math: array parameters can be NULL
[smatch.git] / sparse-llvm.c
blobc7a9fbb7eee161c8cebff6c7f3877def7b29cc63
1 /*
2 * Example usage:
3 * ./sparse-llvm hello.c | llc | as -o hello.o
4 */
6 #include <llvm-c/Core.h>
7 #include <llvm-c/BitWriter.h>
8 #include <llvm-c/Analysis.h>
9 #include <llvm-c/Target.h>
11 #include <stdbool.h>
12 #include <stdio.h>
13 #include <unistd.h>
14 #include <string.h>
15 #include <assert.h>
17 #include "symbol.h"
18 #include "expression.h"
19 #include "linearize.h"
20 #include "flow.h"
22 struct function {
23 LLVMBuilderRef builder;
24 LLVMValueRef fn;
25 LLVMModuleRef module;
28 static LLVMTypeRef symbol_type(struct symbol *sym);
30 static LLVMTypeRef func_return_type(struct symbol *sym)
32 return symbol_type(sym->ctype.base_type);
35 static LLVMTypeRef sym_func_type(struct symbol *sym)
37 int n_arg = symbol_list_size(sym->arguments);
38 LLVMTypeRef *arg_type = calloc(n_arg, sizeof(LLVMTypeRef));
39 LLVMTypeRef ret_type = func_return_type(sym);
40 struct symbol *arg;
41 int idx = 0;
43 FOR_EACH_PTR(sym->arguments, arg) {
44 struct symbol *arg_sym = arg->ctype.base_type;
46 arg_type[idx++] = symbol_type(arg_sym);
47 } END_FOR_EACH_PTR(arg);
49 return LLVMFunctionType(ret_type, arg_type, n_arg, sym->variadic);
52 static LLVMTypeRef sym_array_type(struct symbol *sym)
54 LLVMTypeRef elem_type;
55 struct symbol *base_type;
57 base_type = sym->ctype.base_type;
58 /* empty struct is undefined [6.7.2.1(8)] */
59 assert(base_type->bit_size > 0);
61 elem_type = symbol_type(base_type);
62 if (!elem_type)
63 return NULL;
65 return LLVMArrayType(elem_type, sym->bit_size / base_type->bit_size);
68 #define MAX_STRUCT_MEMBERS 64
70 static LLVMTypeRef sym_struct_type(struct symbol *sym)
72 LLVMTypeRef elem_types[MAX_STRUCT_MEMBERS];
73 struct symbol *member;
74 char buffer[256];
75 LLVMTypeRef ret;
76 unsigned nr = 0;
78 snprintf(buffer, sizeof(buffer), "struct.%s", sym->ident ? sym->ident->name : "anno");
79 ret = LLVMStructCreateNamed(LLVMGetGlobalContext(), buffer);
80 /* set ->aux to avoid recursion */
81 sym->aux = ret;
83 FOR_EACH_PTR(sym->symbol_list, member) {
84 LLVMTypeRef member_type;
86 assert(nr < MAX_STRUCT_MEMBERS);
88 member_type = symbol_type(member);
90 elem_types[nr++] = member_type;
91 } END_FOR_EACH_PTR(member);
93 LLVMStructSetBody(ret, elem_types, nr, 0 /* packed? */);
94 return ret;
97 static LLVMTypeRef sym_union_type(struct symbol *sym)
99 LLVMTypeRef elements;
100 unsigned union_size;
103 * There's no union support in the LLVM API so we treat unions as
104 * opaque structs. The downside is that we lose type information on the
105 * members but as LLVM doesn't care, neither do we.
107 union_size = sym->bit_size / 8;
109 elements = LLVMArrayType(LLVMInt8Type(), union_size);
111 return LLVMStructType(&elements, 1, 0 /* packed? */);
114 static LLVMTypeRef sym_ptr_type(struct symbol *sym)
116 LLVMTypeRef type;
118 /* 'void *' is treated like 'char *' */
119 if (is_void_type(sym->ctype.base_type))
120 type = LLVMInt8Type();
121 else
122 type = symbol_type(sym->ctype.base_type);
124 return LLVMPointerType(type, 0);
127 static LLVMTypeRef sym_basetype_type(struct symbol *sym)
129 LLVMTypeRef ret = NULL;
131 if (is_float_type(sym)) {
132 switch (sym->bit_size) {
133 case 32:
134 ret = LLVMFloatType();
135 break;
136 case 64:
137 ret = LLVMDoubleType();
138 break;
139 case 80:
140 ret = LLVMX86FP80Type();
141 break;
142 default:
143 die("invalid bit size %d for type %d", sym->bit_size, sym->type);
144 break;
146 } else {
147 switch (sym->bit_size) {
148 case -1:
149 ret = LLVMVoidType();
150 break;
151 case 1:
152 ret = LLVMInt1Type();
153 break;
154 case 8:
155 ret = LLVMInt8Type();
156 break;
157 case 16:
158 ret = LLVMInt16Type();
159 break;
160 case 32:
161 ret = LLVMInt32Type();
162 break;
163 case 64:
164 ret = LLVMInt64Type();
165 break;
166 default:
167 die("invalid bit size %d for type %d", sym->bit_size, sym->type);
168 break;
172 return ret;
175 static LLVMTypeRef symbol_type(struct symbol *sym)
177 LLVMTypeRef ret = NULL;
179 /* don't cache the result for SYM_NODE */
180 if (sym->type == SYM_NODE)
181 return symbol_type(sym->ctype.base_type);
183 if (sym->aux)
184 return sym->aux;
186 switch (sym->type) {
187 case SYM_BITFIELD:
188 ret = LLVMIntType(sym->bit_size);
189 break;
190 case SYM_RESTRICT:
191 case SYM_ENUM:
