ac_llvm_build.c revision b8e80941 
1/*
2 * Copyright 2014 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the
6 * "Software"), to deal in the Software without restriction, including
7 * without limitation the rights to use, copy, modify, merge, publish,
8 * distribute, sub license, and/or sell copies of the Software, and to
9 * permit persons to whom the Software is furnished to do so, subject to
10 * the following conditions:
11 *
12 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
13 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
14 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
15 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
16 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
17 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
18 * USE OR OTHER DEALINGS IN THE SOFTWARE.
19 *
20 * The above copyright notice and this permission notice (including the
21 * next paragraph) shall be included in all copies or substantial portions
22 * of the Software.
23 *
24 */
25/* based on pieces from si_pipe.c and radeon_llvm_emit.c */
26#include "ac_llvm_build.h"
27
28#include <llvm-c/Core.h>
29
30#include "c11/threads.h"
31
32#include <assert.h>
33#include <stdio.h>
34
35#include "ac_llvm_util.h"
36#include "ac_exp_param.h"
37#include "util/bitscan.h"
38#include "util/macros.h"
39#include "util/u_atomic.h"
40#include "util/u_math.h"
41#include "sid.h"
42
43#include "shader_enums.h"
44
45#define AC_LLVM_INITIAL_CF_DEPTH 4
46
47/* Data for if/else/endif and bgnloop/endloop control flow structures.
48 */
49struct ac_llvm_flow {
50	/* Loop exit or next part of if/else/endif. */
51	LLVMBasicBlockRef next_block;
52	LLVMBasicBlockRef loop_entry_block;
53};
54
55/* Initialize module-independent parts of the context.
56 *
57 * The caller is responsible for initializing ctx::module and ctx::builder.
58 */
59void
60ac_llvm_context_init(struct ac_llvm_context *ctx,
61		     enum chip_class chip_class, enum radeon_family family)
62{
63	LLVMValueRef args[1];
64
65	ctx->context = LLVMContextCreate();
66
67	ctx->chip_class = chip_class;
68	ctx->family = family;
69	ctx->module = NULL;
70	ctx->builder = NULL;
71
72	ctx->voidt = LLVMVoidTypeInContext(ctx->context);
73	ctx->i1 = LLVMInt1TypeInContext(ctx->context);
74	ctx->i8 = LLVMInt8TypeInContext(ctx->context);
75	ctx->i16 = LLVMIntTypeInContext(ctx->context, 16);
76	ctx->i32 = LLVMIntTypeInContext(ctx->context, 32);
77	ctx->i64 = LLVMIntTypeInContext(ctx->context, 64);
78	ctx->intptr = ctx->i32;
79	ctx->f16 = LLVMHalfTypeInContext(ctx->context);
80	ctx->f32 = LLVMFloatTypeInContext(ctx->context);
81	ctx->f64 = LLVMDoubleTypeInContext(ctx->context);
82	ctx->v2i16 = LLVMVectorType(ctx->i16, 2);
83	ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
84	ctx->v3i32 = LLVMVectorType(ctx->i32, 3);
85	ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
86	ctx->v2f32 = LLVMVectorType(ctx->f32, 2);
87	ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
88	ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
89
90	ctx->i8_0 = LLVMConstInt(ctx->i8, 0, false);
91	ctx->i8_1 = LLVMConstInt(ctx->i8, 1, false);
92	ctx->i16_0 = LLVMConstInt(ctx->i16, 0, false);
93	ctx->i16_1 = LLVMConstInt(ctx->i16, 1, false);
94	ctx->i32_0 = LLVMConstInt(ctx->i32, 0, false);
95	ctx->i32_1 = LLVMConstInt(ctx->i32, 1, false);
96	ctx->i64_0 = LLVMConstInt(ctx->i64, 0, false);
97	ctx->i64_1 = LLVMConstInt(ctx->i64, 1, false);
98	ctx->f16_0 = LLVMConstReal(ctx->f16, 0.0);
99	ctx->f16_1 = LLVMConstReal(ctx->f16, 1.0);
100	ctx->f32_0 = LLVMConstReal(ctx->f32, 0.0);
101	ctx->f32_1 = LLVMConstReal(ctx->f32, 1.0);
102	ctx->f64_0 = LLVMConstReal(ctx->f64, 0.0);
103	ctx->f64_1 = LLVMConstReal(ctx->f64, 1.0);
104
105	ctx->i1false = LLVMConstInt(ctx->i1, 0, false);
106	ctx->i1true = LLVMConstInt(ctx->i1, 1, false);
107
108	ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context,
109						     "range", 5);
110
111	ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context,
112							       "invariant.load", 14);
113
114	ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->context, "fpmath", 6);
115
116	args[0] = LLVMConstReal(ctx->f32, 2.5);
117	ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->context, args, 1);
118
119	ctx->uniform_md_kind = LLVMGetMDKindIDInContext(ctx->context,
120							"amdgpu.uniform", 14);
121
122	ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
123}
124
125void
126ac_llvm_context_dispose(struct ac_llvm_context *ctx)
127{
128	free(ctx->flow);
129	ctx->flow = NULL;
130	ctx->flow_depth_max = 0;
131}
132
133int
134ac_get_llvm_num_components(LLVMValueRef value)
135{
136	LLVMTypeRef type = LLVMTypeOf(value);
137	unsigned num_components = LLVMGetTypeKind(type) == LLVMVectorTypeKind
138	                              ? LLVMGetVectorSize(type)
139	                              : 1;
140	return num_components;
141}
142
143LLVMValueRef
144ac_llvm_extract_elem(struct ac_llvm_context *ac,
145		     LLVMValueRef value,
146		     int index)
147{
148	if (LLVMGetTypeKind(LLVMTypeOf(value)) != LLVMVectorTypeKind) {
149		assert(index == 0);
150		return value;
151	}
152
153	return LLVMBuildExtractElement(ac->builder, value,
154				       LLVMConstInt(ac->i32, index, false), "");
155}
156
157int
158ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type)
159{
160	if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
161		type = LLVMGetElementType(type);
162
163	if (LLVMGetTypeKind(type) == LLVMIntegerTypeKind)
164		return LLVMGetIntTypeWidth(type);
165
166	if (type == ctx->f16)
167		return 16;
168	if (type == ctx->f32)
169		return 32;
170	if (type == ctx->f64)
171		return 64;
172
173	unreachable("Unhandled type kind in get_elem_bits");
174}
175
176unsigned
177ac_get_type_size(LLVMTypeRef type)
178{
179	LLVMTypeKind kind = LLVMGetTypeKind(type);
180
181	switch (kind) {
182	case LLVMIntegerTypeKind:
183		return LLVMGetIntTypeWidth(type) / 8;
184	case LLVMHalfTypeKind:
185		return 2;
186	case LLVMFloatTypeKind:
187		return 4;
188	case LLVMDoubleTypeKind:
189		return 8;
190	case LLVMPointerTypeKind:
191		if (LLVMGetPointerAddressSpace(type) == AC_ADDR_SPACE_CONST_32BIT)
192			return 4;
193		return 8;
194	case LLVMVectorTypeKind:
195		return LLVMGetVectorSize(type) *
196		       ac_get_type_size(LLVMGetElementType(type));
197	case LLVMArrayTypeKind:
198		return LLVMGetArrayLength(type) *
199		       ac_get_type_size(LLVMGetElementType(type));
200	default:
201		assert(0);
202		return 0;
203	}
204}
205
206static LLVMTypeRef to_integer_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
207{
208	if (t == ctx->i8)
209		return ctx->i8;
210	else if (t == ctx->f16 || t == ctx->i16)
211		return ctx->i16;
212	else if (t == ctx->f32 || t == ctx->i32)
213		return ctx->i32;
214	else if (t == ctx->f64 || t == ctx->i64)
215		return ctx->i64;
216	else
217		unreachable("Unhandled integer size");
218}
219
220LLVMTypeRef
221ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
222{
223	if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
224		LLVMTypeRef elem_type = LLVMGetElementType(t);
225		return LLVMVectorType(to_integer_type_scalar(ctx, elem_type),
226		                      LLVMGetVectorSize(t));
227	}
228	if (LLVMGetTypeKind(t) == LLVMPointerTypeKind) {
229		switch (LLVMGetPointerAddressSpace(t)) {
230		case AC_ADDR_SPACE_GLOBAL:
231			return ctx->i64;
232		case AC_ADDR_SPACE_LDS:
233			return ctx->i32;
234		default:
235			unreachable("unhandled address space");
236		}
237	}
238	return to_integer_type_scalar(ctx, t);
239}
240
241LLVMValueRef
242ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v)
243{
244	LLVMTypeRef type = LLVMTypeOf(v);
245	if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
246		return LLVMBuildPtrToInt(ctx->builder, v, ac_to_integer_type(ctx, type), "");
247	}
248	return LLVMBuildBitCast(ctx->builder, v, ac_to_integer_type(ctx, type), "");
249}
250
251LLVMValueRef
252ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v)
253{
254	LLVMTypeRef type = LLVMTypeOf(v);
255	if (LLVMGetTypeKind(type) == LLVMPointerTypeKind)
256		return v;
257	return ac_to_integer(ctx, v);
258}
259
260static LLVMTypeRef to_float_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
261{
262	if (t == ctx->i8)
263		return ctx->i8;
264	else if (t == ctx->i16 || t == ctx->f16)
265		return ctx->f16;
266	else if (t == ctx->i32 || t == ctx->f32)
267		return ctx->f32;
268	else if (t == ctx->i64 || t == ctx->f64)
269		return ctx->f64;
270	else
271		unreachable("Unhandled float size");
272}
273
274LLVMTypeRef
275ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
276{
277	if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
278		LLVMTypeRef elem_type = LLVMGetElementType(t);
279		return LLVMVectorType(to_float_type_scalar(ctx, elem_type),
280		                      LLVMGetVectorSize(t));
281	}
282	return to_float_type_scalar(ctx, t);
283}
284
285LLVMValueRef
286ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v)
287{
288	LLVMTypeRef type = LLVMTypeOf(v);
289	return LLVMBuildBitCast(ctx->builder, v, ac_to_float_type(ctx, type), "");
290}
291
292
293LLVMValueRef
294ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
295		   LLVMTypeRef return_type, LLVMValueRef *params,
296		   unsigned param_count, unsigned attrib_mask)
297{
298	LLVMValueRef function, call;
299	bool set_callsite_attrs = !(attrib_mask & AC_FUNC_ATTR_LEGACY);
300
301	function = LLVMGetNamedFunction(ctx->module, name);
302	if (!function) {
303		LLVMTypeRef param_types[32], function_type;
304		unsigned i;
305
306		assert(param_count <= 32);
307
308		for (i = 0; i < param_count; ++i) {
309			assert(params[i]);
310			param_types[i] = LLVMTypeOf(params[i]);
311		}
312		function_type =
313		    LLVMFunctionType(return_type, param_types, param_count, 0);
314		function = LLVMAddFunction(ctx->module, name, function_type);
315
316		LLVMSetFunctionCallConv(function, LLVMCCallConv);
317		LLVMSetLinkage(function, LLVMExternalLinkage);
318
319		if (!set_callsite_attrs)
320			ac_add_func_attributes(ctx->context, function, attrib_mask);
321	}
322
323	call = LLVMBuildCall(ctx->builder, function, params, param_count, "");
324	if (set_callsite_attrs)
325		ac_add_func_attributes(ctx->context, call, attrib_mask);
326	return call;
327}
328
329/**
330 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
331 * intrinsic names).
332 */
333void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize)
334{
335	LLVMTypeRef elem_type = type;
336
337	assert(bufsize >= 8);
338
339	if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
340		int ret = snprintf(buf, bufsize, "v%u",
341					LLVMGetVectorSize(type));
342		if (ret < 0) {
343			char *type_name = LLVMPrintTypeToString(type);
344			fprintf(stderr, "Error building type name for: %s\n",
345				type_name);
346			return;
347		}
348		elem_type = LLVMGetElementType(type);
349		buf += ret;
350		bufsize -= ret;
351	}
352	switch (LLVMGetTypeKind(elem_type)) {
353	default: break;
354	case LLVMIntegerTypeKind:
355		snprintf(buf, bufsize, "i%d", LLVMGetIntTypeWidth(elem_type));
356		break;
357	case LLVMHalfTypeKind:
358		snprintf(buf, bufsize, "f16");
359		break;
360	case LLVMFloatTypeKind:
361		snprintf(buf, bufsize, "f32");
362		break;
363	case LLVMDoubleTypeKind:
364		snprintf(buf, bufsize, "f64");
365		break;
366	}
367}
368
369/**
370 * Helper function that builds an LLVM IR PHI node and immediately adds
371 * incoming edges.
372 */
373LLVMValueRef
374ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
375	     unsigned count_incoming, LLVMValueRef *values,
376	     LLVMBasicBlockRef *blocks)
377{
378	LLVMValueRef phi = LLVMBuildPhi(ctx->builder, type, "");
379	LLVMAddIncoming(phi, values, blocks, count_incoming);
380	return phi;
381}
382
383void ac_build_s_barrier(struct ac_llvm_context *ctx)
384{
385	ac_build_intrinsic(ctx, "llvm.amdgcn.s.barrier", ctx->voidt, NULL,
386			   0, AC_FUNC_ATTR_CONVERGENT);
387}
388
389/* Prevent optimizations (at least of memory accesses) across the current
390 * point in the program by emitting empty inline assembly that is marked as
391 * having side effects.
392 *
393 * Optionally, a value can be passed through the inline assembly to prevent
394 * LLVM from hoisting calls to ReadNone functions.
395 */
396void
397ac_build_optimization_barrier(struct ac_llvm_context *ctx,
398			      LLVMValueRef *pvgpr)
399{
400	static int counter = 0;
401
402	LLVMBuilderRef builder = ctx->builder;
403	char code[16];
404
405	snprintf(code, sizeof(code), "; %d", p_atomic_inc_return(&counter));
406
407	if (!pvgpr) {
408		LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
409		LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "", true, false);
410		LLVMBuildCall(builder, inlineasm, NULL, 0, "");
411	} else {
412		LLVMTypeRef ftype = LLVMFunctionType(ctx->i32, &ctx->i32, 1, false);
413		LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "=v,0", true, false);
414		LLVMValueRef vgpr = *pvgpr;
415		LLVMTypeRef vgpr_type = LLVMTypeOf(vgpr);
416		unsigned vgpr_size = ac_get_type_size(vgpr_type);
417		LLVMValueRef vgpr0;
418
419		assert(vgpr_size % 4 == 0);
420
421		vgpr = LLVMBuildBitCast(builder, vgpr, LLVMVectorType(ctx->i32, vgpr_size / 4), "");
422		vgpr0 = LLVMBuildExtractElement(builder, vgpr, ctx->i32_0, "");
423		vgpr0 = LLVMBuildCall(builder, inlineasm, &vgpr0, 1, "");
424		vgpr = LLVMBuildInsertElement(builder, vgpr, vgpr0, ctx->i32_0, "");
425		vgpr = LLVMBuildBitCast(builder, vgpr, vgpr_type, "");
426
427		*pvgpr = vgpr;
428	}
429}
430
431LLVMValueRef
432ac_build_shader_clock(struct ac_llvm_context *ctx)
433{
434	LLVMValueRef tmp = ac_build_intrinsic(ctx, "llvm.readcyclecounter",
435					      ctx->i64, NULL, 0, 0);
436	return LLVMBuildBitCast(ctx->builder, tmp, ctx->v2i32, "");
437}
438
439LLVMValueRef
440ac_build_ballot(struct ac_llvm_context *ctx,
441		LLVMValueRef value)
442{
443	LLVMValueRef args[3] = {
444		value,
445		ctx->i32_0,
446		LLVMConstInt(ctx->i32, LLVMIntNE, 0)
447	};
448
449	/* We currently have no other way to prevent LLVM from lifting the icmp
450	 * calls to a dominating basic block.
451	 */
452	ac_build_optimization_barrier(ctx, &args[0]);
453
454	args[0] = ac_to_integer(ctx, args[0]);
455
456	return ac_build_intrinsic(ctx,
457				  "llvm.amdgcn.icmp.i32",
458				  ctx->i64, args, 3,
459				  AC_FUNC_ATTR_NOUNWIND |
460				  AC_FUNC_ATTR_READNONE |
461				  AC_FUNC_ATTR_CONVERGENT);
462}
463
464LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
465				 LLVMValueRef value)
466{
467	LLVMValueRef args[3] = {
468		value,
469		ctx->i1false,
470		LLVMConstInt(ctx->i32, LLVMIntNE, 0),
471	};
472
473	assert(HAVE_LLVM >= 0x0800);
474	return ac_build_intrinsic(ctx, "llvm.amdgcn.icmp.i1", ctx->i64, args, 3,
475				  AC_FUNC_ATTR_NOUNWIND |
476				  AC_FUNC_ATTR_READNONE |
477				  AC_FUNC_ATTR_CONVERGENT);
478}
479
480LLVMValueRef
481ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value)
482{
483	LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
484	LLVMValueRef vote_set = ac_build_ballot(ctx, value);
485	return LLVMBuildICmp(ctx->builder, LLVMIntEQ, vote_set, active_set, "");
486}
487
488LLVMValueRef
489ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value)
490{
491	LLVMValueRef vote_set = ac_build_ballot(ctx, value);
492	return LLVMBuildICmp(ctx->builder, LLVMIntNE, vote_set,
493			     LLVMConstInt(ctx->i64, 0, 0), "");
494}
495
496LLVMValueRef
497ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value)
498{
499	LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
500	LLVMValueRef vote_set = ac_build_ballot(ctx, value);
501
502	LLVMValueRef all = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
503					 vote_set, active_set, "");
504	LLVMValueRef none = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
505					  vote_set,
506					  LLVMConstInt(ctx->i64, 0, 0), "");
507	return LLVMBuildOr(ctx->builder, all, none, "");
508}
509
510LLVMValueRef
511ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
512			       unsigned value_count, unsigned component)
513{
514	LLVMValueRef vec = NULL;
515
516	if (value_count == 1) {
517		return values[component];
518	} else if (!value_count)
519		unreachable("value_count is 0");
520
521	for (unsigned i = component; i < value_count + component; i++) {
522		LLVMValueRef value = values[i];
523
524		if (i == component)
525			vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
526		LLVMValueRef index = LLVMConstInt(ctx->i32, i - component, false);
527		vec = LLVMBuildInsertElement(ctx->builder, vec, value, index, "");
528	}
529	return vec;
530}
531
532LLVMValueRef
533ac_build_gather_values_extended(struct ac_llvm_context *ctx,
534				LLVMValueRef *values,
535				unsigned value_count,
536				unsigned value_stride,
537				bool load,
538				bool always_vector)
539{
540	LLVMBuilderRef builder = ctx->builder;
541	LLVMValueRef vec = NULL;
542	unsigned i;
543
544	if (value_count == 1 && !always_vector) {
545		if (load)
546			return LLVMBuildLoad(builder, values[0], "");
547		return values[0];
548	} else if (!value_count)
549		unreachable("value_count is 0");
550
551	for (i = 0; i < value_count; i++) {
552		LLVMValueRef value = values[i * value_stride];
553		if (load)
554			value = LLVMBuildLoad(builder, value, "");
555
556		if (!i)
557			vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
558		LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
559		vec = LLVMBuildInsertElement(builder, vec, value, index, "");
560	}
561	return vec;
562}
563
564LLVMValueRef
565ac_build_gather_values(struct ac_llvm_context *ctx,
566		       LLVMValueRef *values,
567		       unsigned value_count)
568{
569	return ac_build_gather_values_extended(ctx, values, value_count, 1, false, false);
570}
571
572/* Expand a scalar or vector to <dst_channels x type> by filling the remaining
573 * channels with undef. Extract at most src_channels components from the input.
