1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
29 #include "pipe/p_config.h"
30 #include "pipe/p_compiler.h"
31 #include "util/u_cpu_detect.h"
32 #include "util/u_debug.h"
33 #include "util/u_memory.h"
34 #include "util/u_simple_list.h"
35 #include "os/os_time.h"
37 #include "lp_bld_debug.h"
38 #include "lp_bld_misc.h"
39 #include "lp_bld_init.h"
41 #include <llvm-c/Analysis.h>
42 #include <llvm-c/Transforms/Scalar.h>
43 #include <llvm-c/BitWriter.h>
46 /* Only MCJIT is available as of LLVM SVN r216982 */
47 #if HAVE_LLVM >= 0x0306
55 * AVX is supported in:
56 * - standard JIT from LLVM 3.2 onwards
57 * - MC-JIT from LLVM 3.1
58 * - MC-JIT supports limited OSes (MacOSX and Linux)
59 * - standard JIT in LLVM 3.1, with backports
61 #if defined(PIPE_ARCH_PPC_64) || defined(PIPE_ARCH_S390) || defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
64 #elif HAVE_LLVM >= 0x0302 || (HAVE_LLVM == 0x0301 && defined(HAVE_JIT_AVX_SUPPORT))
67 #elif HAVE_LLVM == 0x0301 && (defined(PIPE_OS_LINUX) || defined(PIPE_OS_APPLE))
75 #endif /* HAVE_LLVM >= 0x0306 */
78 void LLVMLinkInMCJIT();
82 unsigned gallivm_debug
= 0;
84 static const struct debug_named_value lp_bld_debug_flags
[] = {
85 { "tgsi", GALLIVM_DEBUG_TGSI
, NULL
},
86 { "ir", GALLIVM_DEBUG_IR
, NULL
},
87 { "asm", GALLIVM_DEBUG_ASM
, NULL
},
88 { "nopt", GALLIVM_DEBUG_NO_OPT
, NULL
},
89 { "perf", GALLIVM_DEBUG_PERF
, NULL
},
90 { "no_brilinear", GALLIVM_DEBUG_NO_BRILINEAR
, NULL
},
91 { "no_rho_approx", GALLIVM_DEBUG_NO_RHO_APPROX
, NULL
},
92 { "no_quad_lod", GALLIVM_DEBUG_NO_QUAD_LOD
, NULL
},
93 { "gc", GALLIVM_DEBUG_GC
, NULL
},
97 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug
, "GALLIVM_DEBUG", lp_bld_debug_flags
, 0)
101 static boolean gallivm_initialized
= FALSE
;
103 unsigned lp_native_vector_width
;
107 * Optimization values are:
110 * - 2: Default (-O2, -Os)
111 * - 3: Aggressive (-O3)
113 * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
115 enum LLVM_CodeGenOpt_Level
{
124 * Create the LLVM (optimization) pass manager and install
125 * relevant optimization passes.
126 * \return TRUE for success, FALSE for failure
129 create_pass_manager(struct gallivm_state
*gallivm
)
132 assert(!gallivm
->passmgr
);
133 assert(gallivm
->target
);
135 gallivm
->passmgr
= LLVMCreateFunctionPassManagerForModule(gallivm
->module
);
136 if (!gallivm
->passmgr
)
139 // Old versions of LLVM get the DataLayout from the pass manager.
140 LLVMAddTargetData(gallivm
->target
, gallivm
->passmgr
);
142 // New ones from the Module.
143 td_str
= LLVMCopyStringRepOfTargetData(gallivm
->target
);
144 LLVMSetDataLayout(gallivm
->module
, td_str
);
147 if ((gallivm_debug
& GALLIVM_DEBUG_NO_OPT
) == 0) {
148 /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
149 * but there are more on SVN.
150 * TODO: Add more passes.
152 LLVMAddScalarReplAggregatesPass(gallivm
->passmgr
);
153 LLVMAddLICMPass(gallivm
->passmgr
);
154 LLVMAddCFGSimplificationPass(gallivm
->passmgr
);
155 LLVMAddReassociatePass(gallivm
->passmgr
);
156 LLVMAddPromoteMemoryToRegisterPass(gallivm
->passmgr
);
157 LLVMAddConstantPropagationPass(gallivm
->passmgr
);
158 LLVMAddInstructionCombiningPass(gallivm
->passmgr
);
159 LLVMAddGVNPass(gallivm
->passmgr
);
162 /* We need at least this pass to prevent the backends to fail in
165 LLVMAddPromoteMemoryToRegisterPass(gallivm
->passmgr
);
173 * Free gallivm object's LLVM allocations, but not any generated code
174 * nor the gallivm object itself.