192 ret = symbol_type(sym->ctype.base_type);
193 break;
194 case SYM_BASETYPE:
195 ret = sym_basetype_type(sym);
196 break;
197 case SYM_PTR:
198 ret = sym_ptr_type(sym);
199 break;
200 case SYM_UNION:
201 ret = sym_union_type(sym);
202 break;
203 case SYM_STRUCT:
204 ret = sym_struct_type(sym);
205 break;
206 case SYM_ARRAY:
207 ret = sym_array_type(sym);
208 break;
209 case SYM_FN:
210 ret = sym_func_type(sym);
211 break;
212 default:
213 assert(0);
216 /* cache the result */
217 sym->aux = ret;
218 return ret;
221 static LLVMTypeRef insn_symbol_type(struct instruction *insn)
223 if (insn->type)
224 return symbol_type(insn->type);
226 switch (insn->size) {
227 case 8: return LLVMInt8Type();
228 case 16: return LLVMInt16Type();
229 case 32: return LLVMInt32Type();
230 case 64: return LLVMInt64Type();
232 default:
233 die("invalid bit size %d", insn->size);
234 break;
237 return NULL; /* not reached */
240 static LLVMLinkage data_linkage(struct symbol *sym)
242 if (sym->ctype.modifiers & MOD_STATIC)
243 return LLVMPrivateLinkage;
245 return LLVMExternalLinkage;
248 static LLVMLinkage function_linkage(struct symbol *sym)
250 if (sym->ctype.modifiers & MOD_STATIC)
251 return LLVMInternalLinkage;
253 return LLVMExternalLinkage;
256 #define MAX_PSEUDO_NAME 64
258 static const char *pseudo_name(pseudo_t pseudo, char *buf)
260 switch (pseudo->type) {
261 case PSEUDO_REG:
262 snprintf(buf, MAX_PSEUDO_NAME, "R%d.", pseudo->nr);
263 break;
264 case PSEUDO_PHI:
265 snprintf(buf, MAX_PSEUDO_NAME, "PHI%d.", pseudo->nr);
266 break;
267 case PSEUDO_SYM:
268 case PSEUDO_VAL:
269 case PSEUDO_ARG:
270 case PSEUDO_VOID:
271 buf[0] = '\0';
272 break;
273 case PSEUDO_UNDEF:
274 assert(0);
275 break;
276 default:
277 assert(0);
280 return buf;
283 static LLVMValueRef get_sym_value(LLVMModuleRef module, struct symbol *sym)
285 const char *name = show_ident(sym->ident);
286 LLVMTypeRef type = symbol_type(sym);
287 LLVMValueRef result = NULL;
288 struct expression *expr;
290 assert(sym->bb_target == NULL);
292 expr = sym->initializer;
293 if (expr && !sym->ident) {
294 switch (expr->type) {
295 case EXPR_STRING: {
296 const char *s = expr->string->data;
297 LLVMValueRef indices[] = { LLVMConstInt(LLVMInt64Type(), 0, 0), LLVMConstInt(LLVMInt64Type(), 0, 0) };
298 LLVMValueRef data;
300 data = LLVMAddGlobal(module, LLVMArrayType(LLVMInt8Type(), strlen(s) + 1), ".str");
301 LLVMSetLinkage(data, LLVMPrivateLinkage);
302 LLVMSetGlobalConstant(data, 1);
303 LLVMSetInitializer(data, LLVMConstString(strdup(s), strlen(s) + 1, true));
305 result = LLVMConstGEP(data, indices, ARRAY_SIZE(indices));
306 return result;
308 default:
309 break;
313 if (LLVMGetTypeKind(type) == LLVMFunctionTypeKind) {
314 result = LLVMGetNamedFunction(module, name);
315 if (!result)
316 result = LLVMAddFunction(module, name, type);
317 } else {
318 result = LLVMGetNamedGlobal(module, name);
319 if (!result)
320 result = LLVMAddGlobal(module, type, name);
323 return result;
326 static LLVMValueRef constant_value(unsigned long long val, LLVMTypeRef dtype)
328 LLVMValueRef result;
330 switch (LLVMGetTypeKind(dtype)) {
331 case LLVMPointerTypeKind:
332 if (val != 0) { // for example: ... = (void*) 0x123;
333 LLVMTypeRef itype = LLVMIntType(bits_in_pointer);
334 result = LLVMConstInt(itype, val, 1);
335 result = LLVMConstIntToPtr(result, dtype);
336 } else {
337 result = LLVMConstPointerNull(dtype);
339 break;
340 case LLVMIntegerTypeKind:
341 result = LLVMConstInt(dtype, val, 1);
342 break;
343 case LLVMArrayTypeKind:
344 case LLVMStructTypeKind:
345 if (val != 0)
346 return NULL;
347 result = LLVMConstNull(dtype);
348 break;
349 default:
350 return NULL;
352 return result;
355 static LLVMValueRef val_to_value(unsigned long long val, struct symbol *ctype)
357 LLVMValueRef result;
358 LLVMTypeRef dtype;
360 assert(ctype);
361 dtype = symbol_type(ctype);
362 result = constant_value(val, dtype);
363 if (result)
364 return result;
365 sparse_error(ctype->pos, "no value possible for %s", show_typename(ctype));
366 return LLVMGetUndef(symbol_type(ctype));
369 static LLVMValueRef pseudo_to_value(struct function *fn, struct symbol *ctype, pseudo_t pseudo)
371 LLVMValueRef result = NULL;
373 switch (pseudo->type) {
374 case PSEUDO_REG:
375 result = pseudo->priv;
376 break;
377 case PSEUDO_SYM:
