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 isn't recognized as bswap by instcombine (yes, it really is bswap):
229 unsigned long reverse(unsigned v) {
231 t = v ^ ((v << 16) | (v >> 16));
233 v = (v << 24) | (v >> 8);
237 //===---------------------------------------------------------------------===//
239 These idioms should be recognized as popcount (see PR1488):
241 unsigned countbits_slow(unsigned v) {
243 for (c = 0; v; v >>= 1)
247 unsigned countbits_fast(unsigned v){
250 v &= v - 1; // clear the least significant bit set
254 BITBOARD = unsigned long long
255 int PopCnt(register BITBOARD a) {
263 unsigned int popcount(unsigned int input) {
264 unsigned int count = 0;
265 for (unsigned int i = 0; i < 4 * 8; i++)
266 count += (input >> i) & i;
270 //===---------------------------------------------------------------------===//
272 These should turn into single 16-bit (unaligned?) loads on little/big endian
275 unsigned short read_16_le(const unsigned char *adr) {
276 return adr[0] | (adr[1] << 8);
278 unsigned short read_16_be(const unsigned char *adr) {
279 return (adr[0] << 8) | adr[1];
282 //===---------------------------------------------------------------------===//
284 -instcombine should handle this transform:
285 icmp pred (sdiv X / C1 ), C2
286 when X, C1, and C2 are unsigned. Similarly for udiv and signed operands.
288 Currently InstCombine avoids this transform but will do it when the signs of
289 the operands and the sign of the divide match. See the FIXME in
290 InstructionCombining.cpp in the visitSetCondInst method after the switch case
291 for Instruction::UDiv (around line 4447) for more details.
293 The SingleSource/Benchmarks/Shootout-C++/hash and hash2 tests have examples of
296 //===---------------------------------------------------------------------===//
298 viterbi speeds up *significantly* if the various "history" related copy loops
299 are turned into memcpy calls at the source level. We need a "loops to memcpy"
302 //===---------------------------------------------------------------------===//
306 typedef unsigned U32;
307 typedef unsigned long long U64;
308 int test (U32 *inst, U64 *regs) {
311 int r1 = (temp >> 20) & 0xf;
312 int b2 = (temp >> 16) & 0xf;
313 effective_addr2 = temp & 0xfff;
314 if (b2) effective_addr2 += regs[b2];
315 b2 = (temp >> 12) & 0xf;
316 if (b2) effective_addr2 += regs[b2];
317 effective_addr2 &= regs[4];
318 if ((effective_addr2 & 3) == 0)
323 Note that only the low 2 bits of effective_addr2 are used. On 32-bit systems,
324 we don't eliminate the computation of the top half of effective_addr2 because
325 we don't have whole-function selection dags. On x86, this means we use one
326 extra register for the function when effective_addr2 is declared as U64 than
327 when it is declared U32.
329 //===---------------------------------------------------------------------===//
331 LSR should know what GPR types a target has. This code:
333 volatile short X, Y; // globals
337 for (i = 0; i < N; i++) { X = i; Y = i*4; }
340 produces two identical IV's (after promotion) on PPC/ARM:
342 LBB1_1: @bb.preheader
353 add r1, r1, #1 <- [0,+,1]
355 add r2, r2, #1 <- [0,+,1]
360 //===---------------------------------------------------------------------===//
362 Tail call elim should be more aggressive, checking to see if the call is
363 followed by an uncond branch to an exit block.
365 ; This testcase is due to tail-duplication not wanting to copy the return
366 ; instruction into the terminating blocks because there was other code
367 ; optimized out of the function after the taildup happened.
368 ; RUN: llvm-as < %s | opt -tailcallelim | llvm-dis | not grep call
370 define i32 @t4(i32 %a) {
372 %tmp.1 = and i32 %a, 1 ; <i32> [#uses=1]
373 %tmp.2 = icmp ne i32 %tmp.1, 0 ; <i1> [#uses=1]
374 br i1 %tmp.2, label %then.0, label %else.0
376 then.0: ; preds = %entry
377 %tmp.5 = add i32 %a, -1 ; <i32> [#uses=1]
378 %tmp.3 = call i32 @t4( i32 %tmp.5 ) ; <i32> [#uses=1]
381 else.0: ; preds = %entry
382 %tmp.7 = icmp ne i32 %a, 0 ; <i1> [#uses=1]
383 br i1 %tmp.7, label %then.1, label %return
385 then.1: ; preds = %else.0
386 %tmp.11 = add i32 %a, -2 ; <i32> [#uses=1]
387 %tmp.9 = call i32 @t4( i32 %tmp.11 ) ; <i32> [#uses=1]
390 return: ; preds = %then.1, %else.0, %then.0
391 %result.0 = phi i32 [ 0, %else.0 ], [ %tmp.3, %then.0 ],
396 //===---------------------------------------------------------------------===//
398 Tail recursion elimination is not transforming this function, because it is
399 returning n, which fails the isDynamicConstant check in the accumulator
402 long long fib(const long long n) {
408 return fib(n-1) + fib(n-2);
412 //===---------------------------------------------------------------------===//
414 Tail recursion elimination should handle:
419 return 2 * pow2m1 (n - 1) + 1;
422 Also, multiplies can be turned into SHL's, so they should be handled as if
423 they were associative. "return foo() << 1" can be tail recursion eliminated.