177 gallivm_free_ir(struct gallivm_state
*gallivm
)
179 if (gallivm
->passmgr
) {
180 LLVMDisposePassManager(gallivm
->passmgr
);
183 if (gallivm
->engine
) {
184 /* This will already destroy any associated module */
185 LLVMDisposeExecutionEngine(gallivm
->engine
);
186 } else if (gallivm
->module
) {
187 LLVMDisposeModule(gallivm
->module
);
191 /* Don't free the TargetData, it's owned by the exec engine */
193 if (gallivm
->target
) {
194 LLVMDisposeTargetData(gallivm
->target
);
198 if (gallivm
->builder
)
199 LLVMDisposeBuilder(gallivm
->builder
);
201 /* The LLVMContext should be owned by the parent of gallivm. */
203 gallivm
->engine
= NULL
;
204 gallivm
->target
= NULL
;
205 gallivm
->module
= NULL
;
206 gallivm
->passmgr
= NULL
;
207 gallivm
->context
= NULL
;
208 gallivm
->builder
= NULL
;
213 * Free LLVM-generated code. Should be done AFTER gallivm_free_ir().
216 gallivm_free_code(struct gallivm_state
*gallivm
)
218 assert(!gallivm
->module
);
219 assert(!gallivm
->engine
);
220 lp_free_generated_code(gallivm
->code
);
221 gallivm
->code
= NULL
;
222 lp_free_memory_manager(gallivm
->memorymgr
);
223 gallivm
->memorymgr
= NULL
;
228 init_gallivm_engine(struct gallivm_state
*gallivm
)
231 enum LLVM_CodeGenOpt_Level optlevel
;
235 if (gallivm_debug
& GALLIVM_DEBUG_NO_OPT
) {
242 ret
= lp_build_create_jit_compiler_for_module(&gallivm
->engine
,
250 _debug_printf("%s\n", error
);
251 LLVMDisposeMessage(error
);
257 gallivm
->target
= LLVMGetExecutionEngineTargetData(gallivm
->engine
);
258 if (!gallivm
->target
)
263 * Dump the data layout strings.
266 LLVMTargetDataRef target
= LLVMGetExecutionEngineTargetData(gallivm
->engine
);
268 char *engine_data_layout
;
270 data_layout
= LLVMCopyStringRepOfTargetData(gallivm
->target
);
271 engine_data_layout
= LLVMCopyStringRepOfTargetData(target
);
274 debug_printf("module target data = %s\n", data_layout
);
275 debug_printf("engine target data = %s\n", engine_data_layout
);
279 free(engine_data_layout
);
291 * Allocate gallivm LLVM objects.
292 * \return TRUE for success, FALSE for failure
295 init_gallivm_state(struct gallivm_state
*gallivm
, const char *name
,
296 LLVMContextRef context
)
298 assert(!gallivm
->context
);
299 assert(!gallivm
->module
);
301 if (!lp_build_init())
304 gallivm
->context
= context
;
306 if (!gallivm
->context
)
309 gallivm
->module
= LLVMModuleCreateWithNameInContext(name
,
311 if (!gallivm
->module
)
314 gallivm
->builder
= LLVMCreateBuilderInContext(gallivm
->context
);
315 if (!gallivm
->builder
)
318 #if HAVE_LLVM < 0x0306
319 gallivm
->memorymgr
= lp_get_default_memory_manager();
320 if (!gallivm
->memorymgr
)
323 gallivm
->memorymgr
= 0;
326 /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
327 * complete when MC-JIT is created. So defer the MC-JIT engine creation for
331 if (!init_gallivm_engine(gallivm
)) {
336 * MC-JIT engine compiles the module immediately on creation, so we can't
337 * obtain the target data from it. Instead we create a target data layout
340 * The produced layout strings are not precisely the same, but should make
341 * no difference for the kind of optimization passes we run.
343 * For reference this is the layout string on x64:
345 * e-p:64:64:64-S128-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f16:16:16-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-f128:128:128-n8:16:32:64
348 * - http://llvm.org/docs/LangRef.html#datalayout
352 const unsigned pointer_size
= 8 * sizeof(void *);
354 util_snprintf(layout
, sizeof layout
, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
355 #ifdef PIPE_ARCH_LITTLE_ENDIAN
356 'e', // little endian
360 pointer_size
, pointer_size
, pointer_size
, // pointer size, abi alignment, preferred alignment
361 pointer_size
, // aggregate preferred alignment
362 pointer_size
, pointer_size
); // stack objects abi alignment, preferred alignment
364 gallivm
->target
= LLVMCreateTargetData(layout
);
365 if (!gallivm
->target
) {
371 if (!create_pass_manager(gallivm
))
377 gallivm_free_ir(gallivm
);
378 gallivm_free_code(gallivm
);
386 if (gallivm_initialized
)
389 /* XXX: Remove this once lp_bld_misc.cpp has been adapted to the removal
390 * of JITMemoryManager
392 #if HAVE_LLVM >= 0x0306
397 gallivm_debug
= debug_get_option_gallivm_debug();
400 lp_set_target_options();
410 /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
411 * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
412 * actually more efficient to use 4-wide vectors on this processor.