378 result = get_sym_value(fn->module, pseudo->sym);
379 break;
380 case PSEUDO_VAL:
381 result = val_to_value(pseudo->value, ctype);
382 break;
383 case PSEUDO_ARG: {
384 result = LLVMGetParam(fn->fn, pseudo->nr - 1);
385 break;
387 case PSEUDO_PHI:
388 result = pseudo->priv;
389 break;
390 case PSEUDO_VOID:
391 result = NULL;
392 break;
393 case PSEUDO_UNDEF:
394 result = LLVMGetUndef(symbol_type(ctype));
395 break;
396 default:
397 assert(0);
400 return result;
403 static LLVMValueRef pseudo_to_rvalue(struct function *fn, struct symbol *ctype, pseudo_t pseudo)
405 LLVMValueRef val = pseudo_to_value(fn, ctype, pseudo);
406 LLVMTypeRef dtype = symbol_type(ctype);
407 char name[MAX_PSEUDO_NAME];
409 pseudo_name(pseudo, name);
410 return LLVMBuildBitCast(fn->builder, val, dtype, name);
413 static LLVMValueRef value_to_ivalue(struct function *fn, struct symbol *ctype, LLVMValueRef val)
415 const char *name = LLVMGetValueName(val);
416 LLVMTypeRef dtype = symbol_type(ctype);
418 if (LLVMGetTypeKind(LLVMTypeOf(val)) == LLVMPointerTypeKind) {
419 LLVMTypeRef dtype = LLVMIntType(bits_in_pointer);
420 val = LLVMBuildPtrToInt(fn->builder, val, dtype, name);
422 if (ctype && is_int_type(ctype)) {
423 val = LLVMBuildIntCast(fn->builder, val, dtype, name);
425 return val;
428 static LLVMValueRef value_to_pvalue(struct function *fn, struct symbol *ctype, LLVMValueRef val)
430 const char *name = LLVMGetValueName(val);
431 LLVMTypeRef dtype = symbol_type(ctype);
433 assert(is_ptr_type(ctype));
434 switch (LLVMGetTypeKind(LLVMTypeOf(val))) {
435 case LLVMIntegerTypeKind:
436 val = LLVMBuildIntToPtr(fn->builder, val, dtype, name);
437 break;
438 case LLVMPointerTypeKind:
439 val = LLVMBuildBitCast(fn->builder, val, dtype, name);
440 break;
441 default:
442 break;
444 return val;
447 static LLVMValueRef adjust_type(struct function *fn, struct symbol *ctype, LLVMValueRef val)
449 if (is_int_type(ctype))
450 return value_to_ivalue(fn, ctype, val);
451 if (is_ptr_type(ctype))
452 return value_to_pvalue(fn, ctype, val);
453 return val;
457 * Get the LLVMValue corresponding to the pseudo
458 * and force the type corresponding to ctype.
460 static LLVMValueRef get_operand(struct function *fn, struct symbol *ctype, pseudo_t pseudo)
462 LLVMValueRef target = pseudo_to_value(fn, ctype, pseudo);
463 return adjust_type(fn, ctype, target);
467 * Get the LLVMValue corresponding to the pseudo
468 * and force the type corresponding to ctype but
469 * map all pointers to intptr_t.
471 static LLVMValueRef get_ioperand(struct function *fn, struct symbol *ctype, pseudo_t pseudo)
473 LLVMValueRef target = pseudo_to_value(fn, ctype, pseudo);
474 return value_to_ivalue(fn, ctype, target);
477 static LLVMValueRef calc_gep(LLVMBuilderRef builder, LLVMValueRef base, LLVMValueRef off)
479 LLVMTypeRef type = LLVMTypeOf(base);
480 unsigned int as = LLVMGetPointerAddressSpace(type);
481 LLVMTypeRef bytep = LLVMPointerType(LLVMInt8Type(), as);
482 LLVMValueRef addr;
483 const char *name = LLVMGetValueName(off);
485 /* convert base to char* type */
486 base = LLVMBuildPointerCast(builder, base, bytep, name);
487 /* addr = base + off */
488 addr = LLVMBuildInBoundsGEP(builder, base, &off, 1, name);
489 /* convert back to the actual pointer type */
490 addr = LLVMBuildPointerCast(builder, addr, type, name);
491 return addr;
494 static LLVMRealPredicate translate_fop(int opcode)
496 static const LLVMRealPredicate trans_tbl[] = {
497 [OP_FCMP_ORD] = LLVMRealORD,
498 [OP_FCMP_OEQ] = LLVMRealOEQ,
499 [OP_FCMP_ONE] = LLVMRealONE,
500 [OP_FCMP_OLE] = LLVMRealOLE,
501 [OP_FCMP_OGE] = LLVMRealOGE,
502 [OP_FCMP_OLT] = LLVMRealOLT,
503 [OP_FCMP_OGT] = LLVMRealOGT,
504 [OP_FCMP_UEQ] = LLVMRealUEQ,
505 [OP_FCMP_UNE] = LLVMRealUNE,
506 [OP_FCMP_ULE] = LLVMRealULE,
507 [OP_FCMP_UGE] = LLVMRealUGE,
508 [OP_FCMP_ULT] = LLVMRealULT,
509 [OP_FCMP_UGT] = LLVMRealUGT,
510 [OP_FCMP_UNO] = LLVMRealUNO,
513 return trans_tbl[opcode];
516 static LLVMIntPredicate translate_op(int opcode)
518 static const LLVMIntPredicate trans_tbl[] = {
519 [OP_SET_EQ] = LLVMIntEQ,
520 [OP_SET_NE] = LLVMIntNE,
521 [OP_SET_LE] = LLVMIntSLE,
522 [OP_SET_GE] = LLVMIntSGE,
523 [OP_SET_LT] = LLVMIntSLT,
524 [OP_SET_GT] = LLVMIntSGT,
525 [OP_SET_B] = LLVMIntULT,
526 [OP_SET_A] = LLVMIntUGT,
527 [OP_SET_BE] = LLVMIntULE,
528 [OP_SET_AE] = LLVMIntUGE,
531 return trans_tbl[opcode];
