*/
/* based on pieces from si_pipe.c and radeon_llvm_emit.c */
#include "ac_llvm_util.h"
-
+#include "ac_llvm_build.h"
+#include "util/bitscan.h"
#include <llvm-c/Core.h>
-
+#include <llvm-c/Support.h>
+#include <llvm-c/Transforms/IPO.h>
+#include <llvm-c/Transforms/Scalar.h>
+#include <llvm-c/Transforms/Utils.h>
#include "c11/threads.h"
+#include "gallivm/lp_bld_misc.h"
+#include "util/u_math.h"
#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()
{
-#if HAVE_LLVM < 0x0307
- LLVMInitializeR600TargetInfo();
- LLVMInitializeR600Target();
- LLVMInitializeR600TargetMC();
- LLVMInitializeR600AsmPrinter();
-#else
LLVMInitializeAMDGPUTargetInfo();
LLVMInitializeAMDGPUTarget();
LLVMInitializeAMDGPUTargetMC();
LLVMInitializeAMDGPUAsmPrinter();
-#endif
+
+ /* For inline assembly. */
+ LLVMInitializeAMDGPUAsmParser();
+
+ /* Workaround for bug in llvm 4.0 that causes image intrinsics
+ * to disappear.
+ * https://reviews.llvm.org/D26348
+ *
+ * "mesa" is the prefix for error messages.
+ *
+ * -global-isel-abort=2 is a no-op unless global isel has been enabled.
+ * This option tells the backend to fall-back to SelectionDAG and print
+ * a diagnostic message if global isel fails.
+ */
+ const char *argv[3] = { "mesa", "-simplifycfg-sink-common=false", "-global-isel-abort=2" };
+ LLVMParseCommandLineOptions(3, argv, NULL);
}
static once_flag ac_init_llvm_target_once_flag = ONCE_FLAG_INIT;
+void ac_init_llvm_once(void)
+{
+ call_once(&ac_init_llvm_target_once_flag, ac_init_llvm_target);
+}
+
static LLVMTargetRef ac_get_llvm_target(const char *triple)
{
LLVMTargetRef target = NULL;
char *err_message = NULL;
- call_once(&ac_init_llvm_target_once_flag, ac_init_llvm_target);
-
if (LLVMGetTargetFromTriple(triple, &target, &err_message)) {
fprintf(stderr, "Cannot find target for triple %s ", triple);
if (err_message) {
return target;
}
-static const char *ac_get_llvm_processor_name(enum radeon_family family)
+const char *ac_get_llvm_processor_name(enum radeon_family family)
{
switch (family) {
case CHIP_TAHITI:
return "iceland";
case CHIP_CARRIZO:
return "carrizo";
-#if HAVE_LLVM <= 0x0307
- case CHIP_FIJI:
- return "tonga";
- case CHIP_STONEY:
- return "carrizo";
-#else
case CHIP_FIJI:
return "fiji";
case CHIP_STONEY:
return "stoney";
-#endif
-#if HAVE_LLVM <= 0x0308
- case CHIP_POLARIS10:
- return "tonga";
- case CHIP_POLARIS11:
- return "tonga";
-#else
case CHIP_POLARIS10:
return "polaris10";
case CHIP_POLARIS11:
+ case CHIP_POLARIS12:
+ case CHIP_VEGAM:
return "polaris11";
-#endif
+ case CHIP_VEGA10:
+ return "gfx900";
+ case CHIP_RAVEN:
+ return "gfx902";
+ case CHIP_VEGA12:
+ return "gfx904";
+ case CHIP_VEGA20:
+ return "gfx906";
+ case CHIP_RAVEN2:
+ return HAVE_LLVM >= 0x0800 ? "gfx909" : "gfx902";
default:
return "";
}
}
-LLVMTargetMachineRef ac_create_target_machine(enum radeon_family family, bool supports_spill)
+static LLVMTargetMachineRef ac_create_target_machine(enum radeon_family family,
+ enum ac_target_machine_options tm_options,
+ LLVMCodeGenOptLevel level,
+ const char **out_triple)
{
assert(family >= CHIP_TAHITI);
-
- const char *triple = supports_spill ? "amdgcn-mesa-mesa3d" : "amdgcn--";
+ char features[256];
+ const char *triple = (tm_options & AC_TM_SUPPORTS_SPILL) ? "amdgcn-mesa-mesa3d" : "amdgcn--";
LLVMTargetRef target = ac_get_llvm_target(triple);
+
+ snprintf(features, sizeof(features),
+ "+DumpCode,-fp32-denormals,+fp64-denormals%s%s%s%s%s%s",
+ HAVE_LLVM >= 0x0800 ? "" : ",+vgpr-spilling",
+ tm_options & AC_TM_SISCHED ? ",+si-scheduler" : "",
+ tm_options & AC_TM_FORCE_ENABLE_XNACK ? ",+xnack" : "",
+ tm_options & AC_TM_FORCE_DISABLE_XNACK ? ",-xnack" : "",
+ tm_options & AC_TM_PROMOTE_ALLOCA_TO_SCRATCH ? ",-promote-alloca" : "",
+ tm_options & AC_TM_NO_LOAD_STORE_OPT ? ",-load-store-opt" : "");
+
LLVMTargetMachineRef tm = LLVMCreateTargetMachine(
target,
triple,
ac_get_llvm_processor_name(family),
- "+DumpCode,+vgpr-spilling",
- LLVMCodeGenLevelDefault,
+ features,
+ level,
