diff options
Diffstat (limited to 'src/gallium/auxiliary/gallivm')
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_const.h | 18 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_conv.c | 59 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_debug.c | 141 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_debug.cpp | 357 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_debug.h | 10 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_format_aos.c | 81 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_init.c | 60 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_misc.cpp | 62 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_sample_aos.c | 2 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c | 25 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_tgsi.h | 8 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c | 73 |
12 files changed, 595 insertions, 301 deletions
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_const.h b/src/gallium/auxiliary/gallivm/lp_bld_const.h index c749a7a315..69718eb4b3 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_const.h +++ b/src/gallium/auxiliary/gallivm/lp_bld_const.h @@ -125,4 +125,22 @@ lp_build_const_float(struct gallivm_state *gallivm, float x) } +/** Return constant-valued pointer to int */ +static INLINE LLVMValueRef +lp_build_const_int_pointer(struct gallivm_state *gallivm, const void *ptr) +{ + LLVMTypeRef int_type; + LLVMValueRef v; + + /* int type large enough to hold a pointer */ + int_type = LLVMIntTypeInContext(gallivm->context, 8 * sizeof(void *)); + v = LLVMConstInt(int_type, (uintptr_t) ptr, 0); + v = LLVMBuildIntToPtr(gallivm->builder, v, + LLVMPointerType(int_type, 0), + "cast int to ptr"); + return v; +} + + + #endif /* !LP_BLD_CONST_H */ diff --git a/src/gallium/auxiliary/gallivm/lp_bld_conv.c b/src/gallium/auxiliary/gallivm/lp_bld_conv.c index c43ee8ac63..c261d76116 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_conv.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_conv.c @@ -231,44 +231,53 @@ lp_build_unsigned_norm_to_float(struct gallivm_state *gallivm, assert(dst_type.floating); - /* Special-case int8->float, though most cases could be handled - * this way: - */ - if (src_width == 8) { - scale = 1.0/255.0; + mantissa = lp_mantissa(dst_type); + + if (src_width <= (mantissa + 1)) { + /* + * The source width matches fits what can be represented in floating + * point (i.e., mantissa + 1 bits). So do a straight multiplication + * followed by casting. No further rounding is necessary. + */ + + scale = 1.0/(double)((1ULL << src_width) - 1); res = LLVMBuildSIToFP(builder, src, vec_type, ""); res = LLVMBuildFMul(builder, res, lp_build_const_vec(gallivm, dst_type, scale), ""); return res; } + else { + /* + * The source width exceeds what can be represented in floating + * point. So truncate the incoming values. + */ - mantissa = lp_mantissa(dst_type); - - n = MIN2(mantissa, src_width); + n = MIN2(mantissa, src_width); - ubound = ((unsigned long long)1 << n); - mask = ubound - 1; - scale = (double)ubound/mask; - bias = (double)((unsigned long long)1 << (mantissa - n)); + ubound = ((unsigned long long)1 << n); + mask = ubound - 1; + scale = (double)ubound/mask; + bias = (double)((unsigned long long)1 << (mantissa - n)); - res = src; + res = src; - if(src_width > mantissa) { - int shift = src_width - mantissa; - res = LLVMBuildLShr(builder, res, - lp_build_const_int_vec(gallivm, dst_type, shift), ""); - } + if (src_width > mantissa) { + int shift = src_width - mantissa; + res = LLVMBuildLShr(builder, res, + lp_build_const_int_vec(gallivm, dst_type, shift), ""); + } - bias_ = lp_build_const_vec(gallivm, dst_type, bias); + bias_ = lp_build_const_vec(gallivm, dst_type, bias); - res = LLVMBuildOr(builder, - res, - LLVMBuildBitCast(builder, bias_, int_vec_type, ""), ""); + res = LLVMBuildOr(builder, + res, + LLVMBuildBitCast(builder, bias_, int_vec_type, ""), ""); - res = LLVMBuildBitCast(builder, res, vec_type, ""); + res = LLVMBuildBitCast(builder, res, vec_type, ""); - res = LLVMBuildFSub(builder, res, bias_, ""); - res = LLVMBuildFMul(builder, res, lp_build_const_vec(gallivm, dst_type, scale), ""); + res = LLVMBuildFSub(builder, res, bias_, ""); + res = LLVMBuildFMul(builder, res, lp_build_const_vec(gallivm, dst_type, scale), ""); + } return res; } diff --git a/src/gallium/auxiliary/gallivm/lp_bld_debug.c b/src/gallium/auxiliary/gallivm/lp_bld_debug.c deleted file mode 100644 index 93e56553d7..0000000000 --- a/src/gallium/auxiliary/gallivm/lp_bld_debug.c +++ /dev/null @@ -1,141 +0,0 @@ -/************************************************************************** - * - * Copyright 2009 VMware, Inc. - * All Rights Reserved. - * - * 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 above copyright notice and this permission notice (including the - * next paragraph) shall be included in all copies or substantial portions - * of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS - * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF - * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. - * IN NO EVENT SHALL VMWARE 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. - * - **************************************************************************/ - - -#ifdef HAVE_UDIS86 -#include <udis86.h> -#endif - -#include "util/u_math.h" -#include "util/u_debug.h" -#include "lp_bld_debug.h" - - -/** - * Check alignment. - * - * It is important that this check is not implemented as a macro or inlined - * function, as the compiler assumptions in respect to alignment of global - * and stack variables would often make the check a no op, defeating the - * whole purpose of the exercise. - */ -boolean -lp_check_alignment(const void *ptr, unsigned alignment) -{ - assert(util_is_power_of_two(alignment)); - return ((uintptr_t)ptr & (alignment - 1)) == 0; -} - - -void -lp_disassemble(const void* func) -{ -#ifdef HAVE_UDIS86 - ud_t ud_obj; - uint64_t max_jmp_pc; - uint inst_no; - boolean emit_addrs = TRUE, emit_line_nos = FALSE; - - ud_init(&ud_obj); - - ud_set_input_buffer(&ud_obj, (void*)func, 0xffff); - - max_jmp_pc = (uint64_t) (uintptr_t) func; - ud_set_pc(&ud_obj, max_jmp_pc); - -#ifdef PIPE_ARCH_X86 - ud_set_mode(&ud_obj, 32); -#endif -#ifdef PIPE_ARCH_X86_64 - ud_set_mode(&ud_obj, 64); -#endif - - ud_set_syntax(&ud_obj, UD_SYN_ATT); - - while (ud_disassemble(&ud_obj)) { - - if (emit_addrs) { -#ifdef PIPE_ARCH_X86 - debug_printf("0x%08lx:\t", (unsigned long)ud_insn_off(&ud_obj)); -#endif -#ifdef PIPE_ARCH_X86_64 - debug_printf("0x%016llx:\t", (unsigned long long)ud_insn_off(&ud_obj)); -#endif - } - else if (emit_line_nos) { - debug_printf("%6d:\t", inst_no); - inst_no++; - } -#if 0 - debug_printf("%-16s ", ud_insn_hex(&ud_obj)); -#endif - - debug_printf("%s\n", ud_insn_asm(&ud_obj)); - - if(ud_obj.mnemonic != UD_Icall) { - unsigned i; - for(i = 0; i < 3; ++i) { - const struct ud_operand *op = &ud_obj.operand[i]; - if (op->type == UD_OP_JIMM){ - uint64_t pc = ud_obj.pc; - - switch (op->size) { - case 8: - pc += op->lval.sbyte; - break; - case 16: - pc += op->lval.sword; - break; - case 32: - pc += op->lval.sdword; - break; - default: - break; - } - if(pc > max_jmp_pc) - max_jmp_pc = pc; - } - } - } - - if (ud_obj.mnemonic == UD_Iinvalid || - (ud_insn_off(&ud_obj) >= max_jmp_pc && - (ud_obj.mnemonic == UD_Iret || - ud_obj.mnemonic == UD_Ijmp))) - break; - } - -#if 0 - /* Print GDB command, useful to verify udis86 output */ - debug_printf("disassemble %p %p\n", func, (void*)(uintptr_t)ud_obj.pc); -#endif - - debug_printf("\n"); -#else - (void)func; -#endif -} diff --git a/src/gallium/auxiliary/gallivm/lp_bld_debug.cpp b/src/gallium/auxiliary/gallivm/lp_bld_debug.cpp new file mode 100644 index 0000000000..bb2c82fe0e --- /dev/null +++ b/src/gallium/auxiliary/gallivm/lp_bld_debug.cpp @@ -0,0 +1,357 @@ +/************************************************************************** + * + * Copyright 