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 **************************************************************************/
31 * Helpers for emiting intrinsic calls.
33 * LLVM vanilla IR doesn't represent all basic arithmetic operations we care
34 * about, and it is often necessary to resort target-specific intrinsics for
35 * performance, convenience.
37 * Ideally we would like to stay away from target specific intrinsics and
38 * move all the instruction selection logic into upstream LLVM where it belongs.
40 * These functions are also used for calling C functions provided by us from
41 * generated LLVM code.
43 * @author Jose Fonseca <jfonseca@vmware.com>
47 #include "util/u_debug.h"
48 #include "util/u_string.h"
49 #include "util/bitscan.h"
51 #include "lp_bld_const.h"
52 #include "lp_bld_intr.h"
53 #include "lp_bld_type.h"
54 #include "lp_bld_pack.h"
55 #include "lp_bld_debug.h"
59 lp_format_intrinsic(char *name
,
61 const char *name_root
,
68 LLVMTypeKind kind
= LLVMGetTypeKind(type
);
69 if (kind
== LLVMVectorTypeKind
) {
70 length
= LLVMGetVectorSize(type
);
71 type
= LLVMGetElementType(type
);
72 kind
= LLVMGetTypeKind(type
);
76 case LLVMIntegerTypeKind
:
78 width
= LLVMGetIntTypeWidth(type
);
80 case LLVMFloatTypeKind
:
84 case LLVMDoubleTypeKind
:
89 unreachable("unexpected LLVMTypeKind");
93 util_snprintf(name
, size
, "%s.v%u%c%u", name_root
, length
, c
, width
);
95 util_snprintf(name
, size
, "%s.%c%u", name_root
, c
, width
);
101 lp_declare_intrinsic(LLVMModuleRef module
,
103 LLVMTypeRef ret_type
,
104 LLVMTypeRef
*arg_types
,
107 LLVMTypeRef function_type
;
108 LLVMValueRef function
;
110 assert(!LLVMGetNamedFunction(module
, name
));
112 function_type
= LLVMFunctionType(ret_type
, arg_types
, num_args
, 0);
113 function
= LLVMAddFunction(module
, name
, function_type
);
115 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
116 LLVMSetLinkage(function
, LLVMExternalLinkage
);
118 assert(LLVMIsDeclaration(function
));
124 #if HAVE_LLVM < 0x0400
125 static LLVMAttribute
lp_attr_to_llvm_attr(enum lp_func_attr attr
)
128 case LP_FUNC_ATTR_ALWAYSINLINE
: return LLVMAlwaysInlineAttribute
;
129 case LP_FUNC_ATTR_BYVAL
: return LLVMByValAttribute
;
130 case LP_FUNC_ATTR_INREG
: return LLVMInRegAttribute
;
131 case LP_FUNC_ATTR_NOALIAS
: return LLVMNoAliasAttribute
;
132 case LP_FUNC_ATTR_NOUNWIND
: return LLVMNoUnwindAttribute
;
133 case LP_FUNC_ATTR_READNONE
: return LLVMReadNoneAttribute
;
134 case LP_FUNC_ATTR_READONLY
: return LLVMReadOnlyAttribute
;
136 _debug_printf("Unhandled function attribute: %x\n", attr
);
143 static const char *attr_to_str(enum lp_func_attr attr
)
146 case LP_FUNC_ATTR_ALWAYSINLINE
: return "alwaysinline";
147 case LP_FUNC_ATTR_BYVAL
: return "byval";
148 case LP_FUNC_ATTR_INREG
: return "inreg";
149 case LP_FUNC_ATTR_NOALIAS
: return "noalias";
150 case LP_FUNC_ATTR_NOUNWIND
: return "nounwind";
151 case LP_FUNC_ATTR_READNONE
: return "readnone";
152 case LP_FUNC_ATTR_READONLY
: return "readonly";
154 _debug_printf("Unhandled function attribute: %x\n", attr
);
162 lp_add_function_attr(LLVMValueRef function
,
164 enum lp_func_attr attr
)
167 #if HAVE_LLVM < 0x0400
168 LLVMAttribute llvm_attr
= lp_attr_to_llvm_attr(attr
);
169 if (attr_idx
== -1) {
170 LLVMAddFunctionAttr(function
, llvm_attr
);
172 LLVMAddAttribute(LLVMGetParam(function
, attr_idx
- 1), llvm_attr
);
175 LLVMContextRef context
= LLVMGetModuleContext(LLVMGetGlobalParent(function
));
176 const char *attr_name
= attr_to_str(attr
);
177 unsigned kind_id
= LLVMGetEnumAttributeKindForName(attr_name
,
179 LLVMAttributeRef llvm_attr
= LLVMCreateEnumAttribute(context
, kind_id
, 0);
180 LLVMAddAttributeAtIndex(function
, attr_idx
, llvm_attr
);
185 lp_build_intrinsic(LLVMBuilderRef builder
,
187 LLVMTypeRef ret_type
,
192 LLVMModuleRef module
= LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder
)));
193 LLVMValueRef function
;
195 function
= LLVMGetNamedFunction(module
, name
);
197 LLVMTypeRef arg_types
[LP_MAX_FUNC_ARGS
];
200 assert(num_args
<= LP_MAX_FUNC_ARGS
);
202 for(i
= 0; i
< num_args
; ++i
) {
204 arg_types
[i
] = LLVMTypeOf(args
[i
]);
207 function
= lp_declare_intrinsic(module
, name
, ret_type
, arg_types
, num_args
);
209 /* NoUnwind indicates that the intrinsic never raises a C++ exception.
210 * Set it for all intrinsics.
