1 Target Independent Opportunities:
3 //===---------------------------------------------------------------------===//
5 We should recognize idioms for add-with-carry and turn it into the appropriate
6 intrinsics. This example:
8 unsigned add32carry(unsigned sum, unsigned x) {
15 Compiles to: clang t.c -S -o - -O3 -fomit-frame-pointer -m64 -mkernel
17 _add32carry: ## @add32carry
28 leal (%rsi,%rdi), %eax
35 //===---------------------------------------------------------------------===//
37 Dead argument elimination should be enhanced to handle cases when an argument is
38 dead to an externally visible function. Though the argument can't be removed
39 from the externally visible function, the caller doesn't need to pass it in.
40 For example in this testcase:
42 void foo(int X) __attribute__((noinline));
43 void foo(int X) { sideeffect(); }
44 void bar(int A) { foo(A+1); }
48 define void @bar(i32 %A) nounwind ssp {
49 %0 = add nsw i32 %A, 1 ; <i32> [#uses=1]
50 tail call void @foo(i32 %0) nounwind noinline ssp
54 The add is dead, we could pass in 'i32 undef' instead. This occurs for C++
55 templates etc, which usually have linkonce_odr/weak_odr linkage, not internal
58 //===---------------------------------------------------------------------===//
60 With the recent changes to make the implicit def/use set explicit in
61 machineinstrs, we should change the target descriptions for 'call' instructions
62 so that the .td files don't list all the call-clobbered registers as implicit
63 defs. Instead, these should be added by the code generator (e.g. on the dag).
65 This has a number of uses:
67 1. PPC32/64 and X86 32/64 can avoid having multiple copies of call instructions
68 for their different impdef sets.
69 2. Targets with multiple calling convs (e.g. x86) which have different clobber
70 sets don't need copies of call instructions.
71 3. 'Interprocedural register allocation' can be done to reduce the clobber sets
74 //===---------------------------------------------------------------------===//
76 We should recognized various "overflow detection" idioms and translate them into
77 llvm.uadd.with.overflow and similar intrinsics. Here is a multiply idiom:
79 unsigned int mul(unsigned int a,unsigned int b) {
80 if ((unsigned long long)a*b>0xffffffff)
85 //===---------------------------------------------------------------------===//
87 Get the C front-end to expand hypot(x,y) -> llvm.sqrt(x*x+y*y) when errno and
88 precision don't matter (ffastmath). Misc/mandel will like this. :) This isn't
89 safe in general, even on darwin. See the libm implementation of hypot for
90 examples (which special case when x/y are exactly zero to get signed zeros etc
93 //===---------------------------------------------------------------------===//
95 Solve this DAG isel folding deficiency:
113 The problem is the store's chain operand is not the load X but rather
114 a TokenFactor of the load X and load Y, which prevents the folding.
116 There are two ways to fix this:
118 1. The dag combiner can start using alias analysis to realize that y/x
119 don't alias, making the store to X not dependent on the load from Y.
120 2. The generated isel could be made smarter in the case it can't
121 disambiguate the pointers.
123 Number 1 is the preferred solution.
125 This has been "fixed" by a TableGen hack. But that is a short term workaround
126 which will be removed once the proper fix is made.
128 //===---------------------------------------------------------------------===//
130 On targets with expensive 64-bit multiply, we could LSR this:
137 for (i = ...; ++i, tmp+=tmp)
140 This would be a win on ppc32, but not x86 or ppc64.
142 //===---------------------------------------------------------------------===//
144 Shrink: (setlt (loadi32 P), 0) -> (setlt (loadi8 Phi), 0)
146 //===---------------------------------------------------------------------===//
148 Reassociate should turn things like:
150 int factorial(int X) {
151 return X*X*X*X*X*X*X*X;
154 into llvm.powi calls, allowing the code generator to produce balanced
155 multiplication trees.
157 First, the intrinsic needs to be extended to support integers, and second the
158 code generator needs to be enhanced to lower these to multiplication trees.
160 //===---------------------------------------------------------------------===//
162 Interesting? testcase for add/shift/mul reassoc:
164 int bar(int x, int y) {
165 return x*x*x+y+x*x*x*x*x*y*y*y*y;
167 int foo(int z, int n) {
168 return bar(z, n) + bar(2*z, 2*n);
171 This is blocked on not handling X*X*X -> powi(X, 3) (see note above). The issue
172 is that we end up getting t = 2*X s = t*t and don't turn this into 4*X*X,
173 which is the same number of multiplies and is canonical, because the 2*X has
174 multiple uses. Here's a simple example:
176 define i32 @test15(i32 %X1) {
177 %B = mul i32 %X1, 47 ; X1*47
183 //===---------------------------------------------------------------------===//
185 Reassociate should handle the example in GCC PR16157:
187 extern int a0, a1, a2, a3, a4; extern int b0, b1, b2, b3, b4;
188 void f () { /* this can be optimized to four additions... */
189 b4 = a4 + a3 + a2 + a1 + a0;
190 b3 = a3 + a2 + a1 + a0;
195 This requires reassociating to forms of expressions that are already available,
196 something that reassoc doesn't think about yet.
199 //===---------------------------------------------------------------------===//
201 This function: (derived from GCC PR19988)
202 double foo(double x, double y) {
203 return ((x + 0.1234 * y) * (x + -0.1234 * y));
209 mulsd LCPI1_1(%rip), %xmm1
210 mulsd LCPI1_0(%rip), %xmm2
217 Reassociate should be able to turn it into:
219 double foo(double x, double y) {
220 return ((x + 0.1234 * y) * (x - 0.1234 * y));
223 Which allows the multiply by constant to be CSE'd, producing:
226 mulsd LCPI1_0(%rip), %xmm1
233 This doesn't need -ffast-math support at all. This is particularly bad because
234 the llvm-gcc frontend is canonicalizing the later into the former, but clang
235 doesn't have this problem.
237 //===---------------------------------------------------------------------===//
239 These two functions should generate the same code on big-endian systems:
241 int g(int *j,int *l) { return memcmp(j,l,4); }
242 int h(int *j, int *l) { return *j - *l; }
244 this could be done in SelectionDAGISel.cpp, along with other special cases,
247 //===---------------------------------------------------------------------===//
249 It would be nice to revert this patch:
250 http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20060213/031986.html
252 And teach the dag combiner enough to simplify the code expanded before
253 legalize. It seems plausible that this knowledge would let it simplify other
256 //===---------------------------------------------------------------------===//
258 For vector types, TargetData.cpp::getTypeInfo() returns alignment that is equal
259 to the type size. It works but can be overly conservative as the alignment of
260 specific vector types are target dependent.
262 //===---------------------------------------------------------------------===//
264 We should produce an unaligned load from code like this:
266 v4sf example(float *P) {
267 return (v4sf){P[0], P[1], P[2], P[3] };
270 //===---------------------------------------------------------------------===//
272 Add support for conditional increments, and other related patterns. Instead
277 je LBB16_2 #cond_next
288 //===---------------------------------------------------------------------===//
290 Combine: a = sin(x), b = cos(x) into a,b = sincos(x).
292 Expand these to calls of sin/cos and stores:
293 double sincos(double x, double *sin, double *cos);
294 float sincosf(float x, float *sin, float *cos);
295 long double sincosl(long double x, long double *sin, long double *cos);
297 Doing so could allow SROA of the destination pointers. See also:
298 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=17687
300 This is now easily doable with MRVs. We could even make an intrinsic for this
301 if anyone cared enough about sincos.
