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 **************************************************************************/
30 * Helper functions for logical operations.
32 * @author Jose Fonseca <jfonseca@vmware.com>
36 #include "util/u_cpu_detect.h"
38 #include "lp_bld_type.h"
39 #include "lp_bld_const.h"
40 #include "lp_bld_intr.h"
41 #include "lp_bld_logic.h"
45 lp_build_cmp(struct lp_build_context
*bld
,
50 const struct lp_type type
= bld
->type
;
51 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
52 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
53 LLVMValueRef zeros
= LLVMConstNull(int_vec_type
);
54 LLVMValueRef ones
= LLVMConstAllOnes(int_vec_type
);
59 if(func
== PIPE_FUNC_NEVER
)
61 if(func
== PIPE_FUNC_ALWAYS
)
64 /* TODO: optimize the constant case */
66 /* XXX: It is not clear if we should use the ordered or unordered operators */
68 #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
69 if(type
.width
* type
.length
== 128) {
70 if(type
.floating
&& util_cpu_caps
.has_sse
) {
80 case PIPE_FUNC_NOTEQUAL
:
86 case PIPE_FUNC_LEQUAL
:
89 case PIPE_FUNC_GREATER
:
93 case PIPE_FUNC_GEQUAL
:
111 args
[2] = LLVMConstInt(LLVMInt8Type(), cc
, 0);
112 res
= lp_build_intrinsic(bld
->builder
,
113 "llvm.x86.sse.cmp.ps",
116 res
= LLVMBuildBitCast(bld
->builder
, res
, int_vec_type
, "");
119 else if(util_cpu_caps
.has_sse2
) {
120 static const struct {
126 {0, 0, 0, 1}, /* PIPE_FUNC_NEVER */
127 {1, 0, 1, 0}, /* PIPE_FUNC_LESS */
128 {0, 1, 0, 0}, /* PIPE_FUNC_EQUAL */
129 {0, 0, 1, 1}, /* PIPE_FUNC_LEQUAL */
130 {0, 0, 1, 0}, /* PIPE_FUNC_GREATER */
131 {0, 1, 0, 1}, /* PIPE_FUNC_NOTEQUAL */
132 {1, 0, 1, 1}, /* PIPE_FUNC_GEQUAL */
133 {0, 0, 0, 0} /* PIPE_FUNC_ALWAYS */
137 LLVMValueRef args
[2];
140 switch (type
.width
) {
142 pcmpeq
= "llvm.x86.sse2.pcmpeq.b";
143 pcmpgt
= "llvm.x86.sse2.pcmpgt.b";
146 pcmpeq
= "llvm.x86.sse2.pcmpeq.w";
147 pcmpgt
= "llvm.x86.sse2.pcmpgt.w";
150 pcmpeq
= "llvm.x86.sse2.pcmpeq.d";
151 pcmpgt
= "llvm.x86.sse2.pcmpgt.d";
158 /* There are no signed byte and unsigned word/dword comparison
159 * instructions. So flip the sign bit so that the results match.
162 ((type
.width
== 8 && type
.sign
) ||
163 (type
.width
!= 8 && !type
.sign
))) {
164 LLVMValueRef msb
= lp_build_int_const_scalar(type
, (unsigned long long)1 << (type
.width
- 1));
165 a
= LLVMBuildXor(bld
->builder
, a
, msb
, "");
166 b
= LLVMBuildXor(bld
->builder
, b
, msb
, "");
169 if(table
[func
].swap
) {
179 res
= lp_build_intrinsic(bld
->builder
, pcmpeq
, vec_type
, args
, 2);
180 else if (table
[func
].gt
)
181 res
= lp_build_intrinsic(bld
->builder
, pcmpgt
, vec_type
, args
, 2);
183 res
= LLVMConstNull(vec_type
);
186 res
= LLVMBuildNot(bld
->builder
, res
, "");
194 LLVMRealPredicate op
;
196 case PIPE_FUNC_NEVER
:
197 op
= LLVMRealPredicateFalse
;
199 case PIPE_FUNC_ALWAYS
:
200 op
= LLVMRealPredicateTrue
;
202 case PIPE_FUNC_EQUAL
:
205 case PIPE_FUNC_NOTEQUAL
:
211 case PIPE_FUNC_LEQUAL
:
214 case PIPE_FUNC_GREATER
:
217 case PIPE_FUNC_GEQUAL
:
226 /* XXX: Although valid IR, no LLVM target currently support this */
227 cond
= LLVMBuildFCmp(bld
->builder
, op
, a
, b
, "");
228 res
= LLVMBuildSelect(bld
->builder
, cond
, ones
, zeros
, "");
230 debug_printf("%s: warning: using slow element-wise vector comparison\n",
232 res
= LLVMGetUndef(int_vec_type
);
233 for(i
= 0; i
< type
.length
; ++i
) {
234 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
235 cond
= LLVMBuildFCmp(bld
->builder
, op
,
236 LLVMBuildExtractElement(bld
->builder
, a
, index
, ""),
237 LLVMBuildExtractElement(bld