574 */
575static LLVMValueRef
576ac_build_expand(struct ac_llvm_context *ctx,
577		LLVMValueRef value,
578		unsigned src_channels,
579		unsigned dst_channels)
580{
581	LLVMTypeRef elemtype;
582	LLVMValueRef chan[dst_channels];
583
584	if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMVectorTypeKind) {
585		unsigned vec_size = LLVMGetVectorSize(LLVMTypeOf(value));
586
587		if (src_channels == dst_channels && vec_size == dst_channels)
588			return value;
589
590		src_channels = MIN2(src_channels, vec_size);
591
592		for (unsigned i = 0; i < src_channels; i++)
593			chan[i] = ac_llvm_extract_elem(ctx, value, i);
594
595		elemtype = LLVMGetElementType(LLVMTypeOf(value));
596	} else {
597		if (src_channels) {
598			assert(src_channels == 1);
599			chan[0] = value;
600		}
601		elemtype = LLVMTypeOf(value);
602	}
603
604	for (unsigned i = src_channels; i < dst_channels; i++)
605		chan[i] = LLVMGetUndef(elemtype);
606
607	return ac_build_gather_values(ctx, chan, dst_channels);
608}
609
610/* Expand a scalar or vector to <4 x type> by filling the remaining channels
611 * with undef. Extract at most num_channels components from the input.
612 */
613LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
614				     LLVMValueRef value,
615				     unsigned num_channels)
616{
617	return ac_build_expand(ctx, value, num_channels, 4);
618}
619
620LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value)
621{
622	unsigned type_size = ac_get_type_size(LLVMTypeOf(value));
623	const char *name;
624
625	if (type_size == 2)
626		name = "llvm.rint.f16";
627	else if (type_size == 4)
628		name = "llvm.rint.f32";
629	else
630		name = "llvm.rint.f64";
631
632	return ac_build_intrinsic(ctx, name, LLVMTypeOf(value), &value, 1,
633				  AC_FUNC_ATTR_READNONE);
634}
635
636LLVMValueRef
637ac_build_fdiv(struct ac_llvm_context *ctx,
638	      LLVMValueRef num,
639	      LLVMValueRef den)
640{
641	/* If we do (num / den), LLVM >= 7.0 does:
642	 *    return num * v_rcp_f32(den * (fabs(den) > 0x1.0p+96f ? 0x1.0p-32f : 1.0f));
643	 *
644	 * If we do (num * (1 / den)), LLVM does:
645	 *    return num * v_rcp_f32(den);
646	 */
647	LLVMValueRef one = LLVMConstReal(LLVMTypeOf(num), 1.0);
648	LLVMValueRef rcp = LLVMBuildFDiv(ctx->builder, one, den, "");
649	LLVMValueRef ret = LLVMBuildFMul(ctx->builder, num, rcp, "");
650
651	/* Use v_rcp_f32 instead of precise division. */
652	if (!LLVMIsConstant(ret))
653		LLVMSetMetadata(ret, ctx->fpmath_md_kind, ctx->fpmath_md_2p5_ulp);
654	return ret;
655}
656
657/* See fast_idiv_by_const.h. */
658/* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
659LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
660				LLVMValueRef num,
661				LLVMValueRef multiplier,
662				LLVMValueRef pre_shift,
663				LLVMValueRef post_shift,
664				LLVMValueRef increment)
665{
666	LLVMBuilderRef builder = ctx->builder;
667
668	num = LLVMBuildLShr(builder, num, pre_shift, "");
669	num = LLVMBuildMul(builder,
670			   LLVMBuildZExt(builder, num, ctx->i64, ""),
671			   LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
672	num = LLVMBuildAdd(builder, num,
673			   LLVMBuildZExt(builder, increment, ctx->i64, ""), "");
674	num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
675	num = LLVMBuildTrunc(builder, num, ctx->i32, "");
676	return LLVMBuildLShr(builder, num, post_shift, "");
677}
678
679/* See fast_idiv_by_const.h. */
680/* If num != UINT_MAX, this more efficient version can be used. */
681/* Set: increment = util_fast_udiv_info::increment; */
682LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
683				    LLVMValueRef num,
684				    LLVMValueRef multiplier,
685				    LLVMValueRef pre_shift,
686				    LLVMValueRef post_shift,
687				    LLVMValueRef increment)
688{
689	LLVMBuilderRef builder = ctx->builder;
690
691	num = LLVMBuildLShr(builder, num, pre_shift, "");
692	num = LLVMBuildNUWAdd(builder, num, increment, "");
693	num = LLVMBuildMul(builder,
694			   LLVMBuildZExt(builder, num, ctx->i64, ""),
695			   LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
696	num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
697	num = LLVMBuildTrunc(builder, num, ctx->i32, "");
698	return LLVMBuildLShr(builder, num, post_shift, "");
699}
700
701/* See fast_idiv_by_const.h. */
702/* Both operands must fit in 31 bits and the divisor must not be 1. */
703LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
704					      LLVMValueRef num,
705					      LLVMValueRef multiplier,
706					      LLVMValueRef post_shift)
707{
708	LLVMBuilderRef builder = ctx->builder;
709
710	num = LLVMBuildMul(builder,
711			   LLVMBuildZExt(builder, num, ctx->i64, ""),
712			   LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
713	num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
714	num = LLVMBuildTrunc(builder, num, ctx->i32, "");
715	return LLVMBuildLShr(builder, num, post_shift, "");
716}
717
718/* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
719 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
720 * already multiplied by two. id is the cube face number.
721 */
722struct cube_selection_coords {
723	LLVMValueRef stc[2];
724	LLVMValueRef ma;
725	LLVMValueRef id;
726};
727
728static void
729build_cube_intrinsic(struct ac_llvm_context *ctx,
730		     LLVMValueRef in[3],
731		     struct cube_selection_coords *out)
732{
733	LLVMTypeRef f32 = ctx->f32;
734
735	out->stc[1] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubetc",
736					 f32, in, 3, AC_FUNC_ATTR_READNONE);
737	out->stc[0] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubesc",
738					 f32, in, 3, AC_FUNC_ATTR_READNONE);
739	out->ma = ac_build_intrinsic(ctx, "llvm.amdgcn.cubema",
740				     f32, in, 3, AC_FUNC_ATTR_READNONE);
741	out->id = ac_build_intrinsic(ctx, "llvm.amdgcn.cubeid",
742				     f32, in, 3, AC_FUNC_ATTR_READNONE);
743}
744
745/**
746 * Build a manual selection sequence for cube face sc/tc coordinates and
747 * major axis vector (multiplied by 2 for consistency) for the given
748 * vec3 \p coords, for the face implied by \p selcoords.
749 *
750 * For the major axis, we always adjust the sign to be in the direction of
751 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
752 * the selcoords major axis.
753 */
754static void build_cube_select(struct ac_llvm_context *ctx,
755			      const struct cube_selection_coords *selcoords,
756			      const LLVMValueRef *coords,
757			      LLVMValueRef *out_st,
758			      LLVMValueRef *out_ma)
759{
760	LLVMBuilderRef builder = ctx->builder;
761	LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
762	LLVMValueRef is_ma_positive;
763	LLVMValueRef sgn_ma;
764	LLVMValueRef is_ma_z, is_not_ma_z;
765	LLVMValueRef is_ma_y;
766	LLVMValueRef is_ma_x;
767	LLVMValueRef sgn;
768	LLVMValueRef tmp;
769
770	is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
771		selcoords->ma, LLVMConstReal(f32, 0.0), "");
772	sgn_ma = LLVMBuildSelect(builder, is_ma_positive,
773		LLVMConstReal(f32, 1.0), LLVMConstReal(f32, -1.0), "");
774
775	is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
776	is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
777	is_ma_y = LLVMBuildAnd(builder, is_not_ma_z,
778		LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
779	is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
780
781	/* Select sc */
782	tmp = LLVMBuildSelect(builder, is_ma_x, coords[2], coords[0], "");
783	sgn = LLVMBuildSelect(builder, is_ma_y, LLVMConstReal(f32, 1.0),
784		LLVMBuildSelect(builder, is_ma_z, sgn_ma,
785			LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
786	out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
787
788	/* Select tc */
789	tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
790	sgn = LLVMBuildSelect(builder, is_ma_y, sgn_ma,
791		LLVMConstReal(f32, -1.0), "");
792	out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
793
794	/* Select ma */
795	tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
796		LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
797	tmp = ac_build_intrinsic(ctx, "llvm.fabs.f32",
798				 ctx->f32, &tmp, 1, AC_FUNC_ATTR_READNONE);
799	*out_ma = LLVMBuildFMul(builder, tmp, LLVMConstReal(f32, 2.0), "");
800}
801
802void
803ac_prepare_cube_coords(struct ac_llvm_context *ctx,
804		       bool is_deriv, bool is_array, bool is_lod,
805		       LLVMValueRef *coords_arg,
806		       LLVMValueRef *derivs_arg)
807{
808
809	LLVMBuilderRef builder = ctx->builder;
810	struct cube_selection_coords selcoords;
811	LLVMValueRef coords[3];
812	LLVMValueRef invma;
813
814	if (is_array && !is_lod) {
815		LLVMValueRef tmp = ac_build_round(ctx, coords_arg[3]);
816
817		/* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
818		 *
819		 *    "For Array forms, the array layer used will be
820		 *
821		 *       max(0, min(d−1, floor(layer+0.5)))
822		 *
823		 *     where d is the depth of the texture array and layer
824		 *     comes from the component indicated in the tables below.
825		 *     Workaroudn for an issue where the layer is taken from a
826		 *     helper invocation which happens to fall on a different
827		 *     layer due to extrapolation."
828		 *
829		 * VI and earlier attempt to implement this in hardware by
830		 * clamping the value of coords[2] = (8 * layer) + face.
831		 * Unfortunately, this means that the we end up with the wrong
832		 * face when clamping occurs.
833		 *
834		 * Clamp the layer earlier to work around the issue.
835		 */
836		if (ctx->chip_class <= VI) {
837			LLVMValueRef ge0;
838			ge0 = LLVMBuildFCmp(builder, LLVMRealOGE, tmp, ctx->f32_0, "");
839			tmp = LLVMBuildSelect(builder, ge0, tmp, ctx->f32_0, "");
840		}
841
842		coords_arg[3] = tmp;
843	}
844
845	build_cube_intrinsic(ctx, coords_arg, &selcoords);
846
847	invma = ac_build_intrinsic(ctx, "llvm.fabs.f32",
848			ctx->f32, &selcoords.ma, 1, AC_FUNC_ATTR_READNONE);
849	invma = ac_build_fdiv(ctx, LLVMConstReal(ctx->f32, 1.0), invma);
850
851	for (int i = 0; i < 2; ++i)
852		coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
853
854	coords[2] = selcoords.id;
855
856	if (is_deriv && derivs_arg) {
857		LLVMValueRef derivs[4];
858		int axis;
859
860		/* Convert cube derivatives to 2D derivatives. */
861		for (axis = 0; axis < 2; axis++) {
862			LLVMValueRef deriv_st[2];
863			LLVMValueRef deriv_ma;
864
865			/* Transform the derivative alongside the texture
866			 * coordinate. Mathematically, the correct formula is
867			 * as follows. Assume we're projecting onto the +Z face
868			 * and denote by dx/dh the derivative of the (original)
869			 * X texture coordinate with respect to horizontal
870			 * window coordinates. The projection onto the +Z face
871			 * plane is:
872			 *
873			 *   f(x,z) = x/z
874			 *
875			 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
876			 *            = 1/z * dx/dh - x/z * 1/z * dz/dh.
877			 *
878			 * This motivatives the implementation below.
879			 *
880			 * Whether this actually gives the expected results for
881			 * apps that might feed in derivatives obtained via
882			 * finite differences is anyone's guess. The OpenGL spec
883			 * seems awfully quiet about how textureGrad for cube
884			 * maps should be handled.
885			 */
886			build_cube_select(ctx, &selcoords, &derivs_arg[axis * 3],
887					  deriv_st, &deriv_ma);
888
889			deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
890
891			for (int i = 0; i < 2; ++i)
892				derivs[axis * 2 + i] =
893					LLVMBuildFSub(builder,
894						LLVMBuildFMul(builder, deriv_st[i], invma, ""),
895						LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
896		}
897
898		memcpy(derivs_arg, derivs, sizeof(derivs));
899	}
900
901	/* Shift the texture coordinate. This must be applied after the
902	 * derivative calculation.
903	 */
904	for (int i = 0; i < 2; ++i)
905		coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
906
907	if (is_array) {
908		/* for cube arrays coord.z = coord.w(array_index) * 8 + face */
909		/* coords_arg.w component - array_index for cube arrays */
910		coords[2] = ac_build_fmad(ctx, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), coords[2]);
911	}
912
913	memcpy(coords_arg, coords, sizeof(coords));
914}
915
916
917LLVMValueRef
918ac_build_fs_interp(struct ac_llvm_context *ctx,
919		   LLVMValueRef llvm_chan,
920		   LLVMValueRef attr_number,
921		   LLVMValueRef params,
922		   LLVMValueRef i,
923		   LLVMValueRef j)
924{
925	LLVMValueRef args[5];
926	LLVMValueRef p1;
927
928	args[0] = i;
929	args[1] = llvm_chan;
930	args[2] = attr_number;
931	args[3] = params;
932
933	p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1",
934				ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
935
936	args[0] = p1;
937	args[1] = j;
938	args[2] = llvm_chan;
939	args[3] = attr_number;
940	args[4] = params;
941
942	return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2",
943				  ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
944}
945
946LLVMValueRef
947ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
948		       LLVMValueRef llvm_chan,
949		       LLVMValueRef attr_number,
950		       LLVMValueRef params,
951		       LLVMValueRef i,
952		       LLVMValueRef j)
953{
954	LLVMValueRef args[6];
955	LLVMValueRef p1;
956
957	args[0] = i;
958	args[1] = llvm_chan;
959	args[2] = attr_number;
960	args[3] = ctx->i1false;
961	args[4] = params;
962
963	p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1.f16",
964				ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
965
966	args[0] = p1;
967	args[1] = j;
968	args[2] = llvm_chan;
969	args[3] = attr_number;
970	args[4] = ctx->i1false;
971	args[5] = params;
972
973	return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2.f16",
974				  ctx->f16, args, 6, AC_FUNC_ATTR_READNONE);
975}
976
977LLVMValueRef
978ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
979		       LLVMValueRef parameter,
980		       LLVMValueRef llvm_chan,
981		       LLVMValueRef attr_number,
982		       LLVMValueRef params)
983{
984	LLVMValueRef args[4];
985
986	args[0] = parameter;
987	args[1] = llvm_chan;
988	args[2] = attr_number;
989	args[3] = params;
990
991	return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.mov",
992				  ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
993}
994
995LLVMValueRef
996ac_build_gep_ptr(struct ac_llvm_context *ctx,
997	         LLVMValueRef base_ptr,
998	         LLVMValueRef index)
999{
1000	return LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
1001}
1002
1003LLVMValueRef
1004ac_build_gep0(struct ac_llvm_context *ctx,
1005	      LLVMValueRef base_ptr,
1006	      LLVMValueRef index)
1007{
1008	LLVMValueRef indices[2] = {
1009		ctx->i32_0,
1010		index,
1011	};
1012	return LLVMBuildGEP(ctx->builder, base_ptr, indices, 2, "");
1013}
1014
1015LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
1016				  LLVMValueRef index)
1017{
1018	return LLVMBuildPointerCast(ctx->builder,
1019				    ac_build_gep0(ctx, ptr, index),
1020				    LLVMTypeOf(ptr), "");
1021}
1022
1023void
1024ac_build_indexed_store(struct ac_llvm_context *ctx,
1025		       LLVMValueRef base_ptr, LLVMValueRef index,
1026		       LLVMValueRef value)
1027{
1028	LLVMBuildStore(ctx->builder, value,
1029		       ac_build_gep0(ctx, base_ptr, index));
1030}
1031
1032/**
1033 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
1034 * It's equivalent to doing a load from &base_ptr[index].
1035 *
1036 * \param base_ptr  Where the array starts.
1037 * \param index     The element index into the array.
1038 * \param uniform   Whether the base_ptr and index can be assumed to be
1039 *                  dynamically uniform (i.e. load to an SGPR)
1040 * \param invariant Whether the load is invariant (no other opcodes affect it)
1041 * \param no_unsigned_wraparound
1042 *    For all possible re-associations and re-distributions of an expression
1043 *    "base_ptr + index * elemsize" into "addr + offset" (excluding GEPs
1044 *    without inbounds in base_ptr), this parameter is true if "addr + offset"
1045 *    does not result in an unsigned integer wraparound. This is used for
1046 *    optimal code generation of 32-bit pointer arithmetic.
1047 *
1048 *    For example, a 32-bit immediate offset that causes a 32-bit unsigned
1049 *    integer wraparound can't be an imm offset in s_load_dword, because
1050 *    the instruction performs "addr + offset" in 64 bits.