LLVMRelocDefault,
LLVMCodeModelDefault);
+ if (out_triple)
+ *out_triple = triple;
+ if (tm_options & AC_TM_ENABLE_GLOBAL_ISEL)
+ ac_enable_global_isel(tm);
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)
+static LLVMPassManagerRef ac_create_passmgr(LLVMTargetLibraryInfoRef target_library_info,
+ bool check_ir)
{
- 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);
-}
+ LLVMPassManagerRef passmgr = LLVMCreatePassManager();
+ if (!passmgr)
+ return NULL;
-#if HAVE_LLVM < 0x0400
-static LLVMAttribute ac_attr_to_llvm_attr(enum ac_func_attr attr)
-{
- switch (attr) {
- case AC_FUNC_ATTR_ALWAYSINLINE: return LLVMAlwaysInlineAttribute;
- case AC_FUNC_ATTR_BYVAL: return LLVMByValAttribute;
- case AC_FUNC_ATTR_INREG: return LLVMInRegAttribute;
- case AC_FUNC_ATTR_NOALIAS: return LLVMNoAliasAttribute;
- case AC_FUNC_ATTR_NOUNWIND: return LLVMNoUnwindAttribute;
- case AC_FUNC_ATTR_READNONE: return LLVMReadNoneAttribute;
- case AC_FUNC_ATTR_READONLY: return LLVMReadOnlyAttribute;
- default:
- fprintf(stderr, "Unhandled function attribute: %x\n", attr);
- return 0;
- }
+ if (target_library_info)
+ LLVMAddTargetLibraryInfo(target_library_info,
+ passmgr);
+
+ if (check_ir)
+ LLVMAddVerifierPass(passmgr);
+ LLVMAddAlwaysInlinerPass(passmgr);
+ /* Normally, the pass manager runs all passes on one function before
+ * moving onto another. Adding a barrier no-op pass forces the pass
+ * manager to run the inliner on all functions first, which makes sure
+ * that the following passes are only run on the remaining non-inline
+ * function, so it removes useless work done on dead inline functions.
+ */
+ ac_llvm_add_barrier_noop_pass(passmgr);
+ /* This pass should eliminate all the load and store instructions. */
+ LLVMAddPromoteMemoryToRegisterPass(passmgr);
+ LLVMAddScalarReplAggregatesPass(passmgr);
+ LLVMAddLICMPass(passmgr);
+ LLVMAddAggressiveDCEPass(passmgr);
+ LLVMAddCFGSimplificationPass(passmgr);
+ /* This is recommended by the instruction combining pass. */
+ LLVMAddEarlyCSEMemSSAPass(passmgr);
+ LLVMAddInstructionCombiningPass(passmgr);
+ return passmgr;
}
-#else
-
static const char *attr_to_str(enum ac_func_attr attr)
{
switch (attr) {
case AC_FUNC_ATTR_ALWAYSINLINE: return "alwaysinline";
- case AC_FUNC_ATTR_BYVAL: return "byval";
case AC_FUNC_ATTR_INREG: return "inreg";
case AC_FUNC_ATTR_NOALIAS: return "noalias";
case AC_FUNC_ATTR_NOUNWIND: return "nounwind";
case AC_FUNC_ATTR_READNONE: return "readnone";
case AC_FUNC_ATTR_READONLY: return "readonly";
+ case AC_FUNC_ATTR_WRITEONLY: return "writeonly";
+ case AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY: return "inaccessiblememonly";
+ case AC_FUNC_ATTR_CONVERGENT: return "convergent";
default:
fprintf(stderr, "Unhandled function attribute: %x\n", attr);
return 0;
}
}
-#endif
-
void
-ac_add_function_attr(LLVMValueRef function,
- int attr_idx,
- enum ac_func_attr attr)
+ac_add_function_attr(LLVMContextRef ctx, LLVMValueRef function,
+ int attr_idx, enum ac_func_attr attr)
{
-
-#if HAVE_LLVM < 0x0400
- LLVMAttribute llvm_attr = ac_attr_to_llvm_attr(attr);
- if (attr_idx == -1) {
- LLVMAddFunctionAttr(function, llvm_attr);
- } else {
- LLVMAddAttribute(LLVMGetParam(function, attr_idx - 1), llvm_attr);
- }
-#else
- LLVMContextRef context = LLVMGetModuleContext(LLVMGetGlobalParent(function));
const char *attr_name = attr_to_str(attr);
unsigned kind_id = LLVMGetEnumAttributeKindForName(attr_name,
strlen(attr_name));
- LLVMAttributeRef llvm_attr = LLVMCreateEnumAttribute(context, kind_id, 0);
- LLVMAddAttributeAtIndex(function, attr_idx, llvm_attr);
-#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;
+ LLVMAttributeRef llvm_attr = LLVMCreateEnumAttribute(ctx, kind_id, 0);
- 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, "");
+ if (LLVMIsAFunction(function))
+ LLVMAddAttributeAtIndex(function, attr_idx, llvm_attr);
+ else
+ LLVMAddCallSiteAttribute(function, attr_idx, llvm_attr);
}
-LLVMValueRef