303 //===---------------------------------------------------------------------===//
305 quantum_sigma_x in 462.libquantum contains the following loop:
307 for(i=0; i<reg->size; i++)
309 /* Flip the target bit of each basis state */
310 reg->node[i].state ^= ((MAX_UNSIGNED) 1 << target);
313 Where MAX_UNSIGNED/state is a 64-bit int. On a 32-bit platform it would be just
314 so cool to turn it into something like:
316 long long Res = ((MAX_UNSIGNED) 1 << target);
318 for(i=0; i<reg->size; i++)
319 reg->node[i].state ^= Res & 0xFFFFFFFFULL;
321 for(i=0; i<reg->size; i++)
322 reg->node[i].state ^= Res & 0xFFFFFFFF00000000ULL
325 ... which would only do one 32-bit XOR per loop iteration instead of two.
327 It would also be nice to recognize the reg->size doesn't alias reg->node[i], but
330 //===---------------------------------------------------------------------===//
332 This isn't recognized as bswap by instcombine (yes, it really is bswap):
334 unsigned long reverse(unsigned v) {
336 t = v ^ ((v << 16) | (v >> 16));
338 v = (v << 24) | (v >> 8);
342 Neither is this (very standard idiom):
346 return (((n) << 24) | (((n) & 0xff00) << 8)
347 | (((n) >> 8) & 0xff00) | ((n) >> 24));
350 //===---------------------------------------------------------------------===//
354 These idioms should be recognized as popcount (see PR1488):
356 unsigned countbits_slow(unsigned v) {
358 for (c = 0; v; v >>= 1)
362 unsigned countbits_fast(unsigned v){
365 v &= v - 1; // clear the least significant bit set
369 BITBOARD = unsigned long long
370 int PopCnt(register BITBOARD a) {
378 unsigned int popcount(unsigned int input) {
379 unsigned int count = 0;
380 for (unsigned int i = 0; i < 4 * 8; i++)
381 count += (input >> i) & i;
385 This is a form of idiom recognition for loops, the same thing that could be
386 useful for recognizing memset/memcpy. This sort of thing should be added to the
389 //===---------------------------------------------------------------------===//
391 These should turn into single 16-bit (unaligned?) loads on little/big endian
394 unsigned short read_16_le(const unsigned char *adr) {
395 return adr[0] | (adr[1] << 8);
397 unsigned short read_16_be(const unsigned char *adr) {
398 return (adr[0] << 8) | adr[1];
401 //===---------------------------------------------------------------------===//
403 -instcombine should handle this transform:
404 icmp pred (sdiv X / C1 ), C2
405 when X, C1, and C2 are unsigned. Similarly for udiv and signed operands.
407 Currently InstCombine avoids this transform but will do it when the signs of
408 the operands and the sign of the divide match. See the FIXME in
409 InstructionCombining.cpp in the visitSetCondInst method after the switch case
410 for Instruction::UDiv (around line 4447) for more details.
412 The SingleSource/Benchmarks/Shootout-C++/hash and hash2 tests have examples of
415 //===---------------------------------------------------------------------===//
419 SingleSource/Benchmarks/Misc/dt.c shows several interesting optimization
420 opportunities in its double_array_divs_variable function: it needs loop
421 interchange, memory promotion (which LICM already does), vectorization and
422 variable trip count loop unrolling (since it has a constant trip count). ICC
423 apparently produces this very nice code with -ffast-math:
425 ..B1.70: # Preds ..B1.70 ..B1.69
426 mulpd %xmm0, %xmm1 #108.2
427 mulpd %xmm0, %xmm1 #108.2
428 mulpd %xmm0, %xmm1 #108.2
429 mulpd %xmm0, %xmm1 #108.2
431 cmpl $131072, %edx #108.2
432 jb ..B1.70 # Prob 99% #108.2
434 It would be better to count down to zero, but this is a lot better than what we
437 //===---------------------------------------------------------------------===//
441 typedef unsigned U32;
442 typedef unsigned long long U64;
443 int test (U32 *inst, U64 *regs) {
446 int r1 = (temp >> 20) & 0xf;
447 int b2 = (temp >> 16) & 0xf;
448 effective_addr2 = temp & 0xfff;
449 if (b2) effective_addr2 += regs[b2];
450 b2 = (temp >> 12) & 0xf;
451 if (b2) effective_addr2 += regs[b2];
452 effective_addr2 &= regs[4];
453 if ((effective_addr2 & 3) == 0)
458 Note that only the low 2 bits of effective_addr2 are used. On 32-bit systems,
459 we don't eliminate the computation of the top half of effective_addr2 because
460 we don't have whole-function selection dags. On x86, this means we use one
461 extra register for the function when effective_addr2 is declared as U64 than
462 when it is declared U32.
464 PHI Slicing could be extended to do this.
466 //===---------------------------------------------------------------------===//
468 LSR should know what GPR types a target has from TargetData. This code:
470 volatile short X, Y; // globals
474 for (i = 0; i < N; i++) { X = i; Y = i*4; }
477 produces two near identical IV's (after promotion) on PPC/ARM:
487 add r2, r2, #1 <- [0,+,1]
488 sub r0, r0, #1 <- [0,-,1]
492 LSR should reuse the "+" IV for the exit test.
494 //===---------------------------------------------------------------------===//
496 Tail call elim should be more aggressive, checking to see if the call is
497 followed by an uncond branch to an exit block.
499 ; This testcase is due to tail-duplication not wanting to copy the return
500 ; instruction into the terminating blocks because there was other code
501 ; optimized out of the function after the taildup happened.
502 ; RUN: llvm-as < %s | opt -tailcallelim | llvm-dis | not grep call
504 define i32 @t4(i32 %a) {
506 %tmp.1 = and i32 %a, 1 ; <i32> [#uses=1]
507 %tmp.2 = icmp ne i32 %tmp.1, 0 ; <i1> [#uses=1]
508 br i1 %tmp.2, label %then.0, label %else.0
510 then.0: ; preds = %entry
511 %tmp.5 = add i32 %a, -1 ; <i32> [#uses=1]
512 %tmp.3 = call i32 @t4( i32 %tmp.5 ) ; <i32> [#uses=1]
515 else.0: ; preds = %entry
516 %tmp.7 = icmp ne i32 %a, 0 ; <i1> [#uses=1]
517 br i1 %tmp.7, label %then.1, label %return
519 then.1: ; preds = %else.0
520 %tmp.11 = add i32 %a, -2 ; <i32> [#uses=1]
521 %tmp.9 = call i32 @t4( i32 %tmp.11 ) ; <i32> [#uses=1]
524 return: ; preds = %then.1, %else.0, %then.0
525 %result.0 = phi i32 [ 0, %else.0 ], [ %tmp.3, %then.0 ],
530 //===---------------------------------------------------------------------===//
532 Tail recursion elimination should handle:
537 return 2 * pow2m1 (n - 1) + 1;
540 Also, multiplies can be turned into SHL's, so they should be handled as if
541 they were associative. "return foo() << 1" can be tail recursion eliminated.
543 //===---------------------------------------------------------------------===//
545 Argument promotion should promote arguments for recursive functions, like
548 ; RUN: llvm-as < %s | opt -argpromotion | llvm-dis | grep x.val
550 define internal i32 @foo(i32* %x) {
552 %tmp = load i32* %x ; <i32> [#uses=0]
553 %tmp.foo = call i32 @foo( i32* %x ) ; <i32> [#uses=1]
557 define i32 @bar(i32* %x) {
559 %tmp3 = call i32 @foo( i32* %x ) ; <i32> [#uses=1]
563 //===---------------------------------------------------------------------===//
565 We should investigate an instruction sinking pass. Consider this silly
581 je LBB1_2 # cond_true
589 The PIC base computation (call+popl) is only used on one path through the
590 code, but is currently always computed in the entry block. It would be
591 better to sink the picbase computation down into the block for the
592 assertion, as it is the only one that uses it. This happens for a lot of
593 code with early outs.