1051 *
1052 *    Expected usage for bindless textures by chaining GEPs:
1053 *      // possible unsigned wraparound, don't use InBounds:
1054 *      ptr1 = LLVMBuildGEP(base_ptr, index);
1055 *      image = load(ptr1); // becomes "s_load ptr1, 0"
1056 *
1057 *      ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
1058 *      sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
1059 */
1060static LLVMValueRef
1061ac_build_load_custom(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1062		     LLVMValueRef index, bool uniform, bool invariant,
1063		     bool no_unsigned_wraparound)
1064{
1065	LLVMValueRef pointer, result;
1066	LLVMValueRef indices[2] = {ctx->i32_0, index};
1067
1068	if (no_unsigned_wraparound &&
1069	    LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr)) == AC_ADDR_SPACE_CONST_32BIT)
1070		pointer = LLVMBuildInBoundsGEP(ctx->builder, base_ptr, indices, 2, "");
1071	else
1072		pointer = LLVMBuildGEP(ctx->builder, base_ptr, indices, 2, "");
1073
1074	if (uniform)
1075		LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
1076	result = LLVMBuildLoad(ctx->builder, pointer, "");
1077	if (invariant)
1078		LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
1079	return result;
1080}
1081
1082LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1083			   LLVMValueRef index)
1084{
1085	return ac_build_load_custom(ctx, base_ptr, index, false, false, false);
1086}
1087
1088LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
1089				     LLVMValueRef base_ptr, LLVMValueRef index)
1090{
1091	return ac_build_load_custom(ctx, base_ptr, index, false, true, false);
1092}
1093
1094/* This assumes that there is no unsigned integer wraparound during the address
1095 * computation, excluding all GEPs within base_ptr. */
1096LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
1097				   LLVMValueRef base_ptr, LLVMValueRef index)
1098{
1099	return ac_build_load_custom(ctx, base_ptr, index, true, true, true);
1100}
1101
1102/* See ac_build_load_custom() documentation. */
1103LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
1104				   LLVMValueRef base_ptr, LLVMValueRef index)
1105{
1106	return ac_build_load_custom(ctx, base_ptr, index, true, true, false);
1107}
1108
1109static void
1110ac_build_buffer_store_common(struct ac_llvm_context *ctx,
1111			     LLVMValueRef rsrc,
1112			     LLVMValueRef data,
1113			     LLVMValueRef vindex,
1114			     LLVMValueRef voffset,
1115			     unsigned num_channels,
1116			     bool glc,
1117			     bool slc,
1118			     bool writeonly_memory,
1119			     bool use_format)
1120{
1121	LLVMValueRef args[] = {
1122		data,
1123		LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
1124		vindex ? vindex : ctx->i32_0,
1125		voffset,
1126		LLVMConstInt(ctx->i1, glc, 0),
1127		LLVMConstInt(ctx->i1, slc, 0)
1128	};
1129	unsigned func = CLAMP(num_channels, 1, 3) - 1;
1130
1131	const char *type_names[] = {"f32", "v2f32", "v4f32"};
1132	char name[256];
1133
1134	if (use_format) {
1135		snprintf(name, sizeof(name), "llvm.amdgcn.buffer.store.format.%s",
1136			 type_names[func]);
1137	} else {
1138		snprintf(name, sizeof(name), "llvm.amdgcn.buffer.store.%s",
1139			 type_names[func]);
1140	}
1141
1142	ac_build_intrinsic(ctx, name, ctx->voidt, args, ARRAY_SIZE(args),
1143			   ac_get_store_intr_attribs(writeonly_memory));
1144}
1145
1146static void
1147ac_build_llvm8_buffer_store_common(struct ac_llvm_context *ctx,
1148				   LLVMValueRef rsrc,
1149				   LLVMValueRef data,
1150				   LLVMValueRef vindex,
1151				   LLVMValueRef voffset,
1152				   LLVMValueRef soffset,
1153				   unsigned num_channels,
1154				   LLVMTypeRef return_channel_type,
1155				   bool glc,
1156				   bool slc,
1157				   bool writeonly_memory,
1158				   bool use_format,
1159				   bool structurized)
1160{
1161	LLVMValueRef args[6];
1162	int idx = 0;
1163	args[idx++] = data;
1164	args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1165	if (structurized)
1166		args[idx++] = vindex ? vindex : ctx->i32_0;
1167	args[idx++] = voffset ? voffset : ctx->i32_0;
1168	args[idx++] = soffset ? soffset : ctx->i32_0;
1169	args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0);
1170	unsigned func = num_channels == 3 ? 4 : num_channels;
1171	const char *indexing_kind = structurized ? "struct" : "raw";
1172	char name[256], type_name[8];
1173
1174	LLVMTypeRef type = func > 1 ? LLVMVectorType(return_channel_type, func) : return_channel_type;
1175	ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1176
1177	if (use_format) {
1178		snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.format.%s",
1179			 indexing_kind, type_name);
1180	} else {
1181		snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.%s",
1182			 indexing_kind, type_name);
1183	}
1184
1185	ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
1186			   ac_get_store_intr_attribs(writeonly_memory));
1187}
1188
1189void
1190ac_build_buffer_store_format(struct ac_llvm_context *ctx,
1191			     LLVMValueRef rsrc,
1192			     LLVMValueRef data,
1193			     LLVMValueRef vindex,
1194			     LLVMValueRef voffset,
1195			     unsigned num_channels,
1196			     bool glc,
1197			     bool writeonly_memory)
1198{
1199	if (HAVE_LLVM >= 0x800) {
1200		ac_build_llvm8_buffer_store_common(ctx, rsrc, data, vindex,
1201						   voffset, NULL, num_channels,
1202						   ctx->f32, glc, false,
1203						   writeonly_memory, true, true);
1204	} else {
1205		ac_build_buffer_store_common(ctx, rsrc, data, vindex, voffset,
1206					     num_channels, glc, false,
1207					     writeonly_memory, true);
1208	}
1209}
1210
1211/* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1212 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1213 * or v4i32 (num_channels=3,4).
1214 */
1215void
1216ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
1217			    LLVMValueRef rsrc,
1218			    LLVMValueRef vdata,
1219			    unsigned num_channels,
1220			    LLVMValueRef voffset,
1221			    LLVMValueRef soffset,
1222			    unsigned inst_offset,
1223			    bool glc,
1224			    bool slc,
1225			    bool writeonly_memory,
1226			    bool swizzle_enable_hint)
1227{
1228	/* Split 3 channel stores, becase LLVM doesn't support 3-channel
1229	 * intrinsics. */
1230	if (num_channels == 3) {
1231		LLVMValueRef v[3], v01;
1232
1233		for (int i = 0; i < 3; i++) {
1234			v[i] = LLVMBuildExtractElement(ctx->builder, vdata,
1235					LLVMConstInt(ctx->i32, i, 0), "");
1236		}
1237		v01 = ac_build_gather_values(ctx, v, 2);
1238
1239		ac_build_buffer_store_dword(ctx, rsrc, v01, 2, voffset,
1240					    soffset, inst_offset, glc, slc,
1241					    writeonly_memory, swizzle_enable_hint);
1242		ac_build_buffer_store_dword(ctx, rsrc, v[2], 1, voffset,
1243					    soffset, inst_offset + 8,
1244					    glc, slc,
1245					    writeonly_memory, swizzle_enable_hint);
1246		return;
1247	}
1248
1249	/* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
1250	 * (voffset is swizzled, but soffset isn't swizzled).
1251	 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
1252	 */
1253	if (!swizzle_enable_hint) {
1254		LLVMValueRef offset = soffset;
1255
1256		if (inst_offset)
1257			offset = LLVMBuildAdd(ctx->builder, offset,
1258					      LLVMConstInt(ctx->i32, inst_offset, 0), "");
1259
1260		if (HAVE_LLVM >= 0x800) {
1261			ac_build_llvm8_buffer_store_common(ctx, rsrc,
1262							   ac_to_float(ctx, vdata),
1263							   ctx->i32_0,
1264							   voffset, offset,
1265							   num_channels,
1266							   ctx->f32,
1267							   glc, slc,
1268							   writeonly_memory,
1269							   false, false);
1270		} else {
1271			if (voffset)
1272				offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
1273
1274			ac_build_buffer_store_common(ctx, rsrc,
1275						     ac_to_float(ctx, vdata),
1276						     ctx->i32_0, offset,
1277						     num_channels, glc, slc,
1278						     writeonly_memory, false);
1279		}
1280		return;
1281	}
1282
1283	static const unsigned dfmts[] = {
1284		V_008F0C_BUF_DATA_FORMAT_32,
1285		V_008F0C_BUF_DATA_FORMAT_32_32,
1286		V_008F0C_BUF_DATA_FORMAT_32_32_32,
1287		V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1288	};
1289	unsigned dfmt = dfmts[num_channels - 1];
1290	unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1291	LLVMValueRef immoffset = LLVMConstInt(ctx->i32, inst_offset, 0);
1292
1293	ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
1294			           immoffset, num_channels, dfmt, nfmt, glc,
1295				   slc, writeonly_memory);
1296}
1297
1298static LLVMValueRef
1299ac_build_buffer_load_common(struct ac_llvm_context *ctx,
1300			    LLVMValueRef rsrc,
1301			    LLVMValueRef vindex,
1302			    LLVMValueRef voffset,
1303			    unsigned num_channels,
1304			    bool glc,
1305			    bool slc,
1306			    bool can_speculate,
1307			    bool use_format)
1308{
1309	LLVMValueRef args[] = {
1310		LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
1311		vindex ? vindex : ctx->i32_0,
1312		voffset,
1313		LLVMConstInt(ctx->i1, glc, 0),
1314		LLVMConstInt(ctx->i1, slc, 0)
1315	};
1316	unsigned func = CLAMP(num_channels, 1, 3) - 1;
1317
1318	LLVMTypeRef types[] = {ctx->f32, ctx->v2f32, ctx->v4f32};
1319	const char *type_names[] = {"f32", "v2f32", "v4f32"};
1320	char name[256];
1321
1322	if (use_format) {
1323		snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.format.%s",
1324			 type_names[func]);
1325	} else {
1326		snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.%s",
1327			 type_names[func]);
1328	}
1329
1330	return ac_build_intrinsic(ctx, name, types[func], args,
1331				  ARRAY_SIZE(args),
1332				  ac_get_load_intr_attribs(can_speculate));
1333}
1334
1335static LLVMValueRef
1336ac_build_llvm8_buffer_load_common(struct ac_llvm_context *ctx,
1337				  LLVMValueRef rsrc,
1338				  LLVMValueRef vindex,
1339				  LLVMValueRef voffset,
1340				  LLVMValueRef soffset,
1341				  unsigned num_channels,
1342				  LLVMTypeRef channel_type,
1343				  bool glc,
1344				  bool slc,
1345				  bool can_speculate,
1346				  bool use_format,
1347				  bool structurized)
1348{
1349	LLVMValueRef args[5];
1350	int idx = 0;
1351	args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1352	if (structurized)
1353		args[idx++] = vindex ? vindex : ctx->i32_0;
1354	args[idx++] = voffset ? voffset : ctx->i32_0;
1355	args[idx++] = soffset ? soffset : ctx->i32_0;
1356	args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0);
1357	unsigned func = num_channels == 3 ? 4 : num_channels;
1358	const char *indexing_kind = structurized ? "struct" : "raw";
1359	char name[256], type_name[8];
1360
1361	LLVMTypeRef type = func > 1 ? LLVMVectorType(channel_type, func) : channel_type;
1362	ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1363
1364	if (use_format) {
1365		snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.format.%s",
1366			 indexing_kind, type_name);
1367	} else {
1368		snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.%s",
1369			 indexing_kind, type_name);
1370	}
1371
1372	return ac_build_intrinsic(ctx, name, type, args, idx,
1373				  ac_get_load_intr_attribs(can_speculate));
1374}
1375
1376LLVMValueRef
1377ac_build_buffer_load(struct ac_llvm_context *ctx,
1378		     LLVMValueRef rsrc,
1379		     int num_channels,
1380		     LLVMValueRef vindex,
1381		     LLVMValueRef voffset,
1382		     LLVMValueRef soffset,
1383		     unsigned inst_offset,
1384		     unsigned glc,
1385		     unsigned slc,
1386		     bool can_speculate,
1387		     bool allow_smem)
1388{
1389	LLVMValueRef offset = LLVMConstInt(ctx->i32, inst_offset, 0);
1390	if (voffset)
1391		offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
1392	if (soffset)
1393		offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
1394
1395	if (allow_smem && !slc &&
1396	    (!glc || (HAVE_LLVM >= 0x0800 && ctx->chip_class >= VI))) {
1397		assert(vindex == NULL);
1398
1399		LLVMValueRef result[8];
1400
1401		for (int i = 0; i < num_channels; i++) {
1402			if (i) {
1403				offset = LLVMBuildAdd(ctx->builder, offset,
1404						      LLVMConstInt(ctx->i32, 4, 0), "");
1405			}
1406			const char *intrname =
1407				HAVE_LLVM >= 0x0800 ? "llvm.amdgcn.s.buffer.load.f32"
1408						    : "llvm.SI.load.const.v4i32";
1409			unsigned num_args = HAVE_LLVM >= 0x0800 ? 3 : 2;
1410			LLVMValueRef args[3] = {
1411				rsrc,
1412				offset,
1413				glc ? ctx->i32_1 : ctx->i32_0,
1414			};
1415			result[i] = ac_build_intrinsic(ctx, intrname,
1416						       ctx->f32, args, num_args,
1417						       AC_FUNC_ATTR_READNONE |
1418						       (HAVE_LLVM < 0x0800 ? AC_FUNC_ATTR_LEGACY : 0));
1419		}
1420		if (num_channels == 1)
1421			return result[0];
1422
1423		if (num_channels == 3)
1424			result[num_channels++] = LLVMGetUndef(ctx->f32);
1425		return ac_build_gather_values(ctx, result, num_channels);
1426	}
1427
1428	if (HAVE_LLVM >= 0x0800) {
1429		return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex,
1430							 offset, ctx->i32_0,
1431							 num_channels, ctx->f32,
1432							 glc, slc,
1433							 can_speculate, false,
1434							 false);
1435	}
1436
1437	return ac_build_buffer_load_common(ctx, rsrc, vindex, offset,
1438					   num_channels, glc, slc,
1439					   can_speculate, false);
1440}
1441
1442LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
1443					 LLVMValueRef rsrc,
1444					 LLVMValueRef vindex,
1445					 LLVMValueRef voffset,
1446					 unsigned num_channels,
1447					 bool glc,
1448					 bool can_speculate)
1449{
1450	if (HAVE_LLVM >= 0x800) {
1451		return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
1452							 num_channels, ctx->f32,
1453							 glc, false,
1454							 can_speculate, true, true);
1455	}
1456	return ac_build_buffer_load_common(ctx, rsrc, vindex, voffset,
1457					   num_channels, glc, false,
1458					   can_speculate, true);
1459}
1460
1461LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
1462                                                  LLVMValueRef rsrc,
1463                                                  LLVMValueRef vindex,
1464                                                  LLVMValueRef voffset,
1465                                                  unsigned num_channels,
1466                                                  bool glc,
1467                                                  bool can_speculate)
1468{
1469	if (HAVE_LLVM >= 0x800) {
1470		return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
1471							 num_channels, ctx->f32,
1472							 glc, false,
1473							 can_speculate, true, true);
1474	}
1475
1476	LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->builder, rsrc, LLVMConstInt(ctx->i32, 2, 0), "");
1477	LLVMValueRef stride = LLVMBuildExtractElement(ctx->builder, rsrc, ctx->i32_1, "");
1478	stride = LLVMBuildLShr(ctx->builder, stride, LLVMConstInt(ctx->i32, 16, 0), "");
1479
1480	LLVMValueRef new_elem_count = LLVMBuildSelect(ctx->builder,
1481	                                              LLVMBuildICmp(ctx->builder, LLVMIntUGT, elem_count, stride, ""),
1482	                                              elem_count, stride, "");
1483
1484	LLVMValueRef new_rsrc = LLVMBuildInsertElement(ctx->builder, rsrc, new_elem_count,
1485	                                               LLVMConstInt(ctx->i32, 2, 0), "");
1486
1487	return ac_build_buffer_load_common(ctx, new_rsrc, vindex, voffset,
1488	                                   num_channels, glc, false,
1489	                                   can_speculate, true);
1490}
1491
1492static LLVMValueRef
1493ac_build_llvm8_tbuffer_load(struct ac_llvm_context *ctx,
1494			    LLVMValueRef rsrc,
1495			    LLVMValueRef vindex,
1496			    LLVMValueRef voffset,
1497			    LLVMValueRef soffset,
1498			    unsigned num_channels,
1499			    unsigned dfmt,
1500			    unsigned nfmt,
1501			    bool glc,
1502			    bool slc,
1503			    bool can_speculate,
1504			    bool structurized)
1505{
1506	LLVMValueRef args[6];
1507	int idx = 0;
1508	args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1509	if (structurized)
1510		args[idx++] = vindex ? vindex : ctx->i32_0;
1511	args[idx++] = voffset ? voffset : ctx->i32_0;
1512	args[idx++] = soffset ? soffset : ctx->i32_0;
1513	args[idx++] = LLVMConstInt(ctx->i32, dfmt | (nfmt << 4), 0);
1514	args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0);
1515	unsigned func = num_channels == 3 ? 4 : num_channels;
1516	const char *indexing_kind = structurized ? "struct" : "raw";
1517	char name[256], type_name[8];
1518
1519	LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
1520	ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1521
1522	snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.load.%s",
1523		 indexing_kind, type_name);
1524
1525	return ac_build_intrinsic(ctx, name, type, args, idx,
1526				  ac_get_load_intr_attribs(can_speculate));
1527}
1528
1529static LLVMValueRef
1530ac_build_tbuffer_load(struct ac_llvm_context *ctx,
1531			    LLVMValueRef rsrc,
1532			    LLVMValueRef vindex,
1533			    LLVMValueRef voffset,
1534			    LLVMValueRef soffset,
1535			    LLVMValueRef immoffset,
1536			    unsigned num_channels,
1537			    unsigned dfmt,
1538			    unsigned nfmt,
1539			    bool glc,
1540			    bool slc,
1541			    bool can_speculate,
1542			    bool structurized) /* only matters for LLVM 8+ */
1543{
1544	if (HAVE_LLVM >= 0x800) {
1545		voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1546
1547		return ac_build_llvm8_tbuffer_load(ctx, rsrc, vindex, voffset,
1548						   soffset, num_channels,
1549						   dfmt, nfmt, glc, slc,
1550						   can_speculate, structurized);
1551	}
1552
1553	LLVMValueRef args[] = {
1554		rsrc,
1555		vindex ? vindex : ctx->i32_0,
1556		voffset,
1557		soffset,
1558		immoffset,
1559		LLVMConstInt(ctx->i32, dfmt, false),
1560		LLVMConstInt(ctx->i32, nfmt, false),
1561		LLVMConstInt(ctx->i1, glc, false),
1562		LLVMConstInt(ctx->i1, slc, false),
1563	};
1564	unsigned func = CLAMP(num_channels, 1, 3) - 1;
1565	LLVMTypeRef types[] = {ctx->i32, ctx->v2i32, ctx->v4i32};
1566	const char *type_names[] = {"i32", "v2i32", "v4i32"};
1567	char name[256];
1568
1569	snprintf(name, sizeof(name), "llvm.amdgcn.tbuffer.load.%s",
1570		 type_names[func]);
1571
1572	return ac_build_intrinsic(ctx, name, types[func], args, 9,
1573				  ac_get_load_intr_attribs(can_speculate));
1574}
1575
1576LLVMValueRef
1577ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
1578			     LLVMValueRef rsrc,
1579			     LLVMValueRef vindex,
1580			     LLVMValueRef voffset,
1581			     LLVMValueRef soffset,
1582			     LLVMValueRef immoffset,
1583			     unsigned num_channels,
1584			     unsigned dfmt,
1585			     unsigned nfmt,
1586			     bool glc,
1587			     bool slc,
1588			     bool can_speculate)
1589{
1590	return ac_build_tbuffer_load(ctx, rsrc, vindex, voffset, soffset,
1591				     immoffset, num_channels, dfmt, nfmt, glc,
1592				     slc, can_speculate, true);
1593}
1594
1595LLVMValueRef
1596ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
1597			  LLVMValueRef rsrc,