-ac_build_gather_values_extended(struct ac_llvm_context *ctx,
- LLVMValueRef *values,
- unsigned value_count,
- unsigned value_stride,
- bool load)
+void ac_add_func_attributes(LLVMContextRef ctx, LLVMValueRef function,
+ unsigned attrib_mask)
{
- 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;
-}
+ attrib_mask |= AC_FUNC_ATTR_NOUNWIND;
+ attrib_mask &= ~AC_FUNC_ATTR_LEGACY;
-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), "");
+ while (attrib_mask) {
+ enum ac_func_attr attr = 1u << u_bit_scan(&attrib_mask);
+ ac_add_function_attr(ctx, function, -1, attr);
}
}
-/**
- * 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)
{
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)
+ac_llvm_add_target_dep_function_attr(LLVMValueRef F,
+ const char *name, unsigned value)
{
- LLVMBuildStore(ctx->builder, value,
- ac_build_gep0(ctx, base_ptr, index));
-}
+ char str[16];
-/**
- * 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, "");
+ snprintf(str, sizeof(str), "0x%x", value);
+ LLVMAddTargetDependentFunctionAttr(F, name, str);
}
-/**
- * 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)
+unsigned
+ac_count_scratch_private_memory(LLVMValueRef function)
{
- LLVMValueRef result = ac_build_indexed_load(ctx, base_ptr, index, true);
- LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
- return result;
-}
+ unsigned private_mem_vgprs = 0;
-/* 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);
-}
+ /* Process all LLVM instructions. */
+ LLVMBasicBlockRef bb = LLVMGetFirstBasicBlock(function);
+ while (bb) {
+ LLVMValueRef next = LLVMGetFirstInstruction(bb);
-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,
- "");
- }
+ while (next) {
+ LLVMValueRef inst = next;
+ next = LLVMGetNextInstruction(next);
- if (soffset) {
- args[2] = LLVMBuildAdd(ctx->builder, args[2], soffset,
- "");
- }
+ if (LLVMGetInstructionOpcode(inst) != LLVMAlloca)
+ continue;
- 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);
+ LLVMTypeRef type = LLVMGetElementType(LLVMTypeOf(inst));
+ /* No idea why LLVM aligns allocas to 4 elements. */
+ unsigned alignment = LLVMGetAlignment(inst);
+ unsigned dw_size = align(ac_get_type_size(type) / 4, alignment);
+ private_mem_vgprs += dw_size;
}
-
- 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);
+ bb = LLVMGetNextBasicBlock(bb);
}
-}
-/**
- * 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);
+ return private_mem_vgprs;
}
-LLVMValueRef
-ac_get_thread_id(struct ac_llvm_context *ctx)
+bool
+ac_init_llvm_compiler(struct ac_llvm_compiler *compiler,
+ enum radeon_family family,
+ enum ac_target_machine_options tm_options)
{
- 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);
+ const char *triple;
+ memset(compiler, 0, sizeof(*compiler));
+
+ compiler->tm = ac_create_target_machine(family, tm_options,
+ LLVMCodeGenLevelDefault,
+ &triple);
+ if (!compiler->tm)
+ return false;
+
+ if (tm_options & AC_TM_CREATE_LOW_OPT) {
+ compiler->low_opt_tm =
+ ac_create_target_machine(family, tm_options,
+ LLVMCodeGenLevelLess, NULL);
+ if (!compiler->low_opt_tm)
+ goto fail;
}
- set_range_metadata(ctx, tid, 0, 64);
- return tid;
+
+ compiler->target_library_info =
+ ac_create_target_library_info(triple);
+ if (!compiler->target_library_info)
+ goto fail;
+
+ compiler->passmgr = ac_create_passmgr(compiler->target_library_info,
+ tm_options & AC_TM_CHECK_IR);
+ if (!compiler->passmgr)
+ goto fail;
+
+ return true;
+fail:
+ ac_destroy_llvm_compiler(compiler);
+ return false;
}
-/*
- * 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)
+void
+ac_destroy_llvm_compiler(struct ac_llvm_compiler *compiler)
{
- 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;
+ if (compiler->passmgr)
+ LLVMDisposePassManager(compiler->passmgr);
+ if (compiler->target_library_info)
+ ac_dispose_target_library_info(compiler->target_library_info);
+ if (compiler->low_opt_tm)
+ LLVMDisposeTargetMachine(compiler->low_opt_tm);
+ if (compiler->tm)
+ LLVMDisposeTargetMachine(compiler->tm);
}