595 Another example is loads of arguments, which are usually emitted into the
596 entry block on targets like x86. If not used in all paths through a
597 function, they should be sunk into the ones that do.
599 In this case, whole-function-isel would also handle this.
601 //===---------------------------------------------------------------------===//
603 Investigate lowering of sparse switch statements into perfect hash tables:
604 http://burtleburtle.net/bob/hash/perfect.html
606 //===---------------------------------------------------------------------===//
608 We should turn things like "load+fabs+store" and "load+fneg+store" into the
609 corresponding integer operations. On a yonah, this loop:
614 for (b = 0; b < 10000000; b++)
615 for (i = 0; i < 256; i++)
619 is twice as slow as this loop:
624 for (b = 0; b < 10000000; b++)
625 for (i = 0; i < 256; i++)
626 a[i] ^= (1ULL << 63);
629 and I suspect other processors are similar. On X86 in particular this is a
630 big win because doing this with integers allows the use of read/modify/write
633 //===---------------------------------------------------------------------===//
635 DAG Combiner should try to combine small loads into larger loads when
636 profitable. For example, we compile this C++ example:
638 struct THotKey { short Key; bool Control; bool Shift; bool Alt; };
639 extern THotKey m_HotKey;
640 THotKey GetHotKey () { return m_HotKey; }
642 into (-O3 -fno-exceptions -static -fomit-frame-pointer):
647 movb _m_HotKey+3, %cl
648 movb _m_HotKey+4, %dl
649 movb _m_HotKey+2, %ch
664 movzwl _m_HotKey+4, %edx
668 The LLVM IR contains the needed alignment info, so we should be able to
669 merge the loads and stores into 4-byte loads:
671 %struct.THotKey = type { i16, i8, i8, i8 }
672 define void @_Z9GetHotKeyv(%struct.THotKey* sret %agg.result) nounwind {
674 %tmp2 = load i16* getelementptr (@m_HotKey, i32 0, i32 0), align 8
675 %tmp5 = load i8* getelementptr (@m_HotKey, i32 0, i32 1), align 2
676 %tmp8 = load i8* getelementptr (@m_HotKey, i32 0, i32 2), align 1
677 %tmp11 = load i8* getelementptr (@m_HotKey, i32 0, i32 3), align 2
679 Alternatively, we should use a small amount of base-offset alias analysis
680 to make it so the scheduler doesn't need to hold all the loads in regs at
683 //===---------------------------------------------------------------------===//
685 We should add an FRINT node to the DAG to model targets that have legal
686 implementations of ceil/floor/rint.
688 //===---------------------------------------------------------------------===//
693 long long input[8] = {1,0,1,0,1,0,1,0};
697 Clang compiles this into:
699 call void @llvm.memset.p0i8.i64(i8* %tmp, i8 0, i64 64, i32 16, i1 false)
700 %0 = getelementptr [8 x i64]* %input, i64 0, i64 0
701 store i64 1, i64* %0, align 16
702 %1 = getelementptr [8 x i64]* %input, i64 0, i64 2
703 store i64 1, i64* %1, align 16
704 %2 = getelementptr [8 x i64]* %input, i64 0, i64 4
705 store i64 1, i64* %2, align 16
706 %3 = getelementptr [8 x i64]* %input, i64 0, i64 6
707 store i64 1, i64* %3, align 16
709 Which gets codegen'd into:
712 movaps %xmm0, -16(%rbp)
713 movaps %xmm0, -32(%rbp)
714 movaps %xmm0, -48(%rbp)
715 movaps %xmm0, -64(%rbp)
721 It would be better to have 4 movq's of 0 instead of the movaps's.
723 //===---------------------------------------------------------------------===//
725 http://llvm.org/PR717:
727 The following code should compile into "ret int undef". Instead, LLVM
728 produces "ret int 0":
737 //===---------------------------------------------------------------------===//
739 The loop unroller should partially unroll loops (instead of peeling them)
740 when code growth isn't too bad and when an unroll count allows simplification
741 of some code within the loop. One trivial example is:
747 for ( nLoop = 0; nLoop < 1000; nLoop++ ) {
756 Unrolling by 2 would eliminate the '&1' in both copies, leading to a net
757 reduction in code size. The resultant code would then also be suitable for
758 exit value computation.
760 //===---------------------------------------------------------------------===//
762 We miss a bunch of rotate opportunities on various targets, including ppc, x86,
763 etc. On X86, we miss a bunch of 'rotate by variable' cases because the rotate
764 matching code in dag combine doesn't look through truncates aggressively
765 enough. Here are some testcases reduces from GCC PR17886:
767 unsigned long long f(unsigned long long x, int y) {
768 return (x << y) | (x >> 64-y);
770 unsigned f2(unsigned x, int y){
771 return (x << y) | (x >> 32-y);
773 unsigned long long f3(unsigned long long x){
775 return (x << y) | (x >> 64-y);
777 unsigned f4(unsigned x){
779 return (x << y) | (x >> 32-y);
781 unsigned long long f5(unsigned long long x, unsigned long long y) {
782 return (x << 8) | ((y >> 48) & 0xffull);
784 unsigned long long f6(unsigned long long x, unsigned long long y, int z) {
787 return (x << 8) | ((y >> 48) & 0xffull);
789 return (x << 16) | ((y >> 40) & 0xffffull);
791 return (x << 24) | ((y >> 32) & 0xffffffull);
793 return (x << 32) | ((y >> 24) & 0xffffffffull);
795 return (x << 40) | ((y >> 16) & 0xffffffffffull);
799 On X86-64, we only handle f2/f3/f4 right. On x86-32, a few of these
800 generate truly horrible code, instead of using shld and friends. On
801 ARM, we end up with calls to L___lshrdi3/L___ashldi3 in f, which is
802 badness. PPC64 misses f, f5 and f6. CellSPU aborts in isel.
804 //===---------------------------------------------------------------------===//
806 This (and similar related idioms):
808 unsigned int foo(unsigned char i) {
809 return i | (i<<8) | (i<<16) | (i<<24);
814 define i32 @foo(i8 zeroext %i) nounwind readnone ssp noredzone {
816 %conv = zext i8 %i to i32
817 %shl = shl i32 %conv, 8
818 %shl5 = shl i32 %conv, 16
819 %shl9 = shl i32 %conv, 24
820 %or = or i32 %shl9, %conv
821 %or6 = or i32 %or, %shl5
822 %or10 = or i32 %or6, %shl
826 it would be better as:
828 unsigned int bar(unsigned char i) {
829 unsigned int j=i | (i << 8);
835 define i32 @bar(i8 zeroext %i) nounwind readnone ssp noredzone {
837 %conv = zext i8 %i to i32
838 %shl = shl i32 %conv, 8
839 %or = or i32 %shl, %conv
840 %shl5 = shl i32 %or, 16
841 %or6 = or i32 %shl5, %or
845 or even i*0x01010101, depending on the speed of the multiplier. The best way to
846 handle this is to canonicalize it to a multiply in IR and have codegen handle
847 lowering multiplies to shifts on cpus where shifts are faster.