415 * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
418 util_cpu_caps
.has_avx
&&
419 util_cpu_caps
.has_intel
) {
420 lp_native_vector_width
= 256;
422 /* Leave it at 128, even when no SIMD extensions are available.
423 * Really needs to be a multiple of 128 so can fit 4 floats.
425 lp_native_vector_width
= 128;
428 lp_native_vector_width
= debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
429 lp_native_vector_width
);
431 if (lp_native_vector_width
<= 128) {
432 /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
433 * "util_cpu_caps.has_avx" predicate, and lack the
434 * "lp_native_vector_width > 128" predicate. And also to ensure a more
435 * consistent behavior, allowing one to test SSE2 on AVX machines.
436 * XXX: should not play games with util_cpu_caps directly as it might
437 * get used for other things outside llvm too.
439 util_cpu_caps
.has_avx
= 0;
440 util_cpu_caps
.has_avx2
= 0;
445 * note these instructions are VEX-only, so can only emit if we use
446 * avx (don't want to base it on has_avx & has_f16c later as that would
447 * omit it unnecessarily on amd cpus, see above).
449 util_cpu_caps
.has_f16c
= 0;
450 util_cpu_caps
.has_xop
= 0;
453 #ifdef PIPE_ARCH_PPC_64
454 /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
455 * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
456 * that some rounding and half-float to float handling does not round
458 * XXX: should eventually follow same logic on all platforms.
459 * Right now denorms get explicitly disabled (but elsewhere) for x86,
460 * whereas ppc64 explicitly enables them...
462 if (util_cpu_caps
.has_altivec
) {
463 unsigned short mask
[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
464 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
475 gallivm_initialized
= TRUE
;
478 /* For simulating less capable machines */
479 util_cpu_caps
.has_sse3
= 0;
480 util_cpu_caps
.has_ssse3
= 0;
481 util_cpu_caps
.has_sse4_1
= 0;
482 util_cpu_caps
.has_avx
= 0;
483 util_cpu_caps
.has_f16c
= 0;
492 * Create a new gallivm_state object.
494 struct gallivm_state
*
495 gallivm_create(const char *name
, LLVMContextRef context
)
497 struct gallivm_state
*gallivm
;
499 gallivm
= CALLOC_STRUCT(gallivm_state
);
501 if (!init_gallivm_state(gallivm
, name
, context
)) {
512 * Destroy a gallivm_state object.
515 gallivm_destroy(struct gallivm_state
*gallivm
)
517 gallivm_free_ir(gallivm
);
518 gallivm_free_code(gallivm
);
524 * Validate a function.
525 * Verification is only done with debug builds.
528 gallivm_verify_function(struct gallivm_state
*gallivm
,
531 /* Verify the LLVM IR. If invalid, dump and abort */
533 if (LLVMVerifyFunction(func
, LLVMPrintMessageAction
)) {
534 lp_debug_dump_value(func
);
540 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
541 /* Print the LLVM IR to stderr */
542 lp_debug_dump_value(func
);
550 * This does IR optimization on all functions in the module.
553 gallivm_compile_module(struct gallivm_state
*gallivm
)
558 assert(!gallivm
->compiled
);
560 if (gallivm
->builder
) {
561 LLVMDisposeBuilder(gallivm
->builder
);
562 gallivm
->builder
= NULL
;
565 if (gallivm_debug
& GALLIVM_DEBUG_PERF
)
566 time_begin
= os_time_get();
568 /* Run optimization passes */
569 LLVMInitializeFunctionPassManager(gallivm
->passmgr
);
570 func
= LLVMGetFirstFunction(gallivm
->module
);
573 debug_printf("optimizing func %s...\n", LLVMGetValueName(func
));
575 LLVMRunFunctionPassManager(gallivm
->passmgr
, func
);
576 func
= LLVMGetNextFunction(func
);
578 LLVMFinalizeFunctionPassManager(gallivm
->passmgr
);
580 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
581 int64_t time_end
= os_time_get();
582 int time_msec
= (int)(time_end
- time_begin
) / 1000;
583 debug_printf("optimizing module %s took %d msec\n",
584 lp_get_module_id(gallivm
->module
), time_msec
);
587 /* Dump byte code to a file */
589 LLVMWriteBitcodeToFile(gallivm
->module
, "llvmpipe.bc");
590 debug_printf("llvmpipe.bc written\n");
591 debug_printf("Invoke as \"llc -o - llvmpipe.bc\"\n");
595 assert(!gallivm
->engine
);
596 if (!init_gallivm_engine(gallivm
)) {
600 assert(gallivm
->engine
);
608 gallivm_jit_function(struct gallivm_state
*gallivm
,
612 func_pointer jit_func
;
614 assert(gallivm
->compiled
);
615 assert(gallivm
->engine
);
617 code
= LLVMGetPointerToGlobal(gallivm
->engine
, func
);
619 jit_func
= pointer_to_func(code
);
621 if (gallivm_debug
& GALLIVM_DEBUG_ASM
) {
622 lp_disassemble(func
, code
);
626 lp_profile(func
, code
);