534 static void output_op_binary(struct function *fn, struct instruction *insn)
536 LLVMValueRef lhs, rhs, target;
537 char target_name[64];
539 lhs = get_ioperand(fn, insn->type, insn->src1);
540 rhs = get_ioperand(fn, insn->type, insn->src2);
542 pseudo_name(insn->target, target_name);
544 switch (insn->opcode) {
545 /* Binary */
546 case OP_ADD:
547 target = LLVMBuildAdd(fn->builder, lhs, rhs, target_name);
548 break;
549 case OP_SUB:
550 target = LLVMBuildSub(fn->builder, lhs, rhs, target_name);
551 break;
552 case OP_MUL:
553 target = LLVMBuildMul(fn->builder, lhs, rhs, target_name);
554 break;
555 case OP_DIVU:
556 target = LLVMBuildUDiv(fn->builder, lhs, rhs, target_name);
557 break;
558 case OP_DIVS:
559 assert(!is_float_type(insn->type));
560 target = LLVMBuildSDiv(fn->builder, lhs, rhs, target_name);
561 break;
562 case OP_MODU:
563 assert(!is_float_type(insn->type));
564 target = LLVMBuildURem(fn->builder, lhs, rhs, target_name);
565 break;
566 case OP_MODS:
567 assert(!is_float_type(insn->type));
568 target = LLVMBuildSRem(fn->builder, lhs, rhs, target_name);
569 break;
570 case OP_SHL:
571 assert(!is_float_type(insn->type));
572 target = LLVMBuildShl(fn->builder, lhs, rhs, target_name);
573 break;
574 case OP_LSR:
575 assert(!is_float_type(insn->type));
576 target = LLVMBuildLShr(fn->builder, lhs, rhs, target_name);
577 break;
578 case OP_ASR:
579 assert(!is_float_type(insn->type));
580 target = LLVMBuildAShr(fn->builder, lhs, rhs, target_name);
581 break;
583 /* floating-point */
584 case OP_FADD:
585 target = LLVMBuildFAdd(fn->builder, lhs, rhs, target_name);
586 break;
587 case OP_FSUB:
588 target = LLVMBuildFSub(fn->builder, lhs, rhs, target_name);
589 break;
590 case OP_FMUL:
591 target = LLVMBuildFMul(fn->builder, lhs, rhs, target_name);
592 break;
593 case OP_FDIV:
594 target = LLVMBuildFDiv(fn->builder, lhs, rhs, target_name);
595 break;
597 /* Logical */
598 case OP_AND:
599 assert(!is_float_type(insn->type));
600 target = LLVMBuildAnd(fn->builder, lhs, rhs, target_name);
601 break;
602 case OP_OR:
603 assert(!is_float_type(insn->type));
604 target = LLVMBuildOr(fn->builder, lhs, rhs, target_name);
605 break;
606 case OP_XOR:
607 assert(!is_float_type(insn->type));
608 target = LLVMBuildXor(fn->builder, lhs, rhs, target_name);
609 break;
610 default:
611 assert(0);
612 break;
615 target = adjust_type(fn, insn->type, target);
616 insn->target->priv = target;
619 static void output_op_compare(struct function *fn, struct instruction *insn)
621 LLVMValueRef lhs, rhs, target;
622 char target_name[64];
624 lhs = pseudo_to_value(fn, NULL, insn->src1);
625 if (insn->src2->type == PSEUDO_VAL)
626 rhs = constant_value(insn->src2->value, LLVMTypeOf(lhs));
627 else
628 rhs = pseudo_to_value(fn, NULL, insn->src2);
629 if (!rhs)
630 rhs = LLVMGetUndef(symbol_type(insn->type));
632 pseudo_name(insn->target, target_name);
634 LLVMTypeRef dst_type = insn_symbol_type(insn);
636 switch (LLVMGetTypeKind(LLVMTypeOf(lhs))) {
637 case LLVMPointerTypeKind:
638 lhs = value_to_pvalue(fn, &ptr_ctype, lhs);
639 rhs = value_to_pvalue(fn, &ptr_ctype, rhs);
640 /* fall through */
642 case LLVMIntegerTypeKind: {
643 LLVMIntPredicate op = translate_op(insn->opcode);
645 if (LLVMGetTypeKind(LLVMTypeOf(rhs)) == LLVMPointerTypeKind) {
646 LLVMTypeRef ltype = LLVMTypeOf(lhs);
647 rhs = LLVMBuildPtrToInt(fn->builder, rhs, ltype, "");
649 target = LLVMBuildICmp(fn->builder, op, lhs, rhs, target_name);
650 break;
652 case LLVMHalfTypeKind:
653 case LLVMFloatTypeKind:
654 case LLVMDoubleTypeKind:
655 case LLVMX86_FP80TypeKind:
656 case LLVMFP128TypeKind:
657 case LLVMPPC_FP128TypeKind: {
658 LLVMRealPredicate op = translate_fop(insn->opcode);
660 target = LLVMBuildFCmp(fn->builder, op, lhs, rhs, target_name);
661 break;
663 default:
664 assert(0);
667 target = LLVMBuildZExt(fn->builder, target, dst_type, target_name);
669 insn->target->priv = target;
672 static void output_op_ret(struct function *fn, struct instruction *insn)
674 pseudo_t pseudo = insn->src;
676 if (pseudo && pseudo != VOID) {
677 LLVMValueRef result = get_operand(fn, insn->type, pseudo);
678 LLVMBuildRet(fn->builder, result);
679 } else
680 LLVMBuildRetVoid(fn->builder);
683 static LLVMValueRef calc_memop_addr(struct function *fn, struct instruction *insn)
685 LLVMTypeRef int_type, addr_type;
