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) || defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
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 { "no_quad_lod", GALLIVM_DEBUG_NO_QUAD_LOD
, NULL
},
84 { "gc", GALLIVM_DEBUG_GC
, NULL
},
88 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug
, "GALLIVM_DEBUG", lp_bld_debug_flags
, 0)
92 static boolean gallivm_initialized
= FALSE
;
94 unsigned lp_native_vector_width
;
98 * Optimization values are:
101 * - 2: Default (-O2, -Os)
102 * - 3: Aggressive (-O3)
104 * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
106 enum LLVM_CodeGenOpt_Level
{
107 #if HAVE_LLVM >= 0x207
120 #if HAVE_LLVM <= 0x0206
122 * LLVM 2.6 permits only one ExecutionEngine to be created. So use the
123 * same gallivm state everywhere.
125 static struct gallivm_state
*GlobalGallivm
= NULL
;
130 * Create the LLVM (optimization) pass manager and install
131 * relevant optimization passes.
132 * \return TRUE for success, FALSE for failure
135 create_pass_manager(struct gallivm_state
*gallivm
)
137 assert(!gallivm
->passmgr
);
138 assert(gallivm
->target
);
140 gallivm
->passmgr
= LLVMCreateFunctionPassManager(gallivm
->provider
);
141 if (!gallivm
->passmgr
)
144 LLVMAddTargetData(gallivm
->target
, gallivm
->passmgr
);
146 if ((gallivm_debug
& GALLIVM_DEBUG_NO_OPT
) == 0) {
147 /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
148 * but there are more on SVN.
149 * TODO: Add more passes.
151 LLVMAddScalarReplAggregatesPass(gallivm
->passmgr
);
152 LLVMAddLICMPass(gallivm
->passmgr
);
153 LLVMAddCFGSimplificationPass(gallivm
->passmgr
);
154 LLVMAddReassociatePass(gallivm
->passmgr
);
156 if (HAVE_LLVM
>= 0x207 && sizeof(void*) == 4) {
157 /* For LLVM >= 2.7 and 32-bit build, use this order of passes to
158 * avoid generating bad code.
159 * Test with piglit glsl-vs-sqrt-zero test.
161 LLVMAddConstantPropagationPass(gallivm
->passmgr
);
162 LLVMAddPromoteMemoryToRegisterPass(gallivm
->passmgr
);
165 LLVMAddPromoteMemoryToRegisterPass(gallivm
->passmgr
);
166 LLVMAddConstantPropagationPass(gallivm
->passmgr
);
169 if (util_cpu_caps
.has_sse4_1
) {
170 /* FIXME: There is a bug in this pass, whereby the combination
171 * of fptosi and sitofp (necessary for trunc/floor/ceil/round
172 * implementation) somehow becomes invalid code.
174 LLVMAddInstructionCombiningPass(gallivm
->passmgr
);
176 LLVMAddGVNPass(gallivm
->passmgr
);
179 /* We need at least this pass to prevent the backends to fail in
182 LLVMAddPromoteMemoryToRegisterPass(gallivm
->passmgr
);
190 * Free gallivm object's LLVM allocations, but not the gallivm object itself.
193 free_gallivm_state(struct gallivm_state
*gallivm
)
195 #if HAVE_LLVM >= 0x207 /* XXX or 0x208? */
196 /* This leads to crashes w/ some versions of LLVM */
200 if (gallivm
->engine
&& gallivm
->provider
)
201 LLVMRemoveModuleProvider(gallivm
->engine
, gallivm
->provider
,
205 if (gallivm
->passmgr
) {
206 LLVMDisposePassManager(gallivm
->passmgr
);
210 /* XXX this seems to crash with all versions of LLVM */
211 if (gallivm
->provider
)
212 LLVMDisposeModuleProvider(gallivm
->provider
);
215 if (HAVE_LLVM
>= 0x207 && gallivm
->engine
) {
216 /* This will already destroy any associated module */
217 LLVMDisposeExecutionEngine(gallivm
->engine
);
219 LLVMDisposeModule(gallivm
->module
);
223 /* Don't free the TargetData, it's owned by the exec engine */
225 if (gallivm
->target
) {
226 LLVMDisposeTargetData(gallivm
->target
);
230 /* Never free the LLVM context.