2009-2011 VMware, Inc. + * All Rights Reserved. + * + * 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 above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE 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. + * + **************************************************************************/ + +#include <llvm-c/Core.h> +#include <llvm/Target/TargetMachine.h> +#include <llvm/Target/TargetRegistry.h> +#include <llvm/Target/TargetSelect.h> +#include <llvm/Target/TargetInstrInfo.h> +#include <llvm/Support/raw_ostream.h> +#include <llvm/Support/MemoryObject.h> +#include <llvm/System/Host.h> + +#if HAVE_LLVM >= 0x0207 +#include <llvm/MC/MCDisassembler.h> +#include <llvm/MC/MCAsmInfo.h> +#include <llvm/MC/MCInst.h> +#include <llvm/MC/MCInstPrinter.h> +#endif /* HAVE_LLVM >= 0x0207 */ + +#include "util/u_math.h" +#include "util/u_debug.h" + +#include "lp_bld_debug.h" + + + +/** + * Check alignment. + * + * It is important that this check is not implemented as a macro or inlined + * function, as the compiler assumptions in respect to alignment of global + * and stack variables would often make the check a no op, defeating the + * whole purpose of the exercise. + */ +extern "C" boolean +lp_check_alignment(const void *ptr, unsigned alignment) +{ + assert(util_is_power_of_two(alignment)); + return ((uintptr_t)ptr & (alignment - 1)) == 0; +} + + +class raw_debug_ostream : + public llvm::raw_ostream +{ + uint64_t pos; + + void write_impl(const char *Ptr, size_t Size); + uint64_t current_pos() { return pos; } + uint64_t current_pos() const { return pos; } + +#if HAVE_LLVM >= 0x207 + uint64_t preferred_buffer_size() { return 512; } +#else + size_t preferred_buffer_size() { return 512; } +#endif +}; + + +void +raw_debug_ostream::write_impl(const char *Ptr, size_t Size) +{ + if (Size > 0) { + char *lastPtr = (char *)&Ptr[Size]; + char last = *lastPtr; + *lastPtr = 0; + _debug_printf("%*s", Size, Ptr); + *lastPtr = last; + pos += Size; + } +} + + +/** + * Same as LLVMDumpValue, but through our debugging channels. + */ +extern "C" void +lp_debug_dump_value(LLVMValueRef value) +{ +#if (defined(PIPE_OS_WINDOWS) && !defined(PIPE_CC_MSVC)) || defined(PIPE_OS_EMBDDED) + raw_debug_ostream os; + llvm::unwrap(value)->print(os); + os.flush(); +#else + LLVMDumpValue(value); +#endif +} + + +#if HAVE_LLVM >= 0x0207 +/* + * MemoryObject wrapper around a buffer of memory, to be used by MC + * disassembler. + */ +class BufferMemoryObject: + public llvm::MemoryObject +{ +private: + const uint8_t *Bytes; + uint64_t Length; +public: + BufferMemoryObject(const uint8_t *bytes, uint64_t length) : + Bytes(bytes), Length(length) + { + } + + uint64_t getBase() const + { + return 0; + } + + uint64_t getExtent() const + { + return Length; + } + + int readByte(uint64_t addr, uint8_t *byte) const + { + if (addr > getExtent()) + return -1; + *byte = Bytes[addr]; + return 0; + } +}; +#endif /* HAVE_LLVM >= 0x0207 */ + + +/* + * Disassemble a function, using the LLVM MC disassembler. + * + * See also: + * - http://blog.llvm.org/2010/01/x86-disassembler.html + * - http://blog.llvm.org/2010/04/intro-to-llvm-mc-project.html + */ +extern "C" void +lp_disassemble(const void* func) +{ +#if HAVE_LLVM >= 0x0207 + using namespace llvm; + + const uint8_t *bytes = (const uint8_t *)func; + + /* + * Limit disassembly to this extent + */ + const uint64_t extent = 0x10000; + + uint64_t max_pc = 0; + + /* + * Initialize all used objects. + */ + + std::string Triple = sys::getHostTriple(); + + std::string Error; + const Target *T = TargetRegistry::lookupTarget(Triple, Error); + +#if HAVE_LLVM >= 0x0208 + InitializeNativeTargetAsmPrinter(); +#else + InitializeAllAsmPrinters(); +#endif + + InitializeAllDisassemblers(); + + OwningPtr<const MCAsmInfo> AsmInfo(T->createAsmInfo(Triple)); + + if (!AsmInfo) { + debug_printf("error: no assembly info for target %s\n", Triple.c_str()); + return; + } + + OwningPtr<const