686 LLVMValueRef src, off, addr;
687 unsigned int as;
689 /* int type large enough to hold a pointer */
690 int_type = LLVMIntType(bits_in_pointer);
691 off = LLVMConstInt(int_type, insn->offset, 0);
693 /* convert src to the effective pointer type */
694 src = pseudo_to_value(fn, insn->type, insn->src);
695 as = LLVMGetPointerAddressSpace(LLVMTypeOf(src));
696 addr_type = LLVMPointerType(insn_symbol_type(insn), as);
697 src = LLVMBuildPointerCast(fn->builder, src, addr_type, LLVMGetValueName(src));
699 /* addr = src + off */
700 addr = calc_gep(fn->builder, src, off);
701 return addr;
705 static void output_op_load(struct function *fn, struct instruction *insn)
707 LLVMValueRef addr, target;
708 char name[MAX_PSEUDO_NAME];
710 addr = calc_memop_addr(fn, insn);
712 /* perform load */
713 pseudo_name(insn->target, name);
714 target = LLVMBuildLoad(fn->builder, addr, name);
716 insn->target->priv = target;
719 static void output_op_store(struct function *fn, struct instruction *insn)
721 LLVMValueRef addr, target_in;
723 addr = calc_memop_addr(fn, insn);
725 target_in = pseudo_to_rvalue(fn, insn->type, insn->target);
727 /* perform store */
728 LLVMBuildStore(fn->builder, target_in, addr);
731 static LLVMValueRef bool_value(struct function *fn, LLVMValueRef value)
733 if (LLVMTypeOf(value) != LLVMInt1Type())
734 value = LLVMBuildIsNotNull(fn->builder, value, LLVMGetValueName(value));
736 return value;
739 static void output_op_cbr(struct function *fn, struct instruction *br)
741 LLVMValueRef cond = bool_value(fn,
742 pseudo_to_value(fn, NULL, br->cond));
744 LLVMBuildCondBr(fn->builder, cond,
745 br->bb_true->priv,
746 br->bb_false->priv);
749 static void output_op_br(struct function *fn, struct instruction *br)
751 LLVMBuildBr(fn->builder, br->bb_true->priv);
754 static void output_op_sel(struct function *fn, struct instruction *insn)
756 LLVMValueRef target, src1, src2, src3;
757 char name[MAX_PSEUDO_NAME];
759 src1 = bool_value(fn, pseudo_to_value(fn, NULL, insn->src1));
760 src2 = get_operand(fn, insn->type, insn->src2);
761 src3 = get_operand(fn, insn->type, insn->src3);
763 pseudo_name(insn->target, name);
764 target = LLVMBuildSelect(fn->builder, src1, src2, src3, name);
766 insn->target->priv = adjust_type(fn, insn->type, target);
769 static void output_op_switch(struct function *fn, struct instruction *insn)
771 LLVMValueRef sw_val, target;
772 struct basic_block *def = NULL;
773 struct multijmp *jmp;
774 int n_jmp = 0;
776 FOR_EACH_PTR(insn->multijmp_list, jmp) {
777 if (jmp->begin <= jmp->end) {
778 n_jmp += (jmp->end - jmp->begin) + 1;
779 } else /* default case */
780 def = jmp->target;
781 } END_FOR_EACH_PTR(jmp);
783 sw_val = get_ioperand(fn, insn->type, insn->cond);
784 target = LLVMBuildSwitch(fn->builder, sw_val,
785 def ? def->priv : NULL, n_jmp);
787 FOR_EACH_PTR(insn->multijmp_list, jmp) {
788 long long val;
790 for (val = jmp->begin; val <= jmp->end; val++) {
791 LLVMValueRef Val = val_to_value(val, insn->type);
792 LLVMAddCase(target, Val, jmp->target->priv);
794 } END_FOR_EACH_PTR(jmp);
797 static void output_op_call(struct function *fn, struct instruction *insn)
799 LLVMValueRef target, func;
800 struct symbol *ctype;
801 int n_arg = 0, i;
802 struct pseudo *arg;
803 LLVMValueRef *args;
804 char name[64];
806 n_arg = pseudo_list_size(insn->arguments);
807 args = calloc(n_arg, sizeof(LLVMValueRef));
809 PREPARE_PTR_LIST(insn->fntypes, ctype);
810 if (insn->func->type == PSEUDO_REG || insn->func->type == PSEUDO_PHI)
811 func = get_operand(fn, ctype, insn->func);
812 else
813 func = pseudo_to_value(fn, ctype, insn->func);
814 i = 0;
815 FOR_EACH_PTR(insn->arguments, arg) {
816 NEXT_PTR_LIST(ctype);
817 args[i++] = pseudo_to_rvalue(fn, ctype, arg);
818 } END_FOR_EACH_PTR(arg);
819 FINISH_PTR_LIST(ctype);
821 pseudo_name(insn->target, name);
822 target = LLVMBuildCall(fn->builder, func, args, n_arg, name);
824 insn->target->priv = target;
827 static void output_op_phisrc(struct function *fn, struct instruction *insn)
829 LLVMValueRef v;
830 struct instruction *phi;
832 assert(insn->target->priv == NULL);
834 /* target = src */
835 v = get_operand(fn, insn->type, insn->phi_src);
837 FOR_EACH_PTR(insn->phi_users, phi) {
838 LLVMValueRef load, ptr;
840 assert(phi->opcode == OP_PHI);
841 /* phi must be load from alloca */