425 //===---------------------------------------------------------------------===//
427 Argument promotion should promote arguments for recursive functions, like
430 ; RUN: llvm-as < %s | opt -argpromotion | llvm-dis | grep x.val
432 define internal i32 @foo(i32* %x) {
434 %tmp = load i32* %x ; <i32> [#uses=0]
435 %tmp.foo = call i32 @foo( i32* %x ) ; <i32> [#uses=1]
439 define i32 @bar(i32* %x) {
441 %tmp3 = call i32 @foo( i32* %x ) ; <i32> [#uses=1]
445 //===---------------------------------------------------------------------===//
447 "basicaa" should know how to look through "or" instructions that act like add
448 instructions. For example in this code, the x*4+1 is turned into x*4 | 1, and
449 basicaa can't analyze the array subscript, leading to duplicated loads in the
452 void test(int X, int Y, int a[]) {
454 for (i=2; i<1000; i+=4) {
455 a[i+0] = a[i-1+0]*a[i-2+0];
456 a[i+1] = a[i-1+1]*a[i-2+1];
457 a[i+2] = a[i-1+2]*a[i-2+2];
458 a[i+3] = a[i-1+3]*a[i-2+3];
462 BasicAA also doesn't do this for add. It needs to know that &A[i+1] != &A[i].
464 //===---------------------------------------------------------------------===//
466 We should investigate an instruction sinking pass. Consider this silly
482 je LBB1_2 # cond_true
490 The PIC base computation (call+popl) is only used on one path through the
491 code, but is currently always computed in the entry block. It would be
492 better to sink the picbase computation down into the block for the
493 assertion, as it is the only one that uses it. This happens for a lot of
494 code with early outs.
496 Another example is loads of arguments, which are usually emitted into the
497 entry block on targets like x86. If not used in all paths through a
498 function, they should be sunk into the ones that do.
500 In this case, whole-function-isel would also handle this.
502 //===---------------------------------------------------------------------===//
504 Investigate lowering of sparse switch statements into perfect hash tables:
505 http://burtleburtle.net/bob/hash/perfect.html
507 //===---------------------------------------------------------------------===//
509 We should turn things like "load+fabs+store" and "load+fneg+store" into the
510 corresponding integer operations. On a yonah, this loop:
515 for (b = 0; b < 10000000; b++)
516 for (i = 0; i < 256; i++)
520 is twice as slow as this loop:
525 for (b = 0; b < 10000000; b++)
526 for (i = 0; i < 256; i++)
527 a[i] ^= (1ULL << 63);
530 and I suspect other processors are similar. On X86 in particular this is a
531 big win because doing this with integers allows the use of read/modify/write
534 //===---------------------------------------------------------------------===//
536 DAG Combiner should try to combine small loads into larger loads when
537 profitable. For example, we compile this C++ example:
539 struct THotKey { short Key; bool Control; bool Shift; bool Alt; };
540 extern THotKey m_HotKey;
541 THotKey GetHotKey () { return m_HotKey; }
543 into (-O3 -fno-exceptions -static -fomit-frame-pointer):
548 movb _m_HotKey+3, %cl
549 movb _m_HotKey+4, %dl
550 movb _m_HotKey+2, %ch
565 movzwl _m_HotKey+4, %edx
569 The LLVM IR contains the needed alignment info, so we should be able to
570 merge the loads and stores into 4-byte loads:
572 %struct.THotKey = type { i16, i8, i8, i8 }
573 define void @_Z9GetHotKeyv(%struct.THotKey* sret %agg.result) nounwind {
575 %tmp2 = load i16* getelementptr (@m_HotKey, i32 0, i32 0), align 8
576 %tmp5 = load i8* getelementptr (@m_HotKey, i32 0, i32 1), align 2
577 %tmp8 = load i8* getelementptr (@m_HotKey, i32 0, i32 2), align 1
578 %tmp11 = load i8* getelementptr (@m_HotKey, i32 0, i32 3), align 2
580 Alternatively, we should use a small amount of base-offset alias analysis
581 to make it so the scheduler doesn't need to hold all the loads in regs at
584 //===---------------------------------------------------------------------===//
586 We should add an FRINT node to the DAG to model targets that have legal
587 implementations of ceil/floor/rint.
589 //===---------------------------------------------------------------------===//
591 This GCC bug: http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34043
592 contains a testcase that compiles down to:
594 %struct.XMM128 = type { <4 x float> }
596 %src = alloca %struct.XMM128
598 %tmp6263 = bitcast %struct.XMM128* %src to <2 x i64>*
599 %tmp65 = getelementptr %struct.XMM128* %src, i32 0, i32 0
600 store <2 x i64> %tmp5899, <2 x i64>* %tmp6263, align 16
601 %tmp66 = load <4 x float>* %tmp65, align 16
602 %tmp71 = add <4 x float> %tmp66, %tmp66
604 If the mid-level optimizer turned the bitcast of pointer + store of tmp5899
605 into a bitcast of the vector value and a store to the pointer, then the
606 store->load could be easily removed.
