X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTarget%2FREADME.txt;h=6d42d812a00bf3753843531996f3021a0d60c792;hb=867fe8570f299a058f155f98646d85cabc27155b;hp=9831efbdf64d247bb8a0cdb348d0f95ac784171a;hpb=db03983238ac7cebd9f09d453847fcc4ee30b8d2;p=oota-llvm.git diff --git a/lib/Target/README.txt b/lib/Target/README.txt index 9831efbdf64..6d42d812a00 100644 --- a/lib/Target/README.txt +++ b/lib/Target/README.txt @@ -2,10 +2,58 @@ Target Independent Opportunities: //===---------------------------------------------------------------------===// -We should make the various target's "IMPLICIT_DEF" instructions be a single -target-independent opcode like TargetInstrInfo::INLINEASM. This would allow -us to eliminate the TargetInstrDesc::isImplicitDef() method, and would allow -us to avoid having to define this for every target for every register class. +We should recognize idioms for add-with-carry and turn it into the appropriate +intrinsics. This example: + +unsigned add32carry(unsigned sum, unsigned x) { + unsigned z = sum + x; + if (sum + x < x) + z++; + return z; +} + +Compiles to: clang t.c -S -o - -O3 -fomit-frame-pointer -m64 -mkernel + +_add32carry: ## @add32carry + addl %esi, %edi + cmpl %esi, %edi + sbbl %eax, %eax + andl $1, %eax + addl %edi, %eax + ret + +with clang, but to: + +_add32carry: + leal (%rsi,%rdi), %eax + cmpl %esi, %eax + adcl $0, %eax + ret + +with gcc. + +//===---------------------------------------------------------------------===// + +Dead argument elimination should be enhanced to handle cases when an argument is +dead to an externally visible function. Though the argument can't be removed +from the externally visible function, the caller doesn't need to pass it in. +For example in this testcase: + + void foo(int X) __attribute__((noinline)); + void foo(int X) { sideeffect(); } + void bar(int A) { foo(A+1); } + +We compile bar to: + +define void @bar(i32 %A) nounwind ssp { + %0 = add nsw i32 %A, 1 ; [#uses=1] + tail call void @foo(i32 %0) nounwind noinline ssp + ret void +} + +The add is dead, we could pass in 'i32 undef' instead. This occurs for C++ +templates etc, which usually have linkonce_odr/weak_odr linkage, not internal +linkage. //===---------------------------------------------------------------------===// @@ -30,7 +78,10 @@ Make the PPC branch selector target independant //===---------------------------------------------------------------------===// Get the C front-end to expand hypot(x,y) -> llvm.sqrt(x*x+y*y) when errno and -precision don't matter (ffastmath). Misc/mandel will like this. :) +precision don't matter (ffastmath). Misc/mandel will like this. :) This isn't +safe in general, even on darwin. See the libm implementation of hypot for +examples (which special case when x/y are exactly zero to get signed zeros etc +right). //===---------------------------------------------------------------------===// @@ -87,7 +138,17 @@ Shrink: (setlt (loadi32 P), 0) -> (setlt (loadi8 Phi), 0) //===---------------------------------------------------------------------===// -Reassociate should turn: X*X*X*X -> t=(X*X) (t*t) to eliminate a multiply. +Reassociate should turn things like: + +int factorial(int X) { + return X*X*X*X*X*X*X*X; +} + +into llvm.powi calls, allowing the code generator to produce balanced +multiplication trees. + +First, the intrinsic needs to be extended to support integers, and second the +code generator needs to be enhanced to lower these to multiplication trees. //===---------------------------------------------------------------------===// @@ -100,7 +161,71 @@ int foo(int z, int n) { return bar(z, n) + bar(2*z, 2*n); } -Reassociate should handle the example in GCC PR16157. +This is blocked on not handling X*X*X -> powi(X, 3) (see note above). The issue +is that we end up getting t = 2*X s = t*t and don't turn this into 4*X*X, +which is the same number of multiplies and is canonical, because the 2*X has +multiple uses. Here's a simple example: + +define i32 @test15(i32 %X1) { + %B = mul i32 %X1, 47 ; X1*47 + %C = mul i32 %B, %B + ret i32 %C +} + + +//===---------------------------------------------------------------------===// + +Reassociate should handle the example in GCC PR16157: + +extern int a0, a1, a2, a3, a4; extern int b0, b1, b2, b3, b4; +void f () { /* this can be optimized to four additions... */ + b4 = a4 + a3 + a2 + a1 + a0; + b3 = a3 + a2 + a1 + a0; + b2 = a2 + a1 + a0; + b1 = a1 + a0; +} + +This requires reassociating to forms of expressions that are already available, +something that reassoc doesn't think about yet. + + +//===---------------------------------------------------------------------===// + +This function: (derived from GCC PR19988) +double foo(double x, double y) { + return ((x + 0.1234 * y) * (x + -0.1234 * y)); +} + +compiles to: +_foo: + movapd %xmm1, %xmm2 + mulsd LCPI1_1(%rip), %xmm1 + mulsd LCPI1_0(%rip), %xmm2 + addsd %xmm0, %xmm1 + addsd %xmm0, %xmm2 + movapd %xmm1, %xmm0 + mulsd %xmm2, %xmm0 + ret + +Reassociate should be able to turn it into: + +double foo(double x, double y) { + return ((x + 0.1234 * y) * (x - 0.1234 * y)); +} + +Which allows the multiply by constant to be CSE'd, producing: + +_foo: + mulsd LCPI1_0(%rip), %xmm1 + movapd %xmm1, %xmm2 + addsd %xmm0, %xmm2 + subsd %xmm1, %xmm0 + mulsd %xmm2, %xmm0 + ret + +This doesn't need -ffast-math support at all. This is particularly bad because +the llvm-gcc frontend is canonicalizing the later into the former, but clang +doesn't have this problem. //===---------------------------------------------------------------------===// @@ -129,8 +254,7 @@ specific vector types are target dependent. //===---------------------------------------------------------------------===// -We should add 'unaligned load/store' nodes, and produce them from code like -this: +We should produce an unaligned load from code like this: v4sf example(float *P) { return (v4sf){P[0], P[1], P[2], P[3] }; @@ -166,81 +290,8 @@ Expand these to calls of sin/cos and stores: Doing so could allow SROA of the destination pointers. See also: http://gcc.gnu.org/bugzilla/show_bug.cgi?id=17687 -//===---------------------------------------------------------------------===// - -Scalar Repl cannot currently promote this testcase to 'ret long cst': - - %struct.X = type { i32, i32 } - %struct.Y = type { %struct.X } - -define i64 @bar() { - %retval = alloca %struct.Y, align 8 - %tmp12 = getelementptr %struct.Y* %retval, i32 0, i32 0, i32 0 - store i32 0, i32* %tmp12 - %tmp15 = getelementptr %struct.Y* %retval, i32 0, i32 0, i32 1 - store i32 1, i32* %tmp15 - %retval.upgrd.1 = bitcast %struct.Y* %retval to i64* - %retval.upgrd.2 = load i64* %retval.upgrd.1 - ret i64 %retval.upgrd.2 -} - -it should be extended to do so. - -//===---------------------------------------------------------------------===// - --scalarrepl should promote this to be a vector scalar. - - %struct..0anon = type { <4 x float> } - -define void @test1(<4 x float> %V, float* %P) { - %u = alloca %struct..0anon, align 16 - %tmp = getelementptr %struct..0anon* %u, i32 0, i32 0 - store <4 x float> %V, <4 x float>* %tmp - %tmp1 = bitcast %struct..0anon* %u to [4 x float]* - %tmp.upgrd.1 = getelementptr [4 x float]* %tmp1, i32 0, i32 1 - %tmp.upgrd.2 = load float* %tmp.upgrd.1 - %tmp3 = mul float %tmp.upgrd.2, 2.000000e+00 - store float %tmp3, float* %P - ret void -} - -//===---------------------------------------------------------------------===// - -Turn this into a single byte store with no load (the other 3 bytes are -unmodified): - -void %test(uint* %P) { - %tmp = load uint* %P - %tmp14 = or uint %tmp, 3305111552 - %tmp15 = and uint %tmp14, 3321888767 - store uint %tmp15, uint* %P - ret void -} - -//===---------------------------------------------------------------------===// - -dag/inst combine "clz(x)>>5 -> x==0" for 32-bit x. - -Compile: - -int bar(int x) -{ - int t = __builtin_clz(x); - return -(t>>5); -} - -to: - -_bar: addic r3,r3,-1 - subfe r3,r3,r3 - blr - -//===---------------------------------------------------------------------===// - -Legalize should lower ctlz like this: - ctlz(x) = popcnt((x-1) & ~x) - -on targets that have popcnt but not ctlz. itanium, what else? +This is now easily doable with MRVs. We could even make an intrinsic for this +if anyone cared enough about sincos. //===---------------------------------------------------------------------===// @@ -267,7 +318,7 @@ so cool to turn it into something like: ... which would only do one 32-bit XOR per loop iteration instead of two. It would also be nice to recognize the reg->size doesn't alias reg->node[i], but -alas... +this requires TBAA. //===---------------------------------------------------------------------===// @@ -281,8 +332,18 @@ unsigned long reverse(unsigned v) { return v ^ (t >> 8); } +Neither is this (very standard idiom): + +int f(int n) +{ + return (((n) << 24) | (((n) & 0xff00) << 8) + | (((n) >> 8) & 0xff00) | ((n) >> 24)); +} + //===---------------------------------------------------------------------===// +[LOOP RECOGNITION] + These idioms should be recognized as popcount (see PR1488): unsigned countbits_slow(unsigned v) { @@ -314,6 +375,9 @@ unsigned int popcount(unsigned int input) { return count; } +This is a form of idiom recognition for loops, the same thing that could be +useful for recognizing memset/memcpy. + //===---------------------------------------------------------------------===// These should turn into single 16-bit (unaligned?) loads on little/big endian @@ -342,12 +406,36 @@ this construct. //===---------------------------------------------------------------------===// +[LOOP RECOGNITION] + viterbi speeds up *significantly* if the various "history" related copy loops are turned into memcpy calls at the source level. We need a "loops to memcpy" pass. //===---------------------------------------------------------------------===// +[LOOP OPTIMIZATION] + +SingleSource/Benchmarks/Misc/dt.c shows several interesting optimization +opportunities in its double_array_divs_variable function: it needs loop +interchange, memory promotion (which LICM already does), vectorization and +variable trip count loop unrolling (since it has a constant trip count). ICC +apparently produces this very nice code with -ffast-math: + +..B1.70: # Preds ..B1.70 ..B1.69 + mulpd %xmm0, %xmm1 #108.2 + mulpd %xmm0, %xmm1 #108.2 + mulpd %xmm0, %xmm1 #108.2 + mulpd %xmm0, %xmm1 #108.2 + addl $8, %edx # + cmpl $131072, %edx #108.2 + jb ..B1.70 # Prob 99% #108.2 + +It would be better to count down to zero, but this is a lot better than what we +do. + +//===---------------------------------------------------------------------===// + Consider: typedef unsigned U32; @@ -373,14 +461,11 @@ we don't have whole-function selection dags. On x86, this means we use one extra register for the function when effective_addr2 is declared as U64 than when it is declared U32. -//===---------------------------------------------------------------------===// - -Promote for i32 bswap can use i64 bswap + shr. Useful on targets with 64-bit -regs and bswap, like itanium. +PHI Slicing could be extended to do this. //===---------------------------------------------------------------------===// -LSR should know what GPR types a target has. This code: +LSR should know what GPR types a target has from TargetData. This code: volatile short X, Y; // globals @@ -389,25 +474,22 @@ void foo(int N) { for (i = 0; i < N; i++) { X = i; Y = i*4; } } -produces two identical IV's (after promotion) on PPC/ARM: +produces two near identical IV's (after promotion) on PPC/ARM: -LBB1_1: @bb.preheader - mov r3, #0 - mov r2, r3 - mov r1, r3 -LBB1_2: @bb - ldr r12, LCPI1_0 - ldr r12, [r12] - strh r2, [r12] - ldr r12, LCPI1_1 - ldr r12, [r12] - strh r3, [r12] - add r1, r1, #1 <- [0,+,1] - add r3, r3, #4 - add r2, r2, #1 <- [0,+,1] - cmp r1, r0 - bne LBB1_2 @bb +LBB1_2: + ldr r3, LCPI1_0 + ldr r3, [r3] + strh r2, [r3] + ldr r3, LCPI1_1 + ldr r3, [r3] + strh r1, [r3] + add r1, r1, #4 + add r2, r2, #1 <- [0,+,1] + sub r0, r0, #1 <- [0,-,1] + cmp r0, #0 + bne LBB1_2 +LSR should reuse the "+" IV for the exit test. //===---------------------------------------------------------------------===// @@ -447,22 +529,6 @@ return: ; preds = %then.1, %else.0, %then.0 //===---------------------------------------------------------------------===// -Tail recursion elimination is not transforming this function, because it is -returning n, which fails the isDynamicConstant check in the accumulator -recursion checks. - -long long fib(const long long n) { - switch(n) { - case 0: - case 1: - return n; - default: - return fib(n-1) + fib(n-2); - } -} - -//===---------------------------------------------------------------------===// - Tail recursion elimination should handle: int pow2m1(int n) { @@ -496,23 +562,6 @@ entry: //===---------------------------------------------------------------------===// -"basicaa" should know how to look through "or" instructions that act like add -instructions. For example in this code, the x*4+1 is turned into x*4 | 1, and -basicaa can't analyze the array subscript, leading to duplicated loads in the -generated code: - -void test(int X, int Y, int a[]) { -int i; - for (i=2; i<1000; i+=4) { - a[i+0] = a[i-1+0]*a[i-2+0]; - a[i+1] = a[i-1+1]*a[i-2+1]; - a[i+2] = a[i-1+2]*a[i-2+2]; - a[i+3] = a[i-1+3]*a[i-2+3]; - } -} - -//===---------------------------------------------------------------------===// - We should investigate an instruction sinking pass. Consider this silly example in pic mode: @@ -633,56 +682,11 @@ once. //===---------------------------------------------------------------------===// -We should extend parameter attributes to capture more information about -pointer parameters for alias analysis. Some ideas: - -1. Add a "nocapture" attribute, which indicates that the callee does not store - the address of the parameter into a global or any other memory location - visible to the callee. This can be used to make basicaa and other analyses - more powerful. It is true for things like memcpy, strcat, and many other - things, including structs passed by value, most C++ references, etc. -2. Generalize readonly to be set on parameters. This is important mod/ref - info for the function, which is important for basicaa and others. It can - also be used by the inliner to avoid inserting a memcpy for byval - arguments when the function is inlined. - -These functions can be inferred by various analysis passes such as the -globalsmodrefaa pass. Note