MCDisassembler> DisAsm(T->createMCDisassembler()); + if (!DisAsm) { + debug_printf("error: no disassembler for target %s\n", Triple.c_str()); + return; + } + + raw_debug_ostream Out; + + int AsmPrinterVariant = AsmInfo->getAssemblerDialect(); +#if HAVE_LLVM >= 0x0208 + OwningPtr<MCInstPrinter> Printer( + T->createMCInstPrinter(AsmPrinterVariant, *AsmInfo)); +#else + OwningPtr<MCInstPrinter> Printer( + T->createMCInstPrinter(AsmPrinterVariant, *AsmInfo, Out)); +#endif + if (!Printer) { + debug_printf("error: no instruction printer for target %s\n", Triple.c_str()); + return; + } + + TargetMachine *TM = T->createTargetMachine(Triple, ""); + + const TargetInstrInfo *TII = TM->getInstrInfo(); + + /* + * Wrap the data in a MemoryObject + */ + BufferMemoryObject memoryObject((const uint8_t *)bytes, extent); + + uint64_t pc; + pc = 0; + while (true) { + MCInst Inst; + uint64_t Size; + + /* + * Print address. We use addresses relative to the start of the function, + * so that between runs. + */ + + debug_printf("%6lu:\t", (unsigned long)pc); + + if (!DisAsm->getInstruction(Inst, Size, memoryObject, + pc, + nulls())) { + debug_printf("invalid\n"); + pc += 1; + } + + /* + * Output the bytes in hexidecimal format. + */ + + if (0) { + unsigned i; + for (i = 0; i < Size; ++i) { + debug_printf("%02x ", ((const uint8_t*)bytes)[pc + i]); + } + for (; i < 16; ++i) { + debug_printf(" "); + } + } + + /* + * Print the instruction. + */ + +#if HAVE_LLVM >= 0x208 + Printer->printInst(&Inst, Out); +#else + Printer->printInst(&Inst); +#endif + Out.flush(); + + /* + * Advance. + */ + + pc += Size; + + const TargetInstrDesc &TID = TII->get(Inst.getOpcode()); + + /* + * Keep track of forward jumps to a nearby address. + */ + + if (TID.isBranch()) { + for (unsigned i = 0; i < Inst.getNumOperands(); ++i) { + const MCOperand &operand = Inst.getOperand(i); + if (operand.isImm()) { + uint64_t jump; + + /* + * FIXME: Handle both relative and absolute addresses correctly. + * EDInstInfo actually has this info, but operandTypes and + * operandFlags enums are not exposed in the public interface. + */ + + if (1) { + /* + * PC relative addr. + */ + + jump = pc + operand.getImm(); + } else { + /* + * Absolute addr. + */ + + jump = (uint64_t)operand.getImm(); + } + + /* + * Output the address relative to the function start, given + * that MC will print the addresses relative the current pc. + */ + debug_printf("\t\t; %lu", (unsigned long)jump); + + /* + * Ignore far jumps given it could be actually a tail return to + * a random address. + */ + + if (jump > max_pc && + jump < extent) { + max_pc = jump; + } + } + } + } + + debug_printf("\n"); + + /* + * Stop disassembling on return statements, if there is no record of a + * jump to a successive address. + */ + + if (TID.isReturn()) { + if (pc > max_pc) { + break; + } + } + } + + /* + * Print GDB command, useful to verify output. + */ + + if (0) { + debug_printf("disassemble %p %p\n", bytes, bytes + pc); + } + + debug_printf("\n"); +#else /* HAVE_LLVM < 0x0207 */ + (void)func; +#endif /* HAVE_LLVM < 0x0207 */ +} + diff --git a/src/gallium/auxiliary/gallivm/lp_bld_debug.h b/src/gallium/auxiliary/gallivm/lp_bld_debug.h index 8a58f95b78..da873f30b2 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_debug.h +++ b/src/gallium/auxiliary/gallivm/lp_bld_debug.h @@ -45,6 +45,11 @@ #define GALLIVM_DEBUG_GC (1 << 6) +#ifdef __cplusplus +extern "C" { +#endif + + #ifdef DEBUG extern unsigned gallivm_debug; #else @@ -81,4 +86,9 @@ void lp_disassemble(const void* func); +#ifdef __cplusplus +} +#endif + + #endif /* !LP_BLD_DEBUG_H */ diff --git a/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c b/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c index 75d2e666f0..82ab19eda1 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c @@ -36,6 +36,7 @@ #include "util/u_format.h" #include "util/u_memory.h" #include "util/u_math.h" +#include "util/u_pointer.h" #include "util/u_string.h" #include "lp_bld_arit.h" @@ -511,8 +512,6 @@ lp_build_fetch_rgba_aos(struct gallivm_state *gallivm, * or incentive to optimize. */ - LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder))); - char name[256]; LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context); LLVMTypeRef pi8t = LLVMPointerType(i8t, 0); LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context); @@ -522,19 +521,20 @@ lp_build_fetch_rgba_aos(struct gallivm_state *gallivm, LLVMValueRef res; unsigned k; - util_snprintf(name, sizeof name, "util_format_%s_fetch_rgba_8unorm", - format_desc->short_name); - if (gallivm_debug & GALLIVM_DEBUG_PERF) { - debug_printf("%s: falling back to %s\n", __FUNCTION__, name); + debug_printf("%s: falling back to util_format_%s_fetch_rgba_8unorm\n", + __FUNCTION__, format_desc->short_name); } /* * Declare and bind format_desc->fetch_rgba_8unorm(). */ - function = LLVMGetNamedFunction(module, name); - if (!function) { + { + /* + * Function to call looks like: + * fetch(uint8_t *dst, const uint8_t *src, unsigned i, unsigned j) + */ LLVMTypeRef ret_type; LLVMTypeRef arg_types[4]; LLVMTypeRef function_type; @@ -542,17 +542,19 @@ lp_build_fetch_rgba_aos(struct gallivm_state *gallivm, ret_type = LLVMVoidTypeInContext(gallivm->context); arg_types[0] = pi8t; arg_types[1] = pi8t; - arg_types[3] = arg_types[2] = LLVMIntTypeInContext(gallivm->context, sizeof(unsigned) * 8); - function_type = LLVMFunctionType(ret_type, arg_types, Elements(arg_types), 0); - function = LLVMAddFunction(module, name, function_type); - - LLVMSetFunctionCallConv(function, LLVMCCallConv); - LLVMSetLinkage(function, LLVMExternalLinkage); - - assert(LLVMIsDeclaration(function)); - - LLVMAddGlobalMapping(gallivm->engine, function, - func_to_pointer((func_pointer)format_desc->fetch_rgba_8unorm)); + arg_types[2] = i32t; + arg_types[3] = i32t; + function_type = LLVMFunctionType(ret_type, arg_types, + Elements(arg_types), 0); + + /* make const pointer for the C fetch_rgba_8unorm function */ + function = lp_build_const_int_pointer(gallivm, + func_to_pointer((func_pointer) format_desc->fetch_rgba_8unorm)); + + /* cast the callee pointer to the function's type */ + function = LLVMBuildBitCast(builder, function, + LLVMPointerType(function_type, 0), + "cast callee"); } tmp_ptr = lp_build_alloca(gallivm, i32t, ""); @@ -614,48 +616,55 @@ lp_build_fetch_rgba_aos(struct gallivm_state *gallivm, * or incentive to optimize. */ - LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder))); - char name[256]; LLVMTypeRef f32t = LLVMFloatTypeInContext(gallivm->context); LLVMTypeRef f32x4t = LLVMVectorType(f32t, 4); LLVMTypeRef pf32t = LLVMPointerType(f32t, 0); + LLVMTypeRef pi8t = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0); + LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context); LLVMValueRef function; LLVMValueRef tmp_ptr; LLVMValueRef tmps[LP_MAX_VECTOR_LENGTH/4]; LLVMValueRef res; unsigned k; - util_snprintf(name, sizeof name, "util_format_%s_fetch_rgba_float", - format_desc->short_name); - if (gallivm_debug & GALLIVM_DEBUG_PERF) { - debug_printf("%s: falling back to %s\n", __FUNCTION__, name); + debug_printf("%s: falling back to util_format_%s_fetch_rgba_float\n", + __FUNCTION__, format_desc->short_name); } /* * Declare and bind format_desc->fetch_rgba_float(). */ - function = LLVMGetNamedFunction(module, name); - if (!function) { + { + /* + * Function to call looks like: + * fetch(float *dst, const uint8_t *src, unsigned i, unsigned j) + */ LLVMTypeRef ret_type; LLVMTypeRef arg_types[4]; LLVMTypeRef function_type; ret_type = LLVMVoidTypeInContext(gallivm->context); arg_types[0] = pf32t; - arg_types[1] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0); - arg_types[3] = arg_types[2] = LLVMIntTypeInContext(gallivm->context, sizeof(unsigned) * 8); - function_type = LLVMFunctionType(ret_type, arg_types, Elements(arg_types), 0); - function = LLVMAddFunction(module, name, function_type); + arg_types[1] = pi8t; + arg_types[2] = i32t; + arg_types[3] = i32t; + function_type = LLVMFunctionType(ret_type, arg_types, + Elements(arg_types), 0); - LLVMSetFunctionCallConv(function, LLVMCCallConv); - LLVMSetLinkage(function, LLVMExternalLinkage); + /* Note: we're using this casting here instead of LLVMAddGlobalMapping() + * to work around a bug in LLVM 2.6, and for efficiency/simplicity. + */ - assert(LLVMIsDeclaration(function)); + /* make const pointer for the C fetch_rgba_float function */ + function = lp_build_const_int_pointer(gallivm, + func_to_pointer((func_pointer) format_desc->fetch_rgba_float)); - LLVMAddGlobalMapping(gallivm->engine, function, - func_to_pointer((func_pointer)format_desc->fetch_rgba_float)); + /* cast the callee pointer to the function's type */ + function = LLVMBuildBitCast(builder, function, + LLVMPointerType(function_type, 0), + "cast callee"); } tmp_ptr = lp_build_alloca(gallivm, f32x4t, ""); diff --git a/src/gallium/auxiliary/gallivm/lp_bld_init.c b/src/gallium/auxiliary/gallivm/lp_bld_init.c index efe8d38b8f..45addee8fa 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_init.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_init.c @@ -30,6 +30,7 @@ #include "util/u_cpu_detect.h" #include "util/u_debug.h" #include "util/u_memory.h" +#include "util/u_simple_list.h" #include "lp_bld_debug.h" #include "lp_bld_init.h" @@ -221,11 +222,12 @@ free_gallivm_state(struct gallivm_state *gallivm) static boolean init_gallivm_state(struct gallivm_state *gallivm) { - assert(gallivm_initialized); assert(!gallivm->context); assert(!gallivm->module); assert(!gallivm->provider); + lp_build_init(); + gallivm->context = LLVMContextCreate(); if (!gallivm->context) goto fail; @@ -291,12 +293,12 @@ struct callback { garbage_collect_callback_func func; void *cb_data; + struct callback *prev, *next; }; -#define MAX_CALLBACKS 32 -static struct callback Callbacks[MAX_CALLBACKS]; -static unsigned NumCallbacks = 0; +/** list of all garbage collector callbacks */ +static struct callback callback_list = {NULL, NULL, NULL, NULL}; /** @@ -307,20 +309,24 @@ void gallivm_register_garbage_collector_callback(garbage_collect_callback_func func, void *cb_data) { - unsigned i; + struct callback *cb; + + if (!callback_list.prev) { + make_empty_list(&callback_list); + } - for (i = 0; i < NumCallbacks; i++) { - if (Callbacks[i].func == func && Callbacks[i].cb_data == cb_data) { - /* already in list: no-op */ + /* see if already in list */ + foreach(cb, &callback_list) { + if (cb->func == func && cb->cb_data == cb_data) return; - } } - assert(NumCallbacks < MAX_CALLBACKS); - if (NumCallbacks < MAX_CALLBACKS) { - Callbacks[NumCallbacks].func = func; - Callbacks[NumCallbacks].cb_data = cb_data; - NumCallbacks++; + /* add to list */ + cb = CALLOC_STRUCT(callback); + if (cb) { + cb->func = func; + cb->cb_data = cb_data; + insert_at_head(&callback_list, cb); } } @@ -332,15 +338,13 @@ void gallivm_remove_garbage_collector_callback(garbage_collect_callback_func func, void *cb_data) { - unsigned i; - - for (i = 0; i < NumCallbacks; i++) { - if (Callbacks[i].func == func && Callbacks[i].cb_data == cb_data) { - /* found, now remove it */ - NumCallbacks--; - for ( ; i < NumCallbacks; i++) { - Callbacks[i] = Callbacks[i + 1]; - } + struct callback *cb; + + /* search list */ + foreach(cb, &callback_list) { + if (cb->func == func && cb->cb_data == cb_data) { + /* found, remove it */ + remove_from_list(cb); return; } } @@ -354,10 +358,9 @@ gallivm_remove_garbage_collector_callback(garbage_collect_callback_func func, static void call_garbage_collector_callbacks(void) { - unsigned i; - - for (i = 0; i < NumCallbacks; i++) { - Callbacks[i].func(Callbacks[i].cb_data); + struct callback *cb; + foreach(cb, &callback_list) { + cb->func(cb->cb_data); } } @@ -385,6 +388,9 @@ gallivm_garbage_collect(struct gallivm_state *gallivm) void lp_build_init(void) { + if (gallivm_initialized) + return; + #ifdef DEBUG gallivm_debug = debug_get_option_gallivm_debug(); #endif diff --git a/src/gallium/auxiliary/gallivm/lp_bld_misc.cpp b/src/gallium/auxiliary/gallivm/lp_bld_misc.cpp index f56ddee7fd..843a14a500 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_misc.cpp +++ b/src/gallium/auxiliary/gallivm/lp_bld_misc.cpp @@ -46,66 +46,6 @@ #include "util/u_debug.h" -#if (defined(PIPE_OS_WINDOWS) && !defined(PIPE_CC_MSVC)) || defined(PIPE_OS_EMBDDED) - -#include "llvm/Support/raw_ostream.h" - -class raw_debug_ostream : - public llvm::raw_ostream -{ - uint64_t pos; - - void write_impl(const char *Ptr, size_t Size); - uint64_t current_pos() { return pos; } - uint64_t current_pos() const { return pos; } - -#if HAVE_LLVM >= 0x207 - uint64_t preferred_buffer_size() { return 512; } -#else - size_t preferred_buffer_size() { return 512; } -#endif -}; - - -void -raw_debug_ostream::write_impl(const char *Ptr, size_t Size) -{ - if (Size > 0) { - char *lastPtr = (char *)&Ptr[Size]; - char last = *lastPtr; - *lastPtr = 0; - _debug_printf("%*s", Size, Ptr); - *lastPtr = last; - pos += Size; - } -} - - -/** - * Same as LLVMDumpValue, but through our debugging channels. - */ -extern "C" void -lp_debug_dump_value(LLVMValueRef value) -{ - raw_debug_ostream os; - llvm::unwrap(value)->print(os); - os.flush(); -} - - -#else - - -extern "C" void -lp_debug_dump_value(LLVMValueRef value) -{ - LLVMDumpValue(value); -} - - -#endif - - /** * Register the engine with oprofile. * @@ -144,6 +84,7 @@ lp_set_target_options(void) llvm::UnsafeFPMath = true; #endif +#if HAVE_LLVM < 0x0209 /* * LLVM will generate MMX instructions for vectors <= 64 bits, leading to * innefficient code, and in 32bit systems, to the corruption of the FPU @@ -162,6 +103,7 @@ lp_set_target_options(void) llvm::cl::ParseCommandLineOptions(2, const_cast<char**>(options)); first = FALSE; } +#endif /* * By default LLVM adds a signal handler to output a pretty stack trace. diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample_aos.c b/src/gallium/auxiliary/gallivm/lp_bld_sample_aos.c index 991f6fa5ef..e61cf9541e 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_sample_aos.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_sample_aos.c @@ -1098,6 +1098,4 @@ lp_build_sample_aos(struct lp_build_sample_context *bld, texel_out[2] = unswizzled[2]; texel_out[3] = unswizzled[3]; } - - apply_sampler_swizzle(bld, texel_out); } diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c index cf46e2be83..9961ba08f3 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c @@ -187,8 +187,6 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld, border_chan_vec, texel_out[chan]); } } - - apply_sampler_swizzle(bld, texel_out); } @@ -834,14 +832,14 @@ lp_build_sample_mipmap(struct lp_build_sample_context *bld, LLVMValueRef *colors_out) { LLVMBuilderRef builder = bld->gallivm->builder; - LLVMValueRef size0; - LLVMValueRef size1; - LLVMValueRef row_stride0_vec; - LLVMValueRef row_stride1_vec; - LLVMValueRef img_stride0_vec; - LLVMValueRef img_stride1_vec; - LLVMValueRef data_ptr0; - LLVMValueRef data_ptr1; + LLVMValueRef size0 = NULL; + LLVMValueRef size1 = NULL; + LLVMValueRef row_stride0_vec = NULL; + LLVMValueRef row_stride1_vec = NULL; + LLVMValueRef img_stride0_vec = NULL; + LLVMValueRef img_stride1_vec = NULL; + LLVMValueRef data_ptr0 = NULL; + LLVMValueRef data_ptr1 = NULL; LLVMValueRef colors0[4], colors1[4]; unsigned chan; @@ -1110,6 +1108,11 @@ lp_build_sample_compare(struct lp_build_sample_context *bld, coord, tex); } + /* Clamp p coords to [0,1] */ + p = lp_build_clamp(&bld->coord_bld, p, + bld->coord_bld.zero, + bld->coord_bld.one); + /* result = (p FUNC texel) ? 