1598			  LLVMValueRef voffset,
1599			  LLVMValueRef soffset,
1600			  LLVMValueRef immoffset,
1601			  unsigned num_channels,
1602			  unsigned dfmt,
1603			  unsigned nfmt,
1604			  bool glc,
1605			  bool slc,
1606		          bool can_speculate)
1607{
1608	return ac_build_tbuffer_load(ctx, rsrc, NULL, voffset, soffset,
1609				     immoffset, num_channels, dfmt, nfmt, glc,
1610				     slc, can_speculate, false);
1611}
1612
1613LLVMValueRef
1614ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
1615			    LLVMValueRef rsrc,
1616			    LLVMValueRef voffset,
1617			    LLVMValueRef soffset,
1618			    LLVMValueRef immoffset,
1619			    bool glc)
1620{
1621	LLVMValueRef res;
1622
1623	if (HAVE_LLVM >= 0x900) {
1624		voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1625
1626		/* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1627		res = ac_build_llvm8_buffer_load_common(ctx, rsrc, NULL,
1628							voffset, soffset,
1629							1, ctx->i16, glc, false,
1630							false, false, false);
1631	} else {
1632		unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
1633		unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1634
1635		res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
1636						immoffset, 1, dfmt, nfmt, glc, false,
1637						false);
1638
1639		res = LLVMBuildTrunc(ctx->builder, res, ctx->i16, "");
1640	}
1641
1642	return res;
1643}
1644
1645LLVMValueRef
1646ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
1647			   LLVMValueRef rsrc,
1648			   LLVMValueRef voffset,
1649			   LLVMValueRef soffset,
1650			   LLVMValueRef immoffset,
1651			   bool glc)
1652{
1653	LLVMValueRef res;
1654
1655	if (HAVE_LLVM >= 0x900) {
1656		voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1657
1658		/* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1659		res = ac_build_llvm8_buffer_load_common(ctx, rsrc, NULL,
1660							voffset, soffset,
1661							1, ctx->i8, glc, false,
1662							false, false, false);
1663	} else {
1664		unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
1665		unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1666
1667		res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
1668						immoffset, 1, dfmt, nfmt, glc, false,
1669						false);
1670
1671		res = LLVMBuildTrunc(ctx->builder, res, ctx->i8, "");
1672	}
1673
1674	return res;
1675}
1676static void
1677ac_build_llvm8_tbuffer_store(struct ac_llvm_context *ctx,
1678			     LLVMValueRef rsrc,
1679			     LLVMValueRef vdata,
1680			     LLVMValueRef vindex,
1681			     LLVMValueRef voffset,
1682			     LLVMValueRef soffset,
1683			     unsigned num_channels,
1684			     unsigned dfmt,
1685			     unsigned nfmt,
1686			     bool glc,
1687			     bool slc,
1688			     bool writeonly_memory,
1689			     bool structurized)
1690{
1691	LLVMValueRef args[7];
1692	int idx = 0;
1693	args[idx++] = vdata;
1694	args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1695	if (structurized)
1696		args[idx++] = vindex ? vindex : ctx->i32_0;
1697	args[idx++] = voffset ? voffset : ctx->i32_0;
1698	args[idx++] = soffset ? soffset : ctx->i32_0;
1699	args[idx++] = LLVMConstInt(ctx->i32, dfmt | (nfmt << 4), 0);
1700	args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0);
1701	unsigned func = num_channels == 3 ? 4 : num_channels;
1702	const char *indexing_kind = structurized ? "struct" : "raw";
1703	char name[256], type_name[8];
1704
1705	LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
1706	ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1707
1708	snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.store.%s",
1709		 indexing_kind, type_name);
1710
1711	ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
1712			   ac_get_store_intr_attribs(writeonly_memory));
1713}
1714
1715static void
1716ac_build_tbuffer_store(struct ac_llvm_context *ctx,
1717		       LLVMValueRef rsrc,
1718		       LLVMValueRef vdata,
1719		       LLVMValueRef vindex,
1720		       LLVMValueRef voffset,
1721		       LLVMValueRef soffset,
1722		       LLVMValueRef immoffset,
1723		       unsigned num_channels,
1724		       unsigned dfmt,
1725		       unsigned nfmt,
1726		       bool glc,
1727		       bool slc,
1728		       bool writeonly_memory,
1729		       bool structurized) /* only matters for LLVM 8+ */
1730{
1731	if (HAVE_LLVM >= 0x800) {
1732		voffset = LLVMBuildAdd(ctx->builder,
1733				       voffset ? voffset : ctx->i32_0,
1734				       immoffset, "");
1735
1736		ac_build_llvm8_tbuffer_store(ctx, rsrc, vdata, vindex, voffset,
1737					     soffset, num_channels, dfmt, nfmt,
1738					     glc, slc, writeonly_memory,
1739					     structurized);
1740	} else {
1741		LLVMValueRef params[] = {
1742			vdata,
1743			rsrc,
1744			vindex ? vindex : ctx->i32_0,
1745			voffset ? voffset : ctx->i32_0,
1746			soffset ? soffset : ctx->i32_0,
1747			immoffset,
1748			LLVMConstInt(ctx->i32, dfmt, false),
1749			LLVMConstInt(ctx->i32, nfmt, false),
1750			LLVMConstInt(ctx->i1, glc, false),
1751			LLVMConstInt(ctx->i1, slc, false),
1752		};
1753		unsigned func = CLAMP(num_channels, 1, 3) - 1;
1754		const char *type_names[] = {"i32", "v2i32", "v4i32"};
1755		char name[256];
1756
1757		snprintf(name, sizeof(name), "llvm.amdgcn.tbuffer.store.%s",
1758			 type_names[func]);
1759
1760		ac_build_intrinsic(ctx, name, ctx->voidt, params, 10,
1761				   ac_get_store_intr_attribs(writeonly_memory));
1762	}
1763}
1764
1765void
1766ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
1767			      LLVMValueRef rsrc,
1768			      LLVMValueRef vdata,
1769			      LLVMValueRef vindex,
1770			      LLVMValueRef voffset,
1771			      LLVMValueRef soffset,
1772			      LLVMValueRef immoffset,
1773			      unsigned num_channels,
1774			      unsigned dfmt,
1775			      unsigned nfmt,
1776			      bool glc,
1777			      bool slc,
1778			      bool writeonly_memory)
1779{
1780	ac_build_tbuffer_store(ctx, rsrc, vdata, vindex, voffset, soffset,
1781			       immoffset, num_channels, dfmt, nfmt, glc, slc,
1782			       writeonly_memory, true);
1783}
1784
1785void
1786ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
1787			   LLVMValueRef rsrc,
1788			   LLVMValueRef vdata,
1789			   LLVMValueRef voffset,
1790			   LLVMValueRef soffset,
1791			   LLVMValueRef immoffset,
1792			   unsigned num_channels,
1793			   unsigned dfmt,
1794			   unsigned nfmt,
1795			   bool glc,
1796			   bool slc,
1797			   bool writeonly_memory)
1798{
1799	ac_build_tbuffer_store(ctx, rsrc, vdata, NULL, voffset, soffset,
1800			       immoffset, num_channels, dfmt, nfmt, glc, slc,
1801			       writeonly_memory, false);
1802}
1803
1804void
1805ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
1806			     LLVMValueRef rsrc,
1807			     LLVMValueRef vdata,
1808			     LLVMValueRef voffset,
1809			     LLVMValueRef soffset,
1810			     bool glc,
1811			     bool writeonly_memory)
1812{
1813	vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i16, "");
1814
1815	if (HAVE_LLVM >= 0x900) {
1816		/* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1817		ac_build_llvm8_buffer_store_common(ctx, rsrc, vdata, NULL,
1818						   voffset, soffset, 1,
1819						   ctx->i16, glc, false,
1820						   writeonly_memory, false,
1821						   false);
1822	} else {
1823		unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
1824		unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1825
1826		vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
1827
1828		ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
1829					   ctx->i32_0, 1, dfmt, nfmt, glc, false,
1830					   writeonly_memory);
1831	}
1832}
1833
1834void
1835ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
1836			    LLVMValueRef rsrc,
1837			    LLVMValueRef vdata,
1838			    LLVMValueRef voffset,
1839			    LLVMValueRef soffset,
1840			    bool glc,
1841			    bool writeonly_memory)
1842{
1843	vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i8, "");
1844
1845	if (HAVE_LLVM >= 0x900) {
1846		/* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1847		ac_build_llvm8_buffer_store_common(ctx, rsrc, vdata, NULL,
1848						   voffset, soffset, 1,
1849						   ctx->i8, glc, false,
1850						   writeonly_memory, false,
1851						   false);
1852	} else {
1853		unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
1854		unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1855
1856		vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
1857
1858		ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
1859					   ctx->i32_0, 1, dfmt, nfmt, glc, false,
1860					   writeonly_memory);
1861	}
1862}
1863/**
1864 * Set range metadata on an instruction.  This can only be used on load and
1865 * call instructions.  If you know an instruction can only produce the values
1866 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1867 * \p lo is the minimum value inclusive.
1868 * \p hi is the maximum value exclusive.
1869 */
1870static void set_range_metadata(struct ac_llvm_context *ctx,
1871			       LLVMValueRef value, unsigned lo, unsigned hi)
1872{
1873	LLVMValueRef range_md, md_args[2];
1874	LLVMTypeRef type = LLVMTypeOf(value);
1875	LLVMContextRef context = LLVMGetTypeContext(type);
1876
1877	md_args[0] = LLVMConstInt(type, lo, false);
1878	md_args[1] = LLVMConstInt(type, hi, false);
1879	range_md = LLVMMDNodeInContext(context, md_args, 2);
1880	LLVMSetMetadata(value, ctx->range_md_kind, range_md);
1881}
1882
1883LLVMValueRef
1884ac_get_thread_id(struct ac_llvm_context *ctx)
1885{
1886	LLVMValueRef tid;
1887
1888	LLVMValueRef tid_args[2];
1889	tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
1890	tid_args[1] = ctx->i32_0;
1891	tid_args[1] = ac_build_intrinsic(ctx,
1892					 "llvm.amdgcn.mbcnt.lo", ctx->i32,
1893					 tid_args, 2, AC_FUNC_ATTR_READNONE);
1894
1895	tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
1896				 ctx->i32, tid_args,
1897				 2, AC_FUNC_ATTR_READNONE);
1898	set_range_metadata(ctx, tid, 0, 64);
1899	return tid;
1900}
1901
1902/*
1903 * SI implements derivatives using the local data store (LDS)
1904 * All writes to the LDS happen in all executing threads at
1905 * the same time. TID is the Thread ID for the current
1906 * thread and is a value between 0 and 63, representing
1907 * the thread's position in the wavefront.
1908 *
1909 * For the pixel shader threads are grouped into quads of four pixels.
1910 * The TIDs of the pixels of a quad are:
1911 *
1912 *  +------+------+
1913 *  |4n + 0|4n + 1|
1914 *  +------+------+
1915 *  |4n + 2|4n + 3|
1916 *  +------+------+
1917 *
1918 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
1919 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
1920 * the current pixel's column, and masking with 0xfffffffe yields the TID
1921 * of the left pixel of the current pixel's row.
1922 *
1923 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
1924 * adding 2 yields the TID of the pixel below the top pixel.
1925 */
1926LLVMValueRef
1927ac_build_ddxy(struct ac_llvm_context *ctx,
1928	      uint32_t mask,
1929	      int idx,
1930	      LLVMValueRef val)
1931{
1932	unsigned tl_lanes[4], trbl_lanes[4];
1933	char name[32], type[8];
1934	LLVMValueRef tl, trbl;
1935	LLVMTypeRef result_type;
1936	LLVMValueRef result;
1937
1938	result_type = ac_to_float_type(ctx, LLVMTypeOf(val));
1939
1940	if (result_type == ctx->f16)
1941		val = LLVMBuildZExt(ctx->builder, val, ctx->i32, "");
1942
1943	for (unsigned i = 0; i < 4; ++i) {
1944		tl_lanes[i] = i & mask;
1945		trbl_lanes[i] = (i & mask) + idx;
1946	}
1947
1948	tl = ac_build_quad_swizzle(ctx, val,
1949				   tl_lanes[0], tl_lanes[1],
1950				   tl_lanes[2], tl_lanes[3]);
1951	trbl = ac_build_quad_swizzle(ctx, val,
1952				     trbl_lanes[0], trbl_lanes[1],
1953				     trbl_lanes[2], trbl_lanes[3]);
1954
1955	if (result_type == ctx->f16) {
1956		tl = LLVMBuildTrunc(ctx->builder, tl, ctx->i16, "");
1957		trbl = LLVMBuildTrunc(ctx->builder, trbl, ctx->i16, "");
1958	}
1959
1960	tl = LLVMBuildBitCast(ctx->builder, tl, result_type, "");
1961	trbl = LLVMBuildBitCast(ctx->builder, trbl, result_type, "");
1962	result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
1963
1964	ac_build_type_name_for_intr(result_type, type, sizeof(type));
1965	snprintf(name, sizeof(name), "llvm.amdgcn.wqm.%s", type);
1966
1967	return ac_build_intrinsic(ctx, name, result_type, &result, 1, 0);
1968}
1969
1970void
1971ac_build_sendmsg(struct ac_llvm_context *ctx,
1972		 uint32_t msg,
1973		 LLVMValueRef wave_id)
1974{
1975	LLVMValueRef args[2];
1976	args[0] = LLVMConstInt(ctx->i32, msg, false);
1977	args[1] = wave_id;
1978	ac_build_intrinsic(ctx, "llvm.amdgcn.s.sendmsg", ctx->voidt, args, 2, 0);
1979}
1980
1981LLVMValueRef
1982ac_build_imsb(struct ac_llvm_context *ctx,
1983	      LLVMValueRef arg,
1984	      LLVMTypeRef dst_type)
1985{
1986	LLVMValueRef msb = ac_build_intrinsic(ctx, "llvm.amdgcn.sffbh.i32",
1987					      dst_type, &arg, 1,
1988					      AC_FUNC_ATTR_READNONE);
1989
1990	/* The HW returns the last bit index from MSB, but NIR/TGSI wants
1991	 * the index from LSB. Invert it by doing "31 - msb". */
1992	msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
1993			   msb, "");
1994
1995	LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
1996	LLVMValueRef cond = LLVMBuildOr(ctx->builder,
1997					LLVMBuildICmp(ctx->builder, LLVMIntEQ,
1998						      arg, ctx->i32_0, ""),
1999					LLVMBuildICmp(ctx->builder, LLVMIntEQ,
2000						      arg, all_ones, ""), "");
2001
2002	return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
2003}
2004
2005LLVMValueRef
2006ac_build_umsb(struct ac_llvm_context *ctx,
2007	      LLVMValueRef arg,
2008	      LLVMTypeRef dst_type)
2009{
2010	const char *intrin_name;
2011	LLVMTypeRef type;
2012	LLVMValueRef highest_bit;
2013	LLVMValueRef zero;
2014	unsigned bitsize;
2015
2016	bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(arg));
2017	switch (bitsize) {
2018	case 64:
2019		intrin_name = "llvm.ctlz.i64";
2020		type = ctx->i64;
2021		highest_bit = LLVMConstInt(ctx->i64, 63, false);
2022		zero = ctx->i64_0;
2023		break;
2024	case 32:
2025		intrin_name = "llvm.ctlz.i32";
2026		type = ctx->i32;
2027		highest_bit = LLVMConstInt(ctx->i32, 31, false);
2028		zero = ctx->i32_0;
2029		break;
2030	case 16:
2031		intrin_name = "llvm.ctlz.i16";
2032		type = ctx->i16;
2033		highest_bit = LLVMConstInt(ctx->i16, 15, false);
2034		zero = ctx->i16_0;
2035		break;
2036	case 8:
2037		intrin_name = "llvm.ctlz.i8";
2038		type = ctx->i8;
2039		highest_bit = LLVMConstInt(ctx->i8, 7, false);
2040		zero = ctx->i8_0;
2041		break;
2042	default:
2043		unreachable(!"invalid bitsize");
2044		break;
2045	}
2046
2047	LLVMValueRef params[2] = {
2048		arg,
2049		ctx->i1true,
2050	};
2051
2052	LLVMValueRef msb = ac_build_intrinsic(ctx, intrin_name, type,
2053					      params, 2,
2054					      AC_FUNC_ATTR_READNONE);
2055
2056	/* The HW returns the last bit index from MSB, but TGSI/NIR wants
2057	 * the index from LSB. Invert it by doing "31 - msb". */
2058	msb = LLVMBuildSub(ctx->builder, highest_bit, msb, "");
2059
2060	if (bitsize == 64) {
2061		msb = LLVMBuildTrunc(ctx->builder, msb, ctx->i32, "");
2062	} else if (bitsize < 32) {
2063		msb = LLVMBuildSExt(ctx->builder, msb, ctx->i32, "");
2064	}
2065
2066	/* check for zero */
2067	return LLVMBuildSelect(ctx->builder,
2068			       LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg, zero, ""),
2069			       LLVMConstInt(ctx->i32, -1, true), msb, "");
2070}
2071
2072LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
2073			   LLVMValueRef b)
2074{
2075	char name[64];
2076	snprintf(name, sizeof(name), "llvm.minnum.f%d", ac_get_elem_bits(ctx, LLVMTypeOf(a)));
2077	LLVMValueRef args[2] = {a, b};
2078	return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
2079				  AC_FUNC_ATTR_READNONE);
2080}
2081
2082LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
2083			   LLVMValueRef b)
2084{
2085	char name[64];
2086	snprintf(name, sizeof(name), "llvm.maxnum.f%d", ac_get_elem_bits(ctx, LLVMTypeOf(a)));
2087	LLVMValueRef args[2] = {a, b};
2088	return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
2089				  AC_FUNC_ATTR_READNONE);
2090}
2091
2092LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
2093			   LLVMValueRef b)
2094{
2095	LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSLE, a, b, "");
2096	return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2097}
2098
2099LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
2100			   LLVMValueRef b)
2101{
2102	LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, a, b, "");
2103	return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2104}
2105
2106LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a,
2107			   LLVMValueRef b)
2108{
2109	LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntULE, a, b, "");
2110	return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2111}
2112
2113LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a,
2114			   LLVMValueRef b)
2115{
2116	LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, a, b, "");
2117	return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2118}
2119
2120LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value)
2121{
2122	LLVMTypeRef t = LLVMTypeOf(value);
2123	return ac_build_fmin(ctx, ac_build_fmax(ctx, value, LLVMConstReal(t, 0.0)),
2124			     LLVMConstReal(t, 1.0));
2125}
2126
2127void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a)
2128{
2129	LLVMValueRef args[9];
2130
2131	args[0] = LLVMConstInt(ctx->i32, a->target, 0);
2132	args[1] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
2133
2134	if (a->compr) {
2135		LLVMTypeRef i16 = LLVMInt16TypeInContext(ctx->context);
2136		LLVMTypeRef v2i16 = LLVMVectorType(i16, 2);
2137
2138		args[2] = LLVMBuildBitCast(ctx->builder, a->out[0],
2139				v2i16, "");
2140		args[3] = LLVMBuildBitCast(ctx->builder, a->out[1],
2141				v2i16, "");
2142		args[4] = LLVMConstInt(ctx->i1, a->done, 0);
2143		args[5] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
2144
2145		ac_build_intrinsic(ctx, "llvm.amdgcn.exp.compr.v2i16",
2146				   ctx->voidt, args, 6, 0);
2147	} else {
2148		args[2] = a->out[0];
2149		args[3] = a->out[1];
2150		args[4] = a->out[2];
2151		args[5] = a->out[3];
2152		args[6] = LLVMConstInt(ctx->i1, a->done, 0);
2153		args[7] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
2154
2155		ac_build_intrinsic(ctx, "llvm.amdgcn.exp.f32",
2156				   ctx->voidt, args, 8, 0);
2157	}
2158}
2159
2160void ac_build_export_null(struct ac_llvm_context *ctx)
2161{
2162	struct ac_export_args args;
2163
2164	args.enabled_channels = 0x0; /* enabled channels */
2165	args.valid_mask = 1; /* whether the EXEC mask is valid */
2166	args.done = 1; /* DONE bit */
2167	args.target = V_008DFC_SQ_EXP_NULL;
2168	args.compr = 0; /* COMPR flag (0 = 32-bit export) */
2169	args.out[0] = LLVMGetUndef(ctx->f32); /* R */
2170	args.out[1] = LLVMGetUndef(ctx->f32); /* G */
2171	args.out[2] = LLVMGetUndef(ctx->f32); /* B */
2172	args.out[3] = LLVMGetUndef(ctx->f32); /* A */
2173
