-            LLVMValueRef val);
 #ifdef __cplusplus
 }
 #endif
index 566516f..db59c2e 100644 (file)
@@ -22,6 +22,7 @@
  */
 
 #include "ac_nir_to_llvm.h"
+#include "ac_llvm_build.h"
 #include "ac_llvm_util.h"
 #include "ac_binary.h"
 #include "sid.h"
index 05b9868..26cc28d 100644 (file)
@@ -29,6 +29,7 @@
 #include "gallivm/lp_bld_tgsi.h"
 #include "tgsi/tgsi_parse.h"
 #include "ac_llvm_util.h"
+#include "ac_llvm_build.h"
 
 #include <llvm-c/Core.h>
 #include <llvm-c/TargetMachine.h>