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) {
140 /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
141 * but there are more on SVN.
142 * TODO: Add more passes.
144 LLVMAddScalarReplAggregatesPass(gallivm
->passmgr
);
145 LLVMAddLICMPass(gallivm
->passmgr
);
146 LLVMAddCFGSimplificationPass(gallivm
->passmgr
);
147 LLVMAddReassociatePass(gallivm
->passmgr
);
148 LLVMAddPromoteMemoryToRegisterPass(gallivm
->passmgr
);
149 LLVMAddConstantPropagationPass(gallivm
->passmgr
);
150 LLVMAddInstructionCombiningPass(gallivm
->passmgr
);
151 LLVMAddGVNPass(gallivm
->passmgr
);
154 /* We need at least this pass to prevent the backends to fail in
157 LLVMAddPromoteMemoryToRegisterPass(gallivm
->passmgr
);
165 * Free gallivm object's LLVM allocations, but not any generated code
166 * nor the gallivm object itself.
169 gallivm_free_ir(struct gallivm_state
*gallivm
)
171 if (gallivm
->passmgr
) {
172 LLVMDisposePassManager(gallivm
->passmgr
);
175 if (gallivm
->engine
) {
176 /* This will already destroy any associated module */
177 LLVMDisposeExecutionEngine(gallivm
->engine
);
178 } else if (gallivm
->module
) {
179 LLVMDisposeModule(gallivm
->module
);
182 FREE(gallivm
->module_name
);
185 /* Don't free the TargetData, it's owned by the exec engine */
187 if (gallivm
->target
) {
188 LLVMDisposeTargetData(gallivm
->target
);
192 if (gallivm
->builder
)
193 LLVMDisposeBuilder(gallivm
->builder
);
195 /* The LLVMContext should be owned by the parent of gallivm. */
197 gallivm
->engine
= NULL
;
198 gallivm
->target
= NULL
;
199 gallivm
->module
= NULL
;
200 gallivm
->module_name
= NULL
;
201 gallivm
->passmgr
= NULL
;
202 gallivm
->context
= NULL
;
203 gallivm
->builder
= NULL
;
208 * Free LLVM-generated code. Should be done AFTER gallivm_free_ir().
211 gallivm_free_code(struct gallivm_state
*gallivm
)
213 assert(!gallivm
->module
);
214 assert(!gallivm
->engine
);
215 lp_free_generated_code(gallivm
->code
);
216 gallivm
->code
= NULL
;
217 lp_free_memory_manager(gallivm
->memorymgr
);
218 gallivm
->memorymgr
= NULL
;
223 init_gallivm_engine(struct gallivm_state
*gallivm
)
226 enum LLVM_CodeGenOpt_Level optlevel
;
230 if (gallivm_debug
& GALLIVM_DEBUG_NO_OPT
) {
237 ret
= lp_build_create_jit_compiler_for_module(&gallivm
->engine
,
245 _debug_printf("%s\n", error
);
246 LLVMDisposeMessage(error
);
252 gallivm
->target
= LLVMGetExecutionEngineTargetData(gallivm
->engine
);
253 if (!gallivm
->target
)
258 * Dump the data layout strings.
261 LLVMTargetDataRef target
= LLVMGetExecutionEngineTargetData(gallivm
->engine
);
263 char *engine_data_layout
;
265 data_layout
= LLVMCopyStringRepOfTargetData(gallivm
->target
);
266 engine_data_layout
= LLVMCopyStringRepOfTargetData(target
);
269 debug_printf("module target data = %s\n", data_layout
);
270 debug_printf("engine target data = %s\n", engine_data_layout
);
274 free(engine_data_layout
);
286 * Allocate gallivm LLVM objects.
287 * \return TRUE for success, FALSE for failure
290 init_gallivm_state(struct gallivm_state
*gallivm
, const char *name
,
291 LLVMContextRef context
)
293 assert(!gallivm
->context
);
294 assert(!gallivm
->module
);
296 if (!lp_build_init())
299 gallivm
->context
= context
;
301 if (!gallivm
->context
)
304 gallivm
->module_name
= NULL
;
306 size_t size
= strlen(name
) + 1;
307 gallivm
->module_name
= MALLOC(size
);
308 if (gallivm
->module_name
) {
309 memcpy(gallivm
->module_name
, name
, size
);
313 gallivm
->module
= LLVMModuleCreateWithNameInContext(name
,
315 if (!gallivm
->module
)
318 gallivm
->builder
= LLVMCreateBuilderInContext(gallivm
->context
);
319 if (!gallivm
->builder
)
322 gallivm
->memorymgr
= lp_get_default_memory_manager();
323 if (!gallivm
->memorymgr
)
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
)
390 /* LLVMLinkIn* are no-ops at runtime. They just ensure the respective
391 * component is linked at buildtime, which is sufficient for its static
392 * constructors to be called at load time.