608 //===---------------------------------------------------------------------===//
613 long long input[8] = {1,1,1,1,1,1,1,1};
617 We currently compile this into a memcpy from a global array since the
618 initializer is fairly large and not memset'able. This is good, but the memcpy
619 gets lowered to load/stores in the code generator. This is also ok, except
620 that the codegen lowering for memcpy doesn't handle the case when the source
621 is a constant global. This gives us atrocious code like this:
626 movl _C.0.1444-"L1$pb"+32(%eax), %ecx
628 movl _C.0.1444-"L1$pb"+20(%eax), %ecx
630 movl _C.0.1444-"L1$pb"+36(%eax), %ecx
632 movl _C.0.1444-"L1$pb"+44(%eax), %ecx
634 movl _C.0.1444-"L1$pb"+40(%eax), %ecx
636 movl _C.0.1444-"L1$pb"+12(%eax), %ecx
638 movl _C.0.1444-"L1$pb"+4(%eax), %ecx
650 //===---------------------------------------------------------------------===//
652 http://llvm.org/PR717:
654 The following code should compile into "ret int undef". Instead, LLVM
655 produces "ret int 0":
664 //===---------------------------------------------------------------------===//
666 The loop unroller should partially unroll loops (instead of peeling them)
667 when code growth isn't too bad and when an unroll count allows simplification
668 of some code within the loop. One trivial example is:
674 for ( nLoop = 0; nLoop < 1000; nLoop++ ) {
683 Unrolling by 2 would eliminate the '&1' in both copies, leading to a net
684 reduction in code size. The resultant code would then also be suitable for
685 exit value computation.
687 //===---------------------------------------------------------------------===//
689 We miss a bunch of rotate opportunities on various targets, including ppc, x86,
690 etc. On X86, we miss a bunch of 'rotate by variable' cases because the rotate
691 matching code in dag combine doesn't look through truncates aggressively
692 enough. Here are some testcases reduces from GCC PR17886:
694 unsigned long long f(unsigned long long x, int y) {
695 return (x << y) | (x >> 64-y);
697 unsigned f2(unsigned x, int y){
698 return (x << y) | (x >> 32-y);
700 unsigned long long f3(unsigned long long x){
702 return (x << y) | (x >> 64-y);
704 unsigned f4(unsigned x){
706 return (x << y) | (x >> 32-y);
708 unsigned long long f5(unsigned long long x, unsigned long long y) {
709 return (x << 8) | ((y >> 48) & 0xffull);
711 unsigned long long f6(unsigned long long x, unsigned long long y, int z) {
714 return (x << 8) | ((y >> 48) & 0xffull);
716 return (x << 16) | ((y >> 40) & 0xffffull);
718 return (x << 24) | ((y >> 32) & 0xffffffull);
720 return (x << 32) | ((y >> 24) & 0xffffffffull);
722 return (x << 40) | ((y >> 16) & 0xffffffffffull);
726 On X86-64, we only handle f2/f3/f4 right. On x86-32, a few of these
727 generate truly horrible code, instead of using shld and friends. On
728 ARM, we end up with calls to L___lshrdi3/L___ashldi3 in f, which is
729 badness. PPC64 misses f, f5 and f6. CellSPU aborts in isel.
731 //===---------------------------------------------------------------------===//
733 We do a number of simplifications in simplify libcalls to strength reduce
734 standard library functions, but we don't currently merge them together. For
735 example, it is useful to merge memcpy(a,b,strlen(b)) -> strcpy. This can only
736 be done safely if "b" isn't modified between the strlen and memcpy of course.
738 //===---------------------------------------------------------------------===//
740 Reassociate should turn things like:
742 int factorial(int X) {
743 return X*X*X*X*X*X*X*X;
746 into llvm.powi calls, allowing the code generator to produce balanced
747 multiplication trees.
749 //===---------------------------------------------------------------------===//
751 We generate a horrible libcall for llvm.powi. For example, we compile:
754 double f(double a) { return std::pow(a, 4); }
760 movsd 16(%esp), %xmm0
763 call L___powidf2$stub
771 movsd 16(%esp), %xmm0
779 //===---------------------------------------------------------------------===//
781 We compile this program: (from GCC PR11680)
782 http://gcc.gnu.org/bugzilla/attachment.cgi?id=4487
784 Into code that runs the same speed in fast/slow modes, but both modes run 2x
785 slower than when compile with GCC (either 4.0 or 4.2):
787 $ llvm-g++ perf.cpp -O3 -fno-exceptions
789 1.821u 0.003s 0:01.82 100.0% 0+0k 0+0io 0pf+0w
791 $ g++ perf.cpp -O3 -fno-exceptions
793 0.821u 0.001s 0:00.82 100.0% 0+0k 0+0io 0pf+0w
795 It looks like we are making the same inlining decisions, so this may be raw
796 codegen badness or something else (haven't investigated).
798 //===---------------------------------------------------------------------===//
800 We miss some instcombines for stuff like this:
802 void foo (unsigned int a) {
803 /* This one is equivalent to a >= (3 << 2). */
808 A few other related ones are in GCC PR14753.
810 //===---------------------------------------------------------------------===//
812 Divisibility by constant can be simplified (according to GCC PR12849) from
813 being a mulhi to being a mul lo (cheaper). Testcase:
815 void bar(unsigned n) {
820 I think this basically amounts to a dag combine to simplify comparisons against
821 multiply hi's into a comparison against the mullo.
823 //===---------------------------------------------------------------------===//
825 Better mod/ref analysis for scanf would allow us to eliminate the vtable and a
826 bunch of other stuff from this example (see PR1604):
836 std::scanf("%d", &t.val);
837 std::printf("%d\n", t.val);
840 //===---------------------------------------------------------------------===//
842 Instcombine will merge comparisons like (x >= 10) && (x < 20) by producing (x -
843 10) u< 10, but only when the comparisons have matching sign.