1 : 0 */ res = lp_build_cmp(texel_bld, bld->static_state->compare_func, p, texel[chan]); @@ -1268,4 +1271,6 @@ lp_build_sample_soa(struct gallivm_state *gallivm, } lp_build_sample_compare(&bld, r, texel_out); + + apply_sampler_swizzle(&bld, texel_out); } diff --git a/src/gallium/auxiliary/gallivm/lp_bld_tgsi.h b/src/gallium/auxiliary/gallivm/lp_bld_tgsi.h index 40186befb9..9713d10048 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_tgsi.h +++ b/src/gallium/auxiliary/gallivm/lp_bld_tgsi.h @@ -180,6 +180,7 @@ lp_build_tgsi_soa(struct gallivm_state *gallivm, struct lp_type type, struct lp_build_mask_context *mask, LLVMValueRef consts_ptr, + LLVMValueRef system_values_array, const LLVMValueRef *pos, const LLVMValueRef (*inputs)[4], LLVMValueRef (*outputs)[4], @@ -199,4 +200,11 @@ lp_build_tgsi_aos(struct gallivm_state *gallivm, const struct tgsi_shader_info *info); +LLVMValueRef +lp_build_system_values_array(struct gallivm_state *gallivm, + const struct tgsi_shader_info *info, + LLVMValueRef instance_id, + LLVMValueRef facing); + + #endif /* LP_BLD_TGSI_H */ diff --git a/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c index 1b5a8a5903..d1585c8e2b 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c @@ -157,6 +157,8 @@ struct lp_build_tgsi_soa_context */ LLVMValueRef inputs_array; + LLVMValueRef system_values_array; + const struct tgsi_shader_info *info; /** bitmask indicating which register files are accessed indirectly */ unsigned indirect_files; @@ -759,6 +761,23 @@ emit_fetch( } break; + case TGSI_FILE_SYSTEM_VALUE: + assert(!reg->Register.Indirect); + { + LLVMValueRef index; /* index into the system value array */ + LLVMValueRef scalar, scalar_ptr; + + index = lp_build_const_int32(gallivm, + reg->Register.Index * 4 + swizzle); + + scalar_ptr = LLVMBuildGEP(builder, bld->system_values_array, + &index, 1, ""); + scalar = LLVMBuildLoad(builder, scalar_ptr, ""); + + res = lp_build_broadcast_scalar(&bld->base, scalar); + } + break; + default: assert(0 && "invalid src register in emit_fetch()"); return bld->base.undef; @@ -2322,6 +2341,7 @@ lp_build_tgsi_soa(struct gallivm_state *gallivm, struct lp_type type, struct lp_build_mask_context *mask, LLVMValueRef consts_ptr, + LLVMValueRef system_values_array, const LLVMValueRef *pos, const LLVMValueRef (*inputs)[NUM_CHANNELS], LLVMValueRef (*outputs)[NUM_CHANNELS], @@ -2411,6 +2431,8 @@ lp_build_tgsi_soa(struct gallivm_state *gallivm, } } + bld.system_values_array = system_values_array; + tgsi_parse_init( &parse, tokens ); while( !tgsi_parse_end_of_tokens( &parse ) ) { @@ -2512,3 +2534,54 @@ lp_build_tgsi_soa(struct gallivm_state *gallivm, FREE( bld.instructions ); } + +/** + * Build up the system values array out of individual values such as + * the instance ID, front-face, primitive ID, etc. The shader info is + * used to determine which system values are needed and where to put + * them in the system values array. + * + * XXX only instance ID is implemented at this time. + * + * The system values register file is similar to the constants buffer. + * Example declaration: + * DCL SV[0], INSTANCEID + * Example instruction: + * MOVE foo, SV[0].xxxx; + * + * \return LLVM float array (interpreted as float [][4]) + */ +LLVMValueRef +lp_build_system_values_array(struct gallivm_state *gallivm, + const struct tgsi_shader_info *info, + LLVMValueRef instance_id, + LLVMValueRef facing) +{ + LLVMValueRef size = lp_build_const_int32(gallivm, 4 * info->num_system_values); + LLVMTypeRef float_t = LLVMFloatTypeInContext(gallivm->context); + LLVMValueRef array = lp_build_array_alloca(gallivm, float_t, + size, "sysvals_array"); + unsigned i; + + for (i = 0; i < info->num_system_values; i++) { + LLVMValueRef index = lp_build_const_int32(gallivm, i * 4); + LLVMValueRef ptr, value; + + switch (info->system_value_semantic_name[i]) { + case TGSI_SEMANTIC_INSTANCEID: + /* convert instance ID from int to float */ + value = LLVMBuildSIToFP(gallivm->builder, instance_id, float_t, + "sysval_instanceid"); + break; + case TGSI_SEMANTIC_FACE: + /* fall-through */ + default: + assert(0 && "unexpected semantic in build_system_values_array()"); + } + + ptr = LLVMBuildGEP(gallivm->builder, array, &index, 1, ""); + LLVMBuildStore(gallivm->builder, value, ptr); + } + + return array; +} |