1 Target Independent Opportunities:
3 //===---------------------------------------------------------------------===//
5 With the recent changes to make the implicit def/use set explicit in
6 machineinstrs, we should change the target descriptions for 'call' instructions
7 so that the .td files don't list all the call-clobbered registers as implicit
8 defs. Instead, these should be added by the code generator (e.g. on the dag).
10 This has a number of uses:
12 1. PPC32/64 and X86 32/64 can avoid having multiple copies of call instructions
13 for their different impdef sets.
14 2. Targets with multiple calling convs (e.g. x86) which have different clobber
15 sets don't need copies of call instructions.
16 3. 'Interprocedural register allocation' can be done to reduce the clobber sets
19 //===---------------------------------------------------------------------===//
21 Make the PPC branch selector target independant
23 //===---------------------------------------------------------------------===//
25 Get the C front-end to expand hypot(x,y) -> llvm.sqrt(x*x+y*y) when errno and
26 precision don't matter (ffastmath). Misc/mandel will like this. :) This isn't
27 safe in general, even on darwin. See the libm implementation of hypot for
28 examples (which special case when x/y are exactly zero to get signed zeros etc
31 //===---------------------------------------------------------------------===//
33 Solve this DAG isel folding deficiency:
51 The problem is the store's chain operand is not the load X but rather
52 a TokenFactor of the load X and load Y, which prevents the folding.
54 There are two ways to fix this:
56 1. The dag combiner can start using alias analysis to realize that y/x
57 don't alias, making the store to X not dependent on the load from Y.
58 2. The generated isel could be made smarter in the case it can't
59 disambiguate the pointers.
61 Number 1 is the preferred solution.
63 This has been "fixed" by a TableGen hack. But that is a short term workaround
64 which will be removed once the proper fix is made.
66 //===---------------------------------------------------------------------===//
68 On targets with expensive 64-bit multiply, we could LSR this:
75 for (i = ...; ++i, tmp+=tmp)
78 This would be a win on ppc32, but not x86 or ppc64.
80 //===---------------------------------------------------------------------===//
82 Shrink: (setlt (loadi32 P), 0) -> (setlt (loadi8 Phi), 0)
84 //===---------------------------------------------------------------------===//
86 Reassociate should turn: X*X*X*X -> t=(X*X) (t*t) to eliminate a multiply.
88 //===---------------------------------------------------------------------===//
90 Interesting? testcase for add/shift/mul reassoc:
92 int bar(int x, int y) {
93 return x*x*x+y+x*x*x*x*x*y*y*y*y;
95 int foo(int z, int n) {
96 return bar(z, n) + bar(2*z, 2*n);
99 Reassociate should handle the example in GCC PR16157.
101 //===---------------------------------------------------------------------===//
103 These two functions should generate the same code on big-endian systems:
105 int g(int *j,int *l) { return memcmp(j,l,4); }
106 int h(int *j, int *l) { return *j - *l; }
108 this could be done in SelectionDAGISel.cpp, along with other special cases,
111 //===---------------------------------------------------------------------===//
113 It would be nice to revert this patch:
114 http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20060213/031986.html
116 And teach the dag combiner enough to simplify the code expanded before
117 legalize. It seems plausible that this knowledge would let it simplify other
120 //===---------------------------------------------------------------------===//
122 For vector types, TargetData.cpp::getTypeInfo() returns alignment that is equal
123 to the type size. It works but can be overly conservative as the alignment of
124 specific vector types are target dependent.
126 //===---------------------------------------------------------------------===//
128 We should produce an unaligned load from code like this:
130 v4sf example(float *P) {
131 return (v4sf){P[0], P[1], P[2], P[3] };
134 //===---------------------------------------------------------------------===//
136 Add support for conditional increments, and other related patterns. Instead
141 je LBB16_2 #cond_next
152 //===---------------------------------------------------------------------===//
154 Combine: a = sin(x), b = cos(x) into a,b = sincos(x).
156 Expand these to calls of sin/cos and stores:
157 double sincos(double x, double *sin, double *cos);
158 float sincosf(float x, float *sin, float *cos);
159 long double sincosl(long double x, long double *sin, long double *cos);
161 Doing so could allow SROA of the destination pointers. See also:
162 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=17687
164 This is now easily doable with MRVs. We could even make an intrinsic for this
165 if anyone cared enough about sincos.
167 //===---------------------------------------------------------------------===//
169 Turn this into a single byte store with no load (the other 3 bytes are
172 define void @test(i32* %P) {
174 %tmp14 = or i32 %tmp, 3305111552
175 %tmp15 = and i32 %tmp14, 3321888767
176 store i32 %tmp15, i32* %P
180 //===---------------------------------------------------------------------===//
182 dag/inst combine "clz(x)>>5 -> x==0" for 32-bit x.
188 int t = __builtin_clz(x);
198 //===---------------------------------------------------------------------===//
200 quantum_sigma_x in 462.libquantum contains the following loop:
202 for(i=0; i<reg->size; i++)
204 /* Flip the target bit of each basis state */
205 reg->node[i].state ^= ((MAX_UNSIGNED) 1 << target);
208 Where MAX_UNSIGNED/state is a 64-bit int. On a 32-bit platform it would be just
209 so cool to turn it into something like:
211 long long Res = ((MAX_UNSIGNED) 1 << target);
213 for(i=0; i<reg->size; i++)
214 reg->node[i].state ^= Res & 0xFFFFFFFFULL;
216 for(i=0; i<reg->size; i++)
217 reg->node[i].state ^= Res & 0xFFFFFFFF00000000ULL
220 ... which would only do one 32-bit XOR per loop iteration instead of two.
