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/simple_list.h"
35 #include "util/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 #if HAVE_LLVM >= 0x0700
44 #include <llvm-c/Transforms/Utils.h>
46 #include <llvm-c/BitWriter.h>
49 /* Only MCJIT is available as of LLVM SVN r216982 */
50 #if HAVE_LLVM >= 0x0306
52 #elif defined(PIPE_ARCH_PPC_64) || defined(PIPE_ARCH_S390) || defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
56 #if defined(USE_MCJIT)
57 static const bool use_mcjit
= USE_MCJIT
;
59 static bool use_mcjit
= FALSE
;
64 unsigned gallivm_debug
= 0;
66 static const struct debug_named_value lp_bld_debug_flags
[] = {
67 { "tgsi", GALLIVM_DEBUG_TGSI
, NULL
},
68 { "ir", GALLIVM_DEBUG_IR
, NULL
},
69 { "asm", GALLIVM_DEBUG_ASM
, NULL
},
70 { "nopt", GALLIVM_DEBUG_NO_OPT
, NULL
},
71 { "perf", GALLIVM_DEBUG_PERF
, NULL
},
72 { "no_brilinear", GALLIVM_DEBUG_NO_BRILINEAR
, NULL
},
73 { "no_rho_approx", GALLIVM_DEBUG_NO_RHO_APPROX
, NULL
},
74 { "no_quad_lod", GALLIVM_DEBUG_NO_QUAD_LOD
, NULL
},
75 { "gc", GALLIVM_DEBUG_GC
, NULL
},
76 { "dumpbc", GALLIVM_DEBUG_DUMP_BC
, NULL
},
80 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug
, "GALLIVM_DEBUG", lp_bld_debug_flags
, 0)
84 static boolean gallivm_initialized
= FALSE
;
86 unsigned lp_native_vector_width
;
90 * Optimization values are:
93 * - 2: Default (-O2, -Os)
94 * - 3: Aggressive (-O3)
96 * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
98 enum LLVM_CodeGenOpt_Level
{
107 * Create the LLVM (optimization) pass manager and install
108 * relevant optimization passes.
109 * \return TRUE for success, FALSE for failure
112 create_pass_manager(struct gallivm_state
*gallivm
)
114 assert(!gallivm
->passmgr
);
115 assert(gallivm
->target
);
117 gallivm
->passmgr
= LLVMCreateFunctionPassManagerForModule(gallivm
->module
);
118 if (!gallivm
->passmgr
)
121 * TODO: some per module pass manager with IPO passes might be helpful -
122 * the generated texture functions may benefit from inlining if they are
123 * simple, or constant propagation into them, etc.
126 #if HAVE_LLVM < 0x0309
127 // Old versions of LLVM get the DataLayout from the pass manager.
128 LLVMAddTargetData(gallivm
->target
, gallivm
->passmgr
);
133 // New ones from the Module.
134 td_str
= LLVMCopyStringRepOfTargetData(gallivm
->target
);
135 LLVMSetDataLayout(gallivm
->module
, td_str
);
139 if ((gallivm_debug
& GALLIVM_DEBUG_NO_OPT
) == 0) {
141 * TODO: Evaluate passes some more - keeping in mind
142 * both quality of generated code and compile times.
144 LLVMAddScalarReplAggregatesPass(gallivm
->passmgr
);
145 LLVMAddEarlyCSEPass(gallivm
->passmgr
);
146 LLVMAddCFGSimplificationPass(gallivm
->passmgr
);
148 * FIXME: LICM is potentially quite useful. However, for some
149 * rather crazy shaders the compile time can reach _hours_ per shader,
150 * due to licm implying lcssa (since llvm 3.5), which can take forever.
151 * Even for sane shaders, the cost of licm is rather high (and not just
152 * due to lcssa, licm itself too), though mostly only in cases when it
153 * can actually move things, so having to disable it is a pity.
154 * LLVMAddLICMPass(gallivm->passmgr);
156 LLVMAddReassociatePass(gallivm
->passmgr
);
157 LLVMAddPromoteMemoryToRegisterPass(gallivm
->passmgr
);
158 LLVMAddConstantPropagationPass(gallivm
->passmgr
);
159 LLVMAddInstructionCombiningPass(gallivm
->passmgr
);
160 LLVMAddGVNPass(gallivm
->passmgr
);
163 /* We need at least this pass to prevent the backends to fail in
166 LLVMAddPromoteMemoryToRegisterPass(gallivm
->passmgr
);
174 * Free gallivm object's LLVM allocations, but not any generated code
175 * nor the gallivm object itself.