849 //===---------------------------------------------------------------------===//
851 We do a number of simplifications in simplify libcalls to strength reduce
852 standard library functions, but we don't currently merge them together. For
853 example, it is useful to merge memcpy(a,b,strlen(b)) -> strcpy. This can only
854 be done safely if "b" isn't modified between the strlen and memcpy of course.
856 //===---------------------------------------------------------------------===//
858 We compile this program: (from GCC PR11680)
859 http://gcc.gnu.org/bugzilla/attachment.cgi?id=4487
861 Into code that runs the same speed in fast/slow modes, but both modes run 2x
862 slower than when compile with GCC (either 4.0 or 4.2):
864 $ llvm-g++ perf.cpp -O3 -fno-exceptions
866 1.821u 0.003s 0:01.82 100.0% 0+0k 0+0io 0pf+0w
868 $ g++ perf.cpp -O3 -fno-exceptions
870 0.821u 0.001s 0:00.82 100.0% 0+0k 0+0io 0pf+0w
872 It looks like we are making the same inlining decisions, so this may be raw
873 codegen badness or something else (haven't investigated).
875 //===---------------------------------------------------------------------===//
877 We miss some instcombines for stuff like this:
879 void foo (unsigned int a) {
880 /* This one is equivalent to a >= (3 << 2). */
885 A few other related ones are in GCC PR14753.
887 //===---------------------------------------------------------------------===//
889 Divisibility by constant can be simplified (according to GCC PR12849) from
890 being a mulhi to being a mul lo (cheaper). Testcase:
892 void bar(unsigned n) {
897 This is equivalent to the following, where 2863311531 is the multiplicative
898 inverse of 3, and 1431655766 is ((2^32)-1)/3+1:
899 void bar(unsigned n) {
900 if (n * 2863311531U < 1431655766U)
904 The same transformation can work with an even modulo with the addition of a
905 rotate: rotate the result of the multiply to the right by the number of bits
906 which need to be zero for the condition to be true, and shrink the compare RHS
907 by the same amount. Unless the target supports rotates, though, that
908 transformation probably isn't worthwhile.
910 The transformation can also easily be made to work with non-zero equality
911 comparisons: just transform, for example, "n % 3 == 1" to "(n-1) % 3 == 0".
913 //===---------------------------------------------------------------------===//
915 Better mod/ref analysis for scanf would allow us to eliminate the vtable and a
916 bunch of other stuff from this example (see PR1604):
926 std::scanf("%d", &t.val);
927 std::printf("%d\n", t.val);
930 //===---------------------------------------------------------------------===//
932 These functions perform the same computation, but produce different assembly.
934 define i8 @select(i8 %x) readnone nounwind {
935 %A = icmp ult i8 %x, 250
936 %B = select i1 %A, i8 0, i8 1
940 define i8 @addshr(i8 %x) readnone nounwind {
941 %A = zext i8 %x to i9
942 %B = add i9 %A, 6 ;; 256 - 250 == 6
944 %D = trunc i9 %C to i8
948 //===---------------------------------------------------------------------===//
952 f (unsigned long a, unsigned long b, unsigned long c)
954 return ((a & (c - 1)) != 0) || ((b & (c - 1)) != 0);
957 f (unsigned long a, unsigned long b, unsigned long c)
959 return ((a & (c - 1)) != 0) | ((b & (c - 1)) != 0);
961 Both should combine to ((a|b) & (c-1)) != 0. Currently not optimized with
962 "clang -emit-llvm-bc | opt -std-compile-opts".
964 //===---------------------------------------------------------------------===//
967 #define PMD_MASK (~((1UL << 23) - 1))
968 void clear_pmd_range(unsigned long start, unsigned long end)
970 if (!(start & ~PMD_MASK) && !(end & ~PMD_MASK))
973 The expression should optimize to something like
974 "!((start|end)&~PMD_MASK). Currently not optimized with "clang
975 -emit-llvm-bc | opt -std-compile-opts".
977 //===---------------------------------------------------------------------===//
979 unsigned int f(unsigned int i, unsigned int n) {++i; if (i == n) ++i; return
981 unsigned int f2(unsigned int i, unsigned int n) {++i; i += i == n; return i;}
982 These should combine to the same thing. Currently, the first function
983 produces better code on X86.
985 //===---------------------------------------------------------------------===//
988 #define abs(x) x>0?x:-x
991 return (abs(x)) >= 0;
993 This should optimize to x == INT_MIN. (With -fwrapv.) Currently not
994 optimized with "clang -emit-llvm-bc | opt -std-compile-opts".
996 //===---------------------------------------------------------------------===//
1000 rotate_cst (unsigned int a)
1002 a = (a << 10) | (a >> 22);
1007 minus_cst (unsigned int a)
1016 mask_gt (unsigned int a)
1018 /* This is equivalent to a > 15. */
1023 rshift_gt (unsigned int a)
1025 /* This is equivalent to a > 23. */
1029 All should simplify to a single comparison. All of these are
1030 currently not optimized with "clang -emit-llvm-bc | opt
1033 //===---------------------------------------------------------------------===//
1036 int c(int* x) {return (char*)x+2 == (char*)x;}
1037 Should combine to 0. Currently not optimized with "clang
1038 -emit-llvm-bc | opt -std-compile-opts" (although llc can optimize it).
1040 //===---------------------------------------------------------------------===//
1042 int a(unsigned b) {return ((b << 31) | (b << 30)) >> 31;}
1043 Should be combined to "((b >> 1) | b) & 1". Currently not optimized
1044 with "clang -emit-llvm-bc | opt -std-compile-opts".
1046 //===---------------------------------------------------------------------===//
1048 unsigned a(unsigned x, unsigned y) { return x | (y & 1) | (y & 2);}
1049 Should combine to "x | (y & 3)". Currently not optimized with "clang
1050 -emit-llvm-bc | opt -std-compile-opts".
1052 //===---------------------------------------------------------------------===//
1054 int a(int a, int b, int c) {return (~a & c) | ((c|a) & b);}
1055 Should fold to "(~a & c) | (a & b)". Currently not optimized with
1056 "clang -emit-llvm-bc | opt -std-compile-opts".
1058 //===---------------------------------------------------------------------===//
1060 int a(int a,int b) {return (~(a|b))|a;}
1061 Should fold to "a|~b". Currently not optimized with "clang
1062 -emit-llvm-bc | opt -std-compile-opts".
1064 //===---------------------------------------------------------------------===//
1066 int a(int a, int b) {return (a&&b) || (a&&!b);}
1067 Should fold to "a". Currently not optimized with "clang -emit-llvm-bc
1068 | opt -std-compile-opts".
1070 //===---------------------------------------------------------------------===//
1072 int a(int a, int b, int c) {return (a&&b) || (!a&&c);}
1073 Should fold to "a ? b : c", or at least something sane. Currently not
1074 optimized with "clang -emit-llvm-bc | opt -std-compile-opts".
1076 //===---------------------------------------------------------------------===//
1078 int a(int a, int b, int c) {return (a&&b) || (a&&c) || (a&&b&&c);}
1079 Should fold to a && (b || c). Currently not optimized with "clang
1080 -emit-llvm-bc | opt -std-compile-opts".
1082 //===---------------------------------------------------------------------===//
1084 int a(int x) {return x | ((x & 8) ^ 8);}
1085 Should combine to x | 8. Currently not optimized with "clang
1086 -emit-llvm-bc | opt -std-compile-opts".
1088 //===---------------------------------------------------------------------===//
1090 int a(int x) {return x ^ ((x & 8) ^ 8);}
1091 Should also combine to x | 8. Currently not optimized with "clang
1092 -emit-llvm-bc | opt -std-compile-opts".