222 It would also be nice to recognize the reg->size doesn't alias reg->node[i], but
225 //===---------------------------------------------------------------------===//
227 This should be optimized to one 'and' and one 'or', from PR4216:
229 define i32 @test_bitfield(i32 %bf.prev.low) nounwind ssp {
231 %bf.prev.lo.cleared10 = or i32 %bf.prev.low, 32962 ; <i32> [#uses=1]
232 %0 = and i32 %bf.prev.low, -65536 ; <i32> [#uses=1]
233 %1 = and i32 %bf.prev.lo.cleared10, 40186 ; <i32> [#uses=1]
234 %2 = or i32 %1, %0 ; <i32> [#uses=1]
238 //===---------------------------------------------------------------------===//
240 This isn't recognized as bswap by instcombine (yes, it really is bswap):
242 unsigned long reverse(unsigned v) {
244 t = v ^ ((v << 16) | (v >> 16));
246 v = (v << 24) | (v >> 8);
250 //===---------------------------------------------------------------------===//
252 These idioms should be recognized as popcount (see PR1488):
254 unsigned countbits_slow(unsigned v) {
256 for (c = 0; v; v >>= 1)
260 unsigned countbits_fast(unsigned v){
263 v &= v - 1; // clear the least significant bit set
267 BITBOARD = unsigned long long
268 int PopCnt(register BITBOARD a) {
276 unsigned int popcount(unsigned int input) {
277 unsigned int count = 0;
278 for (unsigned int i = 0; i < 4 * 8; i++)
279 count += (input >> i) & i;
283 //===---------------------------------------------------------------------===//
285 These should turn into single 16-bit (unaligned?) loads on little/big endian
288 unsigned short read_16_le(const unsigned char *adr) {
289 return adr[0] | (adr[1] << 8);
291 unsigned short read_16_be(const unsigned char *adr) {
292 return (adr[0] << 8) | adr[1];
295 //===---------------------------------------------------------------------===//
297 -instcombine should handle this transform:
298 icmp pred (sdiv X / C1 ), C2
299 when X, C1, and C2 are unsigned. Similarly for udiv and signed operands.
301 Currently InstCombine avoids this transform but will do it when the signs of
302 the operands and the sign of the divide match. See the FIXME in
303 InstructionCombining.cpp in the visitSetCondInst method after the switch case
304 for Instruction::UDiv (around line 4447) for more details.
306 The SingleSource/Benchmarks/Shootout-C++/hash and hash2 tests have examples of
309 //===---------------------------------------------------------------------===//
311 viterbi speeds up *significantly* if the various "history" related copy loops
312 are turned into memcpy calls at the source level. We need a "loops to memcpy"
315 //===---------------------------------------------------------------------===//
319 typedef unsigned U32;
320 typedef unsigned long long U64;
321 int test (U32 *inst, U64 *regs) {
324 int r1 = (temp >> 20) & 0xf;
325 int b2 = (temp >> 16) & 0xf;
326 effective_addr2 = temp & 0xfff;
327 if (b2) effective_addr2 += regs[b2];
328 b2 = (temp >> 12) & 0xf;
329 if (b2) effective_addr2 += regs[b2];
330 effective_addr2 &= regs[4];
331 if ((effective_addr2 & 3) == 0)
336 Note that only the low 2 bits of effective_addr2 are used. On 32-bit systems,
337 we don't eliminate the computation of the top half of effective_addr2 because
338 we don't have whole-function selection dags. On x86, this means we use one
339 extra register for the function when effective_addr2 is declared as U64 than
340 when it is declared U32.
342 //===---------------------------------------------------------------------===//
344 LSR should know what GPR types a target has. This code:
346 volatile short X, Y; // globals
350 for (i = 0; i < N; i++) { X = i; Y = i*4; }
353 produces two near identical IV's (after promotion) on PPC/ARM:
363 add r2, r2, #1 <- [0,+,1]
364 sub r0, r0, #1 <- [0,-,1]
368 LSR should reuse the "+" IV for the exit test.
371 //===---------------------------------------------------------------------===//
373 Tail call elim should be more aggressive, checking to see if the call is
374 followed by an uncond branch to an exit block.
376 ; This testcase is due to tail-duplication not wanting to copy the return
377 ; instruction into the terminating blocks because there was other code
378 ; optimized out of the function after the taildup happened.
379 ; RUN: llvm-as < %s | opt -tailcallelim | llvm-dis | not grep call
381 define i32 @t4(i32 %a) {
383 %tmp.1 = and i32 %a, 1 ; <i32> [#uses=1]
384 %tmp.2 = icmp ne i32 %tmp.1, 0 ; <i1> [#uses=1]
385 br i1 %tmp.2, label %then.0, label %else.0
387 then.0: ; preds = %entry
388 %tmp.5 = add i32 %a, -1 ; <i32> [#uses=1]
389 %tmp.3 = call i32 @t4( i32 %tmp.5 ) ; <i32> [#uses=1]
392 else.0: ; preds = %entry
393 %tmp.7 = icmp ne i32 %a, 0 ; <i1> [#uses=1]
394 br i1 %tmp.7, label %then.1, label %return
396 then.1: ; preds = %else.0
397 %tmp.11 = add i32 %a, -2 ; <i32> [#uses=1]
398 %tmp.9 = call i32 @t4( i32 %tmp.11 ) ; <i32> [#uses=1]
401 return: ; preds = %then.1, %else.0, %then.0
402 %result.0 = phi i32 [ 0, %else.0 ], [ %tmp.3, %then.0 ],
407 //===---------------------------------------------------------------------===//
409 Tail recursion elimination is not transforming this function, because it is
410 returning n, which fails the isDynamicConstant check in the accumulator
413 long long fib(const long long n) {
419 return fib(n-1) + fib(n-2);
423 //===---------------------------------------------------------------------===//
425 Tail recursion elimination should handle:
430 return 2 * pow2m1 (n - 1) + 1;
433 Also, multiplies can be turned into SHL's, so they should be handled as if
434 they were associative. "return foo() << 1" can be tail recursion eliminated.
436 //===---------------------------------------------------------------------===//
438 Argument promotion should promote arguments for recursive functions, like
441 ; RUN: llvm-as < %s | opt -argpromotion | llvm-dis | grep x.val
443 define internal i32 @foo(i32* %x) {
445 %tmp = load i32* %x ; <i32> [#uses=0]
446 %tmp.foo = call i32 @foo( i32* %x ) ; <i32> [#uses=1]
450 define i32 @bar(i32* %x) {
452 %tmp3 = call i32 @foo( i32* %x ) ; <i32> [#uses=1]
456 //===---------------------------------------------------------------------===//
458 "basicaa" should know how to look through "or" instructions that act like add
459 instructions. For example in this code, the x*4+1 is turned into x*4 | 1, and
460 basicaa can't analyze the array subscript, leading to duplicated loads in the
463 void test(int X, int Y, int a[]) {
465 for (i=2; i<1000; i+=4) {
466 a[i+0] = a[i-1+0]*a[i-2+0];
467 a[i+1] = a[i-1+1]*a[i-2+1];
468 a[i+2] = a[i-1+2]*a[i-2+2];
469 a[i+3] = a[i-1+3]*a[i-2+3];
473 BasicAA also doesn't do this for add. It needs to know that &A[i+1] != &A[i].
475 //===---------------------------------------------------------------------===//
477 We should investigate an instruction sinking pass. Consider this silly
493 je LBB1_2 # cond_true
501 The PIC base computation (call+popl) is only used on one path through the
502 code, but is currently always computed in the entry block. It would be
503 better to sink the picbase computation down into the block for the
504 assertion, as it is the only one that uses it. This happens for a lot of
505 code with early outs.
