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"
36 #include "lp_bld_debug.h"
37 #include "lp_bld_misc.h"
38 #include "lp_bld_init.h"
40 #include <llvm-c/Analysis.h>
41 #include <llvm-c/Transforms/Scalar.h>
42 #include <llvm-c/BitWriter.h>
46 * AVX is supported in:
47 * - standard JIT from LLVM 3.2 onwards
48 * - MC-JIT from LLVM 3.1
49 * - MC-JIT supports limited OSes (MacOSX and Linux)
50 * - standard JIT in LLVM 3.1, with backports
52 #if defined(PIPE_ARCH_PPC_64) || defined(PIPE_ARCH_S390)
55 #elif HAVE_LLVM >= 0x0302 || (HAVE_LLVM == 0x0301 && defined(HAVE_JIT_AVX_SUPPORT))
58 #elif HAVE_LLVM == 0x0301 && (defined(PIPE_OS_LINUX) || defined(PIPE_OS_APPLE))
68 void LLVMLinkInMCJIT();
73 unsigned gallivm_debug
= 0;
75 static const struct debug_named_value lp_bld_debug_flags
[] = {
76 { "tgsi", GALLIVM_DEBUG_TGSI
, NULL
},
77 { "ir", GALLIVM_DEBUG_IR
, NULL
},
78 { "asm", GALLIVM_DEBUG_ASM
, NULL
},
79 { "nopt", GALLIVM_DEBUG_NO_OPT
, NULL
},
80 { "perf", GALLIVM_DEBUG_PERF
, NULL
},
81 { "no_brilinear", GALLIVM_DEBUG_NO_BRILINEAR
, NULL
},
82 { "no_rho_approx", GALLIVM_DEBUG_NO_RHO_APPROX
, NULL
},
83 { "gc", GALLIVM_DEBUG_GC
, NULL
},
87 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug
, "GALLIVM_DEBUG", lp_bld_debug_flags
, 0)
91 static boolean gallivm_initialized
= FALSE
;
93 unsigned lp_native_vector_width
;
97 * Optimization values are:
100 * - 2: Default (-O2, -Os)
101 * - 3: Aggressive (-O3)
103 * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
105 enum LLVM_CodeGenOpt_Level
{
106 #if HAVE_LLVM >= 0x207
119 #if HAVE_LLVM <= 0x0206
121 * LLVM 2.6 permits only one ExecutionEngine to be created. So use the
122 * same gallivm state everywhere.
124 static struct gallivm_state
*GlobalGallivm
= NULL
;
129 * Create the LLVM (optimization) pass manager and install
130 * relevant optimization passes.
131 * \return TRUE for success, FALSE for failure
134 create_pass_manager(struct gallivm_state
*gallivm
)
136 assert(!gallivm
->passmgr
);
137 assert(gallivm
->target
);
139 gallivm
->passmgr
= LLVMCreateFunctionPassManager(gallivm
->provider
);
140 if (!gallivm
->passmgr
)
143 LLVMAddTargetData(gallivm
->target
, gallivm
->passmgr
);
145 if ((gallivm_debug
& GALLIVM_DEBUG_NO_OPT
) == 0) {
146 /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
147 * but there are more on SVN.
148 * TODO: Add more passes.
150 LLVMAddScalarReplAggregatesPass(gallivm
->passmgr
);
151 LLVMAddLICMPass(gallivm
->passmgr
);
152 LLVMAddCFGSimplificationPass(gallivm
->passmgr
);
153 LLVMAddReassociatePass(gallivm
->passmgr
);
155 if (HAVE_LLVM
>= 0x207 && sizeof(void*) == 4) {
156 /* For LLVM >= 2.7 and 32-bit build, use this order of passes to
157 * avoid generating bad code.
158 * Test with piglit glsl-vs-sqrt-zero test.
160 LLVMAddConstantPropagationPass(gallivm
->passmgr
);
161 LLVMAddPromoteMemoryToRegisterPass(gallivm
->passmgr
);
164 LLVMAddPromoteMemoryToRegisterPass(gallivm
->passmgr
);
165 LLVMAddConstantPropagationPass(gallivm
->passmgr
);
168 if (util_cpu_caps
.has_sse4_1
) {
169 /* FIXME: There is a bug in this pass, whereby the combination
170 * of fptosi and sitofp (necessary for trunc/floor/ceil/round
171 * implementation) somehow becomes invalid code.
173 LLVMAddInstructionCombiningPass(gallivm
->passmgr
);
175 LLVMAddGVNPass(gallivm
->passmgr
);
178 /* We need at least this pass to prevent the backends to fail in
181 LLVMAddPromoteMemoryToRegisterPass(gallivm
->passmgr
);
189 * Free gallivm object's LLVM allocations, but not the gallivm object itself.