that getting #2 right is actually really tricky. -Consider this code: - -struct S; S G; -void caller(S byvalarg) { G.field = 1; ... } -void callee() { caller(G); } - -The fact that the caller does not modify byval arg is not enough, we need -to know that it doesn't modify G either. This is very tricky. - -//===---------------------------------------------------------------------===// - We should add an FRINT node to the DAG to model targets that have legal implementations of ceil/floor/rint. //===---------------------------------------------------------------------===// -This GCC bug: http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34043 -contains a testcase that compiles down to: - - %struct.XMM128 = type { <4 x float> } -.. - %src = alloca %struct.XMM128 -.. - %tmp6263 = bitcast %struct.XMM128* %src to <2 x i64>* - %tmp65 = getelementptr %struct.XMM128* %src, i32 0, i32 0 - store <2 x i64> %tmp5899, <2 x i64>* %tmp6263, align 16 - %tmp66 = load <4 x float>* %tmp65, align 16 - %tmp71 = add <4 x float> %tmp66, %tmp66 - -If the mid-level optimizer turned the bitcast of pointer + store of tmp5899 -into a bitcast of the vector value and a store to the pointer, then the -store->load could be easily removed. - -//===---------------------------------------------------------------------===// - Consider: int test() { @@ -799,7 +803,7 @@ unsigned long long f6(unsigned long long x, unsigned long long y, int z) { } } -On X86-64, we only handle f3/f4 right. On x86-32, several of these +On X86-64, we only handle f2/f3/f4 right. On x86-32, a few of these generate truly horrible code, instead of using shld and friends. On ARM, we end up with calls to L___lshrdi3/L___ashldi3 in f, which is badness. PPC64 misses f, f5 and f6. CellSPU aborts in isel. @@ -813,57 +817,6 @@ be done safely if "b" isn't modified between the strlen and memcpy of course. //===---------------------------------------------------------------------===// -We should be able to evaluate this loop: - -int test(int x_offs) { - while (x_offs > 4) - x_offs -= 4; - return x_offs; -} - -//===---------------------------------------------------------------------===// - -Reassociate should turn things like: - -int factorial(int X) { - return X*X*X*X*X*X*X*X; -} - -into llvm.powi calls, allowing the code generator to produce balanced -multiplication trees. - -//===---------------------------------------------------------------------===// - -We generate a horrible libcall for llvm.powi. For example, we compile: - -#include -double f(double a) { return std::pow(a, 4); } - -into: - -__Z1fd: - subl $12, %esp - movsd 16(%esp), %xmm0 - movsd %xmm0, (%esp) - movl $4, 8(%esp) - call L___powidf2$stub - addl $12, %esp - ret - -GCC produces: - -__Z1fd: - subl $12, %esp - movsd 16(%esp), %xmm0 - mulsd %xmm0, %xmm0 - mulsd %xmm0, %xmm0 - movsd %xmm0, (%esp) - fldl (%esp) - addl $12, %esp - ret - -//===---------------------------------------------------------------------===// - We compile this program: (from GCC PR11680) http://gcc.gnu.org/bugzilla/attachment.cgi?id=4487 @@ -903,71 +856,1130 @@ void bar(unsigned n) { true(); } -I think this basically amounts to a dag combine to simplify comparisons against -multiply hi's into a comparison against the mullo. +This is equivalent to the following, where 2863311531 is the multiplicative +inverse of 3, and 1431655766 is ((2^32)-1)/3+1: +void bar(unsigned n) { + if (n * 2863311531U < 1431655766U) + true(); +} + +The same transformation can work with an even modulo with the addition of a +rotate: rotate the result of the multiply to the right by the number of bits +which need to be zero for the condition to be true, and shrink the compare RHS +by the same amount. Unless the target supports rotates, though, that +transformation probably isn't worthwhile. + +The transformation can also easily be made to work with non-zero equality +comparisons: just transform, for example, "n % 3 == 1" to "(n-1) % 3 == 0". //===---------------------------------------------------------------------===// -SROA is not promoting the union on the stack in this example, we should end -up with no allocas. +Better mod/ref analysis for scanf would allow us to eliminate the vtable and a +bunch of other stuff from this example (see PR1604): -union vec2d { - double e[2]; - double v __attribute__((vector_size(16))); +#include +struct test { + int val; + virtual ~test() {} }; -typedef union vec2d vec2d; -static vec2d a={{1,2}}, b={{3,4}}; - -vec2d foo () { - return (vec2d){ .v = a.v + b.v * (vec2d){{5,5}}.v }; +int main() { + test t; + std::scanf("%d", &t.val); + std::printf("%d\n", t.val); } //===---------------------------------------------------------------------===// -This C++ file: -void g(); struct A { int n; int m; A& operator++(void) { ++n; if (n == m) g(); -return *this; } A() : n(0), m(0) { } friend bool operator!