212 attr_mask
|= LP_FUNC_ATTR_NOUNWIND
;
215 enum lp_func_attr attr
= 1 << u_bit_scan(&attr_mask
);
216 lp_add_function_attr(function
, -1, attr
);
219 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
220 lp_debug_dump_value(function
);
224 return LLVMBuildCall(builder
, function
, args
, num_args
, "");
229 lp_build_intrinsic_unary(LLVMBuilderRef builder
,
231 LLVMTypeRef ret_type
,
234 return lp_build_intrinsic(builder
, name
, ret_type
, &a
, 1, 0);
239 lp_build_intrinsic_binary(LLVMBuilderRef builder
,
241 LLVMTypeRef ret_type
,
245 LLVMValueRef args
[2];
250 return lp_build_intrinsic(builder
, name
, ret_type
, args
, 2, 0);
255 * Call intrinsic with arguments adapted to intrinsic vector length.
257 * Split vectors which are too large for the hw, or expand them if they
258 * are too small, so a caller calling a function which might use intrinsics
259 * doesn't need to do splitting/expansion on its own.
260 * This only supports intrinsics where src and dst types match.
263 lp_build_intrinsic_binary_anylength(struct gallivm_state
*gallivm
,
265 struct lp_type src_type
,
271 struct lp_type intrin_type
= src_type
;
272 LLVMBuilderRef builder
= gallivm
->builder
;
273 LLVMValueRef i32undef
= LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
274 LLVMValueRef anative
, bnative
;
275 unsigned intrin_length
= intr_size
/ src_type
.width
;
277 intrin_type
.length
= intrin_length
;
279 if (intrin_length
> src_type
.length
) {
280 LLVMValueRef elems
[LP_MAX_VECTOR_LENGTH
];
281 LLVMValueRef constvec
, tmp
;
283 for (i
= 0; i
< src_type
.length
; i
++) {
284 elems
[i
] = lp_build_const_int32(gallivm
, i
);
286 for (; i
< intrin_length
; i
++) {
289 if (src_type
.length
== 1) {
290 LLVMTypeRef elem_type
= lp_build_elem_type(gallivm
, intrin_type
);
291 a
= LLVMBuildBitCast(builder
, a
, LLVMVectorType(elem_type
, 1), "");
292 b
= LLVMBuildBitCast(builder
, b
, LLVMVectorType(elem_type
, 1), "");
294 constvec
= LLVMConstVector(elems
, intrin_length
);
295 anative
= LLVMBuildShuffleVector(builder
, a
, a
, constvec
, "");
296 bnative
= LLVMBuildShuffleVector(builder
, b
, b
, constvec
, "");
297 tmp
= lp_build_intrinsic_binary(builder
, name
,
298 lp_build_vec_type(gallivm
, intrin_type
),
300 if (src_type
.length
> 1) {
301 constvec
= LLVMConstVector(elems
, src_type
.length
);
302 return LLVMBuildShuffleVector(builder
, tmp
, tmp
, constvec
, "");
305 return LLVMBuildExtractElement(builder
, tmp
, elems
[0], "");
308 else if (intrin_length
< src_type
.length
) {
309 unsigned num_vec
= src_type
.length
/ intrin_length
;
310 LLVMValueRef tmp
[LP_MAX_VECTOR_LENGTH
];
312 /* don't support arbitrary size here as this is so yuck */
313 if (src_type
.length
% intrin_length
) {
314 /* FIXME: This is something which should be supported
315 * but there doesn't seem to be any need for it currently
318 debug_printf("%s: should handle arbitrary vector size\n",
324 for (i
= 0; i
< num_vec
; i
++) {
325 anative
= lp_build_extract_range(gallivm
, a
, i
*intrin_length
,
327 bnative
= lp_build_extract_range(gallivm
, b
, i
*intrin_length
,
329 tmp
[i
] = lp_build_intrinsic_binary(builder
, name
,
330 lp_build_vec_type(gallivm
, intrin_type
),
333 return lp_build_concat(gallivm
, tmp
, intrin_type
, num_vec
);
336 return lp_build_intrinsic_binary(builder
, name
,
337 lp_build_vec_type(gallivm
, src_type
),
344 lp_build_intrinsic_map(struct gallivm_state
*gallivm
,
346 LLVMTypeRef ret_type
,
350 LLVMBuilderRef builder
= gallivm
->builder
;
351 LLVMTypeRef ret_elem_type
= LLVMGetElementType(ret_type
);
352 unsigned n
= LLVMGetVectorSize(ret_type
);
356 assert(num_args
<= LP_MAX_FUNC_ARGS
);
358 res
= LLVMGetUndef(ret_type
);
359 for(i
= 0; i
< n
; ++i
) {
360 LLVMValueRef index
= lp_build_const_int32(gallivm
, i
);
361 LLVMValueRef arg_elems
[LP_MAX_FUNC_ARGS
];
362 LLVMValueRef res_elem
;
363 for(j
= 0; j
< num_args
; ++j
)
364 arg_elems
[j
] = LLVMBuildExtractElement(builder
, args
[j
], index
, "");
365 res_elem
= lp_build_intrinsic(builder
, name
, ret_elem_type
, arg_elems
, num_args
, 0);
366 res
= LLVMBuildInsertElement(builder
, res
, res_elem
, index
, "");
374 lp_build_intrinsic_map_unary(struct gallivm_state
*gallivm
,
376 LLVMTypeRef ret_type
,
379 return lp_build_intrinsic_map(gallivm
, name
, ret_type
, &a
, 1);
384 lp_build_intrinsic_map_binary(struct gallivm_state
*gallivm
,
386 LLVMTypeRef ret_type
,
390 LLVMValueRef args
[2];
395 return lp_build_intrinsic_map(gallivm
, name
, ret_type
, args
, 2);