507 Another example is loads of arguments, which are usually emitted into the
508 entry block on targets like x86. If not used in all paths through a
509 function, they should be sunk into the ones that do.
511 In this case, whole-function-isel would also handle this.
513 //===---------------------------------------------------------------------===//
515 Investigate lowering of sparse switch statements into perfect hash tables:
516 http://burtleburtle.net/bob/hash/perfect.html
518 //===---------------------------------------------------------------------===//
520 We should turn things like "load+fabs+store" and "load+fneg+store" into the
521 corresponding integer operations. On a yonah, this loop:
526 for (b = 0; b < 10000000; b++)
527 for (i = 0; i < 256; i++)
531 is twice as slow as this loop:
536 for (b = 0; b < 10000000; b++)
537 for (i = 0; i < 256; i++)
538 a[i] ^= (1ULL << 63);
541 and I suspect other processors are similar. On X86 in particular this is a
542 big win because doing this with integers allows the use of read/modify/write
545 //===---------------------------------------------------------------------===//
547 DAG Combiner should try to combine small loads into larger loads when
548 profitable. For example, we compile this C++ example:
550 struct THotKey { short Key; bool Control; bool Shift; bool Alt; };
551 extern THotKey m_HotKey;
552 THotKey GetHotKey () { return m_HotKey; }
554 into (-O3 -fno-exceptions -static -fomit-frame-pointer):
559 movb _m_HotKey+3, %cl
560 movb _m_HotKey+4, %dl
561 movb _m_HotKey+2, %ch
576 movzwl _m_HotKey+4, %edx
580 The LLVM IR contains the needed alignment info, so we should be able to
581 merge the loads and stores into 4-byte loads:
583 %struct.THotKey = type { i16, i8, i8, i8 }
584 define void @_Z9GetHotKeyv(%struct.THotKey* sret %agg.result) nounwind {
586 %tmp2 = load i16* getelementptr (@m_HotKey, i32 0, i32 0), align 8
587 %tmp5 = load i8* getelementptr (@m_HotKey, i32 0, i32 1), align 2
588 %tmp8 = load i8* getelementptr (@m_HotKey, i32 0, i32 2), align 1
589 %tmp11 = load i8* getelementptr (@m_HotKey, i32 0, i32 3), align 2
591 Alternatively, we should use a small amount of base-offset alias analysis
592 to make it so the scheduler doesn't need to hold all the loads in regs at
595 //===---------------------------------------------------------------------===//
597 We should add an FRINT node to the DAG to model targets that have legal
598 implementations of ceil/floor/rint.
600 //===---------------------------------------------------------------------===//
605 long long input[8] = {1,1,1,1,1,1,1,1};
609 We currently compile this into a memcpy from a global array since the
610 initializer is fairly large and not memset'able. This is good, but the memcpy
611 gets lowered to load/stores in the code generator. This is also ok, except
612 that the codegen lowering for memcpy doesn't handle the case when the source
613 is a constant global. This gives us atrocious code like this:
618 movl _C.0.1444-"L1$pb"+32(%eax), %ecx
620 movl _C.0.1444-"L1$pb"+20(%eax), %ecx
622 movl _C.0.1444-"L1$pb"+36(%eax), %ecx
624 movl _C.0.1444-"L1$pb"+44(%eax), %ecx
626 movl _C.0.1444-"L1$pb"+40(%eax), %ecx
628 movl _C.0.1444-"L1$pb"+12(%eax), %ecx
630 movl _C.0.1444-"L1$pb"+4(%eax), %ecx
642 //===---------------------------------------------------------------------===//
644 http://llvm.org/PR717:
646 The following code should compile into "ret int undef". Instead, LLVM
647 produces "ret int 0":
656 //===---------------------------------------------------------------------===//
658 The loop unroller should partially unroll loops (instead of peeling them)
659 when code growth isn't too bad and when an unroll count allows simplification
660 of some code within the loop. One trivial example is:
666 for ( nLoop = 0; nLoop < 1000; nLoop++ ) {
675 Unrolling by 2 would eliminate the '&1' in both copies, leading to a net
676 reduction in code size. The resultant code would then also be suitable for
677 exit value computation.
679 //===---------------------------------------------------------------------===//
681 We miss a bunch of rotate opportunities on various targets, including ppc, x86,
682 etc. On X86, we miss a bunch of 'rotate by variable' cases because the rotate
683 matching code in dag combine doesn't look through truncates aggressively
684 enough. Here are some testcases reduces from GCC PR17886:
686 unsigned long long f(unsigned long long x, int y) {
687 return (x << y) | (x >> 64-y);
689 unsigned f2(unsigned x, int y){
690 return (x << y) | (x >> 32-y);
692 unsigned long long f3(unsigned long long x){
694 return (x << y) | (x >> 64-y);
696 unsigned f4(unsigned x){
698 return (x << y) | (x >> 32-y);
700 unsigned long long f5(unsigned long long x, unsigned long long y) {
701 return (x << 8) | ((y >> 48) & 0xffull);
703 unsigned long long f6(unsigned long long x, unsigned long long y, int z) {
706 return (x << 8) | ((y >> 48) & 0xffull);
708 return (x << 16) | ((y >> 40) & 0xffffull);
710 return (x << 24) | ((y >> 32) & 0xffffffull);
712 return (x << 32) | ((y >> 24) & 0xffffffffull);
714 return (x << 40) | ((y >> 16) & 0xffffffffffull);
718 On X86-64, we only handle f2/f3/f4 right. On x86-32, a few of these
719 generate truly horrible code, instead of using shld and friends. On
720 ARM, we end up with calls to L___lshrdi3/L___ashldi3 in f, which is
721 badness. PPC64 misses f, f5 and f6. CellSPU aborts in isel.