2174	ac_build_export(ctx, &args);
2175}
2176
2177static unsigned ac_num_coords(enum ac_image_dim dim)
2178{
2179	switch (dim) {
2180	case ac_image_1d:
2181		return 1;
2182	case ac_image_2d:
2183	case ac_image_1darray:
2184		 return 2;
2185	case ac_image_3d:
2186	case ac_image_cube:
2187	case ac_image_2darray:
2188	case ac_image_2dmsaa:
2189		return 3;
2190	case ac_image_2darraymsaa:
2191		return 4;
2192	default:
2193		unreachable("ac_num_coords: bad dim");
2194	}
2195}
2196
2197static unsigned ac_num_derivs(enum ac_image_dim dim)
2198{
2199	switch (dim) {
2200	case ac_image_1d:
2201	case ac_image_1darray:
2202		return 2;
2203	case ac_image_2d:
2204	case ac_image_2darray:
2205	case ac_image_cube:
2206		return 4;
2207	case ac_image_3d:
2208		return 6;
2209	case ac_image_2dmsaa:
2210	case ac_image_2darraymsaa:
2211	default:
2212		unreachable("derivatives not supported");
2213	}
2214}
2215
2216static const char *get_atomic_name(enum ac_atomic_op op)
2217{
2218	switch (op) {
2219	case ac_atomic_swap: return "swap";
2220	case ac_atomic_add: return "add";
2221	case ac_atomic_sub: return "sub";
2222	case ac_atomic_smin: return "smin";
2223	case ac_atomic_umin: return "umin";
2224	case ac_atomic_smax: return "smax";
2225	case ac_atomic_umax: return "umax";
2226	case ac_atomic_and: return "and";
2227	case ac_atomic_or: return "or";
2228	case ac_atomic_xor: return "xor";
2229	}
2230	unreachable("bad atomic op");
2231}
2232
2233LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
2234				   struct ac_image_args *a)
2235{
2236	const char *overload[3] = { "", "", "" };
2237	unsigned num_overloads = 0;
2238	LLVMValueRef args[18];
2239	unsigned num_args = 0;
2240	enum ac_image_dim dim = a->dim;
2241
2242	assert(!a->lod || a->lod == ctx->i32_0 || a->lod == ctx->f32_0 ||
2243	       !a->level_zero);
2244	assert((a->opcode != ac_image_get_resinfo && a->opcode != ac_image_load_mip &&
2245		a->opcode != ac_image_store_mip) ||
2246	       a->lod);
2247	assert(a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
2248	       (!a->compare && !a->offset));
2249	assert((a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
2250		a->opcode == ac_image_get_lod) ||
2251	       !a->bias);
2252	assert((a->bias ? 1 : 0) +
2253	       (a->lod ? 1 : 0) +
2254	       (a->level_zero ? 1 : 0) +
2255	       (a->derivs[0] ? 1 : 0) <= 1);
2256
2257	if (a->opcode == ac_image_get_lod) {
2258		switch (dim) {
2259		case ac_image_1darray:
2260			dim = ac_image_1d;
2261			break;
2262		case ac_image_2darray:
2263		case ac_image_cube:
2264			dim = ac_image_2d;
2265			break;
2266		default:
2267			break;
2268		}
2269	}
2270
2271	bool sample = a->opcode == ac_image_sample ||
2272		      a->opcode == ac_image_gather4 ||
2273		      a->opcode == ac_image_get_lod;
2274	bool atomic = a->opcode == ac_image_atomic ||
2275		      a->opcode == ac_image_atomic_cmpswap;
2276	LLVMTypeRef coord_type = sample ? ctx->f32 : ctx->i32;
2277
2278	if (atomic || a->opcode == ac_image_store || a->opcode == ac_image_store_mip) {
2279		args[num_args++] = a->data[0];
2280		if (a->opcode == ac_image_atomic_cmpswap)
2281			args[num_args++] = a->data[1];
2282	}
2283
2284	if (!atomic)
2285		args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, false);
2286
2287	if (a->offset)
2288		args[num_args++] = ac_to_integer(ctx, a->offset);
2289	if (a->bias) {
2290		args[num_args++] = ac_to_float(ctx, a->bias);
2291		overload[num_overloads++] = ".f32";
2292	}
2293	if (a->compare)
2294		args[num_args++] = ac_to_float(ctx, a->compare);
2295	if (a->derivs[0]) {
2296		unsigned count = ac_num_derivs(dim);
2297		for (unsigned i = 0; i < count; ++i)
2298			args[num_args++] = ac_to_float(ctx, a->derivs[i]);
2299		overload[num_overloads++] = ".f32";
2300	}
2301	unsigned num_coords =
2302		a->opcode != ac_image_get_resinfo ? ac_num_coords(dim) : 0;
2303	for (unsigned i = 0; i < num_coords; ++i)
2304		args[num_args++] = LLVMBuildBitCast(ctx->builder, a->coords[i], coord_type, "");
2305	if (a->lod)
2306		args[num_args++] = LLVMBuildBitCast(ctx->builder, a->lod, coord_type, "");
2307	overload[num_overloads++] = sample ? ".f32" : ".i32";
2308
2309	args[num_args++] = a->resource;
2310	if (sample) {
2311		args[num_args++] = a->sampler;
2312		args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, false);
2313	}
2314
2315	args[num_args++] = ctx->i32_0; /* texfailctrl */
2316	args[num_args++] = LLVMConstInt(ctx->i32, a->cache_policy, false);
2317
2318	const char *name;
2319	const char *atomic_subop = "";
2320	switch (a->opcode) {
2321	case ac_image_sample: name = "sample"; break;
2322	case ac_image_gather4: name = "gather4"; break;
2323	case ac_image_load: name = "load"; break;
2324	case ac_image_load_mip: name = "load.mip"; break;
2325	case ac_image_store: name = "store"; break;
2326	case ac_image_store_mip: name = "store.mip"; break;
2327	case ac_image_atomic:
2328		name = "atomic.";
2329		atomic_subop = get_atomic_name(a->atomic);
2330		break;
2331	case ac_image_atomic_cmpswap:
2332		name = "atomic.";
2333		atomic_subop = "cmpswap";
2334		break;
2335	case ac_image_get_lod: name = "getlod"; break;
2336	case ac_image_get_resinfo: name = "getresinfo"; break;
2337	default: unreachable("invalid image opcode");
2338	}
2339
2340	const char *dimname;
2341	switch (dim) {
2342	case ac_image_1d: dimname = "1d"; break;
2343	case ac_image_2d: dimname = "2d"; break;
2344	case ac_image_3d: dimname = "3d"; break;
2345	case ac_image_cube: dimname = "cube"; break;
2346	case ac_image_1darray: dimname = "1darray"; break;
2347	case ac_image_2darray: dimname = "2darray"; break;
2348	case ac_image_2dmsaa: dimname = "2dmsaa"; break;
2349	case ac_image_2darraymsaa: dimname = "2darraymsaa"; break;
2350	default: unreachable("invalid dim");
2351	}
2352
2353	bool lod_suffix =
2354		a->lod && (a->opcode == ac_image_sample || a->opcode == ac_image_gather4);
2355	char intr_name[96];
2356	snprintf(intr_name, sizeof(intr_name),
2357		 "llvm.amdgcn.image.%s%s" /* base name */
2358		 "%s%s%s" /* sample/gather modifiers */
2359		 ".%s.%s%s%s%s", /* dimension and type overloads */
2360		 name, atomic_subop,
2361		 a->compare ? ".c" : "",
2362		 a->bias ? ".b" :
2363		 lod_suffix ? ".l" :
2364		 a->derivs[0] ? ".d" :
2365		 a->level_zero ? ".lz" : "",
2366		 a->offset ? ".o" : "",
2367		 dimname,
2368		 atomic ? "i32" : "v4f32",
2369		 overload[0], overload[1], overload[2]);
2370
2371	LLVMTypeRef retty;
2372	if (atomic)
2373		retty = ctx->i32;
2374	else if (a->opcode == ac_image_store || a->opcode == ac_image_store_mip)
2375		retty = ctx->voidt;
2376	else
2377		retty = ctx->v4f32;
2378
2379	LLVMValueRef result =
2380		ac_build_intrinsic(ctx, intr_name, retty, args, num_args,
2381				   a->attributes);
2382	if (!sample && retty == ctx->v4f32) {
2383		result = LLVMBuildBitCast(ctx->builder, result,
2384					  ctx->v4i32, "");
2385	}
2386	return result;
2387}
2388
2389LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
2390				    LLVMValueRef args[2])
2391{
2392	LLVMTypeRef v2f16 =
2393		LLVMVectorType(LLVMHalfTypeInContext(ctx->context), 2);
2394
2395	return ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pkrtz", v2f16,
2396				  args, 2, AC_FUNC_ATTR_READNONE);
2397}
2398
2399LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
2400				     LLVMValueRef args[2])
2401{
2402	LLVMValueRef res =
2403		ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.i16",
2404				   ctx->v2i16, args, 2,
2405				   AC_FUNC_ATTR_READNONE);
2406	return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2407}
2408
2409LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
2410				     LLVMValueRef args[2])
2411{
2412	LLVMValueRef res =
2413		ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.u16",
2414				   ctx->v2i16, args, 2,
2415				   AC_FUNC_ATTR_READNONE);
2416	return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2417}
2418
2419/* The 8-bit and 10-bit clamping is for HW workarounds. */
2420LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
2421				 LLVMValueRef args[2], unsigned bits, bool hi)
2422{
2423	assert(bits == 8 || bits == 10 || bits == 16);
2424
2425	LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2426		bits == 8 ? 127 : bits == 10 ? 511 : 32767, 0);
2427	LLVMValueRef min_rgb = LLVMConstInt(ctx->i32,
2428		bits == 8 ? -128 : bits == 10 ? -512 : -32768, 0);
2429	LLVMValueRef max_alpha =
2430		bits != 10 ? max_rgb : ctx->i32_1;
2431	LLVMValueRef min_alpha =
2432		bits != 10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
2433
2434	/* Clamp. */
2435	if (bits != 16) {
2436		for (int i = 0; i < 2; i++) {
2437			bool alpha = hi && i == 1;
2438			args[i] = ac_build_imin(ctx, args[i],
2439						alpha ? max_alpha : max_rgb);
2440			args[i] = ac_build_imax(ctx, args[i],
2441						alpha ? min_alpha : min_rgb);
2442		}
2443	}
2444
2445	LLVMValueRef res =
2446		ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.i16",
2447				   ctx->v2i16, args, 2,
2448				   AC_FUNC_ATTR_READNONE);
2449	return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2450}
2451
2452/* The 8-bit and 10-bit clamping is for HW workarounds. */
2453LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
2454				 LLVMValueRef args[2], unsigned bits, bool hi)
2455{
2456	assert(bits == 8 || bits == 10 || bits == 16);
2457
2458	LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2459		bits == 8 ? 255 : bits == 10 ? 1023 : 65535, 0);
2460	LLVMValueRef max_alpha =
2461		bits != 10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
2462
2463	/* Clamp. */
2464	if (bits != 16) {
2465		for (int i = 0; i < 2; i++) {
2466			bool alpha = hi && i == 1;
2467			args[i] = ac_build_umin(ctx, args[i],
2468						alpha ? max_alpha : max_rgb);
2469		}
2470	}
2471
2472	LLVMValueRef res =
2473		ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.u16",
2474				   ctx->v2i16, args, 2,
2475				   AC_FUNC_ATTR_READNONE);
2476	return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2477}
2478
2479LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1)
2480{
2481	return ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.vote", ctx->i1,
2482				  &i1, 1, AC_FUNC_ATTR_READNONE);
2483}
2484
2485void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1)
2486{
2487	ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt,
2488			   &i1, 1, 0);
2489}
2490
2491LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
2492			  LLVMValueRef offset, LLVMValueRef width,
2493			  bool is_signed)
2494{
2495	LLVMValueRef args[] = {
2496		input,
2497		offset,
2498		width,
2499	};
2500
2501	return ac_build_intrinsic(ctx,
2502				  is_signed ? "llvm.amdgcn.sbfe.i32" :
2503					      "llvm.amdgcn.ubfe.i32",
2504				  ctx->i32, args, 3,
2505				  AC_FUNC_ATTR_READNONE);
2506}
2507
2508LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2509			   LLVMValueRef s1, LLVMValueRef s2)
2510{
2511	return LLVMBuildAdd(ctx->builder,
2512			    LLVMBuildMul(ctx->builder, s0, s1, ""), s2, "");
2513}
2514
2515LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2516			   LLVMValueRef s1, LLVMValueRef s2)
2517{
2518	return LLVMBuildFAdd(ctx->builder,
2519			     LLVMBuildFMul(ctx->builder, s0, s1, ""), s2, "");
2520}
2521
2522void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned simm16)
2523{
2524	LLVMValueRef args[1] = {
2525		LLVMConstInt(ctx->i32, simm16, false),
2526	};
2527	ac_build_intrinsic(ctx, "llvm.amdgcn.s.waitcnt",
2528			   ctx->voidt, args, 1, 0);
2529}
2530
2531LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
2532			    LLVMValueRef src1, LLVMValueRef src2,
2533			    unsigned bitsize)
2534{
2535	LLVMTypeRef type;
2536	char *intr;
2537
2538	if (bitsize == 16) {
2539		intr = "llvm.amdgcn.fmed3.f16";
2540		type = ctx->f16;
2541	} else if (bitsize == 32) {
2542		intr = "llvm.amdgcn.fmed3.f32";
2543		type = ctx->f32;
2544	} else {
2545		intr = "llvm.amdgcn.fmed3.f64";
2546		type = ctx->f64;
2547	}
2548
2549	LLVMValueRef params[] = {
2550		src0,
2551		src1,
2552		src2,
2553	};
2554	return ac_build_intrinsic(ctx, intr, type, params, 3,
2555				  AC_FUNC_ATTR_READNONE);
2556}
2557
2558LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
2559			    unsigned bitsize)
2560{
2561	LLVMTypeRef type;
2562	char *intr;
2563
2564	if (bitsize == 16) {
2565		intr = "llvm.amdgcn.fract.f16";
2566		type = ctx->f16;
2567	} else if (bitsize == 32) {
2568		intr = "llvm.amdgcn.fract.f32";
2569		type = ctx->f32;
2570	} else {
2571		intr = "llvm.amdgcn.fract.f64";
2572		type = ctx->f64;
2573	}
2574
2575	LLVMValueRef params[] = {
2576		src0,
2577	};
2578	return ac_build_intrinsic(ctx, intr, type, params, 1,
2579				  AC_FUNC_ATTR_READNONE);
2580}
2581
2582LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2583			    unsigned bitsize)
2584{
2585	LLVMTypeRef type = LLVMIntTypeInContext(ctx->context, bitsize);
2586	LLVMValueRef zero = LLVMConstInt(type, 0, false);
2587	LLVMValueRef one = LLVMConstInt(type, 1, false);
2588
2589	LLVMValueRef cmp, val;
2590	cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, src0, zero, "");
2591	val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2592	cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGE, val, zero, "");
2593	val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstInt(type, -1, true), "");
2594	return val;
2595}
2596
2597LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2598			    unsigned bitsize)
2599{
2600	LLVMValueRef cmp, val, zero, one;
2601	LLVMTypeRef type;
2602
2603	if (bitsize == 16) {
2604		type = ctx->f16;
2605		zero = ctx->f16_0;
2606		one = ctx->f16_1;
2607	} else if (bitsize == 32) {
2608		type = ctx->f32;
2609		zero = ctx->f32_0;
2610		one = ctx->f32_1;
2611	} else {
2612		type = ctx->f64;
2613		zero = ctx->f64_0;
2614		one = ctx->f64_1;
2615	}
2616
2617	cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGT, src0, zero, "");
2618	val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2619	cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGE, val, zero, "");
2620	val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstReal(type, -1.0), "");
2621	return val;
2622}
2623
2624LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0)
2625{
2626	LLVMValueRef result;
2627	unsigned bitsize;
2628
2629	bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2630
2631	switch (bitsize) {
2632	case 64:
2633		result = ac_build_intrinsic(ctx, "llvm.ctpop.i64", ctx->i64,
2634					    (LLVMValueRef []) { src0 }, 1,
2635					    AC_FUNC_ATTR_READNONE);
2636
2637		result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
2638		break;
2639	case 32:
2640		result = ac_build_intrinsic(ctx, "llvm.ctpop.i32", ctx->i32,
2641					    (LLVMValueRef []) { src0 }, 1,
2642					    AC_FUNC_ATTR_READNONE);
2643		break;
2644	case 16:
2645		result = ac_build_intrinsic(ctx, "llvm.ctpop.i16", ctx->i16,
2646					    (LLVMValueRef []) { src0 }, 1,
2647					    AC_FUNC_ATTR_READNONE);
2648
2649		result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2650		break;
2651	case 8:
2652		result = ac_build_intrinsic(ctx, "llvm.ctpop.i8", ctx->i8,
2653					    (LLVMValueRef []) { src0 }, 1,
2654					    AC_FUNC_ATTR_READNONE);
2655
2656		result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2657		break;
2658	default:
2659		unreachable(!"invalid bitsize");
2660		break;
2661	}
2662
2663	return result;
2664}
2665
2666LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
2667				       LLVMValueRef src0)
2668{
2669	LLVMValueRef result;
2670	unsigned bitsize;
2671
2672	bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2673
2674	switch (bitsize) {
2675	case 64:
2676		result = ac_build_intrinsic(ctx, "llvm.bitreverse.i64", ctx->i64,
2677					    (LLVMValueRef []) { src0 }, 1,
2678					    AC_FUNC_ATTR_READNONE);
2679
2680		result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
2681		break;
2682	case 32:
2683		result = ac_build_intrinsic(ctx, "llvm.bitreverse.i32", ctx->i32,
2684					    (LLVMValueRef []) { src0 }, 1,
2685					    AC_FUNC_ATTR_READNONE);
2686		break;
2687	case 16:
2688		result = ac_build_intrinsic(ctx, "llvm.bitreverse.i16", ctx->i16,
2689					    (LLVMValueRef []) { src0 }, 1,
2690					    AC_FUNC_ATTR_READNONE);
2691
2692		result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2693		break;
2694	case 8:
2695		result = ac_build_intrinsic(ctx, "llvm.bitreverse.i8", ctx->i8,
2696					    (LLVMValueRef []) { src0 }, 1,
2697					    AC_FUNC_ATTR_READNONE);
2698
2699		result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2700		break;
2701	default:
2702		unreachable(!"invalid bitsize");
2703		break;
2704	}
2705
2706	return result;
2707}
2708
2709#define AC_EXP_TARGET		0
2710#define AC_EXP_ENABLED_CHANNELS 1
2711#define AC_EXP_OUT0		2
2712
2713enum ac_ir_type {
2714	AC_IR_UNDEF,
2715	AC_IR_CONST,
2716	AC_IR_VALUE,
2717};
2718
2719struct ac_vs_exp_chan
2720{
2721	LLVMValueRef value;
2722	float const_float;
2723	enum ac_ir_type type;
2724};
2725
2726struct ac_vs_exp_inst {
2727	unsigned offset;
2728	LLVMValueRef inst;
2729	struct ac_vs_exp_chan chan[4];
2730};
2731
2732struct ac_vs_exports {
2733	unsigned num;
2734	struct ac_vs_exp_inst exp[VARYING_SLOT_MAX];
2735};
2736
2737/* Return true if the PARAM export has been eliminated. */
2738static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset,
2739				      uint32_t num_outputs,
2740				      struct ac_vs_exp_inst *exp)
2741{
2742	unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
2743	bool is_zero[4] = {}, is_one[4] = {};
2744
2745	for (i = 0; i < 4; i++) {
2746		/* It's a constant expression. Undef outputs are eliminated too. */
2747		if (exp->chan[i].type == AC_IR_UNDEF) {
2748			is_zero[i] = true;
2749			is_one[i] = true;
2750		} else if (exp->chan[i].type == AC_IR_CONST) {
2751			if (exp->chan[i].const_float == 0)
2752				is_zero[i] = true;
2753			else if (exp->chan[i].const_float == 1)
2754				is_one[i] = true;
2755			else
2756				return false; /* other constant */
2757		} else
2758			return false;
2759	}
2760
2761	/* Only certain combinations of 0 and 1 can be eliminated. */
2762	if (is_zero[0] && is_zero[1] && is_zero[2])
2763		default_val = is_zero[3] ? 0 : 1;
2764	else if (is_one[0] && is_one[1] && is_one[2])
2765		default_val = is_zero[3] ? 2 : 3;
2766	else
2767		return false;
2768
2769	/* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
2770	LLVMInstructionEraseFromParent(exp->inst);
2771
2772	/* Change OFFSET to DEFAULT_VAL. */
2773	for (i = 0; i < num_outputs; i++) {
2774		if (vs_output_param_offset[i] == exp->offset) {
2775			vs_output_param_offset[i] =
2776				AC_EXP_PARAM_DEFAULT_VAL_0000 + default_val;
2777			break;
2778		}
2779	}
2780	return true;
2781}
2782
2783static bool ac_eliminate_duplicated_output(struct ac_llvm_context *ctx,
2784					   uint8_t *vs_output_param_offset,
2785					   uint32_t num_outputs,
2786					   struct ac_vs_exports *processed,
2787				           struct ac_vs_exp_inst *exp)
2788{
2789	unsigned p, copy_back_channels = 0;
2790
2791	/* See if the output is already in the list of processed outputs.