192 free_gallivm_state(struct gallivm_state
*gallivm
)
194 #if HAVE_LLVM >= 0x207 /* XXX or 0x208? */
195 /* This leads to crashes w/ some versions of LLVM */
199 if (gallivm
->engine
&& gallivm
->provider
)
200 LLVMRemoveModuleProvider(gallivm
->engine
, gallivm
->provider
,
204 if (gallivm
->passmgr
) {
205 LLVMDisposePassManager(gallivm
->passmgr
);
209 /* XXX this seems to crash with all versions of LLVM */
210 if (gallivm
->provider
)
211 LLVMDisposeModuleProvider(gallivm
->provider
);
214 if (HAVE_LLVM
>= 0x207 && gallivm
->engine
) {
215 /* This will already destroy any associated module */
216 LLVMDisposeExecutionEngine(gallivm
->engine
);
218 LLVMDisposeModule(gallivm
->module
);
222 /* Don't free the TargetData, it's owned by the exec engine */
224 if (gallivm
->target
) {
225 LLVMDisposeTargetData(gallivm
->target
);
229 /* Never free the LLVM context.
232 if (gallivm
->context
)
233 LLVMContextDispose(gallivm
->context
);
236 if (gallivm
->builder
)
237 LLVMDisposeBuilder(gallivm
->builder
);
239 gallivm
->engine
= NULL
;
240 gallivm
->target
= NULL
;
241 gallivm
->module
= NULL
;
242 gallivm
->provider
= NULL
;
243 gallivm
->passmgr
= NULL
;
244 gallivm
->context
= NULL
;
245 gallivm
->builder
= NULL
;
250 init_gallivm_engine(struct gallivm_state
*gallivm
)
253 /* We can only create one LLVMExecutionEngine (w/ LLVM 2.6 anyway) */
254 enum LLVM_CodeGenOpt_Level optlevel
;
258 if (gallivm_debug
& GALLIVM_DEBUG_NO_OPT
) {
265 #if HAVE_LLVM >= 0x0301
266 ret
= lp_build_create_jit_compiler_for_module(&gallivm
->engine
,
272 ret
= LLVMCreateJITCompiler(&gallivm
->engine
, gallivm
->provider
,
273 (unsigned) optlevel
, &error
);
276 _debug_printf("%s\n", error
);
277 LLVMDisposeMessage(error
);
282 LLVMAddModuleProvider(gallivm
->engine
, gallivm
->provider
);//new
285 gallivm
->target
= LLVMGetExecutionEngineTargetData(gallivm
->engine
);
286 if (!gallivm
->target
)
291 * Dump the data layout strings.
294 LLVMTargetDataRef target
= LLVMGetExecutionEngineTargetData(gallivm
->engine
);
296 char *engine_data_layout
;
298 data_layout
= LLVMCopyStringRepOfTargetData(gallivm
->target
);
299 engine_data_layout
= LLVMCopyStringRepOfTargetData(target
);
302 debug_printf("module target data = %s\n", data_layout
);
303 debug_printf("engine target data = %s\n", engine_data_layout
);
307 free(engine_data_layout
);
321 * We must never free LLVM contexts, because LLVM has several global caches
322 * which pointing/derived from objects owned by the context, causing false
323 * memory leaks and false cache hits when these objects are destroyed.
325 * TODO: For thread safety on multi-threaded OpenGL we should use one LLVM
326 * context per thread, and put them in a pool when threads are destroyed.
328 static LLVMContextRef gallivm_context
= NULL
;
332 * Allocate gallivm LLVM objects.
333 * \return TRUE for success, FALSE for failure
336 init_gallivm_state(struct gallivm_state
*gallivm
)
338 assert(!gallivm
->context
);
339 assert(!gallivm
->module
);
340 assert(!gallivm
->provider
);
344 if (!gallivm_context
) {
345 gallivm_context
= LLVMContextCreate();
347 gallivm
->context
= gallivm_context
;
348 if (!gallivm
->context
)
351 gallivm
->module
= LLVMModuleCreateWithNameInContext("gallivm",
353 if (!gallivm
->module
)
357 LLVMCreateModuleProviderForExistingModule(gallivm
->module
);
358 if (!gallivm
->provider
)
361 gallivm
->builder
= LLVMCreateBuilderInContext(gallivm
->context
);
362 if (!gallivm
->builder
)
365 /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
366 * complete when MC-JIT is created. So defer the MC-JIT engine creation for
370 if (!init_gallivm_engine(gallivm
)) {
375 * MC-JIT engine compiles the module immediately on creation, so we can't
376 * obtain the target data from it. Instead we create a target data layout
379 * The produced layout strings are not precisely the same, but should make
380 * no difference for the kind of optimization passes we run.
382 * For reference this is the layout string on x64:
384 * 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
387 * - http://llvm.org/docs/LangRef.html#datalayout
391 const unsigned pointer_size
= 8 * sizeof(void *);
393 util_snprintf(layout
, sizeof layout
, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
394 #ifdef PIPE_ARCH_LITTLE_ENDIAN
395 'e', // little endian
399 pointer_size
, pointer_size
, pointer_size
, // pointer size, abi alignment, preferred alignment
400 pointer_size
, // aggregate preferred alignment
401 pointer_size
, pointer_size
); // stack objects abi alignment, preferred alignment
403 gallivm
->target
= LLVMCreateTargetData(layout
);
404 if (!gallivm
->target
) {
410 if (!create_pass_manager(gallivm
))
416 free_gallivm_state(gallivm
);
424 if (gallivm_initialized
)
428 gallivm_debug
= debug_get_option_gallivm_debug();
431 lp_set_target_options();
441 /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
442 * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
443 * actually more efficient to use 4-wide vectors on this processor.