178 gallivm_free_ir(struct gallivm_state
*gallivm
)
180 if (gallivm
->passmgr
) {
181 LLVMDisposePassManager(gallivm
->passmgr
);
184 if (gallivm
->engine
) {
185 /* This will already destroy any associated module */
186 LLVMDisposeExecutionEngine(gallivm
->engine
);
187 } else if (gallivm
->module
) {
188 LLVMDisposeModule(gallivm
->module
);
191 FREE(gallivm
->module_name
);
194 /* Don't free the TargetData, it's owned by the exec engine */
196 if (gallivm
->target
) {
197 LLVMDisposeTargetData(gallivm
->target
);
201 if (gallivm
->builder
)
202 LLVMDisposeBuilder(gallivm
->builder
);
204 /* The LLVMContext should be owned by the parent of gallivm. */
206 gallivm
->engine
= NULL
;
207 gallivm
->target
= NULL
;
208 gallivm
->module
= NULL
;
209 gallivm
->module_name
= NULL
;
210 gallivm
->passmgr
= NULL
;
211 gallivm
->context
= NULL
;
212 gallivm
->builder
= NULL
;
217 * Free LLVM-generated code. Should be done AFTER gallivm_free_ir().
220 gallivm_free_code(struct gallivm_state
*gallivm
)
222 assert(!gallivm
->module
);
223 assert(!gallivm
->engine
);
224 lp_free_generated_code(gallivm
->code
);
225 gallivm
->code
= NULL
;
226 lp_free_memory_manager(gallivm
->memorymgr
);
227 gallivm
->memorymgr
= NULL
;
232 init_gallivm_engine(struct gallivm_state
*gallivm
)
235 enum LLVM_CodeGenOpt_Level optlevel
;
239 if (gallivm_debug
& GALLIVM_DEBUG_NO_OPT
) {
246 ret
= lp_build_create_jit_compiler_for_module(&gallivm
->engine
,
254 _debug_printf("%s\n", error
);
255 LLVMDisposeMessage(error
);
261 gallivm
->target
= LLVMGetExecutionEngineTargetData(gallivm
->engine
);
262 if (!gallivm
->target
)
267 * Dump the data layout strings.
270 LLVMTargetDataRef target
= LLVMGetExecutionEngineTargetData(gallivm
->engine
);
272 char *engine_data_layout
;
274 data_layout
= LLVMCopyStringRepOfTargetData(gallivm
->target
);
275 engine_data_layout
= LLVMCopyStringRepOfTargetData(target
);
278 debug_printf("module target data = %s\n", data_layout
);
279 debug_printf("engine target data = %s\n", engine_data_layout
);
283 free(engine_data_layout
);
295 * Allocate gallivm LLVM objects.
296 * \return TRUE for success, FALSE for failure
299 init_gallivm_state(struct gallivm_state
*gallivm
, const char *name
,
300 LLVMContextRef context
)
302 assert(!gallivm
->context
);
303 assert(!gallivm
->module
);
305 if (!lp_build_init())
308 gallivm
->context
= context
;
310 if (!gallivm
->context
)
313 gallivm
->module_name
= NULL
;
315 size_t size
= strlen(name
) + 1;
316 gallivm
->module_name
= MALLOC(size
);
317 if (gallivm
->module_name
) {
318 memcpy(gallivm
->module_name
, name
, size
);
322 gallivm
->module
= LLVMModuleCreateWithNameInContext(name
,
324 if (!gallivm
->module
)
327 gallivm
->builder
= LLVMCreateBuilderInContext(gallivm
->context
);
328 if (!gallivm
->builder
)
331 gallivm
->memorymgr
= lp_get_default_memory_manager();
332 if (!gallivm
->memorymgr
)
335 /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
336 * complete when MC-JIT is created. So defer the MC-JIT engine creation for
340 if (!init_gallivm_engine(gallivm
)) {
345 * MC-JIT engine compiles the module immediately on creation, so we can't
346 * obtain the target data from it. Instead we create a target data layout
349 * The produced layout strings are not precisely the same, but should make
350 * no difference for the kind of optimization passes we run.
352 * For reference this is the layout string on x64:
354 * 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
357 * - http://llvm.org/docs/LangRef.html#datalayout
361 const unsigned pointer_size
= 8 * sizeof(void *);
363 util_snprintf(layout
, sizeof layout
, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
364 #ifdef PIPE_ARCH_LITTLE_ENDIAN
365 'e', // little endian
369 pointer_size
, pointer_size
, pointer_size
, // pointer size, abi alignment, preferred alignment
370 pointer_size
, // aggregate preferred alignment
371 pointer_size
, pointer_size
); // stack objects abi alignment, preferred alignment
373 gallivm
->target
= LLVMCreateTargetData(layout
);
374 if (!gallivm
->target
) {
380 if (!create_pass_manager(gallivm
))
386 gallivm_free_ir(gallivm
);
387 gallivm_free_code(gallivm
);
395 if (gallivm_initialized
)
399 /* LLVMLinkIn* are no-ops at runtime. They just ensure the respective
400 * component is linked at buildtime, which is sufficient for its static
401 * constructors to be called at load time.