=(A const& a1, -A const& a2) { return a1.n != a2.n; } }; void testfunction(A& iter) { A const -end; while (iter != end) ++iter; } +These functions perform the same computation, but produce different assembly. -Compiles down to: +define i8 @select(i8 %x) readnone nounwind { + %A = icmp ult i8 %x, 250 + %B = select i1 %A, i8 0, i8 1 + ret i8 %B +} -bb: ; preds = %bb3.backedge, %bb.nph - %.rle = phi i32 [ %1, %bb.nph ], [ %7, %bb3.backedge ] ; [#uses=1] - %4 = add i32 %.rle, 1 ; [#uses=2] - store i32 %4, i32* %0, align 4 - %5 = load i32* %3, align 4 ; [#uses=1] - %6 = icmp eq i32 %4, %5 ; [#uses=1] - br i1 %6, label %bb1, label %bb3.backedge +define i8 @addshr(i8 %x) readnone nounwind { + %A = zext i8 %x to i9 + %B = add i9 %A, 6 ;; 256 - 250 == 6 + %C = lshr i9 %B, 8 + %D = trunc i9 %C to i8 + ret i8 %D +} -bb1: ; preds = %bb - tail call void @_Z1gv() - br label %bb3.backedge +//===---------------------------------------------------------------------===// -bb3.backedge: ; preds = %bb, %bb1 - %7 = load i32* %0, align 4 ; [#uses=2] +From gcc bug 24696: +int +f (unsigned long a, unsigned long b, unsigned long c) +{ + return ((a & (c - 1)) != 0) || ((b & (c - 1)) != 0); +} +int +f (unsigned long a, unsigned long b, unsigned long c) +{ + return ((a & (c - 1)) != 0) | ((b & (c - 1)) != 0); +} +Both should combine to ((a|b) & (c-1)) != 0. Currently not optimized with +"clang -emit-llvm-bc | opt -std-compile-opts". +//===---------------------------------------------------------------------===// -The %7 load is partially redundant with the store of %4 to %0, GVN's PRE -should remove it, but it doesn't apply to memory objects. +From GCC Bug 20192: +#define PMD_MASK (~((1UL << 23) - 1)) +void clear_pmd_range(unsigned long start, unsigned long end) +{ + if (!(start & ~PMD_MASK) && !(end & ~PMD_MASK)) + f(); +} +The expression should optimize to something like +"!((start|end)&~PMD_MASK). Currently not optimized with "clang +-emit-llvm-bc | opt -std-compile-opts". //===---------------------------------------------------------------------===// -Better mod/ref analysis for scanf would allow us to eliminate the vtable and a -bunch of other stuff from this example (see PR1604): +void a(int variable) +{ + if (variable == 4 || variable == 6) + bar(); +} +This should optimize to "if ((variable | 2) == 6)". Currently not +optimized with "clang -emit-llvm-bc | opt -std-compile-opts | llc". -#include -struct test { - int val; - virtual ~test() {} -}; +//===---------------------------------------------------------------------===// -int main() { - test t; - std::scanf("%d", &t.val); - std::printf("%d\n", t.val); +unsigned int f(unsigned int i, unsigned int n) {++i; if (i == n) ++i; return +i;} +unsigned int f2(unsigned int i, unsigned int n) {++i; i += i == n; return i;} +These should combine to the same thing. Currently, the first function +produces better code on X86. + +//===---------------------------------------------------------------------===// + +From GCC Bug 15784: +#define abs(x) x>0?x:-x +int f(int x, int y) +{ + return (abs(x)) >= 0; +} +This should optimize to x == INT_MIN. (With -fwrapv.) Currently not +optimized with "clang -emit-llvm-bc | opt -std-compile-opts". + +//===---------------------------------------------------------------------===// + +From GCC Bug 14753: +void +rotate_cst (unsigned int a) +{ + a = (a << 10) | (a >> 22); + if (a == 123) + bar (); +} +void +minus_cst (unsigned int a) +{ + unsigned int tem; + + tem = 20 - a; + if (tem == 5) + bar (); +} +void +mask_gt (unsigned int a) +{ + /* This is equivalent to a > 15. */ + if ((a & ~7) > 8) + bar (); } +void +rshift_gt (unsigned int a) +{ + /* This is equivalent to a > 23. */ + if ((a >> 2) > 5) + bar (); +} +All should simplify to a single comparison. All of these are +currently not optimized with "clang -emit-llvm-bc | opt +-std-compile-opts". + +//===---------------------------------------------------------------------===// + +From GCC Bug 32605: +int c(int* x) {return (char*)x+2 == (char*)x;} +Should combine to 0. Currently not optimized with "clang +-emit-llvm-bc | opt -std-compile-opts" (although llc can optimize it). + +//===---------------------------------------------------------------------===// + +int a(unsigned b) {return ((b << 31) | (b << 30)) >> 31;} +Should be combined to "((b >> 1) | b) & 1". Currently not optimized +with "clang -emit-llvm-bc | opt -std-compile-opts". + +//===---------------------------------------------------------------------===// + +unsigned a(unsigned x, unsigned y) { return x | (y & 1) | (y & 2);} +Should combine to "x | (y & 3)". Currently not optimized with "clang +-emit-llvm-bc | opt -std-compile-opts". + +//===---------------------------------------------------------------------===// + +int a(int a, int b, int c) {return (~a & c) | ((c|a) & b);} +Should fold to "(~a & c) | (a & b)". Currently not optimized with +"clang -emit-llvm-bc | opt -std-compile-opts". + +//===---------------------------------------------------------------------===// + +int a(int a,int b) {return (~(a|b))|a;} +Should fold to "a|~b". Currently not optimized with "clang +-emit-llvm-bc | opt -std-compile-opts". + +//===---------------------------------------------------------------------===// + +int a(int a, int b) {return (a&&b) || (a&&!b);} +Should fold to "a". Currently not optimized with "clang -emit-llvm-bc +| opt -std-compile-opts". //===---------------------------------------------------------------------===// +int a(int a, int b, int c) {return (a&&b) || (!a&&c);} +Should fold to "a ? b : c", or at least something sane. Currently not +optimized with "clang -emit-llvm-bc | opt -std-compile-opts". + +//===---------------------------------------------------------------------===// + +int a(int a, int b, int c) {return (a&&b) || (a&&c) || (a&&b&&c);} +Should fold to a && (b || c). Currently not optimized with "clang +-emit-llvm-bc | opt -std-compile-opts". + +//===---------------------------------------------------------------------===// + +int a(int x) {return x | ((x & 8) ^ 8);} +Should combine to x | 8. Currently not optimized with "clang +-emit-llvm-bc | opt -std-compile-opts". + +//===---------------------------------------------------------------------===// + +int a(int x) {return x ^ ((x & 8) ^ 8);} +Should also combine to x | 8. Currently not optimized with "clang +-emit-llvm-bc | opt -std-compile-opts". + +//===---------------------------------------------------------------------===// + +int a(int x) {return (x & 8) == 0 ? -1 : -9;} +Should combine to (x | -9) ^ 8. Currently not optimized with "clang +-emit-llvm-bc | opt -std-compile-opts". + +//===---------------------------------------------------------------------===// + +int a(int x) {return (x & 8) == 0 ? -9 : -1;} +Should combine to x | -9. Currently not optimized with "clang +-emit-llvm-bc | opt -std-compile-opts". + +//===---------------------------------------------------------------------===// + +int a(int x) {return ((x | -9) ^ 8) & x;} +Should combine to x & -9. Currently not optimized with "clang +-emit-llvm-bc | opt -std-compile-opts". + +//===---------------------------------------------------------------------===// + +unsigned a(unsigned a) {return a * 0x11111111 >> 28 & 1;} +Should combine to "a * 0x88888888 >> 31". Currently not optimized +with "clang -emit-llvm-bc | opt -std-compile-opts". + +//===---------------------------------------------------------------------===// + +unsigned a(char* x) {if ((*x & 32) == 0) return b();} +There's an unnecessary zext in the generated code with "clang +-emit-llvm-bc | opt -std-compile-opts". + +//===---------------------------------------------------------------------===// + +unsigned a(unsigned long long x) {return 40 * (x >> 1);} +Should combine to "20 * (((unsigned)x) & -2)". Currently not +optimized with "clang -emit-llvm-bc | opt -std-compile-opts". + +//===---------------------------------------------------------------------===// + +This was noticed in the entryblock for grokdeclarator in 403.gcc: + + %tmp = icmp eq i32 %decl_context, 4 + %decl_context_addr.0 = select i1 %tmp, i32 3, i32 %decl_context + %tmp1 = icmp eq i32 %decl_context_addr.0, 1 + %decl_context_addr.1 = select i1 %tmp1, i32 0, i32 %decl_context_addr.0 + +tmp1 should be simplified to something like: + (!tmp || decl_context == 1) + +This allows recursive simplifications, tmp1 is used all over the place in +the function, e.g. by: + + %tmp23 = icmp eq i32 %decl_context_addr.1, 0 ; [#uses=1] + %tmp24 = xor i1 %tmp1, true ; [#uses=1] + %or.cond8 = and i1 %tmp23, %tmp24 ; [#uses=1] + +later. + +//===---------------------------------------------------------------------===// + +[STORE SINKING] + +Store sinking: This code: + +void f (int n, int *cond, int *res) { + int i; + *res = 0; + for (i = 0; i < n; i++) + if (*cond) + *res ^= 234; /* (*) */ +} + +On this function GVN hoists the fully redundant value of *res, but nothing +moves the store out. This gives us this code: + +bb: ; preds = %bb2, %entry + %.rle = phi i32 [ 0, %entry ], [ %.rle6, %bb2 ] + %i.05 = phi i32 [ 0, %entry ], [ %indvar.next, %bb2 ] + %1 = load i32* %cond, align 4 + %2 = icmp eq i32 %1, 0 + br i1 %2, label %bb2, label %bb1 + +bb1: ; preds = %bb + %3 = xor i32 %.rle, 234 + store i32 %3, i32* %res, align 4 + br label %bb2 + +bb2: ; preds = %bb, %bb1 + %.rle6 = phi i32 [ %3, %bb1 ], [ %.rle, %bb ] + %indvar.next = add i32 %i.05, 1 + %exitcond = icmp eq i32 %indvar.next, %n + br i1 %exitcond, label %return, label %bb + +DSE should sink partially dead stores to get the store out of the loop. + +Here's another partial dead case: +http://gcc.gnu.org/bugzilla/show_bug.cgi?id=12395 + +//===---------------------------------------------------------------------===// + +Scalar PRE hoists the mul in the common block up to the else: + +int test (int a, int b, int c, int g) { + int d, e; + if (a) + d = b * c; + else + d = b - c; + e = b * c + g; + return d + e; +} + +It would be better to do the mul once to reduce codesize above the if. +This is GCC PR38204. + +//===---------------------------------------------------------------------===// + +[STORE SINKING] + +GCC PR37810 is an interesting case where we should sink load/store reload +into the if block and outside the loop, so we don't reload/store it on the +non-call path. + +for () { + *P += 1; + if () + call(); + else + ... +-> +tmp = *P +for () { + tmp += 1; + if () { + *P = tmp; + call(); + tmp = *P; + } else ... +} +*P = tmp; + +We now hoist the reload after the call (Transforms/GVN/lpre-call-wrap.ll), but +we don't sink the store. We need partially dead store sinking. + +//===---------------------------------------------------------------------===// + +[LOAD PRE CRIT EDGE SPLITTING] + +GCC PR37166: Sinking of loads prevents SROA'ing the "g" struct on the stack +leading to excess stack traffic. This could be handled by GVN with some crazy +symbolic phi translation. The code we get looks like (g is on the stack): + +bb2: ; preds = %bb1 +.. + %9 = getelementptr %struct.f* %g, i32 0, i32 0 + store i32 %8, i32* %9, align bel %bb3 + +bb3: ; preds = %bb1, %bb2, %bb + %c_addr.0 = phi %struct.f* [ %g, %bb2 ], [ %c, %bb ], [ %c, %bb1 ] + %b_addr.0 = phi %struct.f* [ %b, %bb2 ], [ %g, %bb ], [ %b, %bb1 ] + %10 = getelementptr %struct.f* %c_addr.0, i32 0, i32 0 + %11 = load i32* %10, align 4 + +%11 is partially redundant, an in BB2 it should have the value %8. + +GCC PR33344 and PR35287 are similar cases. + + +//===---------------------------------------------------------------------===// + +[LOAD PRE] + +There are many load PRE testcases in testsuite/gcc.dg/tree-ssa/loadpre* in the +GCC testsuite, ones we don't get yet are (checked through loadpre25): + +[CRIT EDGE BREAKING] +loadpre3.c predcom-4.c + +[PRE OF READONLY CALL] +loadpre5.c + +[TURN SELECT INTO BRANCH] +loadpre14.c loadpre15.c + +actually a conditional increment: loadpre18.c loadpre19.c + + +//===---------------------------------------------------------------------===// + +[SCALAR PRE] +There are many PRE testcases in testsuite/gcc.dg/tree-ssa/ssa-pre-*.c in the +GCC testsuite. + +//===---------------------------------------------------------------------===// + +There are some interesting cases in testsuite/gcc.dg/tree-ssa/pred-comm* in the +GCC testsuite. For example, we get the first example in predcom-1.c, but +miss the second one: + +unsigned fib[1000]; +unsigned avg[1000]; + +__attribute__ ((noinline)) +void count_averages(int n) { + int i; + for (i = 1; i < n; i++) + avg[i] = (((unsigned long) fib[i - 1] + fib[i] + fib[i + 1]) / 3) & 0xffff; +} + +which compiles into two loads instead of one in the loop. + +predcom-2.c is the same as predcom-1.c + +predcom-3.c is very similar but needs loads feeding each other instead of +store->load. + + +//===---------------------------------------------------------------------===// + +[ALIAS ANALYSIS] + +Type based alias analysis: +http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14705 + +We should do better analysis of posix_memalign. At the least it should +no-capture its pointer argument, at best, we should know that the out-value +result doesn't point to anything (like malloc). One example of this is in +SingleSource/Benchmarks/Misc/dt.c + +//===---------------------------------------------------------------------===// + +A/B get pinned to the stack because we turn an if/then into a select instead +of PRE'ing the load/store. This may be fixable in instcombine: +http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37892 + +struct X { int i; }; +int foo (int x) { + struct X a; + struct X b; + struct X *p; + a.i = 1; + b.i = 2; + if (x) + p = &a; + else + p = &b; + return p->i; +} + +//===---------------------------------------------------------------------===// + +Interesting missed case because of control flow flattening (should be 2 loads): +http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26629 +With: llvm-gcc t2.c -S -o - -O0 -emit-llvm | llvm-as | + opt -mem2reg -gvn -instcombine | llvm-dis +we miss it because we need 1) CRIT EDGE 2) MULTIPLE DIFFERENT +VALS PRODUCED BY ONE BLOCK OVER DIFFERENT PATHS + +//===---------------------------------------------------------------------===// + +http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19633 +We could eliminate the branch condition here, loading from null is undefined: + +struct S { int w, x, y, z; }; +struct T { int r; struct S s; }; +void bar (struct S, int); +void