845 This could be converted with a similiar technique. (PR1941)
847 define i1 @test(i8 %x) {
848 %A = icmp uge i8 %x, 5
849 %B = icmp slt i8 %x, 20
854 //===---------------------------------------------------------------------===//
856 These functions perform the same computation, but produce different assembly.
858 define i8 @select(i8 %x) readnone nounwind {
859 %A = icmp ult i8 %x, 250
860 %B = select i1 %A, i8 0, i8 1
864 define i8 @addshr(i8 %x) readnone nounwind {
865 %A = zext i8 %x to i9
866 %B = add i9 %A, 6 ;; 256 - 250 == 6
868 %D = trunc i9 %C to i8
872 //===---------------------------------------------------------------------===//
876 f (unsigned long a, unsigned long b, unsigned long c)
878 return ((a & (c - 1)) != 0) || ((b & (c - 1)) != 0);
881 f (unsigned long a, unsigned long b, unsigned long c)
883 return ((a & (c - 1)) != 0) | ((b & (c - 1)) != 0);
885 Both should combine to ((a|b) & (c-1)) != 0. Currently not optimized with
886 "clang -emit-llvm-bc | opt -std-compile-opts".
888 //===---------------------------------------------------------------------===//
891 #define PMD_MASK (~((1UL << 23) - 1))
892 void clear_pmd_range(unsigned long start, unsigned long end)
894 if (!(start & ~PMD_MASK) && !(end & ~PMD_MASK))
897 The expression should optimize to something like
898 "!((start|end)&~PMD_MASK). Currently not optimized with "clang
899 -emit-llvm-bc | opt -std-compile-opts".
901 //===---------------------------------------------------------------------===//
905 foo (unsigned int a, unsigned int b)
907 if (a <= 7 && b <= 7)
910 Should combine to "(a|b) <= 7". Currently not optimized with "clang
911 -emit-llvm-bc | opt -std-compile-opts".
913 //===---------------------------------------------------------------------===//
919 return (n >= 0 ? 1 : -1);
921 Should combine to (n >> 31) | 1. Currently not optimized with "clang
922 -emit-llvm-bc | opt -std-compile-opts | llc".
924 //===---------------------------------------------------------------------===//
927 int test(int a, int b)
934 Should combine to "a <= b". Currently not optimized with "clang
935 -emit-llvm-bc | opt -std-compile-opts | llc".
937 //===---------------------------------------------------------------------===//
941 if (variable == 4 || variable == 6)
944 This should optimize to "if ((variable | 2) == 6)". Currently not
945 optimized with "clang -emit-llvm-bc | opt -std-compile-opts | llc".
947 //===---------------------------------------------------------------------===//
949 unsigned int f(unsigned int i, unsigned int n) {++i; if (i == n) ++i; return
951 unsigned int f2(unsigned int i, unsigned int n) {++i; i += i == n; return i;}
952 These should combine to the same thing. Currently, the first function
953 produces better code on X86.
955 //===---------------------------------------------------------------------===//
958 #define abs(x) x>0?x:-x
961 return (abs(x)) >= 0;
963 This should optimize to x == INT_MIN. (With -fwrapv.) Currently not
964 optimized with "clang -emit-llvm-bc | opt -std-compile-opts".
966 //===---------------------------------------------------------------------===//
970 rotate_cst (unsigned int a)
972 a = (a << 10) | (a >> 22);
977 minus_cst (unsigned int a)
986 mask_gt (unsigned int a)
988 /* This is equivalent to a > 15. */
993 rshift_gt (unsigned int a)
995 /* This is equivalent to a > 23. */
999 All should simplify to a single comparison. All of these are
1000 currently not optimized with "clang -emit-llvm-bc | opt
1003 //===---------------------------------------------------------------------===//
1006 int c(int* x) {return (char*)x+2 == (char*)x;}
1007 Should combine to 0. Currently not optimized with "clang
1008 -emit-llvm-bc | opt -std-compile-opts" (although llc can optimize it).
1010 //===---------------------------------------------------------------------===//
1012 int a(unsigned char* b) {return *b > 99;}
1013 There's an unnecessary zext in the generated code with "clang
1014 -emit-llvm-bc | opt -std-compile-opts".
1016 //===---------------------------------------------------------------------===//
1018 int a(unsigned b) {return ((b << 31) | (b << 30)) >> 31;}
1019 Should be combined to "((b >> 1) | b) & 1". Currently not optimized
1020 with "clang -emit-llvm-bc | opt -std-compile-opts".
1022 //===---------------------------------------------------------------------===//
1024 unsigned a(unsigned x, unsigned y) { return x | (y & 1) | (y & 2);}
1025 Should combine to "x | (y & 3)". Currently not optimized with "clang
1026 -emit-llvm-bc | opt -std-compile-opts".
1028 //===---------------------------------------------------------------------===//
1030 unsigned a(unsigned a) {return ((a | 1) & 3) | (a & -4);}
1031 Should combine to "a | 1". Currently not optimized with "clang
1032 -emit-llvm-bc | opt -std-compile-opts".
1034 //===---------------------------------------------------------------------===//
1036 int a(int a, int b, int c) {return (~a & c) | ((c|a) & b);}
1037 Should fold to "(~a & c) | (a & b)". Currently not optimized with
1038 "clang -emit-llvm-bc | opt -std-compile-opts".
1040 //===---------------------------------------------------------------------===//
1042 int a(int a,int b) {return (~(a|b))|a;}
1043 Should fold to "a|~b". Currently not optimized with "clang
1044 -emit-llvm-bc | opt -std-compile-opts".
1046 //===---------------------------------------------------------------------===//
1048 int a(int a, int b) {return (a&&b) || (a&&!b);}
1049 Should fold to "a". Currently not optimized with "clang -emit-llvm-bc
1050 | opt -std-compile-opts".
1052 //===---------------------------------------------------------------------===//
1054 int a(int a, int b, int c) {return (a&&b) || (!a&&c);}
1055 Should fold to "a ? b : c", or at least something sane. Currently not
1056 optimized with "clang -emit-llvm-bc | opt -std-compile-opts".