394 #if defined(USE_MCJIT)
401 use_mcjit
= debug_get_bool_option("GALLIVM_MCJIT", FALSE
);
407 gallivm_debug
= debug_get_option_gallivm_debug();
410 lp_set_target_options();
414 /* For simulating less capable machines */
416 if (debug_get_bool_option("LP_FORCE_SSE2", FALSE
)) {
417 assert(util_cpu_caps
.has_sse2
);
418 util_cpu_caps
.has_sse3
= 0;
419 util_cpu_caps
.has_ssse3
= 0;
420 util_cpu_caps
.has_sse4_1
= 0;
421 util_cpu_caps
.has_sse4_2
= 0;
422 util_cpu_caps
.has_avx
= 0;
423 util_cpu_caps
.has_avx2
= 0;
424 util_cpu_caps
.has_f16c
= 0;
425 util_cpu_caps
.has_fma
= 0;
429 /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
430 * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
431 * actually more efficient to use 4-wide vectors on this processor.
434 * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
436 if (util_cpu_caps
.has_avx
&&
437 util_cpu_caps
.has_intel
) {
438 lp_native_vector_width
= 256;
440 /* Leave it at 128, even when no SIMD extensions are available.
441 * Really needs to be a multiple of 128 so can fit 4 floats.
443 lp_native_vector_width
= 128;
446 lp_native_vector_width
= debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
447 lp_native_vector_width
);
449 if (lp_native_vector_width
<= 128) {
450 /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
451 * "util_cpu_caps.has_avx" predicate, and lack the
452 * "lp_native_vector_width > 128" predicate. And also to ensure a more
453 * consistent behavior, allowing one to test SSE2 on AVX machines.
454 * XXX: should not play games with util_cpu_caps directly as it might
455 * get used for other things outside llvm too.
457 util_cpu_caps
.has_avx
= 0;
458 util_cpu_caps
.has_avx2
= 0;
459 util_cpu_caps
.has_f16c
= 0;
460 util_cpu_caps
.has_fma
= 0;
462 if (HAVE_LLVM
< 0x0304 || !use_mcjit
) {
463 /* AVX2 support has only been tested with LLVM 3.4, and it requires
465 util_cpu_caps
.has_avx2
= 0;
468 #ifdef PIPE_ARCH_PPC_64
469 /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
470 * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
471 * that some rounding and half-float to float handling does not round
473 * XXX: should eventually follow same logic on all platforms.
474 * Right now denorms get explicitly disabled (but elsewhere) for x86,
475 * whereas ppc64 explicitly enables them...
477 if (util_cpu_caps
.has_altivec
) {
478 unsigned short mask
[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
479 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
490 gallivm_initialized
= TRUE
;
498 * Create a new gallivm_state object.
500 struct gallivm_state
*
501 gallivm_create(const char *name
, LLVMContextRef context
)
503 struct gallivm_state
*gallivm
;
505 gallivm
= CALLOC_STRUCT(gallivm_state
);
507 if (!init_gallivm_state(gallivm
, name
, context
)) {
518 * Destroy a gallivm_state object.
521 gallivm_destroy(struct gallivm_state
*gallivm
)
523 gallivm_free_ir(gallivm
);
524 gallivm_free_code(gallivm
);
530 * Validate a function.
531 * Verification is only done with debug builds.
534 gallivm_verify_function(struct gallivm_state
*gallivm
,
537 /* Verify the LLVM IR. If invalid, dump and abort */
539 if (LLVMVerifyFunction(func
, LLVMPrintMessageAction
)) {
540 lp_debug_dump_value(func
);
546 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
547 /* Print the LLVM IR to stderr */
548 lp_debug_dump_value(func
);
556 * This does IR optimization on all functions in the module.