403 #if defined(USE_MCJIT)
410 use_mcjit
= debug_get_bool_option("GALLIVM_MCJIT", FALSE
);
416 gallivm_debug
= debug_get_option_gallivm_debug();
419 lp_set_target_options();
423 /* For simulating less capable machines */
425 if (debug_get_bool_option("LP_FORCE_SSE2", FALSE
)) {
426 assert(util_cpu_caps
.has_sse2
);
427 util_cpu_caps
.has_sse3
= 0;
428 util_cpu_caps
.has_ssse3
= 0;
429 util_cpu_caps
.has_sse4_1
= 0;
430 util_cpu_caps
.has_sse4_2
= 0;
431 util_cpu_caps
.has_avx
= 0;
432 util_cpu_caps
.has_avx2
= 0;
433 util_cpu_caps
.has_f16c
= 0;
434 util_cpu_caps
.has_fma
= 0;
438 /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
439 * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
440 * actually more efficient to use 4-wide vectors on this processor.
443 * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
445 if (util_cpu_caps
.has_avx
&&
446 util_cpu_caps
.has_intel
) {
447 lp_native_vector_width
= 256;
449 /* Leave it at 128, even when no SIMD extensions are available.
450 * Really needs to be a multiple of 128 so can fit 4 floats.
452 lp_native_vector_width
= 128;
455 lp_native_vector_width
= debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
456 lp_native_vector_width
);
458 if (lp_native_vector_width
<= 128) {
459 /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
460 * "util_cpu_caps.has_avx" predicate, and lack the
461 * "lp_native_vector_width > 128" predicate. And also to ensure a more
462 * consistent behavior, allowing one to test SSE2 on AVX machines.
463 * XXX: should not play games with util_cpu_caps directly as it might
464 * get used for other things outside llvm too.
466 util_cpu_caps
.has_avx
= 0;
467 util_cpu_caps
.has_avx2
= 0;
468 util_cpu_caps
.has_f16c
= 0;
469 util_cpu_caps
.has_fma
= 0;
471 if (HAVE_LLVM
< 0x0304 || !use_mcjit
) {
472 /* AVX2 support has only been tested with LLVM 3.4, and it requires
474 util_cpu_caps
.has_avx2
= 0;
477 #ifdef PIPE_ARCH_PPC_64
478 /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
479 * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
480 * that some rounding and half-float to float handling does not round
482 * XXX: should eventually follow same logic on all platforms.
483 * Right now denorms get explicitly disabled (but elsewhere) for x86,
484 * whereas ppc64 explicitly enables them...
486 if (util_cpu_caps
.has_altivec
) {
487 unsigned short mask
[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
488 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
499 gallivm_initialized
= TRUE
;
507 * Create a new gallivm_state object.
509 struct gallivm_state
*
510 gallivm_create(const char *name
, LLVMContextRef context
)
512 struct gallivm_state
*gallivm
;
514 gallivm
= CALLOC_STRUCT(gallivm_state
);
516 if (!init_gallivm_state(gallivm
, name
, context
)) {
527 * Destroy a gallivm_state object.
530 gallivm_destroy(struct gallivm_state
*gallivm
)
532 gallivm_free_ir(gallivm
);
533 gallivm_free_code(gallivm
);
539 * Validate a function.
540 * Verification is only done with debug builds.
543 gallivm_verify_function(struct gallivm_state
*gallivm
,
546 /* Verify the LLVM IR. If invalid, dump and abort */
548 if (LLVMVerifyFunction(func
, LLVMPrintMessageAction
)) {
549 lp_debug_dump_value(func
);
555 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
556 /* Print the LLVM IR to stderr */
557 lp_debug_dump_value(func
);
565 * This does IR optimization on all functions in the module.