723 //===---------------------------------------------------------------------===//
725 We do a number of simplifications in simplify libcalls to strength reduce
726 standard library functions, but we don't currently merge them together. For
727 example, it is useful to merge memcpy(a,b,strlen(b)) -> strcpy. This can only
728 be done safely if "b" isn't modified between the strlen and memcpy of course.
730 //===---------------------------------------------------------------------===//
732 Reassociate should turn things like:
734 int factorial(int X) {
735 return X*X*X*X*X*X*X*X;
738 into llvm.powi calls, allowing the code generator to produce balanced
739 multiplication trees.
741 //===---------------------------------------------------------------------===//
743 We generate a horrible libcall for llvm.powi. For example, we compile:
746 double f(double a) { return std::pow(a, 4); }
752 movsd 16(%esp), %xmm0
755 call L___powidf2$stub
763 movsd 16(%esp), %xmm0
771 //===---------------------------------------------------------------------===//
773 We compile this program: (from GCC PR11680)
774 http://gcc.gnu.org/bugzilla/attachment.cgi?id=4487
776 Into code that runs the same speed in fast/slow modes, but both modes run 2x
777 slower than when compile with GCC (either 4.0 or 4.2):
779 $ llvm-g++ perf.cpp -O3 -fno-exceptions
781 1.821u 0.003s 0:01.82 100.0% 0+0k 0+0io 0pf+0w
783 $ g++ perf.cpp -O3 -fno-exceptions
785 0.821u 0.001s 0:00.82 100.0% 0+0k 0+0io 0pf+0w
787 It looks like we are making the same inlining decisions, so this may be raw
788 codegen badness or something else (haven't investigated).
790 //===---------------------------------------------------------------------===//
792 We miss some instcombines for stuff like this:
794 void foo (unsigned int a) {
795 /* This one is equivalent to a >= (3 << 2). */
800 A few other related ones are in GCC PR14753.
802 //===---------------------------------------------------------------------===//
804 Divisibility by constant can be simplified (according to GCC PR12849) from
805 being a mulhi to being a mul lo (cheaper). Testcase:
807 void bar(unsigned n) {
812 I think this basically amounts to a dag combine to simplify comparisons against
813 multiply hi's into a comparison against the mullo.
815 //===---------------------------------------------------------------------===//
817 Better mod/ref analysis for scanf would allow us to eliminate the vtable and a
818 bunch of other stuff from this example (see PR1604):
828 std::scanf("%d", &t.val);
829 std::printf("%d\n", t.val);
832 //===---------------------------------------------------------------------===//
834 Instcombine will merge comparisons like (x >= 10) && (x < 20) by producing (x -
835 10) u< 10, but only when the comparisons have matching sign.
837 This could be converted with a similiar technique. (PR1941)
839 define i1 @test(i8 %x) {
840 %A = icmp uge i8 %x, 5
841 %B = icmp slt i8 %x, 20
846 //===---------------------------------------------------------------------===//
848 These functions perform the same computation, but produce different assembly.
850 define i8 @select(i8 %x) readnone nounwind {
851 %A = icmp ult i8 %x, 250
852 %B = select i1 %A, i8 0, i8 1
856 define i8 @addshr(i8 %x) readnone nounwind {
857 %A = zext i8 %x to i9
858 %B = add i9 %A, 6 ;; 256 - 250 == 6
860 %D = trunc i9 %C to i8
864 //===---------------------------------------------------------------------===//
868 f (unsigned long a, unsigned long b, unsigned long c)
870 return ((a & (c - 1)) != 0) || ((b & (c - 1)) != 0);
873 f (unsigned long a, unsigned long b, unsigned long c)
875 return ((a & (c - 1)) != 0) | ((b & (c - 1)) != 0);
877 Both should combine to ((a|b) & (c-1)) != 0. Currently not optimized with
878 "clang -emit-llvm-bc | opt -std-compile-opts".
880 //===---------------------------------------------------------------------===//
883 #define PMD_MASK (~((1UL << 23) - 1))
884 void clear_pmd_range(unsigned long start, unsigned long end)
886 if (!(start & ~PMD_MASK) && !(end & ~PMD_MASK))
889 The expression should optimize to something like
890 "!((start|end)&~PMD_MASK). Currently not optimized with "clang
891 -emit-llvm-bc | opt -std-compile-opts".
893 //===---------------------------------------------------------------------===//
897 foo (unsigned int a, unsigned int b)
899 if (a <= 7 && b <= 7)
902 Should combine to "(a|b) <= 7". Currently not optimized with "clang
903 -emit-llvm-bc | opt -std-compile-opts".
905 //===---------------------------------------------------------------------===//
911 return (n >= 0 ? 1 : -1);
913 Should combine to (n >> 31) | 1. Currently not optimized with "clang
914 -emit-llvm-bc | opt -std-compile-opts | llc".
916 //===---------------------------------------------------------------------===//
919 int test(int a, int b)
926 Should combine to "a <= b". Currently not optimized with "clang
927 -emit-llvm-bc | opt -std-compile-opts | llc".
929 //===---------------------------------------------------------------------===//
933 if (variable == 4 || variable == 6)
936 This should optimize to "if ((variable | 2) == 6)". Currently not
937 optimized with "clang -emit-llvm-bc | opt -std-compile-opts | llc".
939 //===---------------------------------------------------------------------===//
941 unsigned int f(unsigned int i, unsigned int n) {++i; if (i == n) ++i; return
943 unsigned int f2(unsigned int i, unsigned int n) {++i; i += i == n; return i;}
944 These should combine to the same thing. Currently, the first function
945 produces better code on X86.
947 //===---------------------------------------------------------------------===//
950 #define abs(x) x>0?x:-x
953 return (abs(x)) >= 0;
955 This should optimize to x == INT_MIN. (With -fwrapv.) Currently not
956 optimized with "clang -emit-llvm-bc | opt -std-compile-opts".