559 gallivm_compile_module(struct gallivm_state
*gallivm
)
562 int64_t time_begin
= 0;
564 assert(!gallivm
->compiled
);
566 if (gallivm
->builder
) {
567 LLVMDisposeBuilder(gallivm
->builder
);
568 gallivm
->builder
= NULL
;
571 if (gallivm_debug
& GALLIVM_DEBUG_PERF
)
572 time_begin
= os_time_get();
574 /* Run optimization passes */
575 LLVMInitializeFunctionPassManager(gallivm
->passmgr
);
576 func
= LLVMGetFirstFunction(gallivm
->module
);
579 debug_printf("optimizing func %s...\n", LLVMGetValueName(func
));
582 /* Disable frame pointer omission on debug/profile builds */
583 /* XXX: And workaround http://llvm.org/PR21435 */
584 #if HAVE_LLVM >= 0x0307 && \
585 (defined(DEBUG) || defined(PROFILE) || \
586 defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64))
587 LLVMAddTargetDependentFunctionAttr(func
, "no-frame-pointer-elim", "true");
588 LLVMAddTargetDependentFunctionAttr(func
, "no-frame-pointer-elim-non-leaf", "true");
591 LLVMRunFunctionPassManager(gallivm
->passmgr
, func
);
592 func
= LLVMGetNextFunction(func
);
594 LLVMFinalizeFunctionPassManager(gallivm
->passmgr
);
596 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
597 int64_t time_end
= os_time_get();
598 int time_msec
= (int)(time_end
- time_begin
) / 1000;
599 assert(gallivm
->module_name
);
600 debug_printf("optimizing module %s took %d msec\n",
601 gallivm
->module_name
, time_msec
);
604 /* Dump byte code to a file */
605 if (gallivm_debug
& GALLIVM_DEBUG_DUMP_BC
) {
607 assert(gallivm
->module_name
);
608 util_snprintf(filename
, sizeof(filename
), "ir_%s.bc", gallivm
->module_name
);
609 LLVMWriteBitcodeToFile(gallivm
->module
, filename
);
610 debug_printf("%s written\n", filename
);
611 debug_printf("Invoke as \"llc %s%s -o - %s\"\n",
612 (HAVE_LLVM
>= 0x0305) ? "[-mcpu=<-mcpu option>] " : "",
613 "[-mattr=<-mattr option(s)>]",
618 /* Setting the module's DataLayout to an empty string will cause the
619 * ExecutionEngine to copy to the DataLayout string from its target
620 * machine to the module. As of LLVM 3.8 the module and the execution
621 * engine are required to have the same DataLayout.
623 * We must make sure we do this after running the optimization passes,
624 * because those passes need a correct datalayout string. For example,
625 * if those optimization passes see an empty datalayout, they will assume
626 * this is a little endian target and will do optimizations that break big
629 * TODO: This is just a temporary work-around. The correct solution is
630 * for gallivm_init_state() to create a TargetMachine and pull the
631 * DataLayout from there. Currently, the TargetMachine used by llvmpipe
632 * is being implicitly created by the EngineBuilder in
633 * lp_build_create_jit_compiler_for_module()
635 LLVMSetDataLayout(gallivm
->module
, "");
636 assert(!gallivm
->engine
);
637 if (!init_gallivm_engine(gallivm
)) {
641 assert(gallivm
->engine
);
645 if (gallivm_debug
& GALLIVM_DEBUG_ASM
) {
646 LLVMValueRef llvm_func
= LLVMGetFirstFunction(gallivm
->module
);
650 * Need to filter out functions which don't have an implementation,
651 * such as the intrinsics. May not be sufficient in case of IPO?
652 * LLVMGetPointerToGlobal() will abort otherwise.
654 if (!LLVMIsDeclaration(llvm_func
)) {
655 void *func_code
= LLVMGetPointerToGlobal(gallivm
->engine
, llvm_func
);
656 lp_disassemble(llvm_func
, func_code
);
658 llvm_func
= LLVMGetNextFunction(llvm_func
);
664 LLVMValueRef llvm_func
= LLVMGetFirstFunction(gallivm
->module
);
667 if (!LLVMIsDeclaration(llvm_func
)) {
668 void *func_code
= LLVMGetPointerToGlobal(gallivm
->engine
, llvm_func
);
669 lp_profile(llvm_func
, func_code
);
671 llvm_func
= LLVMGetNextFunction(llvm_func
);
680 gallivm_jit_function(struct gallivm_state
*gallivm
,
684 func_pointer jit_func
;
685 int64_t time_begin
= 0;
687 assert(gallivm
->compiled
);
688 assert(gallivm
->engine
);
690 if (gallivm_debug
& GALLIVM_DEBUG_PERF
)
691 time_begin
= os_time_get();
693 code
= LLVMGetPointerToGlobal(gallivm
->engine
, func
);
695 jit_func
= pointer_to_func(code
);
697 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
698 int64_t time_end
= os_time_get();
699 int time_msec
= (int)(time_end
- time_begin
) / 1000;
700 debug_printf(" jitting func %s took %d msec\n",
701 LLVMGetValueName(func
), time_msec
);