1058 //===---------------------------------------------------------------------===//
1060 int a(int a, int b, int c) {return (a&&b) || (a&&c) || (a&&b&&c);}
1061 Should fold to a && (b || c). Currently not optimized with "clang
1062 -emit-llvm-bc | opt -std-compile-opts".
1064 //===---------------------------------------------------------------------===//
1066 int a(int x) {return x | ((x & 8) ^ 8);}
1067 Should combine to x | 8. Currently not optimized with "clang
1068 -emit-llvm-bc | opt -std-compile-opts".
1070 //===---------------------------------------------------------------------===//
1072 int a(int x) {return x ^ ((x & 8) ^ 8);}
1073 Should also combine to x | 8. Currently not optimized with "clang
1074 -emit-llvm-bc | opt -std-compile-opts".
1076 //===---------------------------------------------------------------------===//
1078 int a(int x) {return (x & 8) == 0 ? -1 : -9;}
1079 Should combine to (x | -9) ^ 8. Currently not optimized with "clang
1080 -emit-llvm-bc | opt -std-compile-opts".
1082 //===---------------------------------------------------------------------===//
1084 int a(int x) {return (x & 8) == 0 ? -9 : -1;}
1085 Should combine to x | -9. Currently not optimized with "clang
1086 -emit-llvm-bc | opt -std-compile-opts".
1088 //===---------------------------------------------------------------------===//
1090 int a(int x) {return ((x | -9) ^ 8) & x;}
1091 Should combine to x & -9. Currently not optimized with "clang
1092 -emit-llvm-bc | opt -std-compile-opts".
1094 //===---------------------------------------------------------------------===//
1096 unsigned a(unsigned a) {return a * 0x11111111 >> 28 & 1;}
1097 Should combine to "a * 0x88888888 >> 31". Currently not optimized
1098 with "clang -emit-llvm-bc | opt -std-compile-opts".
1100 //===---------------------------------------------------------------------===//
1102 unsigned a(char* x) {if ((*x & 32) == 0) return b();}
1103 There's an unnecessary zext in the generated code with "clang
1104 -emit-llvm-bc | opt -std-compile-opts".
1106 //===---------------------------------------------------------------------===//
1108 unsigned a(unsigned long long x) {return 40 * (x >> 1);}
1109 Should combine to "20 * (((unsigned)x) & -2)". Currently not
1110 optimized with "clang -emit-llvm-bc | opt -std-compile-opts".
1112 //===---------------------------------------------------------------------===//
1114 This was noticed in the entryblock for grokdeclarator in 403.gcc:
1116 %tmp = icmp eq i32 %decl_context, 4
1117 %decl_context_addr.0 = select i1 %tmp, i32 3, i32 %decl_context
1118 %tmp1 = icmp eq i32 %decl_context_addr.0, 1
1119 %decl_context_addr.1 = select i1 %tmp1, i32 0, i32 %decl_context_addr.0
1121 tmp1 should be simplified to something like:
1122 (!tmp || decl_context == 1)
1124 This allows recursive simplifications, tmp1 is used all over the place in
1125 the function, e.g. by:
1127 %tmp23 = icmp eq i32 %decl_context_addr.1, 0 ; <i1> [#uses=1]
1128 %tmp24 = xor i1 %tmp1, true ; <i1> [#uses=1]
1129 %or.cond8 = and i1 %tmp23, %tmp24 ; <i1> [#uses=1]
1133 //===---------------------------------------------------------------------===//
1135 Store sinking: This code:
1137 void f (int n, int *cond, int *res) {
1140 for (i = 0; i < n; i++)
1142 *res ^= 234; /* (*) */
1145 On this function GVN hoists the fully redundant value of *res, but nothing
1146 moves the store out. This gives us this code:
1148 bb: ; preds = %bb2, %entry
1149 %.rle = phi i32 [ 0, %entry ], [ %.rle6, %bb2 ]
1150 %i.05 = phi i32 [ 0, %entry ], [ %indvar.next, %bb2 ]
1151 %1 = load i32* %cond, align 4
1152 %2 = icmp eq i32 %1, 0
1153 br i1 %2, label %bb2, label %bb1
1156 %3 = xor i32 %.rle, 234
1157 store i32 %3, i32* %res, align 4
1160 bb2: ; preds = %bb, %bb1
1161 %.rle6 = phi i32 [ %3, %bb1 ], [ %.rle, %bb ]
1162 %indvar.next = add i32 %i.05, 1
1163 %exitcond = icmp eq i32 %indvar.next, %n
1164 br i1 %exitcond, label %return, label %bb
1166 DSE should sink partially dead stores to get the store out of the loop.
1168 Here's another partial dead case:
1169 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=12395
1171 //===---------------------------------------------------------------------===//
1173 Scalar PRE hoists the mul in the common block up to the else:
1175 int test (int a, int b, int c, int g) {
1185 It would be better to do the mul once to reduce codesize above the if.
1186 This is GCC PR38204.
1188 //===---------------------------------------------------------------------===//
1190 GCC PR37810 is an interesting case where we should sink load/store reload
1191 into the if block and outside the loop, so we don't reload/store it on the
1212 We now hoist the reload after the call (Transforms/GVN/lpre-call-wrap.ll), but
1213 we don't sink the store. We need partially dead store sinking.