1094 //===---------------------------------------------------------------------===//
1096 int a(int x) {return ((x | -9) ^ 8) & x;}
1097 Should combine to x & -9. Currently not optimized with "clang
1098 -emit-llvm-bc | opt -std-compile-opts".
1100 //===---------------------------------------------------------------------===//
1102 unsigned a(unsigned a) {return a * 0x11111111 >> 28 & 1;}
1103 Should combine to "a * 0x88888888 >> 31". Currently not optimized
1104 with "clang -emit-llvm-bc | opt -std-compile-opts".
1106 //===---------------------------------------------------------------------===//
1108 unsigned a(char* x) {if ((*x & 32) == 0) return b();}
1109 There's an unnecessary zext in the generated code with "clang
1110 -emit-llvm-bc | opt -std-compile-opts".
1112 //===---------------------------------------------------------------------===//
1114 unsigned a(unsigned long long x) {return 40 * (x >> 1);}
1115 Should combine to "20 * (((unsigned)x) & -2)". Currently not
1116 optimized with "clang -emit-llvm-bc | opt -std-compile-opts".
1118 //===---------------------------------------------------------------------===//
1120 This was noticed in the entryblock for grokdeclarator in 403.gcc:
1122 %tmp = icmp eq i32 %decl_context, 4
1123 %decl_context_addr.0 = select i1 %tmp, i32 3, i32 %decl_context
1124 %tmp1 = icmp eq i32 %decl_context_addr.0, 1
1125 %decl_context_addr.1 = select i1 %tmp1, i32 0, i32 %decl_context_addr.0
1127 tmp1 should be simplified to something like:
1128 (!tmp || decl_context == 1)
1130 This allows recursive simplifications, tmp1 is used all over the place in
1131 the function, e.g. by:
1133 %tmp23 = icmp eq i32 %decl_context_addr.1, 0 ; <i1> [#uses=1]
1134 %tmp24 = xor i1 %tmp1, true ; <i1> [#uses=1]
1135 %or.cond8 = and i1 %tmp23, %tmp24 ; <i1> [#uses=1]
1139 //===---------------------------------------------------------------------===//
1143 Store sinking: This code:
1145 void f (int n, int *cond, int *res) {
1148 for (i = 0; i < n; i++)
1150 *res ^= 234; /* (*) */
1153 On this function GVN hoists the fully redundant value of *res, but nothing
1154 moves the store out. This gives us this code:
1156 bb: ; preds = %bb2, %entry
1157 %.rle = phi i32 [ 0, %entry ], [ %.rle6, %bb2 ]
1158 %i.05 = phi i32 [ 0, %entry ], [ %indvar.next, %bb2 ]
1159 %1 = load i32* %cond, align 4
1160 %2 = icmp eq i32 %1, 0
1161 br i1 %2, label %bb2, label %bb1
1164 %3 = xor i32 %.rle, 234
1165 store i32 %3, i32* %res, align 4
1168 bb2: ; preds = %bb, %bb1
1169 %.rle6 = phi i32 [ %3, %bb1 ], [ %.rle, %bb ]
1170 %indvar.next = add i32 %i.05, 1
1171 %exitcond = icmp eq i32 %indvar.next, %n
1172 br i1 %exitcond, label %return, label %bb
1174 DSE should sink partially dead stores to get the store out of the loop.
1176 Here's another partial dead case:
1177 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=12395
1179 //===---------------------------------------------------------------------===//
1181 Scalar PRE hoists the mul in the common block up to the else:
1183 int test (int a, int b, int c, int g) {
1193 It would be better to do the mul once to reduce codesize above the if.
1194 This is GCC PR38204.
1196 //===---------------------------------------------------------------------===//
1200 GCC PR37810 is an interesting case where we should sink load/store reload
1201 into the if block and outside the loop, so we don't reload/store it on the
1222 We now hoist the reload after the call (Transforms/GVN/lpre-call-wrap.ll), but
1223 we don't sink the store. We need partially dead store sinking.
1225 //===---------------------------------------------------------------------===//
1227 [LOAD PRE CRIT EDGE SPLITTING]
1229 GCC PR37166: Sinking of loads prevents SROA'ing the "g" struct on the stack
1230 leading to excess stack traffic. This could be handled by GVN with some crazy
1231 symbolic phi translation. The code we get looks like (g is on the stack):
1235 %9 = getelementptr %struct.f* %g, i32 0, i32 0
1236 store i32 %8, i32* %9, align bel %bb3
1238 bb3: ; preds = %bb1, %bb2, %bb
1239 %c_addr.0 = phi %struct.f* [ %g, %bb2 ], [ %c, %bb ], [ %c, %bb1 ]
1240 %b_addr.0 = phi %struct.f* [ %b, %bb2 ], [ %g, %bb ], [ %b, %bb1 ]
1241 %10 = getelementptr %struct.f* %c_addr.0, i32 0, i32 0
1242 %11 = load i32* %10, align 4
1244 %11 is partially redundant, an in BB2 it should have the value %8.
1246 GCC PR33344 and PR35287 are similar cases.
1249 //===---------------------------------------------------------------------===//
1253 There are many load PRE testcases in testsuite/gcc.dg/tree-ssa/loadpre* in the
1254 GCC testsuite, ones we don't get yet are (checked through loadpre25):
1256 [CRIT EDGE BREAKING]
1257 loadpre3.c predcom-4.c
1259 [PRE OF READONLY CALL]
1262 [TURN SELECT INTO BRANCH]
1263 loadpre14.c loadpre15.c
1265 actually a conditional increment: loadpre18.c loadpre19.c
1267 //===---------------------------------------------------------------------===//
1269 [LOAD PRE / STORE SINKING / SPEC HACK]
1271 This is a chunk of code from 456.hmmer:
1273 int f(int M, int *mc, int *mpp, int *tpmm, int *ip, int *tpim, int *dpp,
1274 int *tpdm, int xmb, int *bp, int *ms) {
1276 for (k = 1; k <= M; k++) {
1277 mc[k] = mpp[k-1] + tpmm[k-1];
1278 if ((sc = ip[k-1] + tpim[k-1]) > mc[k]) mc[k] = sc;
1279 if ((sc = dpp[k-1] + tpdm[k-1]) > mc[k]) mc[k] = sc;
1280 if ((sc = xmb + bp[k]) > mc[k]) mc[k] = sc;
1285 It is very profitable for this benchmark to turn the conditional stores to mc[k]
1286 into a conditional move (select instr in IR) and allow the final store to do the
1287 store. See GCC PR27313 for more details. Note that this is valid to xform even
1288 with the new C++ memory model, since mc[k] is previously loaded and later
1291 //===---------------------------------------------------------------------===//
1294 There are many PRE testcases in testsuite/gcc.dg/tree-ssa/ssa-pre-*.c in the
1297 //===---------------------------------------------------------------------===//
1299 There are some interesting cases in testsuite/gcc.dg/tree-ssa/pred-comm* in the
1300 GCC testsuite. For example, we get the first example in predcom-1.c, but
1301 miss the second one:
1306 __attribute__ ((noinline))
1307 void count_averages(int n) {
1309 for (i = 1; i < n; i++)
1310 avg[i] = (((unsigned long) fib[i - 1] + fib[i] + fib[i + 1]) / 3) & 0xffff;
1313 which compiles into two loads instead of one in the loop.