2792	 * The LLVMValueRef comparison relies on SSA.
2793	 */
2794	for (p = 0; p < processed->num; p++) {
2795		bool different = false;
2796
2797		for (unsigned j = 0; j < 4; j++) {
2798			struct ac_vs_exp_chan *c1 = &processed->exp[p].chan[j];
2799			struct ac_vs_exp_chan *c2 = &exp->chan[j];
2800
2801			/* Treat undef as a match. */
2802			if (c2->type == AC_IR_UNDEF)
2803				continue;
2804
2805			/* If c1 is undef but c2 isn't, we can copy c2 to c1
2806			 * and consider the instruction duplicated.
2807			 */
2808			if (c1->type == AC_IR_UNDEF) {
2809				copy_back_channels |= 1 << j;
2810				continue;
2811			}
2812
2813			/* Test whether the channels are not equal. */
2814			if (c1->type != c2->type ||
2815			    (c1->type == AC_IR_CONST &&
2816			     c1->const_float != c2->const_float) ||
2817			    (c1->type == AC_IR_VALUE &&
2818			     c1->value != c2->value)) {
2819				different = true;
2820				break;
2821			}
2822		}
2823		if (!different)
2824			break;
2825
2826		copy_back_channels = 0;
2827	}
2828	if (p == processed->num)
2829		return false;
2830
2831	/* If a match was found, but the matching export has undef where the new
2832	 * one has a normal value, copy the normal value to the undef channel.
2833	 */
2834	struct ac_vs_exp_inst *match = &processed->exp[p];
2835
2836	/* Get current enabled channels mask. */
2837	LLVMValueRef arg = LLVMGetOperand(match->inst, AC_EXP_ENABLED_CHANNELS);
2838	unsigned enabled_channels = LLVMConstIntGetZExtValue(arg);
2839
2840	while (copy_back_channels) {
2841		unsigned chan = u_bit_scan(&copy_back_channels);
2842
2843		assert(match->chan[chan].type == AC_IR_UNDEF);
2844		LLVMSetOperand(match->inst, AC_EXP_OUT0 + chan,
2845			       exp->chan[chan].value);
2846		match->chan[chan] = exp->chan[chan];
2847
2848		/* Update number of enabled channels because the original mask
2849		 * is not always 0xf.
2850		 */
2851		enabled_channels |= (1 << chan);
2852		LLVMSetOperand(match->inst, AC_EXP_ENABLED_CHANNELS,
2853			       LLVMConstInt(ctx->i32, enabled_channels, 0));
2854	}
2855
2856	/* The PARAM export is duplicated. Kill it. */
2857	LLVMInstructionEraseFromParent(exp->inst);
2858
2859	/* Change OFFSET to the matching export. */
2860	for (unsigned i = 0; i < num_outputs; i++) {
2861		if (vs_output_param_offset[i] == exp->offset) {
2862			vs_output_param_offset[i] = match->offset;
2863			break;
2864		}
2865	}
2866	return true;
2867}
2868
2869void ac_optimize_vs_outputs(struct ac_llvm_context *ctx,
2870			    LLVMValueRef main_fn,
2871			    uint8_t *vs_output_param_offset,
2872			    uint32_t num_outputs,
2873			    uint8_t *num_param_exports)
2874{
2875	LLVMBasicBlockRef bb;
2876	bool removed_any = false;
2877	struct ac_vs_exports exports;
2878
2879	exports.num = 0;
2880
2881	/* Process all LLVM instructions. */
2882	bb = LLVMGetFirstBasicBlock(main_fn);
2883	while (bb) {
2884		LLVMValueRef inst = LLVMGetFirstInstruction(bb);
2885
2886		while (inst) {
2887			LLVMValueRef cur = inst;
2888			inst = LLVMGetNextInstruction(inst);
2889			struct ac_vs_exp_inst exp;
2890
2891			if (LLVMGetInstructionOpcode(cur) != LLVMCall)
2892				continue;
2893
2894			LLVMValueRef callee = ac_llvm_get_called_value(cur);
2895
2896			if (!ac_llvm_is_function(callee))
2897				continue;
2898
2899			const char *name = LLVMGetValueName(callee);
2900			unsigned num_args = LLVMCountParams(callee);
2901
2902			/* Check if this is an export instruction. */
2903			if ((num_args != 9 && num_args != 8) ||
2904			    (strcmp(name, "llvm.SI.export") &&
2905			     strcmp(name, "llvm.amdgcn.exp.f32")))
2906				continue;
2907
2908			LLVMValueRef arg = LLVMGetOperand(cur, AC_EXP_TARGET);
2909			unsigned target = LLVMConstIntGetZExtValue(arg);
2910
2911			if (target < V_008DFC_SQ_EXP_PARAM)
2912				continue;
2913
2914			target -= V_008DFC_SQ_EXP_PARAM;
2915
2916			/* Parse the instruction. */
2917			memset(&exp, 0, sizeof(exp));
2918			exp.offset = target;
2919			exp.inst = cur;
2920
2921			for (unsigned i = 0; i < 4; i++) {
2922				LLVMValueRef v = LLVMGetOperand(cur, AC_EXP_OUT0 + i);
2923
2924				exp.chan[i].value = v;
2925
2926				if (LLVMIsUndef(v)) {
2927					exp.chan[i].type = AC_IR_UNDEF;
2928				} else if (LLVMIsAConstantFP(v)) {
2929					LLVMBool loses_info;
2930					exp.chan[i].type = AC_IR_CONST;
2931					exp.chan[i].const_float =
2932						LLVMConstRealGetDouble(v, &loses_info);
2933				} else {
2934					exp.chan[i].type = AC_IR_VALUE;
2935				}
2936			}
2937
2938			/* Eliminate constant and duplicated PARAM exports. */
2939			if (ac_eliminate_const_output(vs_output_param_offset,
2940						      num_outputs, &exp) ||
2941			    ac_eliminate_duplicated_output(ctx,
2942							   vs_output_param_offset,
2943							   num_outputs, &exports,
2944							   &exp)) {
2945				removed_any = true;
2946			} else {
2947				exports.exp[exports.num++] = exp;
2948			}
2949		}
2950		bb = LLVMGetNextBasicBlock(bb);
2951	}
2952
2953	/* Remove holes in export memory due to removed PARAM exports.
2954	 * This is done by renumbering all PARAM exports.
2955	 */
2956	if (removed_any) {
2957		uint8_t old_offset[VARYING_SLOT_MAX];
2958		unsigned out, i;
2959
2960		/* Make a copy of the offsets. We need the old version while
2961		 * we are modifying some of them. */
2962		memcpy(old_offset, vs_output_param_offset,
2963		       sizeof(old_offset));
2964
2965		for (i = 0; i < exports.num; i++) {
2966			unsigned offset = exports.exp[i].offset;
2967
2968			/* Update vs_output_param_offset. Multiple outputs can
2969			 * have the same offset.
2970			 */
2971			for (out = 0; out < num_outputs; out++) {
2972				if (old_offset[out] == offset)
2973					vs_output_param_offset[out] = i;
2974			}
2975
2976			/* Change the PARAM offset in the instruction. */
2977			LLVMSetOperand(exports.exp[i].inst, AC_EXP_TARGET,
2978				       LLVMConstInt(ctx->i32,
2979						    V_008DFC_SQ_EXP_PARAM + i, 0));
2980		}
2981		*num_param_exports = exports.num;
2982	}
2983}
2984
2985void ac_init_exec_full_mask(struct ac_llvm_context *ctx)
2986{
2987	LLVMValueRef full_mask = LLVMConstInt(ctx->i64, ~0ull, 0);
2988	ac_build_intrinsic(ctx,
2989			   "llvm.amdgcn.init.exec", ctx->voidt,
2990			   &full_mask, 1, AC_FUNC_ATTR_CONVERGENT);
2991}
2992
2993void ac_declare_lds_as_pointer(struct ac_llvm_context *ctx)
2994{
2995	unsigned lds_size = ctx->chip_class >= CIK ? 65536 : 32768;
2996	ctx->lds = LLVMBuildIntToPtr(ctx->builder, ctx->i32_0,
2997				     LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), AC_ADDR_SPACE_LDS),
2998				     "lds");
2999}
3000
3001LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
3002			 LLVMValueRef dw_addr)
3003{
3004	return ac_build_load(ctx, ctx->lds, dw_addr);
3005}
3006
3007void ac_lds_store(struct ac_llvm_context *ctx,
3008		  LLVMValueRef dw_addr,
3009		  LLVMValueRef value)
3010{
3011	value = ac_to_integer(ctx, value);
3012	ac_build_indexed_store(ctx, ctx->lds,
3013			       dw_addr, value);
3014}
3015
3016LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
3017			 LLVMTypeRef dst_type,
3018			 LLVMValueRef src0)
3019{
3020	unsigned src0_bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
3021	const char *intrin_name;
3022	LLVMTypeRef type;
3023	LLVMValueRef zero;
3024
3025	switch (src0_bitsize) {
3026	case 64:
3027		intrin_name = "llvm.cttz.i64";
3028		type = ctx->i64;
3029		zero = ctx->i64_0;
3030		break;
3031	case 32:
3032		intrin_name = "llvm.cttz.i32";
3033		type = ctx->i32;
3034		zero = ctx->i32_0;
3035		break;
3036	case 16:
3037		intrin_name = "llvm.cttz.i16";
3038		type = ctx->i16;
3039		zero = ctx->i16_0;
3040		break;
3041	case 8:
3042		intrin_name = "llvm.cttz.i8";
3043		type = ctx->i8;
3044		zero = ctx->i8_0;
3045		break;
3046	default:
3047		unreachable(!"invalid bitsize");
3048	}
3049
3050	LLVMValueRef params[2] = {
3051		src0,
3052
3053		/* The value of 1 means that ffs(x=0) = undef, so LLVM won't
3054		 * add special code to check for x=0. The reason is that
3055		 * the LLVM behavior for x=0 is different from what we
3056		 * need here. However, LLVM also assumes that ffs(x) is
3057		 * in [0, 31], but GLSL expects that ffs(0) = -1, so
3058		 * a conditional assignment to handle 0 is still required.
3059		 *
3060		 * The hardware already implements the correct behavior.
3061		 */
3062		ctx->i1true,
3063	};
3064
3065	LLVMValueRef lsb = ac_build_intrinsic(ctx, intrin_name, type,
3066					      params, 2,
3067					      AC_FUNC_ATTR_READNONE);
3068
3069	if (src0_bitsize == 64) {
3070		lsb = LLVMBuildTrunc(ctx->builder, lsb, ctx->i32, "");
3071	} else if (src0_bitsize < 32) {
3072		lsb = LLVMBuildSExt(ctx->builder, lsb, ctx->i32, "");
3073	}
3074
3075	/* TODO: We need an intrinsic to skip this conditional. */
3076	/* Check for zero: */
3077	return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder,
3078							   LLVMIntEQ, src0,
3079							   zero, ""),
3080			       LLVMConstInt(ctx->i32, -1, 0), lsb, "");
3081}
3082
3083LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type)
3084{
3085	return LLVMPointerType(LLVMArrayType(elem_type, 0),
3086			       AC_ADDR_SPACE_CONST);
3087}
3088
3089LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type)
3090{
3091	return LLVMPointerType(LLVMArrayType(elem_type, 0),
3092			       AC_ADDR_SPACE_CONST_32BIT);
3093}
3094
3095static struct ac_llvm_flow *
3096get_current_flow(struct ac_llvm_context *ctx)
3097{
3098	if (ctx->flow_depth > 0)
3099		return &ctx->flow[ctx->flow_depth - 1];
3100	return NULL;
3101}
3102
3103static struct ac_llvm_flow *
3104get_innermost_loop(struct ac_llvm_context *ctx)
3105{
3106	for (unsigned i = ctx->flow_depth; i > 0; --i) {
3107		if (ctx->flow[i - 1].loop_entry_block)
3108			return &ctx->flow[i - 1];
3109	}
3110	return NULL;
3111}
3112
3113static struct ac_llvm_flow *
3114push_flow(struct ac_llvm_context *ctx)
3115{
3116	struct ac_llvm_flow *flow;
3117
3118	if (ctx->flow_depth >= ctx->flow_depth_max) {
3119		unsigned new_max = MAX2(ctx->flow_depth << 1,
3120					AC_LLVM_INITIAL_CF_DEPTH);
3121
3122		ctx->flow = realloc(ctx->flow, new_max * sizeof(*ctx->flow));
3123		ctx->flow_depth_max = new_max;
3124	}
3125
3126	flow = &ctx->flow[ctx->flow_depth];
3127	ctx->flow_depth++;
3128
3129	flow->next_block = NULL;
3130	flow->loop_entry_block = NULL;
3131	return flow;
3132}
3133
3134static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base,
3135				int label_id)
3136{
3137	char buf[32];
3138	snprintf(buf, sizeof(buf), "%s%d", base, label_id);
3139	LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
3140}
3141
3142/* Append a basic block at the level of the parent flow.
3143 */
3144static LLVMBasicBlockRef append_basic_block(struct ac_llvm_context *ctx,
3145					    const char *name)
3146{
3147	assert(ctx->flow_depth >= 1);
3148
3149	if (ctx->flow_depth >= 2) {
3150		struct ac_llvm_flow *flow = &ctx->flow[ctx->flow_depth - 2];
3151
3152		return LLVMInsertBasicBlockInContext(ctx->context,
3153						     flow->next_block, name);
3154	}
3155
3156	LLVMValueRef main_fn =
3157		LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->builder));
3158	return LLVMAppendBasicBlockInContext(ctx->context, main_fn, name);
3159}
3160
3161/* Emit a branch to the given default target for the current block if
3162 * applicable -- that is, if the current block does not already contain a
3163 * branch from a break or continue.