958 //===---------------------------------------------------------------------===//
962 rotate_cst (unsigned int a)
964 a = (a << 10) | (a >> 22);
969 minus_cst (unsigned int a)
978 mask_gt (unsigned int a)
980 /* This is equivalent to a > 15. */
985 rshift_gt (unsigned int a)
987 /* This is equivalent to a > 23. */
991 All should simplify to a single comparison. All of these are
992 currently not optimized with "clang -emit-llvm-bc | opt
995 //===---------------------------------------------------------------------===//
998 int c(int* x) {return (char*)x+2 == (char*)x;}
999 Should combine to 0. Currently not optimized with "clang
1000 -emit-llvm-bc | opt -std-compile-opts" (although llc can optimize it).
1002 //===---------------------------------------------------------------------===//
1004 int a(unsigned char* b) {return *b > 99;}
1005 There's an unnecessary zext in the generated code with "clang
1006 -emit-llvm-bc | opt -std-compile-opts".
1008 //===---------------------------------------------------------------------===//
1010 int a(unsigned b) {return ((b << 31) | (b << 30)) >> 31;}
1011 Should be combined to "((b >> 1) | b) & 1". Currently not optimized
1012 with "clang -emit-llvm-bc | opt -std-compile-opts".
1014 //===---------------------------------------------------------------------===//
1016 unsigned a(unsigned x, unsigned y) { return x | (y & 1) | (y & 2);}
1017 Should combine to "x | (y & 3)". Currently not optimized with "clang
1018 -emit-llvm-bc | opt -std-compile-opts".
1020 //===---------------------------------------------------------------------===//
1022 unsigned a(unsigned a) {return ((a | 1) & 3) | (a & -4);}
1023 Should combine to "a | 1". Currently not optimized with "clang
1024 -emit-llvm-bc | opt -std-compile-opts".
1026 //===---------------------------------------------------------------------===//
1028 int a(int a, int b, int c) {return (~a & c) | ((c|a) & b);}
1029 Should fold to "(~a & c) | (a & b)". Currently not optimized with
1030 "clang -emit-llvm-bc | opt -std-compile-opts".
1032 //===---------------------------------------------------------------------===//
1034 int a(int a,int b) {return (~(a|b))|a;}
1035 Should fold to "a|~b". Currently not optimized with "clang
1036 -emit-llvm-bc | opt -std-compile-opts".
1038 //===---------------------------------------------------------------------===//
1040 int a(int a, int b) {return (a&&b) || (a&&!b);}
1041 Should fold to "a". Currently not optimized with "clang -emit-llvm-bc
1042 | opt -std-compile-opts".
1044 //===---------------------------------------------------------------------===//
1046 int a(int a, int b, int c) {return (a&&b) || (!a&&c);}
1047 Should fold to "a ? b : c", or at least something sane. Currently not
1048 optimized with "clang -emit-llvm-bc | opt -std-compile-opts".
1050 //===---------------------------------------------------------------------===//
1052 int a(int a, int b, int c) {return (a&&b) || (a&&c) || (a&&b&&c);}
1053 Should fold to a && (b || c). Currently not optimized with "clang
1054 -emit-llvm-bc | opt -std-compile-opts".
1056 //===---------------------------------------------------------------------===//
1058 int a(int x) {return x | ((x & 8) ^ 8);}
1059 Should combine to x | 8. Currently not optimized with "clang
1060 -emit-llvm-bc | opt -std-compile-opts".
1062 //===---------------------------------------------------------------------===//
1064 int a(int x) {return x ^ ((x & 8) ^ 8);}
1065 Should also combine to x | 8. Currently not optimized with "clang
1066 -emit-llvm-bc | opt -std-compile-opts".
1068 //===---------------------------------------------------------------------===//
1070 int a(int x) {return (x & 8) == 0 ? -1 : -9;}
1071 Should combine to (x | -9) ^ 8. Currently not optimized with "clang
1072 -emit-llvm-bc | opt -std-compile-opts".
1074 //===---------------------------------------------------------------------===//
1076 int a(int x) {return (x & 8) == 0 ? -9 : -1;}
1077 Should combine to x | -9. Currently not optimized with "clang
1078 -emit-llvm-bc | opt -std-compile-opts".
1080 //===---------------------------------------------------------------------===//
1082 int a(int x) {return ((x | -9) ^ 8) & x;}
1083 Should combine to x & -9. Currently not optimized with "clang
1084 -emit-llvm-bc | opt -std-compile-opts".
1086 //===---------------------------------------------------------------------===//
1088 unsigned a(unsigned a) {return a * 0x11111111 >> 28 & 1;}
1089 Should combine to "a * 0x88888888 >> 31". Currently not optimized
1090 with "clang -emit-llvm-bc | opt -std-compile-opts".
1092 //===---------------------------------------------------------------------===//
1094 unsigned a(char* x) {if ((*x & 32) == 0) return b();}
1095 There's an unnecessary zext in the generated code with "clang
1096 -emit-llvm-bc | opt -std-compile-opts".
1098 //===---------------------------------------------------------------------===//
1100 unsigned a(unsigned long long x) {return 40 * (x >> 1);}
1101 Should combine to "20 * (((unsigned)x) & -2)". Currently not
1102 optimized with "clang -emit-llvm-bc | opt -std-compile-opts".
1104 //===---------------------------------------------------------------------===//
1106 This was noticed in the entryblock for grokdeclarator in 403.gcc:
1108 %tmp = icmp eq i32 %decl_context, 4
1109 %decl_context_addr.0 = select i1 %tmp, i32 3, i32 %decl_context
1110 %tmp1 = icmp eq i32 %decl_context_addr.0, 1
1111 %decl_context_addr.1 = select i1 %tmp1, i32 0, i32 %decl_context_addr.0
1113 tmp1 should be simplified to something like:
1114 (!tmp || decl_context == 1)
1116 This allows recursive simplifications, tmp1 is used all over the place in
1117 the function, e.g. by:
1119 %tmp23 = icmp eq i32 %decl_context_addr.1, 0 ; <i1> [#uses=1]
1120 %tmp24 = xor i1 %tmp1, true ; <i1> [#uses=1]
1121 %or.cond8 = and i1 %tmp23, %tmp24 ; <i1> [#uses=1]
1125 //===---------------------------------------------------------------------===//
1127 Store sinking: This code:
1129 void f (int n, int *cond, int *res) {
1132 for (i = 0; i < n; i++)
1134 *res ^= 234; /* (*) */
1137 On this function GVN hoists the fully redundant value of *res, but nothing
1138 moves the store out. This gives us this code:
1140 bb: ; preds = %bb2, %entry
1141 %.rle = phi i32 [ 0, %entry ], [ %.rle6, %bb2 ]
1142 %i.05 = phi i32 [ 0, %entry ], [ %indvar.next, %bb2 ]
1143 %1 = load i32* %cond, align 4
1144 %2 = icmp eq i32 %1, 0
1145 br i1 %2, label %bb2, label %bb1
1148 %3 = xor i32 %.rle, 234
1149 store i32 %3, i32* %res, align 4
1152 bb2: ; preds = %bb, %bb1
1153 %.rle6 = phi i32 [ %3, %bb1 ], [ %.rle, %bb ]