842 load = phi->target->priv;
843 assert(LLVMGetInstructionOpcode(load) == LLVMLoad);
844 ptr = LLVMGetOperand(load, 0);
845 /* store v to alloca */
846 LLVMBuildStore(fn->builder, v, ptr);
847 } END_FOR_EACH_PTR(phi);
850 static void output_op_phi(struct function *fn, struct instruction *insn)
852 LLVMValueRef load = insn->target->priv;
854 /* forward load */
855 assert(LLVMGetInstructionOpcode(load) == LLVMLoad);
856 /* forward load has no parent block */
857 assert(!LLVMGetInstructionParent(load));
858 /* finalize load in current block */
859 LLVMInsertIntoBuilder(fn->builder, load);
862 static void output_op_ptrcast(struct function *fn, struct instruction *insn)
864 LLVMValueRef src, target;
865 LLVMTypeRef dtype;
866 struct symbol *otype = insn->orig_type;
867 LLVMOpcode op;
868 char target_name[64];
870 src = get_operand(fn, otype, insn->src);
871 pseudo_name(insn->target, target_name);
873 dtype = symbol_type(insn->type);
874 switch (insn->opcode) {
875 case OP_UTPTR:
876 case OP_SEXT: // FIXME
877 assert(is_int_type(otype));
878 assert(is_ptr_type(insn->type));
879 op = LLVMIntToPtr;
880 break;
881 case OP_PTRTU:
882 assert(is_ptr_type(otype));
883 assert(is_int_type(insn->type));
884 op = LLVMPtrToInt;
885 break;
886 case OP_PTRCAST:
887 case OP_ZEXT: // FIXME
888 assert(is_ptr_type(otype));
889 assert(is_ptr_type(insn->type));
890 op = LLVMBitCast;
891 break;
892 default:
893 assert(0);
896 target = LLVMBuildCast(fn->builder, op, src, dtype, target_name);
897 insn->target->priv = target;
900 static void output_op_cast(struct function *fn, struct instruction *insn, LLVMOpcode op)
902 LLVMValueRef src, target;
903 LLVMTypeRef dtype;
904 struct symbol *otype = insn->orig_type;
905 char target_name[64];
907 if (is_ptr_type(insn->type)) // cast to void* is OP_CAST ...
908 return output_op_ptrcast(fn, insn);
910 assert(is_int_type(insn->type));
912 src = get_operand(fn, otype, insn->src);
913 pseudo_name(insn->target, target_name);
915 dtype = symbol_type(insn->type);
916 if (is_ptr_type(otype)) {
917 op = LLVMPtrToInt;
918 } else if (is_float_type(otype)) {
919 assert(op == LLVMFPToUI || op == LLVMFPToSI);
920 } else if (is_int_type(otype)) {
921 unsigned int width = otype->bit_size;
922 if (insn->size < width)
923 op = LLVMTrunc;
924 else if (insn->size == width)
925 op = LLVMBitCast;
926 } else {
927 assert(0);
930 target = LLVMBuildCast(fn->builder, op, src, dtype, target_name);
931 insn->target->priv = target;
934 static void output_op_fpcast(struct function *fn, struct instruction *insn)
936 LLVMTypeRef dtype = symbol_type(insn->type);
937 LLVMValueRef src, target;
938 struct symbol *otype = insn->orig_type;
939 char name[64];
941 assert(is_float_type(insn->type));
943 pseudo_name(insn->target, name);
944 src = get_operand(fn, otype, insn->src);
945 switch (insn->opcode) {
946 case OP_FCVTF:
947 target = LLVMBuildFPCast(fn->builder, src, dtype, name);
948 break;
949 case OP_SCVTF:
950 target = LLVMBuildSIToFP(fn->builder, src, dtype, name);
951 break;
952 case OP_UCVTF:
953 target = LLVMBuildUIToFP(fn->builder, src, dtype, name);
954 break;
955 default:
956 assert(0);
958 insn->target->priv = target;
961 static void output_op_setval(struct function *fn, struct instruction *insn)
963 struct expression *val = insn->val;
964 LLVMValueRef target;
966 switch (val->type) {
967 case EXPR_LABEL:
968 target = LLVMBlockAddress(fn->fn, val->symbol->bb_target->priv);
969 break;
970 default:
971 assert(0);
974 insn->target->priv = target;
977 static void output_op_setfval(struct function *fn, struct instruction *insn)
979 LLVMTypeRef dtype = symbol_type(insn->type);
980 LLVMValueRef target;
982 target = LLVMConstReal(dtype, insn->fvalue);
983 insn->target->priv = target;
986 static void output_insn(struct function *fn, struct instruction *insn)
988 switch (insn->opcode) {
989 case OP_RET:
990 output_op_ret(fn, insn);
991 break;
992 case OP_BR:
993 output_op_br(fn, insn);
994 break;
995 case OP_CBR:
996 output_op_cbr(fn, insn);
997 break;
998 case OP_SYMADDR:
999 assert(0);
1000 break;
1001 case OP_SETVAL:
1002 output_op_setval(fn, insn);
1003 break;
1004 case OP_SETFVAL:
1005 output_op_setfval(fn, insn);
1006 break;
1007 case OP_SWITCH:
1008 output_op_switch(fn, insn);
1009 break;
1010 case OP_COMPUTEDGOTO:
1011 assert(0);
1012 break;
1013 case OP_PHISOURCE:
1014 output_op_phisrc(fn, insn);
1015 break;
1016 case OP_PHI:
1017 output_op_phi(fn, insn);
1018 break;
1019 case OP_LOAD:
1020 output_op_load(fn, insn);
1021 break;
1022 case OP_STORE:
1023 output_op_store(fn, insn);
1024 break;
1025 case OP_INLINED_CALL:
1026 break;
1027 case OP_CALL:
1028 output_op_call(fn, insn);
1029 break;
1030 case OP_ZEXT:
1031 output_op_cast(fn, insn, LLVMZExt);
1032 break;
1033 case OP_SEXT:
1034 output_op_cast(fn, insn, LLVMSExt);
1035 break;
1036 case OP_TRUNC:
1037 output_op_cast(fn, insn, LLVMTrunc);
1038 break;
1039 case OP_FCVTU:
1040 output_op_cast(fn, insn, LLVMFPToUI);
1041 break;
1042 case OP_FCVTS:
1043 output_op_cast(fn, insn, LLVMFPToSI);
1044 break;
1045 case OP_UCVTF: case OP_SCVTF:
1046 case OP_FCVTF:
1047 output_op_fpcast(fn, insn);
1048 break;
1049 case OP_UTPTR:
1050 case OP_PTRTU:
1051 case OP_PTRCAST:
1052 output_op_ptrcast(fn, insn);
1053 break;
1054 case OP_BINARY ... OP_BINARY_END:
1055 output_op_binary(fn, insn);
1056 break;
1057 case OP_FPCMP ... OP_BINCMP_END:
1058 output_op_compare(fn, insn);
1059 break;
1060 case OP_SEL:
1061 output_op_sel(fn, insn);
1062 break;
1063 case OP_SLICE:
1064 assert(0);
1065 break;
1066 case OP_NOT: {
1067 LLVMValueRef src, target;
1068 char target_name[64];
1070 src = pseudo_to_value(fn, insn->type, insn->src);
1072 pseudo_name(insn->target, target_name);
1074 target = LLVMBuildNot(fn->builder, src, target_name);
1076 insn->target->priv = target;
1077 break;
1079 case OP_FNEG:
1080 case OP_NEG: {
1081 LLVMValueRef src, target;
1082 char target_name[64];
1084 src = pseudo_to_value(fn, insn->type, insn->src);
1086 pseudo_name(insn->target, target_name);
1088 if (insn->opcode == OP_FNEG)
1089 target = LLVMBuildFNeg(fn->builder, src, target_name);
1090 else
1091 target = LLVMBuildNeg(fn->builder, src, target_name);
1093 insn->target->priv = target;
1094 break;
1096 case OP_CONTEXT:
1097 assert(0);
1098 break;
1099 case OP_RANGE:
1100 assert(0);
1101 break;
1102 case OP_NOP:
1103 assert(0);
1104 break;
1105 case OP_DEATHNOTE:
1106 break;
1107 case OP_ASM:
1108 assert(0);
1109 break;
1110 case OP_COPY:
1111 assert(0);
1112 break;
1113 default:
1114 break;
1118 static void output_bb(struct function *fn, struct basic_block *bb)
1120 struct instruction *insn;
1122 FOR_EACH_PTR(bb->insns, insn) {
1123 if (!insn->bb)
1124 continue;
1126 output_insn(fn, insn);
1128 END_FOR_EACH_PTR(insn);
1131 #define MAX_ARGS 64
1133 static void output_fn(LLVMModuleRef module, struct entrypoint *ep)
1135 struct symbol *sym = ep->name;
1136 struct symbol *base_type = sym->ctype.base_type;
1137 struct function function = { .module = module };
1138 struct basic_block *bb;
1139 int nr_args = 0;
1140 int i;
1142 function.fn = get_sym_value(module, sym);
1143 LLVMSetFunctionCallConv(function.fn, LLVMCCallConv);
1144 LLVMSetLinkage(function.fn, function_linkage(sym));
1146 function.builder = LLVMCreateBuilder();
1148 /* give a name to each argument */
1149 nr_args = symbol_list_size(base_type->arguments);
1150 for (i = 0; i < nr_args; i++) {
1151 char name[MAX_PSEUDO_NAME];
1152 LLVMValueRef arg;
1154 arg = LLVMGetParam(function.fn, i);
1155 snprintf(name, sizeof(name), "ARG%d.", i+1);
1156 LLVMSetValueName(arg, name);
1159 /* create the BBs */
1160 FOR_EACH_PTR(ep->bbs, bb) {
1161 static int nr_bb;
1162 LLVMBasicBlockRef bbr;
1163 char bbname[32];
1164 struct instruction *insn;
1166 sprintf(bbname, "L%d", nr_bb++);
1167 bbr = LLVMAppendBasicBlock(function.fn, bbname);
1169 bb->priv = bbr;
1171 /* allocate alloca for each phi */
1172 FOR_EACH_PTR(bb->insns, insn) {
1173 LLVMBasicBlockRef entrybbr;
1174 LLVMTypeRef phi_type;
1175 LLVMValueRef ptr;
1177 if (!insn->bb || insn->opcode != OP_PHI)
1178 continue;
1179 /* insert alloca into entry block */
1180 entrybbr = LLVMGetEntryBasicBlock(function.fn);
1181 LLVMPositionBuilderAtEnd(function.builder, entrybbr);
1182 phi_type = insn_symbol_type(insn);
1183 ptr = LLVMBuildAlloca(function.builder, phi_type, "");
1184 /* emit forward load for phi */
1185 LLVMClearInsertionPosition(function.builder);
1186 insn->target->priv = LLVMBuildLoad(function.builder, ptr, "phi");
1187 } END_FOR_EACH_PTR(insn);
1189 END_FOR_EACH_PTR(bb);
1191 FOR_EACH_PTR(ep->bbs, bb) {
1192 LLVMPositionBuilderAtEnd(function.builder, bb->priv);
1194 output_bb(&function, bb);
1196 END_FOR_EACH_PTR(bb);