1215 //===---------------------------------------------------------------------===//
1217 [PHI TRANSLATE GEPs]
1219 GCC PR37166: Sinking of loads prevents SROA'ing the "g" struct on the stack
1220 leading to excess stack traffic. This could be handled by GVN with some crazy
1221 symbolic phi translation. The code we get looks like (g is on the stack):
1225 %9 = getelementptr %struct.f* %g, i32 0, i32 0
1226 store i32 %8, i32* %9, align bel %bb3
1228 bb3: ; preds = %bb1, %bb2, %bb
1229 %c_addr.0 = phi %struct.f* [ %g, %bb2 ], [ %c, %bb ], [ %c, %bb1 ]
1230 %b_addr.0 = phi %struct.f* [ %b, %bb2 ], [ %g, %bb ], [ %b, %bb1 ]
1231 %10 = getelementptr %struct.f* %c_addr.0, i32 0, i32 0
1232 %11 = load i32* %10, align 4
1234 %11 is fully redundant, an in BB2 it should have the value %8.
1236 GCC PR33344 is a similar case.
1238 //===---------------------------------------------------------------------===//
1240 There are many load PRE testcases in testsuite/gcc.dg/tree-ssa/loadpre* in the
1241 GCC testsuite. There are many pre testcases as ssa-pre-*.c
1243 //===---------------------------------------------------------------------===//
1245 There are some interesting cases in testsuite/gcc.dg/tree-ssa/pred-comm* in the
1246 GCC testsuite. For example, predcom-1.c is:
1248 for (i = 2; i < 1000; i++)
1249 fib[i] = (fib[i-1] + fib[i - 2]) & 0xffff;
1251 which compiles into:
1253 bb1: ; preds = %bb1, %bb1.thread
1254 %indvar = phi i32 [ 0, %bb1.thread ], [ %0, %bb1 ]
1255 %i.0.reg2mem.0 = add i32 %indvar, 2
1256 %0 = add i32 %indvar, 1 ; <i32> [#uses=3]
1257 %1 = getelementptr [1000 x i32]* @fib, i32 0, i32 %0
1258 %2 = load i32* %1, align 4 ; <i32> [#uses=1]
1259 %3 = getelementptr [1000 x i32]* @fib, i32 0, i32 %indvar
1260 %4 = load i32* %3, align 4 ; <i32> [#uses=1]
1261 %5 = add i32 %4, %2 ; <i32> [#uses=1]
1262 %6 = and i32 %5, 65535 ; <i32> [#uses=1]
1263 %7 = getelementptr [1000 x i32]* @fib, i32 0, i32 %i.0.reg2mem.0
1264 store i32 %6, i32* %7, align 4
1265 %exitcond = icmp eq i32 %0, 998 ; <i1> [#uses=1]
1266 br i1 %exitcond, label %return, label %bb1
1273 instead of handling this as a loop or other xform, all we'd need to do is teach
1274 load PRE to phi translate the %0 add (i+1) into the predecessor as (i'+1+1) =
1275 (i'+2) (where i' is the previous iteration of i). This would find the store
1278 predcom-2.c is apparently the same as predcom-1.c
1279 predcom-3.c is very similar but needs loads feeding each other instead of
1281 predcom-4.c seems the same as the rest.
1284 //===---------------------------------------------------------------------===//
1286 Other simple load PRE cases:
1287 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35287 [LPRE crit edge splitting]
1289 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34677 (licm does this, LPRE crit edge)
1290 llvm-gcc t2.c -S -o - -O0 -emit-llvm | llvm-as | opt -mem2reg -simplifycfg -gvn | llvm-dis
1292 //===---------------------------------------------------------------------===//
1294 Type based alias analysis:
1295 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14705
1297 //===---------------------------------------------------------------------===//
1299 When GVN/PRE finds a store of float* to a must aliases pointer when expecting
1300 an int*, it should turn it into a bitcast. This is a nice generalization of
1301 the SROA hack that would apply to other cases, e.g.:
1303 int foo(int C, int *P, float X) {
1314 One example (that requires crazy phi translation) is:
1315 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16799 [BITCAST PHI TRANS]
1317 //===---------------------------------------------------------------------===//
1319 A/B get pinned to the stack because we turn an if/then into a select instead
1320 of PRE'ing the load/store. This may be fixable in instcombine:
1321 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37892
1325 Interesting missed case because of control flow flattening (should be 2 loads):
1326 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26629
1327 With: llvm-gcc t2.c -S -o - -O0 -emit-llvm | llvm-as |
1328 opt -mem2reg -gvn -instcombine | llvm-dis
1329 we miss it because we need 1) GEP PHI TRAN, 2) CRIT EDGE 3) MULTIPLE DIFFERENT
1330 VALS PRODUCED BY ONE BLOCK OVER DIFFERENT PATHS
1332 //===---------------------------------------------------------------------===//
1334 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19633
1335 We could eliminate the branch condition here, loading from null is undefined:
1337 struct S { int w, x, y, z; };
1338 struct T { int r; struct S s; };
1339 void bar (struct S, int);
1340 void foo (int a, struct T b)
1348 //===---------------------------------------------------------------------===//
1350 simplifylibcalls should do several optimizations for strspn/strcspn:
1352 strcspn(x, "") -> strlen(x)
1355 strspn(x, "") -> strlen(x)
1356 strspn(x, "a") -> strchr(x, 'a')-x
1358 strcspn(x, "a") -> inlined loop for up to 3 letters (similarly for strspn):
1360 size_t __strcspn_c3 (__const char *__s, int __reject1, int __reject2,
1362 register size_t __result = 0;
1363 while (__s[__result] != '\0' && __s[__result] != __reject1 &&
1364 __s[__result] != __reject2 && __s[__result] != __reject3)
1369 This should turn into a switch on the character. See PR3253 for some notes on
1372 456.hmmer apparently uses strcspn and strspn a lot. 471.omnetpp uses strspn.