1154 %indvar.next = add i32 %i.05, 1
1155 %exitcond = icmp eq i32 %indvar.next, %n
1156 br i1 %exitcond, label %return, label %bb
1158 DSE should sink partially dead stores to get the store out of the loop.
1160 Here's another partial dead case:
1161 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=12395
1163 //===---------------------------------------------------------------------===//
1165 Scalar PRE hoists the mul in the common block up to the else:
1167 int test (int a, int b, int c, int g) {
1177 It would be better to do the mul once to reduce codesize above the if.
1178 This is GCC PR38204.
1180 //===---------------------------------------------------------------------===//
1182 GCC PR37810 is an interesting case where we should sink load/store reload
1183 into the if block and outside the loop, so we don't reload/store it on the
1204 We now hoist the reload after the call (Transforms/GVN/lpre-call-wrap.ll), but
1205 we don't sink the store. We need partially dead store sinking.
1207 //===---------------------------------------------------------------------===//
1209 [PHI TRANSLATE GEPs]
1211 GCC PR37166: Sinking of loads prevents SROA'ing the "g" struct on the stack
1212 leading to excess stack traffic. This could be handled by GVN with some crazy
1213 symbolic phi translation. The code we get looks like (g is on the stack):
1217 %9 = getelementptr %struct.f* %g, i32 0, i32 0
1218 store i32 %8, i32* %9, align bel %bb3
1220 bb3: ; preds = %bb1, %bb2, %bb
1221 %c_addr.0 = phi %struct.f* [ %g, %bb2 ], [ %c, %bb ], [ %c, %bb1 ]
1222 %b_addr.0 = phi %struct.f* [ %b, %bb2 ], [ %g, %bb ], [ %b, %bb1 ]
1223 %10 = getelementptr %struct.f* %c_addr.0, i32 0, i32 0
1224 %11 = load i32* %10, align 4
1226 %11 is fully redundant, an in BB2 it should have the value %8.
1228 GCC PR33344 is a similar case.
1230 //===---------------------------------------------------------------------===//
1232 There are many load PRE testcases in testsuite/gcc.dg/tree-ssa/loadpre* in the
1233 GCC testsuite. There are many pre testcases as ssa-pre-*.c
1235 //===---------------------------------------------------------------------===//
1237 There are some interesting cases in testsuite/gcc.dg/tree-ssa/pred-comm* in the
1238 GCC testsuite. For example, predcom-1.c is:
1240 for (i = 2; i < 1000; i++)
1241 fib[i] = (fib[i-1] + fib[i - 2]) & 0xffff;
1243 which compiles into:
1245 bb1: ; preds = %bb1, %bb1.thread
1246 %indvar = phi i32 [ 0, %bb1.thread ], [ %0, %bb1 ]
1247 %i.0.reg2mem.0 = add i32 %indvar, 2
1248 %0 = add i32 %indvar, 1 ; <i32> [#uses=3]
1249 %1 = getelementptr [1000 x i32]* @fib, i32 0, i32 %0
1250 %2 = load i32* %1, align 4 ; <i32> [#uses=1]
1251 %3 = getelementptr [1000 x i32]* @fib, i32 0, i32 %indvar
1252 %4 = load i32* %3, align 4 ; <i32> [#uses=1]
1253 %5 = add i32 %4, %2 ; <i32> [#uses=1]
1254 %6 = and i32 %5, 65535 ; <i32> [#uses=1]
1255 %7 = getelementptr [1000 x i32]* @fib, i32 0, i32 %i.0.reg2mem.0
1256 store i32 %6, i32* %7, align 4
1257 %exitcond = icmp eq i32 %0, 998 ; <i1> [#uses=1]
1258 br i1 %exitcond, label %return, label %bb1
1265 instead of handling this as a loop or other xform, all we'd need to do is teach
1266 load PRE to phi translate the %0 add (i+1) into the predecessor as (i'+1+1) =
1267 (i'+2) (where i' is the previous iteration of i). This would find the store
1270 predcom-2.c is apparently the same as predcom-1.c
1271 predcom-3.c is very similar but needs loads feeding each other instead of
1273 predcom-4.c seems the same as the rest.
1276 //===---------------------------------------------------------------------===//
1278 Other simple load PRE cases:
1279 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35287 [LPRE crit edge splitting]
1281 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34677 (licm does this, LPRE crit edge)
1282 llvm-gcc t2.c -S -o - -O0 -emit-llvm | llvm-as | opt -mem2reg -simplifycfg -gvn | llvm-dis
1284 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16799 [BITCAST PHI TRANS]
1286 //===---------------------------------------------------------------------===//
1288 Type based alias analysis:
1289 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14705
1291 //===---------------------------------------------------------------------===//
1293 A/B get pinned to the stack because we turn an if/then into a select instead
1294 of PRE'ing the load/store. This may be fixable in instcombine:
1295 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37892
1297 struct X { int i; };
1311 //===---------------------------------------------------------------------===//
1313 Interesting missed case because of control flow flattening (should be 2 loads):
1314 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26629
1315 With: llvm-gcc t2.c -S -o - -O0 -emit-llvm | llvm-as |
1316 opt -mem2reg -gvn -instcombine | llvm-dis
1317 we miss it because we need 1) GEP PHI TRAN, 2) CRIT EDGE 3) MULTIPLE DIFFERENT
1318 VALS PRODUCED BY ONE BLOCK OVER DIFFERENT PATHS
1320 //===---------------------------------------------------------------------===//
1322 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19633
1323 We could eliminate the branch condition here, loading from null is undefined:
1325 struct S { int w, x, y, z; };
1326 struct T { int r; struct S s; };
1327 void bar (struct S, int);
1328 void foo (int a, struct T b)
1336 //===---------------------------------------------------------------------===//
1338 simplifylibcalls should do several optimizations for strspn/strcspn:
1340 strcspn(x, "") -> strlen(x)
1343 strspn(x, "") -> strlen(x)
1344 strspn(x, "a") -> strchr(x, 'a')-x
1346 strcspn(x, "a") -> inlined loop for up to 3 letters (similarly for strspn):
1348 size_t __strcspn_c3 (__const char *__s, int __reject1, int __reject2,
1350 register size_t __result = 0;
1351 while (__s[__result] != '\0' && __s[__result] != __reject1 &&
1352 __s[__result] != __reject2 && __s[__result] != __reject3)
1357 This should turn into a switch on the character. See PR3253 for some notes on
1360 456.hmmer apparently uses strcspn and strspn a lot. 471.omnetpp uses strspn.