233 if (gallivm
->context
)
234 LLVMContextDispose(gallivm
->context
);
237 if (gallivm
->builder
)
238 LLVMDisposeBuilder(gallivm
->builder
);
240 gallivm
->engine
= NULL
;
241 gallivm
->target
= NULL
;
242 gallivm
->module
= NULL
;
243 gallivm
->provider
= NULL
;
244 gallivm
->passmgr
= NULL
;
245 gallivm
->context
= NULL
;
246 gallivm
->builder
= NULL
;
251 init_gallivm_engine(struct gallivm_state
*gallivm
)
254 /* We can only create one LLVMExecutionEngine (w/ LLVM 2.6 anyway) */
255 enum LLVM_CodeGenOpt_Level optlevel
;
259 if (gallivm_debug
& GALLIVM_DEBUG_NO_OPT
) {
266 #if HAVE_LLVM >= 0x0301
267 ret
= lp_build_create_jit_compiler_for_module(&gallivm
->engine
,
273 ret
= LLVMCreateJITCompiler(&gallivm
->engine
, gallivm
->provider
,
274 (unsigned) optlevel
, &error
);
277 _debug_printf("%s\n", error
);
278 LLVMDisposeMessage(error
);
283 LLVMAddModuleProvider(gallivm
->engine
, gallivm
->provider
);//new
286 gallivm
->target
= LLVMGetExecutionEngineTargetData(gallivm
->engine
);
287 if (!gallivm
->target
)
292 * Dump the data layout strings.
295 LLVMTargetDataRef target
= LLVMGetExecutionEngineTargetData(gallivm
->engine
);
297 char *engine_data_layout
;
299 data_layout
= LLVMCopyStringRepOfTargetData(gallivm
->target
);
300 engine_data_layout
= LLVMCopyStringRepOfTargetData(target
);
303 debug_printf("module target data = %s\n", data_layout
);
304 debug_printf("engine target data = %s\n", engine_data_layout
);
308 free(engine_data_layout
);
322 * We must never free LLVM contexts, because LLVM has several global caches
323 * which pointing/derived from objects owned by the context, causing false
324 * memory leaks and false cache hits when these objects are destroyed.
326 * TODO: For thread safety on multi-threaded OpenGL we should use one LLVM
327 * context per thread, and put them in a pool when threads are destroyed.
329 static LLVMContextRef gallivm_context
= NULL
;
333 * Allocate gallivm LLVM objects.
334 * \return TRUE for success, FALSE for failure
337 init_gallivm_state(struct gallivm_state
*gallivm
)
339 assert(!gallivm
->context
);
340 assert(!gallivm
->module
);
341 assert(!gallivm
->provider
);
345 if (!gallivm_context
) {
346 gallivm_context
= LLVMContextCreate();
348 gallivm
->context
= gallivm_context
;
349 if (!gallivm
->context
)
352 gallivm
->module
= LLVMModuleCreateWithNameInContext("gallivm",
354 if (!gallivm
->module
)
358 LLVMCreateModuleProviderForExistingModule(gallivm
->module
);
359 if (!gallivm
->provider
)
362 gallivm
->builder
= LLVMCreateBuilderInContext(gallivm
->context
);
363 if (!gallivm
->builder
)
366 /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
367 * complete when MC-JIT is created. So defer the MC-JIT engine creation for
371 if (!init_gallivm_engine(gallivm
)) {
376 * MC-JIT engine compiles the module immediately on creation, so we can't
377 * obtain the target data from it. Instead we create a target data layout
380 * The produced layout strings are not precisely the same, but should make
381 * no difference for the kind of optimization passes we run.
383 * For reference this is the layout string on x64:
385 * 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
388 * - http://llvm.org/docs/LangRef.html#datalayout
392 const unsigned pointer_size
= 8 * sizeof(void *);
394 util_snprintf(layout
, sizeof layout
, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
395 #ifdef PIPE_ARCH_LITTLE_ENDIAN
396 'e', // little endian
400 pointer_size
, pointer_size
, pointer_size
, // pointer size, abi alignment, preferred alignment
401 pointer_size
, // aggregate preferred alignment
402 pointer_size
, pointer_size
); // stack objects abi alignment, preferred alignment
404 gallivm
->target
= LLVMCreateTargetData(layout
);
405 if (!gallivm
->target
) {
411 if (!create_pass_manager(gallivm
))
417 free_gallivm_state(gallivm
);
425 if (gallivm_initialized
)
429 gallivm_debug
= debug_get_option_gallivm_debug();
432 lp_set_target_options();
442 /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
443 * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
444 * actually more efficient to use 4-wide vectors on this processor.
447 * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
450 util_cpu_caps
.has_avx
&&
451 util_cpu_caps
.has_intel
) {
452 lp_native_vector_width
= 256;
454 /* Leave it at 128, even when no SIMD extensions are available.
455 * Really needs to be a multiple of 128 so can fit 4 floats.