1315 predcom-2.c is the same as predcom-1.c
1317 predcom-3.c is very similar but needs loads feeding each other instead of
1321 //===---------------------------------------------------------------------===//
1325 Type based alias analysis:
1326 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14705
1328 We should do better analysis of posix_memalign. At the least it should
1329 no-capture its pointer argument, at best, we should know that the out-value
1330 result doesn't point to anything (like malloc). One example of this is in
1331 SingleSource/Benchmarks/Misc/dt.c
1333 //===---------------------------------------------------------------------===//
1335 A/B get pinned to the stack because we turn an if/then into a select instead
1336 of PRE'ing the load/store. This may be fixable in instcombine:
1337 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37892
1339 struct X { int i; };
1353 //===---------------------------------------------------------------------===//
1355 Interesting missed case because of control flow flattening (should be 2 loads):
1356 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26629
1357 With: llvm-gcc t2.c -S -o - -O0 -emit-llvm | llvm-as |
1358 opt -mem2reg -gvn -instcombine | llvm-dis
1359 we miss it because we need 1) CRIT EDGE 2) MULTIPLE DIFFERENT
1360 VALS PRODUCED BY ONE BLOCK OVER DIFFERENT PATHS
1362 //===---------------------------------------------------------------------===//
1364 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19633
1365 We could eliminate the branch condition here, loading from null is undefined:
1367 struct S { int w, x, y, z; };
1368 struct T { int r; struct S s; };
1369 void bar (struct S, int);
1370 void foo (int a, struct T b)
1378 //===---------------------------------------------------------------------===//
1380 simplifylibcalls should do several optimizations for strspn/strcspn:
1382 strcspn(x, "a") -> inlined loop for up to 3 letters (similarly for strspn):
1384 size_t __strcspn_c3 (__const char *__s, int __reject1, int __reject2,
1386 register size_t __result = 0;
1387 while (__s[__result] != '\0' && __s[__result] != __reject1 &&
1388 __s[__result] != __reject2 && __s[__result] != __reject3)
1393 This should turn into a switch on the character. See PR3253 for some notes on
1396 456.hmmer apparently uses strcspn and strspn a lot. 471.omnetpp uses strspn.
1398 //===---------------------------------------------------------------------===//
1400 "gas" uses this idiom:
1401 else if (strchr ("+-/*%|&^:[]()~", *intel_parser.op_string))
1403 else if (strchr ("<>", *intel_parser.op_string)
1405 Those should be turned into a switch.
1407 //===---------------------------------------------------------------------===//
1409 252.eon contains this interesting code:
1411 %3072 = getelementptr [100 x i8]* %tempString, i32 0, i32 0
1412 %3073 = call i8* @strcpy(i8* %3072, i8* %3071) nounwind
1413 %strlen = call i32 @strlen(i8* %3072) ; uses = 1
1414 %endptr = getelementptr [100 x i8]* %tempString, i32 0, i32 %strlen
1415 call void @llvm.memcpy.i32(i8* %endptr,
1416 i8* getelementptr ([5 x i8]* @"\01LC42", i32 0, i32 0), i32 5, i32 1)
1417 %3074 = call i32 @strlen(i8* %endptr) nounwind readonly
1419 This is interesting for a couple reasons. First, in this:
1421 The memcpy+strlen strlen can be replaced with:
1423 %3074 = call i32 @strlen([5 x i8]* @"\01LC42") nounwind readonly
1425 Because the destination was just copied into the specified memory buffer. This,
1426 in turn, can be constant folded to "4".
1428 In other code, it contains:
1430 %endptr6978 = bitcast i8* %endptr69 to i32*
1431 store i32 7107374, i32* %endptr6978, align 1
1432 %3167 = call i32 @strlen(i8* %endptr69) nounwind readonly
1434 Which could also be constant folded. Whatever is producing this should probably
1435 be fixed to leave this as a memcpy from a string.
1437 Further, eon also has an interesting partially redundant strlen call:
1439 bb8: ; preds = %_ZN18eonImageCalculatorC1Ev.exit
1440 %682 = getelementptr i8** %argv, i32 6 ; <i8**> [#uses=2]
1441 %683 = load i8** %682, align 4 ; <i8*> [#uses=4]
1442 %684 = load i8* %683, align 1 ; <i8> [#uses=1]
1443 %685 = icmp eq i8 %684, 0 ; <i1> [#uses=1]
1444 br i1 %685, label %bb10, label %bb9
1447 %686 = call i32 @strlen(i8* %683) nounwind readonly
1448 %687 = icmp ugt i32 %686, 254 ; <i1> [#uses=1]
1449 br i1 %687, label %bb10, label %bb11
1451 bb10: ; preds = %bb9, %bb8
1452 %688 = call i32 @strlen(i8* %683) nounwind readonly
1454 This could be eliminated by doing the strlen once in bb8, saving code size and
1455 improving perf on the bb8->9->10 path.
1457 //===---------------------------------------------------------------------===//
1459 I see an interesting fully redundant call to strlen left in 186.crafty:InputMove
1461 %movetext11 = getelementptr [128 x i8]* %movetext, i32 0, i32 0
1464 bb62: ; preds = %bb55, %bb53
1465 %promote.0 = phi i32 [ %169, %bb55 ], [ 0, %bb53 ]
1466 %171 = call i32 @strlen(i8* %movetext11) nounwind readonly align 1
1467 %172 = add i32 %171, -1 ; <i32> [#uses=1]
1468 %173 = getelementptr [128 x i8]* %movetext, i32 0, i32 %172
1471 br i1 %or.cond, label %bb65, label %bb72
1473 bb65: ; preds = %bb62
1474 store i8 0, i8* %173, align 1
1477 bb72: ; preds = %bb65, %bb62
1478 %trank.1 = phi i32 [ %176, %bb65 ], [ -1, %bb62 ]
1479 %177 = call i32 @strlen(i8* %movetext11) nounwind readonly align 1
1481 Note that on the bb62->bb72 path, that the %177 strlen call is partially
1482 redundant with the %171 call. At worst, we could shove the %177 strlen call
1483 up into the bb65 block moving it out of the bb62->bb72 path. However, note
1484 that bb65 stores to the string, zeroing out the last byte. This means that on
1485 that path the value of %177 is actually just %171-1. A sub is cheaper than a
1488 This pattern repeats several times, basically doing:
1493 where it is "obvious" that B = A-1.
1495 //===---------------------------------------------------------------------===//
1497 186.crafty has this interesting pattern with the "out.4543" variable:
1499 call void @llvm.memcpy.i32(
1500 i8* getelementptr ([10 x i8]* @out.4543, i32 0, i32 0),
1501 i8* getelementptr ([7 x i8]* @"\01LC28700", i32 0, i32 0), i32 7, i32 1)
1502 %101 = call@printf(i8* ... @out.4543, i32 0, i32 0)) nounwind
1504 It is basically doing:
1506 memcpy(globalarray, "string");
1507 printf(..., globalarray);
1509 Anyway, by knowing that printf just reads the memory and forward substituting
1510 the string directly into the printf, this eliminates reads from globalarray.
1511 Since this pattern occurs frequently in crafty (due to the "DisplayTime" and
1512 other similar functions) there are many stores to "out". Once all the printfs
1513 stop using "out", all that is left is the memcpy's into it. This should allow
1514 globalopt to remove the "stored only" global.
1516 //===---------------------------------------------------------------------===//
1520 define inreg i32 @foo(i8* inreg %p) nounwind {
1522 %tmp1 = ashr i8 %tmp0, 5
1523 %tmp2 = sext i8 %tmp1 to i32
1527 could be dagcombine'd to a sign-extending load with a shift.