3164 */
3165static void emit_default_branch(LLVMBuilderRef builder,
3166				LLVMBasicBlockRef target)
3167{
3168	if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
3169		 LLVMBuildBr(builder, target);
3170}
3171
3172void ac_build_bgnloop(struct ac_llvm_context *ctx, int label_id)
3173{
3174	struct ac_llvm_flow *flow = push_flow(ctx);
3175	flow->loop_entry_block = append_basic_block(ctx, "LOOP");
3176	flow->next_block = append_basic_block(ctx, "ENDLOOP");
3177	set_basicblock_name(flow->loop_entry_block, "loop", label_id);
3178	LLVMBuildBr(ctx->builder, flow->loop_entry_block);
3179	LLVMPositionBuilderAtEnd(ctx->builder, flow->loop_entry_block);
3180}
3181
3182void ac_build_break(struct ac_llvm_context *ctx)
3183{
3184	struct ac_llvm_flow *flow = get_innermost_loop(ctx);
3185	LLVMBuildBr(ctx->builder, flow->next_block);
3186}
3187
3188void ac_build_continue(struct ac_llvm_context *ctx)
3189{
3190	struct ac_llvm_flow *flow = get_innermost_loop(ctx);
3191	LLVMBuildBr(ctx->builder, flow->loop_entry_block);
3192}
3193
3194void ac_build_else(struct ac_llvm_context *ctx, int label_id)
3195{
3196	struct ac_llvm_flow *current_branch = get_current_flow(ctx);
3197	LLVMBasicBlockRef endif_block;
3198
3199	assert(!current_branch->loop_entry_block);
3200
3201	endif_block = append_basic_block(ctx, "ENDIF");
3202	emit_default_branch(ctx->builder, endif_block);
3203
3204	LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
3205	set_basicblock_name(current_branch->next_block, "else", label_id);
3206
3207	current_branch->next_block = endif_block;
3208}
3209
3210void ac_build_endif(struct ac_llvm_context *ctx, int label_id)
3211{
3212	struct ac_llvm_flow *current_branch = get_current_flow(ctx);
3213
3214	assert(!current_branch->loop_entry_block);
3215
3216	emit_default_branch(ctx->builder, current_branch->next_block);
3217	LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
3218	set_basicblock_name(current_branch->next_block, "endif", label_id);
3219
3220	ctx->flow_depth--;
3221}
3222
3223void ac_build_endloop(struct ac_llvm_context *ctx, int label_id)
3224{
3225	struct ac_llvm_flow *current_loop = get_current_flow(ctx);
3226
3227	assert(current_loop->loop_entry_block);
3228
3229	emit_default_branch(ctx->builder, current_loop->loop_entry_block);
3230
3231	LLVMPositionBuilderAtEnd(ctx->builder, current_loop->next_block);
3232	set_basicblock_name(current_loop->next_block, "endloop", label_id);
3233	ctx->flow_depth--;
3234}
3235
3236void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id)
3237{
3238	struct ac_llvm_flow *flow = push_flow(ctx);
3239	LLVMBasicBlockRef if_block;
3240
3241	if_block = append_basic_block(ctx, "IF");
3242	flow->next_block = append_basic_block(ctx, "ELSE");
3243	set_basicblock_name(if_block, "if", label_id);
3244	LLVMBuildCondBr(ctx->builder, cond, if_block, flow->next_block);
3245	LLVMPositionBuilderAtEnd(ctx->builder, if_block);
3246}
3247
3248void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
3249		 int label_id)
3250{
3251	LLVMValueRef cond = LLVMBuildFCmp(ctx->builder, LLVMRealUNE,
3252					  value, ctx->f32_0, "");
3253	ac_build_ifcc(ctx, cond, label_id);
3254}
3255
3256void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
3257		  int label_id)
3258{
3259	LLVMValueRef cond = LLVMBuildICmp(ctx->builder, LLVMIntNE,
3260					  ac_to_integer(ctx, value),
3261					  ctx->i32_0, "");
3262	ac_build_ifcc(ctx, cond, label_id);
3263}
3264
3265LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
3266			     const char *name)
3267{
3268	LLVMBuilderRef builder = ac->builder;
3269	LLVMBasicBlockRef current_block = LLVMGetInsertBlock(builder);
3270	LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
3271	LLVMBasicBlockRef first_block = LLVMGetEntryBasicBlock(function);
3272	LLVMValueRef first_instr = LLVMGetFirstInstruction(first_block);
3273	LLVMBuilderRef first_builder = LLVMCreateBuilderInContext(ac->context);
3274	LLVMValueRef res;
3275
3276	if (first_instr) {
3277		LLVMPositionBuilderBefore(first_builder, first_instr);
3278	} else {
3279		LLVMPositionBuilderAtEnd(first_builder, first_block);
3280	}
3281
3282	res = LLVMBuildAlloca(first_builder, type, name);
3283	LLVMDisposeBuilder(first_builder);
3284	return res;
3285}
3286
3287LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac,
3288				   LLVMTypeRef type, const char *name)
3289{
3290	LLVMValueRef ptr = ac_build_alloca_undef(ac, type, name);
3291	LLVMBuildStore(ac->builder, LLVMConstNull(type), ptr);
3292	return ptr;
3293}
3294
3295LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
3296                         LLVMTypeRef type)
3297{
3298	int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
3299	return LLVMBuildBitCast(ctx->builder, ptr,
3300	                        LLVMPointerType(type, addr_space), "");
3301}
3302
3303LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
3304			    unsigned count)
3305{
3306	unsigned num_components = ac_get_llvm_num_components(value);
3307	if (count == num_components)
3308		return value;
3309
3310	LLVMValueRef masks[MAX2(count, 2)];
3311	masks[0] = ctx->i32_0;
3312	masks[1] = ctx->i32_1;
3313	for (unsigned i = 2; i < count; i++)
3314		masks[i] = LLVMConstInt(ctx->i32, i, false);
3315
3316	if (count == 1)
3317		return LLVMBuildExtractElement(ctx->builder, value, masks[0],
3318		                               "");
3319
3320	LLVMValueRef swizzle = LLVMConstVector(masks, count);
3321	return LLVMBuildShuffleVector(ctx->builder, value, value, swizzle, "");
3322}
3323
3324LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
3325			     unsigned rshift, unsigned bitwidth)
3326{
3327	LLVMValueRef value = param;
3328	if (rshift)
3329		value = LLVMBuildLShr(ctx->builder, value,
3330				      LLVMConstInt(ctx->i32, rshift, false), "");
3331
3332	if (rshift + bitwidth < 32) {
3333		unsigned mask = (1 << bitwidth) - 1;
3334		value = LLVMBuildAnd(ctx->builder, value,
3335				     LLVMConstInt(ctx->i32, mask, false), "");
3336	}
3337	return value;
3338}
3339
3340/* Adjust the sample index according to FMASK.
3341 *
3342 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
3343 * which is the identity mapping. Each nibble says which physical sample
3344 * should be fetched to get that sample.
3345 *
3346 * For example, 0x11111100 means there are only 2 samples stored and
3347 * the second sample covers 3/4 of the pixel. When reading samples 0
3348 * and 1, return physical sample 0 (determined by the first two 0s
3349 * in FMASK), otherwise return physical sample 1.
3350 *
3351 * The sample index should be adjusted as follows:
3352 *   addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
3353 */
3354void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
3355			      LLVMValueRef *addr, bool is_array_tex)
3356{
3357	struct ac_image_args fmask_load = {};
3358	fmask_load.opcode = ac_image_load;
3359	fmask_load.resource = fmask;
3360	fmask_load.dmask = 0xf;
3361	fmask_load.dim = is_array_tex ? ac_image_2darray : ac_image_2d;
3362	fmask_load.attributes = AC_FUNC_ATTR_READNONE;
3363
3364	fmask_load.coords[0] = addr[0];
3365	fmask_load.coords[1] = addr[1];
3366	if (is_array_tex)
3367		fmask_load.coords[2] = addr[2];
3368
3369	LLVMValueRef fmask_value = ac_build_image_opcode(ac, &fmask_load);
3370	fmask_value = LLVMBuildExtractElement(ac->builder, fmask_value,
3371					      ac->i32_0, "");
3372
3373	/* Apply the formula. */
3374	unsigned sample_chan = is_array_tex ? 3 : 2;
3375	LLVMValueRef final_sample;
3376	final_sample = LLVMBuildMul(ac->builder, addr[sample_chan],
3377				    LLVMConstInt(ac->i32, 4, 0), "");
3378	final_sample = LLVMBuildLShr(ac->builder, fmask_value, final_sample, "");
3379	/* Mask the sample index by 0x7, because 0x8 means an unknown value
3380	 * with EQAA, so those will map to 0. */
3381	final_sample = LLVMBuildAnd(ac->builder, final_sample,
3382				    LLVMConstInt(ac->i32, 0x7, 0), "");
3383
3384	/* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
3385	 * resource descriptor is 0 (invalid).
3386	 */
3387	LLVMValueRef tmp;
3388	tmp = LLVMBuildBitCast(ac->builder, fmask, ac->v8i32, "");
3389	tmp = LLVMBuildExtractElement(ac->builder, tmp, ac->i32_1, "");
3390	tmp = LLVMBuildICmp(ac->builder, LLVMIntNE, tmp, ac->i32_0, "");
3391
3392	/* Replace the MSAA sample index. */
3393	addr[sample_chan] = LLVMBuildSelect(ac->builder, tmp, final_sample,
3394					    addr[sample_chan], "");
3395}
3396
3397static LLVMValueRef
3398_ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
3399{
3400	ac_build_optimization_barrier(ctx, &src);
3401	return ac_build_intrinsic(ctx,
3402			lane == NULL ? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
3403			LLVMTypeOf(src), (LLVMValueRef []) {
3404			src, lane },
3405			lane == NULL ? 1 : 2,
3406			AC_FUNC_ATTR_READNONE |
3407			AC_FUNC_ATTR_CONVERGENT);
3408}
3409
3410/**
3411 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
3412 * @param ctx
3413 * @param src
3414 * @param lane - id of the lane or NULL for the first active lane
3415 * @return value of the lane
3416 */
3417LLVMValueRef
3418ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
3419{
3420	LLVMTypeRef src_type = LLVMTypeOf(src);
3421	src = ac_to_integer(ctx, src);
3422	unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3423	LLVMValueRef ret;
3424
3425	if (bits == 32) {
3426		ret = _ac_build_readlane(ctx, src, lane);
3427	} else {
3428		assert(bits % 32 == 0);
3429		LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3430		LLVMValueRef src_vector =
3431			LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3432		ret = LLVMGetUndef(vec_type);
3433		for (unsigned i = 0; i < bits / 32; i++) {
3434			src = LLVMBuildExtractElement(ctx->builder, src_vector,
3435						LLVMConstInt(ctx->i32, i, 0), "");
3436			LLVMValueRef ret_comp = _ac_build_readlane(ctx, src, lane);
3437			ret = LLVMBuildInsertElement(ctx->builder, ret, ret_comp,
3438						LLVMConstInt(ctx->i32, i, 0), "");
3439		}
3440	}
3441	if (LLVMGetTypeKind(src_type) == LLVMPointerTypeKind)
3442		return LLVMBuildIntToPtr(ctx->builder, ret, src_type, "");
3443	return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3444}
3445
3446LLVMValueRef
3447ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane)
3448{
3449	/* TODO: Use the actual instruction when LLVM adds an intrinsic for it.
3450	 */
3451	LLVMValueRef pred = LLVMBuildICmp(ctx->builder, LLVMIntEQ, lane,
3452					  ac_get_thread_id(ctx), "");
3453	return LLVMBuildSelect(ctx->builder, pred, value, src, "");
3454}
3455
3456LLVMValueRef
3457ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask)
3458{
3459	LLVMValueRef mask_vec = LLVMBuildBitCast(ctx->builder, mask,
3460						 LLVMVectorType(ctx->i32, 2),
3461						 "");
3462	LLVMValueRef mask_lo = LLVMBuildExtractElement(ctx->builder, mask_vec,
3463						       ctx->i32_0, "");
3464	LLVMValueRef mask_hi = LLVMBuildExtractElement(ctx->builder, mask_vec,
3465						       ctx->i32_1, "");
3466	LLVMValueRef val =
3467		ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
3468				   (LLVMValueRef []) { mask_lo, ctx->i32_0 },
3469				   2, AC_FUNC_ATTR_READNONE);
3470	val = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi", ctx->i32,
3471				 (LLVMValueRef []) { mask_hi, val },
3472				 2, AC_FUNC_ATTR_READNONE);
3473	return val;
3474}
3475
3476enum dpp_ctrl {
3477	_dpp_quad_perm = 0x000,
3478	_dpp_row_sl = 0x100,
3479	_dpp_row_sr = 0x110,
3480	_dpp_row_rr = 0x120,
3481	dpp_wf_sl1 = 0x130,
3482	dpp_wf_rl1 = 0x134,
3483	dpp_wf_sr1 = 0x138,
3484	dpp_wf_rr1 = 0x13C,
3485	dpp_row_mirror = 0x140,
3486	dpp_row_half_mirror = 0x141,
3487	dpp_row_bcast15 = 0x142,
3488	dpp_row_bcast31 = 0x143
3489};
3490
3491static inline enum dpp_ctrl
3492dpp_quad_perm(unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
3493{
3494	assert(lane0 < 4 && lane1 < 4 && lane2 < 4 && lane3 < 4);
3495	return _dpp_quad_perm | lane0 | (lane1 << 2) | (lane2 << 4) | (lane3 << 6);
3496}
3497
3498static inline enum dpp_ctrl
3499dpp_row_sl(unsigned amount)
3500{
3501	assert(amount > 0 && amount < 16);
3502	return _dpp_row_sl | amount;
3503}
3504
3505static inline enum dpp_ctrl
3506dpp_row_sr(unsigned amount)
3507{
3508	assert(amount > 0 && amount < 16);
3509	return _dpp_row_sr | amount;
3510}
3511
3512static LLVMValueRef
3513_ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3514	      enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3515	      bool bound_ctrl)
3516{
3517	return ac_build_intrinsic(ctx, "llvm.amdgcn.update.dpp.i32",
3518					LLVMTypeOf(old),
3519					(LLVMValueRef[]) {
3520						old, src,
3521						LLVMConstInt(ctx->i32, dpp_ctrl, 0),
3522						LLVMConstInt(ctx->i32, row_mask, 0),
3523						LLVMConstInt(ctx->i32, bank_mask, 0),
3524						LLVMConstInt(ctx->i1, bound_ctrl, 0) },
3525					6, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3526}
3527
3528static LLVMValueRef
3529ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3530	     enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3531	     bool bound_ctrl)
3532{
3533	LLVMTypeRef src_type = LLVMTypeOf(src);
3534	src = ac_to_integer(ctx, src);
3535	old = ac_to_integer(ctx, old);
3536	unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3537	LLVMValueRef ret;
3538	if (bits == 32) {
3539		ret = _ac_build_dpp(ctx, old, src, dpp_ctrl, row_mask,
3540				    bank_mask, bound_ctrl);
3541	} else {
3542		assert(bits % 32 == 0);
3543		LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3544		LLVMValueRef src_vector =
3545			LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3546		LLVMValueRef old_vector =
3547			LLVMBuildBitCast(ctx->builder, old, vec_type, "");
3548		ret = LLVMGetUndef(vec_type);
3549		for (unsigned i = 0; i < bits / 32; i++) {
3550			src = LLVMBuildExtractElement(ctx->builder, src_vector,
3551						      LLVMConstInt(ctx->i32, i,
3552								   0), "");
3553			old = LLVMBuildExtractElement(ctx->builder, old_vector,
3554						      LLVMConstInt(ctx->i32, i,
3555								   0), "");
3556			LLVMValueRef ret_comp = _ac_build_dpp(ctx, old, src,
3557							      dpp_ctrl,
3558							      row_mask,
3559							      bank_mask,
3560							      bound_ctrl);
3561			ret = LLVMBuildInsertElement(ctx->builder, ret,
3562						     ret_comp,
3563						     LLVMConstInt(ctx->i32, i,
3564								  0), "");
3565		}
3566	}
3567	return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3568}
3569
3570static inline unsigned
3571ds_pattern_bitmode(unsigned and_mask, unsigned or_mask, unsigned xor_mask)
3572{
3573	assert(and_mask < 32 && or_mask < 32 && xor_mask < 32);
3574	return and_mask | (or_mask << 5) | (xor_mask << 10);
3575}
3576
3577static LLVMValueRef
3578_ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
3579{
3580	return ac_build_intrinsic(ctx, "llvm.amdgcn.ds.swizzle",
3581				   LLVMTypeOf(src), (LLVMValueRef []) {
3582					src, LLVMConstInt(ctx->i32, mask, 0) },
3583				   2, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3584}
3585
3586LLVMValueRef
3587ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
3588{
3589	LLVMTypeRef src_type = LLVMTypeOf(src);
3590	src = ac_to_integer(ctx, src);
3591	unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3592	LLVMValueRef ret;
3593	if (bits == 32) {
3594		ret = _ac_build_ds_swizzle(ctx, src, mask);
3595	} else {
3596		assert(bits % 32 == 0);
3597		LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3598		LLVMValueRef src_vector =
3599			LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3600		ret = LLVMGetUndef(vec_type);
3601		for (unsigned i = 0; i < bits / 32; i++) {
3602			src = LLVMBuildExtractElement(ctx->builder, src_vector,
3603						      LLVMConstInt(ctx->i32, i,
3604								   0), "");
3605			LLVMValueRef ret_comp = _ac_build_ds_swizzle(ctx, src,
3606								     mask);
3607			ret = LLVMBuildInsertElement(ctx->builder, ret,
3608						     ret_comp,
3609						     LLVMConstInt(ctx->i32, i,
3610								  0), "");
3611		}
3612	}
3613	return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3614}
3615
3616static LLVMValueRef
3617ac_build_wwm(struct ac_llvm_context *ctx, LLVMValueRef src)
3618{
3619	char name[32], type[8];
3620	ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
3621	snprintf(name, sizeof(name), "llvm.amdgcn.wwm.%s", type);
3622	return ac_build_intrinsic(ctx, name, LLVMTypeOf(src),
3623				  (LLVMValueRef []) { src }, 1,
3624				  AC_FUNC_ATTR_READNONE);
3625}
3626
3627static LLVMValueRef
3628ac_build_set_inactive(struct ac_llvm_context *ctx, LLVMValueRef src,
3629		      LLVMValueRef inactive)
3630{
3631	char name[33], type[8];
3632	LLVMTypeRef src_type = LLVMTypeOf(src);
3633	src = ac_to_integer(ctx, src);
3634	inactive = ac_to_integer(ctx, inactive);
3635	ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
3636	snprintf(name, sizeof(name), "llvm.amdgcn.set.inactive.%s", type);
3637	LLVMValueRef ret =
3638		ac_build_intrinsic(ctx, name,
3639					LLVMTypeOf(src), (LLVMValueRef []) {
3640					src, inactive }, 2,
3641					AC_FUNC_ATTR_READNONE |
3642					AC_FUNC_ATTR_CONVERGENT);
3643	return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3644}
3645
3646static LLVMValueRef
3647get_reduction_identity(struct ac_llvm_context *ctx, nir_op op, unsigned type_size)
3648{
3649	if (type_size == 4) {
3650		switch (op) {
3651		case nir_op_iadd: return ctx->i32_0;
3652		case nir_op_fadd: return ctx->f32_0;
3653		case nir_op_imul: return ctx->i32_1;
3654		case nir_op_fmul: return ctx->f32_1;
3655		case nir_op_imin: return LLVMConstInt(ctx->i32, INT32_MAX, 0);
3656		case nir_op_umin: return LLVMConstInt(ctx->i32, UINT32_MAX, 0);
3657		case nir_op_fmin: return LLVMConstReal(ctx->f32, INFINITY);
3658		case nir_op_imax: return LLVMConstInt(ctx->i32, INT32_MIN, 0);
3659		case nir_op_umax: return ctx->i32_0;
3660		case nir_op_fmax: return LLVMConstReal(ctx->f32, -INFINITY);
3661		case nir_op_iand: return LLVMConstInt(ctx->i32, -1, 0);
3662		case nir_op_ior: return ctx->i32_0;
3663		case nir_op_ixor: return ctx->i32_0;
3664		default:
3665			unreachable("bad reduction intrinsic");
3666		}
3667	} else { /* type_size == 64bit */
3668		switch (op) {
3669		case nir_op_iadd: return ctx->i64_0;
3670		case nir_op_fadd: return ctx->f64_0;
3671		case nir_op_imul: return ctx->i64_1;
3672		case nir_op_fmul: return ctx->f64_1;
3673		case nir_op_imin: return LLVMConstInt(ctx->i64, INT64_MAX, 0);
3674		case nir_op_umin: return LLVMConstInt(ctx->i64, UINT64_MAX, 0);
3675		case nir_op_fmin: return LLVMConstReal(ctx->f64, INFINITY);
3676		case nir_op_imax: return LLVMConstInt(ctx->i64, INT64_MIN, 0);
3677		case nir_op_umax: return ctx->i64_0;
3678		case nir_op_fmax: return LLVMConstReal(ctx->f64, -INFINITY);
3679		case nir_op_iand: return LLVMConstInt(ctx->i64, -1, 0);
3680		case nir_op_ior: return ctx->i64_0;
3681		case nir_op_ixor: return ctx->i64_0;
3682		default:
3683			unreachable("bad reduction intrinsic");
3684		}
3685	}
3686}
3687
3688static LLVMValueRef
3689ac_build_alu_op(struct ac_llvm_context *ctx, LLVMValueRef lhs, LLVMValueRef rhs, nir_op op)
3690{
3691	bool _64bit = ac_get_type_size(LLVMTypeOf(lhs)) == 8;
3692	switch (op) {
3693	case nir_op_iadd: return LLVMBuildAdd(ctx->builder, lhs, rhs, "");
3694	case nir_op_fadd: return LLVMBuildFAdd(ctx->builder, lhs, rhs, "");
3695	case nir_op_imul: return LLVMBuildMul(ctx->builder, lhs, rhs, "");
3696	case nir_op_fmul: return LLVMBuildFMul(ctx->builder, lhs, rhs, "");
3697	case nir_op_imin: return LLVMBuildSelect(ctx->builder,
3698					LLVMBuildICmp(ctx->builder, LLVMIntSLT, lhs, rhs, ""),
3699					lhs, rhs, "");
3700	case nir_op_umin: return LLVMBuildSelect(ctx->builder,
3701					LLVMBuildICmp(ctx->builder, LLVMIntULT, lhs, rhs, ""),
3702					lhs, rhs, "");
3703	case nir_op_fmin: return ac_build_intrinsic(ctx,
3704					_64bit ? "llvm.minnum.f64" : "llvm.minnum.f32",
3705					_64bit ? ctx->f64 : ctx->f32,
3706					(LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
3707	case nir_op_imax: return LLVMBuildSelect(ctx->builder,
3708					LLVMBuildICmp(ctx->builder, LLVMIntSGT, lhs, rhs, ""),
3709					lhs, rhs, "");
3710	case nir_op_umax: return LLVMBuildSelect(ctx->builder,
3711					LLVMBuildICmp(ctx->builder, LLVMIntUGT, lhs, rhs, ""),
3712					lhs, rhs, "");
3713	case nir_op_fmax: return ac_build_intrinsic(ctx,
3714					_64bit ? "llvm.maxnum.f64" : "llvm.maxnum.f32",
3715					_64bit ? ctx->f64 : ctx->f32,
3716					(LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
3717	case nir_op_iand: return LLVMBuildAnd(ctx->builder, lhs, rhs, "");
3718	case nir_op_ior: return LLVMBuildOr(ctx->builder, lhs, rhs, "");
3719	case nir_op_ixor: return LLVMBuildXor(ctx->builder, lhs, rhs, "");
3720	default:
3721		unreachable("bad reduction intrinsic");
3722	}
3723}
3724
3725/**
3726 * \param maxprefix specifies that the result only needs to be correct for a
3727 *     prefix of this many threads
3728 *
3729 * TODO: add inclusive and excluse scan functions for SI chip class.