446 * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
449 util_cpu_caps
.has_avx
&&
450 util_cpu_caps
.has_intel
) {
451 lp_native_vector_width
= 256;
453 /* Leave it at 128, even when no SIMD extensions are available.
454 * Really needs to be a multiple of 128 so can fit 4 floats.
456 lp_native_vector_width
= 128;
459 lp_native_vector_width
= debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
460 lp_native_vector_width
);
462 if (lp_native_vector_width
<= 128) {
463 /* Hide AVX support, as often LLVM AVX instrinsics are only guarded by
464 * "util_cpu_caps.has_avx" predicate, and lack the
465 * "lp_native_vector_width > 128" predicate. And also to ensure a more
466 * consistent behavior, allowing one to test SSE2 on AVX machines.
468 util_cpu_caps
.has_avx
= 0;
473 * note these instructions are VEX-only, so can only emit if we use
474 * avx (don't want to base it on has_avx & has_f16c later as that would
475 * omit it unnecessarily on amd cpus, see above).
477 util_cpu_caps
.has_f16c
= 0;
480 #ifdef PIPE_ARCH_PPC_64
481 /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
482 * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This garantees
483 * that some rounding and half-float to float handling does not round
486 if (util_cpu_caps
.has_altivec
) {
487 unsigned short mask
[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
488 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
499 gallivm_initialized
= TRUE
;
502 /* For simulating less capable machines */
503 util_cpu_caps
.has_sse3
= 0;
504 util_cpu_caps
.has_ssse3
= 0;
505 util_cpu_caps
.has_sse4_1
= 0;
506 util_cpu_caps
.has_avx
= 0;
507 util_cpu_caps
.has_f16c
= 0;
514 * Create a new gallivm_state object.
515 * Note that we return a singleton.
517 struct gallivm_state
*
520 struct gallivm_state
*gallivm
;
522 #if HAVE_LLVM <= 0x206
524 return GlobalGallivm
;
528 gallivm
= CALLOC_STRUCT(gallivm_state
);
530 if (!init_gallivm_state(gallivm
)) {
536 #if HAVE_LLVM <= 0x206
537 GlobalGallivm
= gallivm
;
545 * Destroy a gallivm_state object.
548 gallivm_destroy(struct gallivm_state
*gallivm
)
550 #if HAVE_LLVM <= 0x0206
551 /* No-op: don't destroy the singleton */
554 free_gallivm_state(gallivm
);
561 * Validate and optimze a function.
564 gallivm_optimize_function(struct gallivm_state
*gallivm
,
568 debug_printf("optimizing %s...\n", LLVMGetValueName(func
));
571 assert(gallivm
->passmgr
);
573 /* Apply optimizations to LLVM IR */
574 LLVMRunFunctionPassManager(gallivm
->passmgr
, func
);
577 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
578 /* Print the LLVM IR to stderr */
579 lp_debug_dump_value(func
);
587 * Validate a function.
590 gallivm_verify_function(struct gallivm_state
*gallivm
,
593 /* Verify the LLVM IR. If invalid, dump and abort */
595 if (LLVMVerifyFunction(func
, LLVMPrintMessageAction
)) {
596 lp_debug_dump_value(func
);
602 gallivm_optimize_function(gallivm
, func
);
604 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
605 /* Print the LLVM IR to stderr */
606 lp_debug_dump_value(func
);
613 gallivm_compile_module(struct gallivm_state
*gallivm
)
615 #if HAVE_LLVM > 0x206
616 assert(!gallivm
->compiled
);
619 /* Dump byte code to a file */
621 LLVMWriteBitcodeToFile(gallivm
->module
, "llvmpipe.bc");
622 debug_printf("llvmpipe.bc written\n");
623 debug_printf("Invoke as \"llc -o - llvmpipe.bc\"\n");
627 assert(!gallivm
->engine
);
628 if (!init_gallivm_engine(gallivm
)) {
632 assert(gallivm
->engine
);
640 gallivm_jit_function(struct gallivm_state
*gallivm
,
644 func_pointer jit_func
;
646 assert(gallivm
->compiled
);
647 assert(gallivm
->engine
);
649 code
= LLVMGetPointerToGlobal(gallivm
->engine
, func
);
651 jit_func
= pointer_to_func(code
);
653 if (gallivm_debug
& GALLIVM_DEBUG_ASM
) {
654 lp_disassemble(func
, code
);
658 lp_profile(func
, code
);
661 /* Free the function body to save memory */
662 lp_func_delete_body(func
);
669 * Free the function (and its machine code).
672 gallivm_free_function(struct gallivm_state
*gallivm
,
678 LLVMFreeMachineCodeForFunction(gallivm
->engine
, func
);
681 LLVMDeleteFunction(func
);