1528 For example, on x86 this currently gets this:
1534 while it could get this:
1539 //===---------------------------------------------------------------------===//
1543 int test(int x) { return 1-x == x; } // --> return false
1544 int test2(int x) { return 2-x == x; } // --> return x == 1 ?
1546 Always foldable for odd constants, what is the rule for even?
1548 //===---------------------------------------------------------------------===//
1550 PR 3381: GEP to field of size 0 inside a struct could be turned into GEP
1551 for next field in struct (which is at same address).
1553 For example: store of float into { {{}}, float } could be turned into a store to
1556 //===---------------------------------------------------------------------===//
1558 The arg promotion pass should make use of nocapture to make its alias analysis
1559 stuff much more precise.
1561 //===---------------------------------------------------------------------===//
1563 The following functions should be optimized to use a select instead of a
1564 branch (from gcc PR40072):
1566 char char_int(int m) {if(m>7) return 0; return m;}
1567 int int_char(char m) {if(m>7) return 0; return m;}
1569 //===---------------------------------------------------------------------===//
1571 int func(int a, int b) { if (a & 0x80) b |= 0x80; else b &= ~0x80; return b; }
1575 define i32 @func(i32 %a, i32 %b) nounwind readnone ssp {
1577 %0 = and i32 %a, 128 ; <i32> [#uses=1]
1578 %1 = icmp eq i32 %0, 0 ; <i1> [#uses=1]
1579 %2 = or i32 %b, 128 ; <i32> [#uses=1]
1580 %3 = and i32 %b, -129 ; <i32> [#uses=1]
1581 %b_addr.0 = select i1 %1, i32 %3, i32 %2 ; <i32> [#uses=1]
1585 However, it's functionally equivalent to:
1587 b = (b & ~0x80) | (a & 0x80);
1589 Which generates this:
1591 define i32 @func(i32 %a, i32 %b) nounwind readnone ssp {
1593 %0 = and i32 %b, -129 ; <i32> [#uses=1]
1594 %1 = and i32 %a, 128 ; <i32> [#uses=1]
1595 %2 = or i32 %0, %1 ; <i32> [#uses=1]
1599 This can be generalized for other forms:
1601 b = (b & ~0x80) | (a & 0x40) << 1;
1603 //===---------------------------------------------------------------------===//
1605 These two functions produce different code. They shouldn't:
1609 uint8_t p1(uint8_t b, uint8_t a) {
1610 b = (b & ~0xc0) | (a & 0xc0);
1614 uint8_t p2(uint8_t b, uint8_t a) {
1615 b = (b & ~0x40) | (a & 0x40);
1616 b = (b & ~0x80) | (a & 0x80);
1620 define zeroext i8 @p1(i8 zeroext %b, i8 zeroext %a) nounwind readnone ssp {
1622 %0 = and i8 %b, 63 ; <i8> [#uses=1]
1623 %1 = and i8 %a, -64 ; <i8> [#uses=1]
1624 %2 = or i8 %1, %0 ; <i8> [#uses=1]
1628 define zeroext i8 @p2(i8 zeroext %b, i8 zeroext %a) nounwind readnone ssp {
1630 %0 = and i8 %b, 63 ; <i8> [#uses=1]
1631 %.masked = and i8 %a, 64 ; <i8> [#uses=1]
1632 %1 = and i8 %a, -128 ; <i8> [#uses=1]
1633 %2 = or i8 %1, %0 ; <i8> [#uses=1]
1634 %3 = or i8 %2, %.masked ; <i8> [#uses=1]
1638 //===---------------------------------------------------------------------===//
1640 IPSCCP does not currently propagate argument dependent constants through
1641 functions where it does not not all of the callers. This includes functions
1642 with normal external linkage as well as templates, C99 inline functions etc.
1643 Specifically, it does nothing to:
1645 define i32 @test(i32 %x, i32 %y, i32 %z) nounwind {
1647 %0 = add nsw i32 %y, %z
1650 %3 = add nsw i32 %1, %2
1654 define i32 @test2() nounwind {
1656 %0 = call i32 @test(i32 1, i32 2, i32 4) nounwind
1660 It would be interesting extend IPSCCP to be able to handle simple cases like
1661 this, where all of the arguments to a call are constant. Because IPSCCP runs
1662 before inlining, trivial templates and inline functions are not yet inlined.
1663 The results for a function + set of constant arguments should be memoized in a
1666 //===---------------------------------------------------------------------===//
1668 The libcall constant folding stuff should be moved out of SimplifyLibcalls into
1669 libanalysis' constantfolding logic. This would allow IPSCCP to be able to
1670 handle simple things like this:
1672 static int foo(const char *X) { return strlen(X); }
1673 int bar() { return foo("abcd"); }
1675 //===---------------------------------------------------------------------===//
1677 InstCombine should use SimplifyDemandedBits to remove the or instruction:
1679 define i1 @test(i8 %x, i8 %y) {
1681 %B = icmp ugt i8 %A, 3
1685 Currently instcombine calls SimplifyDemandedBits with either all bits or just
1686 the sign bit, if the comparison is obviously a sign test. In this case, we only
1687 need all but the bottom two bits from %A, and if we gave that mask to SDB it
1688 would delete the or instruction for us.
1690 //===---------------------------------------------------------------------===//
1692 functionattrs doesn't know much about memcpy/memset. This function should be
1693 marked readnone rather than readonly, since it only twiddles local memory, but
1694 functionattrs doesn't handle memset/memcpy/memmove aggressively:
1696 struct X { int *p; int *q; };
1703 p = __builtin_memcpy (&x, &y, sizeof (int *));
1707 This can be seen at:
1708 $ clang t.c -S -o - -mkernel -O0 -emit-llvm | opt -functionattrs -S
1711 //===---------------------------------------------------------------------===//
1713 Missed instcombine transformation:
1714 define i1 @a(i32 %x) nounwind readnone {
1716 %cmp = icmp eq i32 %x, 30
1717 %sub = add i32 %x, -30
1718 %cmp2 = icmp ugt i32 %sub, 9
1719 %or = or i1 %cmp, %cmp2
1722 This should be optimized to a single compare. Testcase derived from gcc.
1724 //===---------------------------------------------------------------------===//
1726 Missed instcombine or reassociate transformation:
1727 int a(int a, int b) { return (a==12)&(b>47)&(b<58); }
1729 The sgt and slt should be combined into a single comparison. Testcase derived
1732 //===---------------------------------------------------------------------===//
1734 Missed instcombine transformation:
1736 %382 = srem i32 %tmp14.i, 64 ; [#uses=1]
1737 %383 = zext i32 %382 to i64 ; [#uses=1]
1738 %384 = shl i64 %381, %383 ; [#uses=1]
1739 %385 = icmp slt i32 %tmp14.i, 64 ; [#uses=1]
1741 The srem can be transformed to an and because if %tmp14.i is negative, the
1742 shift is undefined. Testcase derived from 403.gcc.
1744 //===---------------------------------------------------------------------===//
1746 This is a range comparison on a divided result (from 403.gcc):
1748 %1337 = sdiv i32 %1336, 8 ; [#uses=1]
1749 %.off.i208 = add i32 %1336, 7 ; [#uses=1]
1750 %1338 = icmp ult i32 %.off.i208, 15 ; [#uses=1]
1752 We already catch this (removing the sdiv) if there isn't an add, we should
1753 handle the 'add' as well. This is a common idiom with it's builtin_alloca code.