3730 */
3731static LLVMValueRef
3732ac_build_scan(struct ac_llvm_context *ctx, nir_op op, LLVMValueRef src, LLVMValueRef identity,
3733	      unsigned maxprefix)
3734{
3735	LLVMValueRef result, tmp;
3736	result = src;
3737	if (maxprefix <= 1)
3738		return result;
3739	tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
3740	result = ac_build_alu_op(ctx, result, tmp, op);
3741	if (maxprefix <= 2)
3742		return result;
3743	tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(2), 0xf, 0xf, false);
3744	result = ac_build_alu_op(ctx, result, tmp, op);
3745	if (maxprefix <= 3)
3746		return result;
3747	tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(3), 0xf, 0xf, false);
3748	result = ac_build_alu_op(ctx, result, tmp, op);
3749	if (maxprefix <= 4)
3750		return result;
3751	tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(4), 0xf, 0xe, false);
3752	result = ac_build_alu_op(ctx, result, tmp, op);
3753	if (maxprefix <= 8)
3754		return result;
3755	tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(8), 0xf, 0xc, false);
3756	result = ac_build_alu_op(ctx, result, tmp, op);
3757	if (maxprefix <= 16)
3758		return result;
3759	tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
3760	result = ac_build_alu_op(ctx, result, tmp, op);
3761	if (maxprefix <= 32)
3762		return result;
3763	tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
3764	result = ac_build_alu_op(ctx, result, tmp, op);
3765	return result;
3766}
3767
3768LLVMValueRef
3769ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
3770{
3771	LLVMValueRef result;
3772
3773	if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
3774		LLVMBuilderRef builder = ctx->builder;
3775		src = LLVMBuildZExt(builder, src, ctx->i32, "");
3776		result = ac_build_ballot(ctx, src);
3777		result = ac_build_mbcnt(ctx, result);
3778		result = LLVMBuildAdd(builder, result, src, "");
3779		return result;
3780	}
3781
3782	ac_build_optimization_barrier(ctx, &src);
3783
3784	LLVMValueRef identity =
3785		get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
3786	result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
3787				  LLVMTypeOf(identity), "");
3788	result = ac_build_scan(ctx, op, result, identity, 64);
3789
3790	return ac_build_wwm(ctx, result);
3791}
3792
3793LLVMValueRef
3794ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
3795{
3796	LLVMValueRef result;
3797
3798	if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
3799		LLVMBuilderRef builder = ctx->builder;
3800		src = LLVMBuildZExt(builder, src, ctx->i32, "");
3801		result = ac_build_ballot(ctx, src);
3802		result = ac_build_mbcnt(ctx, result);
3803		return result;
3804	}
3805
3806	ac_build_optimization_barrier(ctx, &src);
3807
3808	LLVMValueRef identity =
3809		get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
3810	result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
3811				  LLVMTypeOf(identity), "");
3812	result = ac_build_dpp(ctx, identity, result, dpp_wf_sr1, 0xf, 0xf, false);
3813	result = ac_build_scan(ctx, op, result, identity, 64);
3814
3815	return ac_build_wwm(ctx, result);
3816}
3817
3818LLVMValueRef
3819ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size)
3820{
3821	if (cluster_size == 1) return src;
3822	ac_build_optimization_barrier(ctx, &src);
3823	LLVMValueRef result, swap;
3824	LLVMValueRef identity = get_reduction_identity(ctx, op,
3825								ac_get_type_size(LLVMTypeOf(src)));
3826	result = LLVMBuildBitCast(ctx->builder,
3827								ac_build_set_inactive(ctx, src, identity),
3828								LLVMTypeOf(identity), "");
3829	swap = ac_build_quad_swizzle(ctx, result, 1, 0, 3, 2);
3830	result = ac_build_alu_op(ctx, result, swap, op);
3831	if (cluster_size == 2) return ac_build_wwm(ctx, result);
3832
3833	swap = ac_build_quad_swizzle(ctx, result, 2, 3, 0, 1);
3834	result = ac_build_alu_op(ctx, result, swap, op);
3835	if (cluster_size == 4) return ac_build_wwm(ctx, result);
3836
3837	if (ctx->chip_class >= VI)
3838		swap = ac_build_dpp(ctx, identity, result, dpp_row_half_mirror, 0xf, 0xf, false);
3839	else
3840		swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x04));
3841	result = ac_build_alu_op(ctx, result, swap, op);
3842	if (cluster_size == 8) return ac_build_wwm(ctx, result);
3843
3844	if (ctx->chip_class >= VI)
3845		swap = ac_build_dpp(ctx, identity, result, dpp_row_mirror, 0xf, 0xf, false);
3846	else
3847		swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x08));
3848	result = ac_build_alu_op(ctx, result, swap, op);
3849	if (cluster_size == 16) return ac_build_wwm(ctx, result);
3850
3851	if (ctx->chip_class >= VI && cluster_size != 32)
3852		swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
3853	else
3854		swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x10));
3855	result = ac_build_alu_op(ctx, result, swap, op);
3856	if (cluster_size == 32) return ac_build_wwm(ctx, result);
3857
3858	if (ctx->chip_class >= VI) {
3859		swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
3860		result = ac_build_alu_op(ctx, result, swap, op);
3861		result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 63, 0));
3862		return ac_build_wwm(ctx, result);
3863	} else {
3864		swap = ac_build_readlane(ctx, result, ctx->i32_0);
3865		result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 32, 0));
3866		result = ac_build_alu_op(ctx, result, swap, op);
3867		return ac_build_wwm(ctx, result);
3868	}
3869}
3870
3871/**
3872 * "Top half" of a scan that reduces per-wave values across an entire
3873 * workgroup.
3874 *
3875 * The source value must be present in the highest lane of the wave, and the
3876 * highest lane must be live.
3877 */
3878void
3879ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3880{
3881	if (ws->maxwaves <= 1)
3882		return;
3883
3884	const LLVMValueRef i32_63 = LLVMConstInt(ctx->i32, 63, false);
3885	LLVMBuilderRef builder = ctx->builder;
3886	LLVMValueRef tid = ac_get_thread_id(ctx);
3887	LLVMValueRef tmp;
3888
3889	tmp = LLVMBuildICmp(builder, LLVMIntEQ, tid, i32_63, "");
3890	ac_build_ifcc(ctx, tmp, 1000);
3891	LLVMBuildStore(builder, ws->src, LLVMBuildGEP(builder, ws->scratch, &ws->waveidx, 1, ""));
3892	ac_build_endif(ctx, 1000);
3893}
3894
3895/**
3896 * "Bottom half" of a scan that reduces per-wave values across an entire
3897 * workgroup.
3898 *
3899 * The caller must place a barrier between the top and bottom halves.
3900 */
3901void
3902ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3903{
3904	const LLVMTypeRef type = LLVMTypeOf(ws->src);
3905	const LLVMValueRef identity =
3906		get_reduction_identity(ctx, ws->op, ac_get_type_size(type));
3907
3908	if (ws->maxwaves <= 1) {
3909		ws->result_reduce = ws->src;
3910		ws->result_inclusive = ws->src;
3911		ws->result_exclusive = identity;
3912		return;
3913	}
3914	assert(ws->maxwaves <= 32);
3915
3916	LLVMBuilderRef builder = ctx->builder;
3917	LLVMValueRef tid = ac_get_thread_id(ctx);
3918	LLVMBasicBlockRef bbs[2];
3919	LLVMValueRef phivalues_scan[2];
3920	LLVMValueRef tmp, tmp2;
3921
3922	bbs[0] = LLVMGetInsertBlock(builder);
3923	phivalues_scan[0] = LLVMGetUndef(type);
3924
3925	if (ws->enable_reduce)
3926		tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->numwaves, "");
3927	else if (ws->enable_inclusive)
3928		tmp = LLVMBuildICmp(builder, LLVMIntULE, tid, ws->waveidx, "");
3929	else
3930		tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->waveidx, "");
3931	ac_build_ifcc(ctx, tmp, 1001);
3932	{
3933		tmp = LLVMBuildLoad(builder, LLVMBuildGEP(builder, ws->scratch, &tid, 1, ""), "");
3934
3935		ac_build_optimization_barrier(ctx, &tmp);
3936
3937		bbs[1] = LLVMGetInsertBlock(builder);
3938		phivalues_scan[1] = ac_build_scan(ctx, ws->op, tmp, identity, ws->maxwaves);
3939	}
3940	ac_build_endif(ctx, 1001);
3941
3942	const LLVMValueRef scan = ac_build_phi(ctx, type, 2, phivalues_scan, bbs);
3943
3944	if (ws->enable_reduce) {
3945		tmp = LLVMBuildSub(builder, ws->numwaves, ctx->i32_1, "");
3946		ws->result_reduce = ac_build_readlane(ctx, scan, tmp);
3947	}
3948	if (ws->enable_inclusive)
3949		ws->result_inclusive = ac_build_readlane(ctx, scan, ws->waveidx);
3950	if (ws->enable_exclusive) {
3951		tmp = LLVMBuildSub(builder, ws->waveidx, ctx->i32_1, "");
3952		tmp = ac_build_readlane(ctx, scan, tmp);
3953		tmp2 = LLVMBuildICmp(builder, LLVMIntEQ, ws->waveidx, ctx->i32_0, "");
3954		ws->result_exclusive = LLVMBuildSelect(builder, tmp2, identity, tmp, "");
3955	}
3956}
3957
3958/**
3959 * Inclusive scan of a per-wave value across an entire workgroup.
3960 *
3961 * This implies an s_barrier instruction.
3962 *
3963 * Unlike ac_build_inclusive_scan, the caller \em must ensure that all threads
3964 * of the workgroup are live. (This requirement cannot easily be relaxed in a
3965 * useful manner because of the barrier in the algorithm.)
3966 */
3967void
3968ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3969{
3970	ac_build_wg_wavescan_top(ctx, ws);
3971	ac_build_s_barrier(ctx);
3972	ac_build_wg_wavescan_bottom(ctx, ws);
3973}
3974
3975/**
3976 * "Top half" of a scan that reduces per-thread values across an entire
3977 * workgroup.
3978 *
3979 * All lanes must be active when this code runs.
3980 */
3981void
3982ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3983{
3984	if (ws->enable_exclusive) {
3985		ws->extra = ac_build_exclusive_scan(ctx, ws->src, ws->op);
3986		if (LLVMTypeOf(ws->src) == ctx->i1 && ws->op == nir_op_iadd)
3987			ws->src = LLVMBuildZExt(ctx->builder, ws->src, ctx->i32, "");
3988		ws->src = ac_build_alu_op(ctx, ws->extra, ws->src, ws->op);
3989	} else {
3990		ws->src = ac_build_inclusive_scan(ctx, ws->src, ws->op);
3991	}
3992
3993	bool enable_inclusive = ws->enable_inclusive;
3994	bool enable_exclusive = ws->enable_exclusive;
3995	ws->enable_inclusive = false;
3996	ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
3997	ac_build_wg_wavescan_top(ctx, ws);
3998	ws->enable_inclusive = enable_inclusive;
3999	ws->enable_exclusive = enable_exclusive;
4000}
4001
4002/**
4003 * "Bottom half" of a scan that reduces per-thread values across an entire
4004 * workgroup.
4005 *
4006 * The caller must place a barrier between the top and bottom halves.
4007 */
4008void
4009ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4010{
4011	bool enable_inclusive = ws->enable_inclusive;
4012	bool enable_exclusive = ws->enable_exclusive;
4013	ws->enable_inclusive = false;
4014	ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
4015	ac_build_wg_wavescan_bottom(ctx, ws);
4016	ws->enable_inclusive = enable_inclusive;
4017	ws->enable_exclusive = enable_exclusive;
4018
4019	/* ws->result_reduce is already the correct value */
4020	if (ws->enable_inclusive)
4021		ws->result_inclusive = ac_build_alu_op(ctx, ws->result_inclusive, ws->src, ws->op);
4022	if (ws->enable_exclusive)
4023		ws->result_exclusive = ac_build_alu_op(ctx, ws->result_exclusive, ws->extra, ws->op);
4024}
4025
4026/**
4027 * A scan that reduces per-thread values across an entire workgroup.
4028 *
4029 * The caller must ensure that all lanes are active when this code runs
4030 * (WWM is insufficient!), because there is an implied barrier.
4031 */
4032void
4033ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4034{
4035	ac_build_wg_scan_top(ctx, ws);
4036	ac_build_s_barrier(ctx);
4037	ac_build_wg_scan_bottom(ctx, ws);
4038}
4039
4040LLVMValueRef
4041ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
4042		unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
4043{
4044	unsigned mask = dpp_quad_perm(lane0, lane1, lane2, lane3);
4045	if (ctx->chip_class >= VI) {
4046		return ac_build_dpp(ctx, src, src, mask, 0xf, 0xf, false);
4047	} else {
4048		return ac_build_ds_swizzle(ctx, src, (1 << 15) | mask);
4049	}
4050}
4051
4052LLVMValueRef
4053ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index)
4054{
4055	index = LLVMBuildMul(ctx->builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
4056	return ac_build_intrinsic(ctx,
4057		  "llvm.amdgcn.ds.bpermute", ctx->i32,
4058		  (LLVMValueRef []) {index, src}, 2,
4059		  AC_FUNC_ATTR_READNONE |
4060		  AC_FUNC_ATTR_CONVERGENT);
4061}
4062
4063LLVMValueRef
4064ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
4065		   unsigned bitsize)
4066{
4067	LLVMTypeRef type;
4068	char *intr;
4069
4070	if (bitsize == 16) {
4071		intr = "llvm.amdgcn.frexp.exp.i16.f16";
4072		type = ctx->i16;
4073	} else if (bitsize == 32) {
4074		intr = "llvm.amdgcn.frexp.exp.i32.f32";
4075		type = ctx->i32;
4076	} else {
4077		intr = "llvm.amdgcn.frexp.exp.i32.f64";
4078		type = ctx->i32;
4079	}
4080
4081	LLVMValueRef params[] = {
4082		src0,
4083	};
4084	return ac_build_intrinsic(ctx, intr, type, params, 1,
4085				  AC_FUNC_ATTR_READNONE);
4086}
4087LLVMValueRef
4088ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
4089		    unsigned bitsize)
4090{
4091	LLVMTypeRef type;
4092	char *intr;
4093
4094	if (bitsize == 16) {
4095		intr = "llvm.amdgcn.frexp.mant.f16";
4096		type = ctx->f16;
4097	} else if (bitsize == 32) {
4098		intr = "llvm.amdgcn.frexp.mant.f32";
4099		type = ctx->f32;
4100	} else {
4101		intr = "llvm.amdgcn.frexp.mant.f64";
4102		type = ctx->f64;
4103	}
4104
4105	LLVMValueRef params[] = {
4106		src0,
4107	};
4108	return ac_build_intrinsic(ctx, intr, type, params, 1,
4109				  AC_FUNC_ATTR_READNONE);
4110}
4111
4112/*
4113 * this takes an I,J coordinate pair,
4114 * and works out the X and Y derivatives.
4115 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
4116 */
4117LLVMValueRef
4118ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij)
4119{
4120	LLVMValueRef result[4], a;
4121	unsigned i;
4122
4123	for (i = 0; i < 2; i++) {
4124		a = LLVMBuildExtractElement(ctx->builder, interp_ij,
4125					    LLVMConstInt(ctx->i32, i, false), "");
4126		result[i] = ac_build_ddxy(ctx, AC_TID_MASK_TOP_LEFT, 1, a);
4127		result[2+i] = ac_build_ddxy(ctx, AC_TID_MASK_TOP_LEFT, 2, a);
4128	}
4129	return ac_build_gather_values(ctx, result, 4);
4130}
4131
4132LLVMValueRef
4133ac_build_load_helper_invocation(struct ac_llvm_context *ctx)
4134{
4135	LLVMValueRef result = ac_build_intrinsic(ctx, "llvm.amdgcn.ps.live",
4136						 ctx->i1, NULL, 0,
4137						 AC_FUNC_ATTR_READNONE);
4138	result = LLVMBuildNot(ctx->builder, result, "");
4139	return LLVMBuildSExt(ctx->builder, result, ctx->i32, "");
4140}
4141