1374 //===---------------------------------------------------------------------===//
1376 "gas" uses this idiom:
1377 else if (strchr ("+-/*%|&^:[]()~", *intel_parser.op_string))
1379 else if (strchr ("<>", *intel_parser.op_string)
1381 Those should be turned into a switch.
1383 //===---------------------------------------------------------------------===//
1385 252.eon contains this interesting code:
1387 %3072 = getelementptr [100 x i8]* %tempString, i32 0, i32 0
1388 %3073 = call i8* @strcpy(i8* %3072, i8* %3071) nounwind
1389 %strlen = call i32 @strlen(i8* %3072) ; uses = 1
1390 %endptr = getelementptr [100 x i8]* %tempString, i32 0, i32 %strlen
1391 call void @llvm.memcpy.i32(i8* %endptr,
1392 i8* getelementptr ([5 x i8]* @"\01LC42", i32 0, i32 0), i32 5, i32 1)
1393 %3074 = call i32 @strlen(i8* %endptr) nounwind readonly
1395 This is interesting for a couple reasons. First, in this:
1397 %3073 = call i8* @strcpy(i8* %3072, i8* %3071) nounwind
1398 %strlen = call i32 @strlen(i8* %3072)
1400 The strlen could be replaced with: %strlen = sub %3072, %3073, because the
1401 strcpy call returns a pointer to the end of the string. Based on that, the
1402 endptr GEP just becomes equal to 3073, which eliminates a strlen call and GEP.
1404 Second, the memcpy+strlen strlen can be replaced with:
1406 %3074 = call i32 @strlen([5 x i8]* @"\01LC42") nounwind readonly
1408 Because the destination was just copied into the specified memory buffer. This,
1409 in turn, can be constant folded to "4".
1411 In other code, it contains:
1413 %endptr6978 = bitcast i8* %endptr69 to i32*
1414 store i32 7107374, i32* %endptr6978, align 1
1415 %3167 = call i32 @strlen(i8* %endptr69) nounwind readonly
1417 Which could also be constant folded. Whatever is producing this should probably
1418 be fixed to leave this as a memcpy from a string.
1420 Further, eon also has an interesting partially redundant strlen call:
1422 bb8: ; preds = %_ZN18eonImageCalculatorC1Ev.exit
1423 %682 = getelementptr i8** %argv, i32 6 ; <i8**> [#uses=2]
1424 %683 = load i8** %682, align 4 ; <i8*> [#uses=4]
1425 %684 = load i8* %683, align 1 ; <i8> [#uses=1]
1426 %685 = icmp eq i8 %684, 0 ; <i1> [#uses=1]
1427 br i1 %685, label %bb10, label %bb9
1430 %686 = call i32 @strlen(i8* %683) nounwind readonly
1431 %687 = icmp ugt i32 %686, 254 ; <i1> [#uses=1]
1432 br i1 %687, label %bb10, label %bb11
1434 bb10: ; preds = %bb9, %bb8
1435 %688 = call i32 @strlen(i8* %683) nounwind readonly
1437 This could be eliminated by doing the strlen once in bb8, saving code size and
1438 improving perf on the bb8->9->10 path.
1440 //===---------------------------------------------------------------------===//
1442 I see an interesting fully redundant call to strlen left in 186.crafty:InputMove
1444 %movetext11 = getelementptr [128 x i8]* %movetext, i32 0, i32 0
1447 bb62: ; preds = %bb55, %bb53
1448 %promote.0 = phi i32 [ %169, %bb55 ], [ 0, %bb53 ]
1449 %171 = call i32 @strlen(i8* %movetext11) nounwind readonly align 1
1450 %172 = add i32 %171, -1 ; <i32> [#uses=1]
1451 %173 = getelementptr [128 x i8]* %movetext, i32 0, i32 %172
1454 br i1 %or.cond, label %bb65, label %bb72
1456 bb65: ; preds = %bb62
1457 store i8 0, i8* %173, align 1
1460 bb72: ; preds = %bb65, %bb62
1461 %trank.1 = phi i32 [ %176, %bb65 ], [ -1, %bb62 ]
1462 %177 = call i32 @strlen(i8* %movetext11) nounwind readonly align 1
1464 Note that on the bb62->bb72 path, that the %177 strlen call is partially
1465 redundant with the %171 call. At worst, we could shove the %177 strlen call
1466 up into the bb65 block moving it out of the bb62->bb72 path. However, note
1467 that bb65 stores to the string, zeroing out the last byte. This means that on
1468 that path the value of %177 is actually just %171-1. A sub is cheaper than a
1471 This pattern repeats several times, basically doing:
1476 where it is "obvious" that B = A-1.