1756 int a(int x) { return (unsigned)(x/16+7) < 15; }
1758 Another similar case involves truncations on 64-bit targets:
1760 %361 = sdiv i64 %.046, 8 ; [#uses=1]
1761 %362 = trunc i64 %361 to i32 ; [#uses=2]
1763 %367 = icmp eq i32 %362, 0 ; [#uses=1]
1765 //===---------------------------------------------------------------------===//
1767 Missed instcombine/dagcombine transformation:
1768 define void @lshift_lt(i8 zeroext %a) nounwind {
1770 %conv = zext i8 %a to i32
1771 %shl = shl i32 %conv, 3
1772 %cmp = icmp ult i32 %shl, 33
1773 br i1 %cmp, label %if.then, label %if.end
1776 tail call void @bar() nounwind
1782 declare void @bar() nounwind
1784 The shift should be eliminated. Testcase derived from gcc.
1786 //===---------------------------------------------------------------------===//
1788 These compile into different code, one gets recognized as a switch and the
1789 other doesn't due to phase ordering issues (PR6212):
1791 int test1(int mainType, int subType) {
1794 else if (mainType == 9)
1796 else if (mainType == 11)
1801 int test2(int mainType, int subType) {
1811 //===---------------------------------------------------------------------===//
1813 The following test case (from PR6576):
1815 define i32 @mul(i32 %a, i32 %b) nounwind readnone {
1817 %cond1 = icmp eq i32 %b, 0 ; <i1> [#uses=1]
1818 br i1 %cond1, label %exit, label %bb.nph
1819 bb.nph: ; preds = %entry
1820 %tmp = mul i32 %b, %a ; <i32> [#uses=1]
1822 exit: ; preds = %entry
1826 could be reduced to:
1828 define i32 @mul(i32 %a, i32 %b) nounwind readnone {
1830 %tmp = mul i32 %b, %a
1834 //===---------------------------------------------------------------------===//
1836 We should use DSE + llvm.lifetime.end to delete dead vtable pointer updates.
1839 Another interesting case is that something related could be used for variables
1840 that go const after their ctor has finished. In these cases, globalopt (which
1841 can statically run the constructor) could mark the global const (so it gets put
1842 in the readonly section). A testcase would be:
1845 using namespace std;
1846 const complex<char> should_be_in_rodata (42,-42);
1847 complex<char> should_be_in_data (42,-42);
1848 complex<char> should_be_in_bss;
1850 Where we currently evaluate the ctors but the globals don't become const because
1851 the optimizer doesn't know they "become const" after the ctor is done. See
1852 GCC PR4131 for more examples.
1854 //===---------------------------------------------------------------------===//
1859 return x > 1 ? x : 1;
1862 LLVM emits a comparison with 1 instead of 0. 0 would be equivalent
1863 and cheaper on most targets.
1865 LLVM prefers comparisons with zero over non-zero in general, but in this
1866 case it choses instead to keep the max operation obvious.
1868 //===---------------------------------------------------------------------===//
1870 Take the following testcase on x86-64 (similar testcases exist for all targets
1873 define void @a(i64* nocapture %s, i64* nocapture %t, i64 %a, i64 %b,
1876 %0 = zext i64 %a to i128 ; <i128> [#uses=1]
1877 %1 = zext i64 %b to i128 ; <i128> [#uses=1]
1878 %2 = add i128 %1, %0 ; <i128> [#uses=2]
1879 %3 = zext i64 %c to i128 ; <i128> [#uses=1]
1880 %4 = shl i128 %3, 64 ; <i128> [#uses=1]
1881 %5 = add i128 %4, %2 ; <i128> [#uses=1]
1882 %6 = lshr i128 %5, 64 ; <i128> [#uses=1]
1883 %7 = trunc i128 %6 to i64 ; <i64> [#uses=1]
1884 store i64 %7, i64* %s, align 8
1885 %8 = trunc i128 %2 to i64 ; <i64> [#uses=1]
1886 store i64 %8, i64* %t, align 8
1906 The generated SelectionDAG has an ADD of an ADDE, where both operands of the
1907 ADDE are zero. Replacing one of the operands of the ADDE with the other operand
1908 of the ADD, and replacing the ADD with the ADDE, should give the desired result.
1910 (That said, we are doing a lot better than gcc on this testcase. :) )
1912 //===---------------------------------------------------------------------===//
1914 Switch lowering generates less than ideal code for the following switch:
1915 define void @a(i32 %x) nounwind {
1917 switch i32 %x, label %if.end [
1918 i32 0, label %if.then
1919 i32 1, label %if.then
1920 i32 2, label %if.then
1921 i32 3, label %if.then
1922 i32 5, label %if.then
1925 tail call void @foo() nounwind
1932 Generated code on x86-64 (other platforms give similar results):
1945 The movl+movl+btq+jb could be simplified to a cmpl+jne.
1947 Or, if we wanted to be really clever, we could simplify the whole thing to
1948 something like the following, which eliminates a branch:
1955 //===---------------------------------------------------------------------===//
1956 Given a branch where the two target blocks are identical ("ret i32 %b" in
1957 both), simplifycfg will simplify them away. But not so for a switch statement:
1959 define i32 @f(i32 %a, i32 %b) nounwind readnone {
1961 switch i32 %a, label %bb3 [
1966 bb: ; preds = %entry, %entry
1969 bb3: ; preds = %entry
1972 //===---------------------------------------------------------------------===//
1974 clang -O3 fails to devirtualize this virtual inheritance case: (GCC PR45875)
1975 Looks related to PR3100
1979 virtual void foo ();
1981 struct c11 : c10, c1{
1984 struct c28 : virtual c11{
1993 //===---------------------------------------------------------------------===//
1997 int foo(int a) { return (a & (~15)) / 16; }
2001 define i32 @foo(i32 %a) nounwind readnone ssp {
2003 %and = and i32 %a, -16
2004 %div = sdiv i32 %and, 16
2008 but this code (X & -A)/A is X >> log2(A) when A is a power of 2, so this case
2009 should be instcombined into just "a >> 4".
2011 We do get this at the codegen level, so something knows about it, but
2012 instcombine should catch it earlier:
2020 //===---------------------------------------------------------------------===//
2022 This code (from GCC PR28685):
2024 int test(int a, int b) {
2034 define i32 @test(i32 %a, i32 %b) nounwind readnone ssp {
2036 %cmp = icmp slt i32 %a, %b
2037 br i1 %cmp, label %return, label %if.end
2039 if.end: ; preds = %entry
2040 %cmp5 = icmp eq i32 %a, %b
2041 %conv6 = zext i1 %cmp5 to i32
2044 return: ; preds = %entry
2050 define i32 @test__(i32 %a, i32 %b) nounwind readnone ssp {
2052 %0 = icmp sle i32 %a, %b
2053 %retval = zext i1 %0 to i32
2057 //===---------------------------------------------------------------------===//
2059 This compare could fold to false:
2061 define i1 @g(i32 a) nounwind readnone {
2062 %add = shl i32 %a, 1
2063 %mul = shl i32 %a, 1
2064 %cmp = icmp ugt i32 %add, %mul
2068 //===---------------------------------------------------------------------===//
2070 This code can be seen in viterbi:
2072 %64 = call noalias i8* @malloc(i64 %62) nounwind
2074 %67 = call i64 @llvm.objectsize.i64(i8* %64, i1 false) nounwind
2075 %68 = call i8* @__memset_chk(i8* %64, i32 0, i64 %62, i64 %67) nounwind
2077 llvm.objectsize.i64 should be taught about malloc/calloc, allowing it to
2078 fold to %62. This is a security win (overflows of malloc will get caught)
2079 and also a performance win by exposing more memsets to the optimizer.
2081 This occurs several times in viterbi.
2083 //===---------------------------------------------------------------------===//