->builder
, b
, index
, ""),
239 cond
= LLVMBuildSelect(bld
->builder
, cond
,
240 LLVMConstExtractElement(ones
, index
),
241 LLVMConstExtractElement(zeros
, index
),
243 res
= LLVMBuildInsertElement(bld
->builder
, res
, cond
, index
, "");
250 case PIPE_FUNC_EQUAL
:
253 case PIPE_FUNC_NOTEQUAL
:
257 op
= type
.sign
? LLVMIntSLT
: LLVMIntULT
;
259 case PIPE_FUNC_LEQUAL
:
260 op
= type
.sign
? LLVMIntSLE
: LLVMIntULE
;
262 case PIPE_FUNC_GREATER
:
263 op
= type
.sign
? LLVMIntSGT
: LLVMIntUGT
;
265 case PIPE_FUNC_GEQUAL
:
266 op
= type
.sign
? LLVMIntSGE
: LLVMIntUGE
;
274 /* XXX: Although valid IR, no LLVM target currently support this */
275 cond
= LLVMBuildICmp(bld
->builder
, op
, a
, b
, "");
276 res
= LLVMBuildSelect(bld
->builder
, cond
, ones
, zeros
, "");
278 debug_printf("%s: warning: using slow element-wise vector comparison\n",
280 res
= LLVMGetUndef(int_vec_type
);
281 for(i
= 0; i
< type
.length
; ++i
) {
282 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), i
, 0);
283 cond
= LLVMBuildICmp(bld
->builder
, op
,
284 LLVMBuildExtractElement(bld
->builder
, a
, index
, ""),
285 LLVMBuildExtractElement(bld
->builder
, b
, index
, ""),
287 cond
= LLVMBuildSelect(bld
->builder
, cond
,
288 LLVMConstExtractElement(ones
, index
),
289 LLVMConstExtractElement(zeros
, index
),
291 res
= LLVMBuildInsertElement(bld
->builder
, res
, cond
, index
, "");
301 lp_build_select(struct lp_build_context
*bld
,
306 struct lp_type type
= bld
->type
;
313 LLVMTypeRef int_vec_type
= lp_build_int_vec_type(type
);
314 a
= LLVMBuildBitCast(bld
->builder
, a
, int_vec_type
, "");
315 b
= LLVMBuildBitCast(bld
->builder
, b
, int_vec_type
, "");
318 a
= LLVMBuildAnd(bld
->builder
, a
, mask
, "");
320 /* This often gets translated to PANDN, but sometimes the NOT is
321 * pre-computed and stored in another constant. The best strategy depends
322 * on available registers, so it is not a big deal -- hopefully LLVM does
323 * the right decision attending the rest of the program.
325 b
= LLVMBuildAnd(bld
->builder
, b
, LLVMBuildNot(bld
->builder
, mask
, ""), "");
327 res
= LLVMBuildOr(bld
->builder
, a
, b
, "");
330 LLVMTypeRef vec_type
= lp_build_vec_type(type
);
331 res
= LLVMBuildBitCast(bld
->builder
, res
, vec_type
, "");
339 lp_build_select_aos(struct lp_build_context
*bld
,
342 const boolean cond
[4])
344 const struct lp_type type
= bld
->type
;
345 const unsigned n
= type
.length
;
350 if(cond
[0] && cond
[1] && cond
[2] && cond
[3])
352 if(!cond
[0] && !cond
[1] && !cond
[2] && !cond
[3])
354 if(a
== bld
->undef
|| b
== bld
->undef
)
358 * There are three major ways of accomplishing this:
361 * - or with a bit mask.
363 * Select isn't supported for vector types yet.
364 * The flip between these is empirical and might need to be.
370 LLVMTypeRef elem_type
= LLVMInt32Type();
371 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
373 for(j
= 0; j
< n
; j
+= 4)
374 for(i
= 0; i
< 4; ++i
)
375 shuffles
[j
+ i
] = LLVMConstInt(elem_type
, (cond
[i
] ? 0 : n
) + j
+ i
, 0);
377 return LLVMBuildShuffleVector(bld
->builder
, a
, b
, LLVMConstVector(shuffles
, n
), "");
381 /* XXX: Unfortunately select of vectors do not work */
383 LLVMTypeRef elem_type
= LLVMInt1Type();
384 LLVMValueRef cond
[LP_MAX_VECTOR_LENGTH
];
386 for(j
= 0; j
< n
; j
+= 4)
387 for(i
= 0; i
< 4; ++i
)
388 cond
[j
+ i
] = LLVMConstInt(elem_type
, cond
[i
] ? 1 : 0, 0);
390 return LLVMBuildSelect(bld
->builder
, LLVMConstVector(cond
, n
), a
, b
, "");
392 LLVMValueRef mask
= lp_build_const_mask_aos(type
, cond
);
393 return lp_build_select(bld
, mask
, a
, b
);