457 lp_native_vector_width
= 128;
460 lp_native_vector_width
= debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
461 lp_native_vector_width
);
463 if (lp_native_vector_width
<= 128) {
464 /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
465 * "util_cpu_caps.has_avx" predicate, and lack the
466 * "lp_native_vector_width > 128" predicate. And also to ensure a more
467 * consistent behavior, allowing one to test SSE2 on AVX machines.
468 * XXX: should not play games with util_cpu_caps directly as it might
469 * get used for other things outside llvm too.
471 util_cpu_caps
.has_avx
= 0;
472 util_cpu_caps
.has_avx2
= 0;
477 * note these instructions are VEX-only, so can only emit if we use
478 * avx (don't want to base it on has_avx & has_f16c later as that would
479 * omit it unnecessarily on amd cpus, see above).
481 util_cpu_caps
.has_f16c
= 0;
482 util_cpu_caps
.has_xop
= 0;
485 #ifdef PIPE_ARCH_PPC_64
486 /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
487 * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
488 * that some rounding and half-float to float handling does not round
490 * XXX: should eventually follow same logic on all platforms.
491 * Right now denorms get explicitly disabled (but elsewhere) for x86,
492 * whereas ppc64 explicitly enables them...
494 if (util_cpu_caps
.has_altivec
) {
495 unsigned short mask
[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
496 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
507 gallivm_initialized
= TRUE
;
510 /* For simulating less capable machines */
511 util_cpu_caps
.has_sse3
= 0;
512 util_cpu_caps
.has_ssse3
= 0;
513 util_cpu_caps
.has_sse4_1
= 0;
514 util_cpu_caps
.has_avx
= 0;
515 util_cpu_caps
.has_f16c
= 0;
522 * Create a new gallivm_state object.
523 * Note that we return a singleton.
525 struct gallivm_state
*
528 struct gallivm_state
*gallivm
;
530 #if HAVE_LLVM <= 0x206
532 return GlobalGallivm
;
536 gallivm
= CALLOC_STRUCT(gallivm_state
);
538 if (!init_gallivm_state(gallivm
)) {
544 #if HAVE_LLVM <= 0x206
545 GlobalGallivm
= gallivm
;
553 * Destroy a gallivm_state object.
556 gallivm_destroy(struct gallivm_state
*gallivm
)
558 #if HAVE_LLVM <= 0x0206
559 /* No-op: don't destroy the singleton */
562 free_gallivm_state(gallivm
);
569 * Validate and optimze a function.
572 gallivm_optimize_function(struct gallivm_state
*gallivm
,
576 debug_printf("optimizing %s...\n", LLVMGetValueName(func
));
579 assert(gallivm
->passmgr
);
581 /* Apply optimizations to LLVM IR */
582 LLVMRunFunctionPassManager(gallivm
->passmgr
, func
);
585 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
586 /* Print the LLVM IR to stderr */
587 lp_debug_dump_value(func
);
595 * Validate a function.
598 gallivm_verify_function(struct gallivm_state
*gallivm
,
601 /* Verify the LLVM IR. If invalid, dump and abort */
603 if (LLVMVerifyFunction(func
, LLVMPrintMessageAction
)) {
604 lp_debug_dump_value(func
);
610 gallivm_optimize_function(gallivm
, func
);
612 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
613 /* Print the LLVM IR to stderr */
614 lp_debug_dump_value(func
);
621 gallivm_compile_module(struct gallivm_state
*gallivm
)
623 #if HAVE_LLVM > 0x206
624 assert(!gallivm
->compiled
);
627 /* Dump byte code to a file */
629 LLVMWriteBitcodeToFile(gallivm
->module
, "llvmpipe.bc");
630 debug_printf("llvmpipe.bc written\n");
631 debug_printf("Invoke as \"llc -o - llvmpipe.bc\"\n");
635 assert(!gallivm
->engine
);
636 if (!init_gallivm_engine(gallivm
)) {
640 assert(gallivm
->engine
);
648 gallivm_jit_function(struct gallivm_state
*gallivm
,
652 func_pointer jit_func
;
654 assert(gallivm
->compiled
);
655 assert(gallivm
->engine
);
657 code
= LLVMGetPointerToGlobal(gallivm
->engine
, func
);
659 jit_func
= pointer_to_func(code
);
661 if (gallivm_debug
& GALLIVM_DEBUG_ASM
) {
662 lp_disassemble(func
, code
);
666 lp_profile(func
, code
);
669 /* Free the function body to save memory */
670 lp_func_delete_body(func
);
677 * Free the function (and its machine code).
680 gallivm_free_function(struct gallivm_state
*gallivm
,
686 LLVMFreeMachineCodeForFunction(gallivm
->engine
, func
);
689 LLVMDeleteFunction(func
);