1362 //===---------------------------------------------------------------------===//
1364 "gas" uses this idiom:
1365 else if (strchr ("+-/*%|&^:[]()~", *intel_parser.op_string))
1367 else if (strchr ("<>", *intel_parser.op_string)
1369 Those should be turned into a switch.
1371 //===---------------------------------------------------------------------===//
1373 252.eon contains this interesting code:
1375 %3072 = getelementptr [100 x i8]* %tempString, i32 0, i32 0
1376 %3073 = call i8* @strcpy(i8* %3072, i8* %3071) nounwind
1377 %strlen = call i32 @strlen(i8* %3072) ; uses = 1
1378 %endptr = getelementptr [100 x i8]* %tempString, i32 0, i32 %strlen
1379 call void @llvm.memcpy.i32(i8* %endptr,
1380 i8* getelementptr ([5 x i8]* @"\01LC42", i32 0, i32 0), i32 5, i32 1)
1381 %3074 = call i32 @strlen(i8* %endptr) nounwind readonly
1383 This is interesting for a couple reasons. First, in this:
1385 %3073 = call i8* @strcpy(i8* %3072, i8* %3071) nounwind
1386 %strlen = call i32 @strlen(i8* %3072)
1388 The strlen could be replaced with: %strlen = sub %3072, %3073, because the
1389 strcpy call returns a pointer to the end of the string. Based on that, the
1390 endptr GEP just becomes equal to 3073, which eliminates a strlen call and GEP.
1392 Second, the memcpy+strlen strlen can be replaced with:
1394 %3074 = call i32 @strlen([5 x i8]* @"\01LC42") nounwind readonly
1396 Because the destination was just copied into the specified memory buffer. This,
1397 in turn, can be constant folded to "4".
1399 In other code, it contains:
1401 %endptr6978 = bitcast i8* %endptr69 to i32*
1402 store i32 7107374, i32* %endptr6978, align 1
1403 %3167 = call i32 @strlen(i8* %endptr69) nounwind readonly
1405 Which could also be constant folded. Whatever is producing this should probably
1406 be fixed to leave this as a memcpy from a string.
1408 Further, eon also has an interesting partially redundant strlen call:
1410 bb8: ; preds = %_ZN18eonImageCalculatorC1Ev.exit
1411 %682 = getelementptr i8** %argv, i32 6 ; <i8**> [#uses=2]
1412 %683 = load i8** %682, align 4 ; <i8*> [#uses=4]
1413 %684 = load i8* %683, align 1 ; <i8> [#uses=1]
1414 %685 = icmp eq i8 %684, 0 ; <i1> [#uses=1]
1415 br i1 %685, label %bb10, label %bb9
1418 %686 = call i32 @strlen(i8* %683) nounwind readonly
1419 %687 = icmp ugt i32 %686, 254 ; <i1> [#uses=1]
1420 br i1 %687, label %bb10, label %bb11
1422 bb10: ; preds = %bb9, %bb8
1423 %688 = call i32 @strlen(i8* %683) nounwind readonly
1425 This could be eliminated by doing the strlen once in bb8, saving code size and
1426 improving perf on the bb8->9->10 path.
1428 //===---------------------------------------------------------------------===//
1430 I see an interesting fully redundant call to strlen left in 186.crafty:InputMove
1432 %movetext11 = getelementptr [128 x i8]* %movetext, i32 0, i32 0
1435 bb62: ; preds = %bb55, %bb53
1436 %promote.0 = phi i32 [ %169, %bb55 ], [ 0, %bb53 ]
1437 %171 = call i32 @strlen(i8* %movetext11) nounwind readonly align 1
1438 %172 = add i32 %171, -1 ; <i32> [#uses=1]
1439 %173 = getelementptr [128 x i8]* %movetext, i32 0, i32 %172
1442 br i1 %or.cond, label %bb65, label %bb72
1444 bb65: ; preds = %bb62
1445 store i8 0, i8* %173, align 1
1448 bb72: ; preds = %bb65, %bb62
1449 %trank.1 = phi i32 [ %176, %bb65 ], [ -1, %bb62 ]
1450 %177 = call i32 @strlen(i8* %movetext11) nounwind readonly align 1
1452 Note that on the bb62->bb72 path, that the %177 strlen call is partially
1453 redundant with the %171 call. At worst, we could shove the %177 strlen call
1454 up into the bb65 block moving it out of the bb62->bb72 path. However, note
1455 that bb65 stores to the string, zeroing out the last byte. This means that on
1456 that path the value of %177 is actually just %171-1. A sub is cheaper than a
1459 This pattern repeats several times, basically doing:
1464 where it is "obvious" that B = A-1.