568 gallivm_compile_module(struct gallivm_state
*gallivm
)
571 int64_t time_begin
= 0;
573 assert(!gallivm
->compiled
);
575 if (gallivm
->builder
) {
576 LLVMDisposeBuilder(gallivm
->builder
);
577 gallivm
->builder
= NULL
;
580 if (gallivm_debug
& GALLIVM_DEBUG_PERF
)
581 time_begin
= os_time_get();
583 /* Run optimization passes */
584 LLVMInitializeFunctionPassManager(gallivm
->passmgr
);
585 func
= LLVMGetFirstFunction(gallivm
->module
);
588 debug_printf("optimizing func %s...\n", LLVMGetValueName(func
));
591 /* Disable frame pointer omission on debug/profile builds */
592 /* XXX: And workaround http://llvm.org/PR21435 */
593 #if HAVE_LLVM >= 0x0307 && \
594 (defined(DEBUG) || defined(PROFILE) || \
595 defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64))
596 LLVMAddTargetDependentFunctionAttr(func
, "no-frame-pointer-elim", "true");
597 LLVMAddTargetDependentFunctionAttr(func
, "no-frame-pointer-elim-non-leaf", "true");
600 LLVMRunFunctionPassManager(gallivm
->passmgr
, func
);
601 func
= LLVMGetNextFunction(func
);
603 LLVMFinalizeFunctionPassManager(gallivm
->passmgr
);
605 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
606 int64_t time_end
= os_time_get();
607 int time_msec
= (int)((time_end
- time_begin
) / 1000);
608 assert(gallivm
->module_name
);
609 debug_printf("optimizing module %s took %d msec\n",
610 gallivm
->module_name
, time_msec
);
613 /* Dump byte code to a file */
614 if (gallivm_debug
& GALLIVM_DEBUG_DUMP_BC
) {
616 assert(gallivm
->module_name
);
617 util_snprintf(filename
, sizeof(filename
), "ir_%s.bc", gallivm
->module_name
);
618 LLVMWriteBitcodeToFile(gallivm
->module
, filename
);
619 debug_printf("%s written\n", filename
);
620 debug_printf("Invoke as \"llc %s%s -o - %s\"\n",
621 (HAVE_LLVM
>= 0x0305) ? "[-mcpu=<-mcpu option>] " : "",
622 "[-mattr=<-mattr option(s)>]",
627 /* Setting the module's DataLayout to an empty string will cause the
628 * ExecutionEngine to copy to the DataLayout string from its target
629 * machine to the module. As of LLVM 3.8 the module and the execution
630 * engine are required to have the same DataLayout.
632 * We must make sure we do this after running the optimization passes,
633 * because those passes need a correct datalayout string. For example,
634 * if those optimization passes see an empty datalayout, they will assume
635 * this is a little endian target and will do optimizations that break big
638 * TODO: This is just a temporary work-around. The correct solution is
639 * for gallivm_init_state() to create a TargetMachine and pull the
640 * DataLayout from there. Currently, the TargetMachine used by llvmpipe
641 * is being implicitly created by the EngineBuilder in
642 * lp_build_create_jit_compiler_for_module()
644 LLVMSetDataLayout(gallivm
->module
, "");
645 assert(!gallivm
->engine
);
646 if (!init_gallivm_engine(gallivm
)) {
650 assert(gallivm
->engine
);
654 if (gallivm_debug
& GALLIVM_DEBUG_ASM
) {
655 LLVMValueRef llvm_func
= LLVMGetFirstFunction(gallivm
->module
);
659 * Need to filter out functions which don't have an implementation,
660 * such as the intrinsics. May not be sufficient in case of IPO?
661 * LLVMGetPointerToGlobal() will abort otherwise.
663 if (!LLVMIsDeclaration(llvm_func
)) {
664 void *func_code
= LLVMGetPointerToGlobal(gallivm
->engine
, llvm_func
);
665 lp_disassemble(llvm_func
, func_code
);
667 llvm_func
= LLVMGetNextFunction(llvm_func
);
673 LLVMValueRef llvm_func
= LLVMGetFirstFunction(gallivm
->module
);
676 if (!LLVMIsDeclaration(llvm_func
)) {
677 void *func_code
= LLVMGetPointerToGlobal(gallivm
->engine
, llvm_func
);
678 lp_profile(llvm_func
, func_code
);
680 llvm_func
= LLVMGetNextFunction(llvm_func
);
689 gallivm_jit_function(struct gallivm_state
*gallivm
,
693 func_pointer jit_func
;
694 int64_t time_begin
= 0;
696 assert(gallivm
->compiled
);
697 assert(gallivm
->engine
);
699 if (gallivm_debug
& GALLIVM_DEBUG_PERF
)
700 time_begin
= os_time_get();
702 code
= LLVMGetPointerToGlobal(gallivm
->engine
, func
);
704 jit_func
= pointer_to_func(code
);
706 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
707 int64_t time_end
= os_time_get();
708 int time_msec
= (int)(time_end
- time_begin
) / 1000;
709 debug_printf(" jitting func %s took %d msec\n",
710 LLVMGetValueName(func
), time_msec
);