1478 //===---------------------------------------------------------------------===//
1480 186.crafty contains this interesting pattern:
1482 %77 = call i8* @strstr(i8* getelementptr ([6 x i8]* @"\01LC5", i32 0, i32 0),
1484 %phitmp648 = icmp eq i8* %77, getelementptr ([6 x i8]* @"\01LC5", i32 0, i32 0)
1485 br i1 %phitmp648, label %bb70, label %bb76
1487 bb70: ; preds = %OptionMatch.exit91, %bb69
1488 %78 = call i32 @strlen(i8* %30) nounwind readonly align 1 ; <i32> [#uses=1]
1492 if (strstr(cststr, P) == cststr) {
1496 The strstr call would be significantly cheaper written as:
1499 if (memcmp(P, str, strlen(P)))
1502 This is memcmp+strlen instead of strstr. This also makes the strlen fully
1505 //===---------------------------------------------------------------------===//
1507 186.crafty also contains this code:
1509 %1906 = call i32 @strlen(i8* getelementptr ([32 x i8]* @pgn_event, i32 0,i32 0))
1510 %1907 = getelementptr [32 x i8]* @pgn_event, i32 0, i32 %1906
1511 %1908 = call i8* @strcpy(i8* %1907, i8* %1905) nounwind align 1
1512 %1909 = call i32 @strlen(i8* getelementptr ([32 x i8]* @pgn_event, i32 0,i32 0))
1513 %1910 = getelementptr [32 x i8]* @pgn_event, i32 0, i32 %1909
1515 The last strlen is computable as 1908-@pgn_event, which means 1910=1908.
1517 //===---------------------------------------------------------------------===//
1519 186.crafty has this interesting pattern with the "out.4543" variable:
1521 call void @llvm.memcpy.i32(
1522 i8* getelementptr ([10 x i8]* @out.4543, i32 0, i32 0),
1523 i8* getelementptr ([7 x i8]* @"\01LC28700", i32 0, i32 0), i32 7, i32 1)
1524 %101 = call@printf(i8* ... @out.4543, i32 0, i32 0)) nounwind
1526 It is basically doing:
1528 memcpy(globalarray, "string");
1529 printf(..., globalarray);
1531 Anyway, by knowing that printf just reads the memory and forward substituting
1532 the string directly into the printf, this eliminates reads from globalarray.
1533 Since this pattern occurs frequently in crafty (due to the "DisplayTime" and
1534 other similar functions) there are many stores to "out". Once all the printfs
1535 stop using "out", all that is left is the memcpy's into it. This should allow
1536 globalopt to remove the "stored only" global.
1538 //===---------------------------------------------------------------------===//
1542 define inreg i32 @foo(i8* inreg %p) nounwind {
1544 %tmp1 = ashr i8 %tmp0, 5
1545 %tmp2 = sext i8 %tmp1 to i32
1549 could be dagcombine'd to a sign-extending load with a shift.
1550 For example, on x86 this currently gets this:
1556 while it could get this:
1561 //===---------------------------------------------------------------------===//
1565 int test(int x) { return 1-x == x; } // --> return false
1566 int test2(int x) { return 2-x == x; } // --> return x == 1 ?
1568 Always foldable for odd constants, what is the rule for even?
1570 //===---------------------------------------------------------------------===//
1572 PR 3381: GEP to field of size 0 inside a struct could be turned into GEP
1573 for next field in struct (which is at same address).
1575 For example: store of float into { {{}}, float } could be turned into a store to
1578 //===---------------------------------------------------------------------===//
1581 double foo(double a) { return sin(a); }
1583 This compiles into this on x86-64 Linux:
1594 //===---------------------------------------------------------------------===//
1596 The arg promotion pass should make use of nocapture to make its alias analysis
1597 stuff much more precise.
1599 //===---------------------------------------------------------------------===//
1601 The following functions should be optimized to use a select instead of a
1602 branch (from gcc PR40072):
1604 char char_int(int m) {if(m>7) return 0; return m;}
1605 int int_char(char m) {if(m>7) return 0; return m;}
1607 //===---------------------------------------------------------------------===//
1609 Instcombine should replace the load with a constant in:
1611 static const char x[4] = {'a', 'b', 'c', 'd'};
1613 unsigned int y(void) {
1614 return *(unsigned int *)x;
1617 It currently only does this transformation when the size of the constant
1618 is the same as the size of the integer (so, try x[5]) and the last byte
1619 is a null (making it a C string). There's no need for these restrictions.
1621 //===---------------------------------------------------------------------===//
1623 InstCombine's "turn load from constant into constant" optimization should be
1624 more aggressive in the presence of bitcasts. For example, because of unions,
1629 double v __attribute__((vector_size(16)));
1631 typedef union vec2d vec2d;
1633 static vec2d a={{1,2}}, b={{3,4}};
1636 return (vec2d){ .v = a.v + b.v * (vec2d){{5,5}}.v };
1641 @a = internal constant %0 { [2 x double]
1642 [double 1.000000e+00, double 2.000000e+00] }, align 16
1643 @b = internal constant %0 { [2 x double]
1644 [double 3.000000e+00, double 4.000000e+00] }, align 16
1646 define void @foo(%struct.vec2d* noalias nocapture sret %agg.result) nounwind {
1648 %0 = load <2 x double>* getelementptr (%struct.vec2d*
1649 bitcast (%0* @a to %struct.vec2d*), i32 0, i32 0), align 16
1650 %1 = load <2 x double>* getelementptr (%struct.vec2d*
1651 bitcast (%0* @b to %struct.vec2d*), i32 0, i32 0), align 16
1654 Instcombine should be able to optimize away the loads (and thus the globals).
1658 //===---------------------------------------------------------------------===//
1660 I saw this constant expression in real code after llvm-g++ -O2:
1662 declare extern_weak i32 @0(i64)
1664 define void @foo() {
1665 br i1 icmp eq (i32 zext (i1 icmp ne (i32 (i64)* @0, i32 (i64)* null) to i32),
1666 i32 0), label %cond_true, label %cond_false
1673 That branch expression should be reduced to:
1675 i1 icmp eq (i32 (i64)* @0, i32 (i64)* null)
1677 It's probably not a perf issue, I just happened to see it while examining
1678 something else and didn't want to forget about it.
1680 //===---------------------------------------------------------------------===//