1466 //===---------------------------------------------------------------------===//
1468 186.crafty contains this interesting pattern:
1470 %77 = call i8* @strstr(i8* getelementptr ([6 x i8]* @"\01LC5", i32 0, i32 0),
1472 %phitmp648 = icmp eq i8* %77, getelementptr ([6 x i8]* @"\01LC5", i32 0, i32 0)
1473 br i1 %phitmp648, label %bb70, label %bb76
1475 bb70: ; preds = %OptionMatch.exit91, %bb69
1476 %78 = call i32 @strlen(i8* %30) nounwind readonly align 1 ; <i32> [#uses=1]
1480 if (strstr(cststr, P) == cststr) {
1484 The strstr call would be significantly cheaper written as:
1487 if (memcmp(P, str, strlen(P)))
1490 This is memcmp+strlen instead of strstr. This also makes the strlen fully
1493 //===---------------------------------------------------------------------===//
1495 186.crafty also contains this code:
1497 %1906 = call i32 @strlen(i8* getelementptr ([32 x i8]* @pgn_event, i32 0,i32 0))
1498 %1907 = getelementptr [32 x i8]* @pgn_event, i32 0, i32 %1906
1499 %1908 = call i8* @strcpy(i8* %1907, i8* %1905) nounwind align 1
1500 %1909 = call i32 @strlen(i8* getelementptr ([32 x i8]* @pgn_event, i32 0,i32 0))
1501 %1910 = getelementptr [32 x i8]* @pgn_event, i32 0, i32 %1909
1503 The last strlen is computable as 1908-@pgn_event, which means 1910=1908.
1505 //===---------------------------------------------------------------------===//
1507 186.crafty has this interesting pattern with the "out.4543" variable:
1509 call void @llvm.memcpy.i32(
1510 i8* getelementptr ([10 x i8]* @out.4543, i32 0, i32 0),
1511 i8* getelementptr ([7 x i8]* @"\01LC28700", i32 0, i32 0), i32 7, i32 1)
1512 %101 = call@printf(i8* ... @out.4543, i32 0, i32 0)) nounwind
1514 It is basically doing:
1516 memcpy(globalarray, "string");
1517 printf(..., globalarray);
1519 Anyway, by knowing that printf just reads the memory and forward substituting
1520 the string directly into the printf, this eliminates reads from globalarray.
1521 Since this pattern occurs frequently in crafty (due to the "DisplayTime" and
1522 other similar functions) there are many stores to "out". Once all the printfs
1523 stop using "out", all that is left is the memcpy's into it. This should allow
1524 globalopt to remove the "stored only" global.
1526 //===---------------------------------------------------------------------===//
1530 define inreg i32 @foo(i8* inreg %p) nounwind {
1532 %tmp1 = ashr i8 %tmp0, 5
1533 %tmp2 = sext i8 %tmp1 to i32
1537 could be dagcombine'd to a sign-extending load with a shift.
1538 For example, on x86 this currently gets this:
1544 while it could get this:
1549 //===---------------------------------------------------------------------===//
1553 int test(int x) { return 1-x == x; } // --> return false
1554 int test2(int x) { return 2-x == x; } // --> return x == 1 ?
1556 Always foldable for odd constants, what is the rule for even?
1558 //===---------------------------------------------------------------------===//
1560 PR 3381: GEP to field of size 0 inside a struct could be turned into GEP
1561 for next field in struct (which is at same address).
1563 For example: store of float into { {{}}, float } could be turned into a store to
1566 //===---------------------------------------------------------------------===//
1569 double foo(double a) { return sin(a); }
1571 This compiles into this on x86-64 Linux:
1582 //===---------------------------------------------------------------------===//
1584 The arg promotion pass should make use of nocapture to make its alias analysis
1585 stuff much more precise.
1587 //===---------------------------------------------------------------------===//
1589 The following functions should be optimized to use a select instead of a
1590 branch (from gcc PR40072):
1592 char char_int(int m) {if(m>7) return 0; return m;}
1593 int int_char(char m) {if(m>7) return 0; return m;}
1595 //===---------------------------------------------------------------------===//
1597 IPSCCP is propagating elements of first class aggregates, but is not propagating
1598 the entire aggregate itself. This leads it to miss opportunities, for example
1599 in test/Transforms/SCCP/ipsccp-basic.ll:test5b.
1601 //===---------------------------------------------------------------------===//
1603 int func(int a, int b) { if (a & 0x80) b |= 0x80; else b &= ~0x80; return b; }
1607 define i32 @func(i32 %a, i32 %b) nounwind readnone ssp {
1609 %0 = and i32 %a, 128 ; <i32> [#uses=1]
1610 %1 = icmp eq i32 %0, 0 ; <i1> [#uses=1]
1611 %2 = or i32 %b, 128 ; <i32> [#uses=1]
1612 %3 = and i32 %b, -129 ; <i32> [#uses=1]
1613 %b_addr.0 = select i1 %1, i32 %3, i32 %2 ; <i32> [#uses=1]
1617 However, it's functionally equivalent to:
1619 b = (b & ~0x80) | (a & 0x80);
1621 Which generates this:
1623 define i32 @func(i32 %a, i32 %b) nounwind readnone ssp {
1625 %0 = and i32 %b, -129 ; <i32> [#uses=1]
1626 %1 = and i32 %a, 128 ; <i32> [#uses=1]
1627 %2 = or i32 %0, %1 ; <i32> [#uses=1]
1631 This can be generalized for other forms:
1633 b = (b & ~0x80) | (a & 0x40) << 1;
1635 //===---------------------------------------------------------------------===//
1637 These two functions produce different code. They shouldn't:
1641 uint8_t p1(uint8_t b, uint8_t a) {
1642 b = (b & ~0xc0) | (a & 0xc0);
1646 uint8_t p2(uint8_t b, uint8_t a) {
1647 b = (b & ~0x40) | (a & 0x40);
1648 b = (b & ~0x80) | (a & 0x80);
1652 define zeroext i8 @p1(i8 zeroext %b, i8 zeroext %a) nounwind readnone ssp {
1654 %0 = and i8 %b, 63 ; <i8> [#uses=1]
1655 %1 = and i8 %a, -64 ; <i8> [#uses=1]
1656 %2 = or i8 %1, %0 ; <i8> [#uses=1]
1660 define zeroext i8 @p2(i8 zeroext %b, i8 zeroext %a) nounwind readnone ssp {
1662 %0 = and i8 %b, 63 ; <i8> [#uses=1]
1663 %.masked = and i8 %a, 64 ; <i8> [#uses=1]
1664 %1 = and i8 %a, -128 ; <i8> [#uses=1]
1665 %2 = or i8 %1, %0 ; <i8> [#uses=1]
1666 %3 = or i8 %2, %.masked ; <i8> [#uses=1]
1670 //===---------------------------------------------------------------------===//