1199 static LLVMValueRef output_data(LLVMModuleRef module, struct symbol *sym)
1201 struct expression *initializer = sym->initializer;
1202 LLVMValueRef initial_value;
1203 LLVMValueRef data;
1204 const char *name;
1206 if (initializer) {
1207 switch (initializer->type) {
1208 case EXPR_VALUE:
1209 initial_value = LLVMConstInt(symbol_type(sym), initializer->value, 1);
1210 break;
1211 case EXPR_FVALUE:
1212 initial_value = LLVMConstReal(symbol_type(sym), initializer->fvalue);
1213 break;
1214 case EXPR_SYMBOL: {
1215 struct symbol *sym = initializer->symbol;
1217 initial_value = LLVMGetNamedGlobal(module, show_ident(sym->ident));
1218 if (!initial_value)
1219 initial_value = output_data(module, sym);
1220 break;
1222 case EXPR_STRING: {
1223 const char *s = initializer->string->data;
1225 initial_value = LLVMConstString(strdup(s), strlen(s) + 1, true);
1226 break;
1228 default:
1229 warning(initializer->pos, "can't initialize type: %s", show_typename(sym));
1230 initial_value = NULL;
1231 break;
1233 } else {
1234 LLVMTypeRef type = symbol_type(sym);
1236 initial_value = LLVMConstNull(type);
1239 if (!initial_value)
1240 return NULL;
1242 name = sym->ident ? show_ident(sym->ident) : "" ;
1244 data = LLVMAddGlobal(module, LLVMTypeOf(initial_value), name);
1246 LLVMSetLinkage(data, data_linkage(sym));
1247 if (sym->ctype.modifiers & MOD_CONST)
1248 LLVMSetGlobalConstant(data, 1);
1249 if (sym->ctype.modifiers & MOD_TLS)
1250 LLVMSetThreadLocal(data, 1);
1251 if (sym->ctype.alignment)
1252 LLVMSetAlignment(data, sym->ctype.alignment);
1254 if (!(sym->ctype.modifiers & MOD_EXTERN))
1255 LLVMSetInitializer(data, initial_value);
1257 return data;
1260 static int is_prototype(struct symbol *sym)
1262 if (sym->type == SYM_NODE)
1263 sym = sym->ctype.base_type;
1264 return sym && sym->type == SYM_FN && !sym->stmt;
1267 static int compile(LLVMModuleRef module, struct symbol_list *list)
1269 struct symbol *sym;
1271 FOR_EACH_PTR(list, sym) {
1272 struct entrypoint *ep;
1273 expand_symbol(sym);
1275 if (is_prototype(sym)) {
1276 // this will do the LLVMAddFunction() we want
1277 get_sym_value(module, sym);
1278 continue;
1281 ep = linearize_symbol(sym);
1282 if (ep)
1283 output_fn(module, ep);
1284 else
1285 output_data(module, sym);
1287 END_FOR_EACH_PTR(sym);
1289 return 0;
1292 #ifndef LLVM_DEFAULT_TARGET_TRIPLE
1293 #define LLVM_DEFAULT_TARGET_TRIPLE LLVM_HOSTTRIPLE
1294 #endif
1296 #define X86_LINUX_LAYOUT \
1297 "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-" \
1298 "i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-" \
1299 "a0:0:64-f80:32:32-n8:16:32-S128"
1301 #define X86_64_LINUX_LAYOUT \
1302 "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-" \
1303 "i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-" \
1304 "a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
1306 static void set_target(LLVMModuleRef module)
1308 char target[] = LLVM_DEFAULT_TARGET_TRIPLE;
1309 const char *arch, *vendor, *os, *env, *layout = NULL;
1310 char triple[256];
1312 arch = strtok(target, "-");
1313 vendor = strtok(NULL, "-");
1314 os = strtok(NULL, "-");
1315 env = strtok(NULL, "-");
1317 if (!os)
1318 return;
1319 if (!env)
1320 env = "unknown";
1322 if (!strcmp(arch, "x86_64") && !strcmp(os, "linux")) {
1323 if (arch_m64) {
1324 layout = X86_64_LINUX_LAYOUT;
1325 } else {
1326 arch = "i386";
1327 layout = X86_LINUX_LAYOUT;
1331 /* unsupported target */
1332 if (!layout)
1333 return;
1335 snprintf(triple, sizeof(triple), "%s-%s-%s-%s", arch, vendor, os, env);
1336 LLVMSetTarget(module, triple);
1337 LLVMSetDataLayout(module, layout);
1340 int main(int argc, char **argv)
1342 struct string_list *filelist = NULL;
1343 struct symbol_list *symlist;
1344 LLVMModuleRef module;
1345 char *file;
1347 symlist = sparse_initialize(argc, argv, &filelist);
1349 module = LLVMModuleCreateWithName("sparse");
1350 set_target(module);
1352 compile(module, symlist);
1354 /* need ->phi_users */
1355 dbg_dead = 1;
1356 FOR_EACH_PTR(filelist, file) {
1357 symlist = sparse(file);
1358 if (die_if_error)
1359 return 1;
1360 compile(module, symlist);
1361 } END_FOR_EACH_PTR(file);
1363 LLVMVerifyModule(module, LLVMPrintMessageAction, NULL);
1365 LLVMWriteBitcodeToFD(module, STDOUT_FILENO, 0, 0);
1367 LLVMDisposeModule(module);
1369 report_stats();
1370 return 0;