foo (int a, struct T b) +{ + struct S *c = 0; + if (a) + c = &b.s; + bar (*c, a); +} + +//===---------------------------------------------------------------------===// + +simplifylibcalls should do several optimizations for strspn/strcspn: + +strcspn(x, "a") -> inlined loop for up to 3 letters (similarly for strspn): + +size_t __strcspn_c3 (__const char *__s, int __reject1, int __reject2, + int __reject3) { + register size_t __result = 0; + while (__s[__result] != '\0' && __s[__result] != __reject1 && + __s[__result] != __reject2 && __s[__result] != __reject3) + ++__result; + return __result; +} + +This should turn into a switch on the character. See PR3253 for some notes on +codegen. + +456.hmmer apparently uses strcspn and strspn a lot. 471.omnetpp uses strspn. + +//===---------------------------------------------------------------------===// + +"gas" uses this idiom: + else if (strchr ("+-/*%|&^:[]()~", *intel_parser.op_string)) +.. + else if (strchr ("<>", *intel_parser.op_string) + +Those should be turned into a switch. + +//===---------------------------------------------------------------------===// + +252.eon contains this interesting code: + + %3072 = getelementptr [100 x i8]* %tempString, i32 0, i32 0 + %3073 = call i8* @strcpy(i8* %3072, i8* %3071) nounwind + %strlen = call i32 @strlen(i8* %3072) ; uses = 1 + %endptr = getelementptr [100 x i8]* %tempString, i32 0, i32 %strlen + call void @llvm.memcpy.i32(i8* %endptr, + i8* getelementptr ([5 x i8]* @"\01LC42", i32 0, i32 0), i32 5, i32 1) + %3074 = call i32 @strlen(i8* %endptr) nounwind readonly + +This is interesting for a couple reasons. First, in this: + + %3073 = call i8* @strcpy(i8* %3072, i8* %3071) nounwind + %strlen = call i32 @strlen(i8* %3072) + +The strlen could be replaced with: %strlen = sub %3072, %3073, because the +strcpy call returns a pointer to the end of the string. Based on that, the +endptr GEP just becomes equal to 3073, which eliminates a strlen call and GEP. + +Second, the memcpy+strlen strlen can be replaced with: + + %3074 = call i32 @strlen([5 x i8]* @"\01LC42") nounwind readonly + +Because the destination was just copied into the specified memory buffer. This, +in turn, can be constant folded to "4". + +In other code, it contains: + + %endptr6978 = bitcast i8* %endptr69 to i32* + store i32 7107374, i32* %endptr6978, align 1 + %3167 = call i32 @strlen(i8* %endptr69) nounwind readonly + +Which could also be constant folded. Whatever is producing this should probably +be fixed to leave this as a memcpy from a string. + +Further, eon also has an interesting partially redundant strlen call: + +bb8: ; preds = %_ZN18eonImageCalculatorC1Ev.exit + %682 = getelementptr i8** %argv, i32 6 ; [#uses=2] + %683 = load i8** %682, align 4 ; [#uses=4] + %684 = load i8* %683, align 1 ; [#uses=1] + %685 = icmp eq i8 %684, 0 ; [#uses=1] + br i1 %685, label %bb10, label %bb9 + +bb9: ; preds = %bb8 + %686 = call i32 @strlen(i8* %683) nounwind readonly + %687 = icmp ugt i32 %686, 254 ; [#uses=1] + br i1 %687, label %bb10, label %bb11 + +bb10: ; preds = %bb9, %bb8 + %688 = call i32 @strlen(i8* %683) nounwind readonly + +This could be eliminated by doing the strlen once in bb8, saving code size and +improving perf on the bb8->9->10 path. + +//===---------------------------------------------------------------------===// + +I see an interesting fully redundant call to strlen left in 186.crafty:InputMove +which looks like: + %movetext11 = getelementptr [128 x i8]* %movetext, i32 0, i32 0 + + +bb62: ; preds = %bb55, %bb53 + %promote.0 = phi i32 [ %169, %bb55 ], [ 0, %bb53 ] + %171 = call i32 @strlen(i8* %movetext11) nounwind readonly align 1 + %172 = add i32 %171, -1 ; [#uses=1] + %173 = getelementptr [128 x i8]* %movetext, i32 0, i32 %172 + +... no stores ... + br i1 %or.cond, label %bb65, label %bb72 + +bb65: ; preds = %bb62 + store i8 0, i8* %173, align 1 + br label %bb72 + +bb72: ; preds = %bb65, %bb62 + %trank.1 = phi i32 [ %176, %bb65 ], [ -1, %bb62 ] + %177 = call i32 @strlen(i8* %movetext11) nounwind readonly align 1 + +Note that on the bb62->bb72 path, that the %177 strlen call is partially +redundant with the %171 call. At worst, we could shove the %177 strlen call +up into the bb65 block moving it out of the bb62->bb72 path. However, note +that bb65 stores to the string, zeroing out the last byte. This means that on +that path the value of %177 is actually just %171-1. A sub is cheaper than a +strlen! + +This pattern repeats several times, basically doing: + + A = strlen(P); + P[A-1] = 0; + B = strlen(P); + where it is "obvious" that B = A-1. + +//===---------------------------------------------------------------------===// + +186.crafty also contains this code: + +%1906 = call i32 @strlen(i8* getelementptr ([32 x i8]* @pgn_event, i32 0,i32 0)) +%1907 = getelementptr [32 x i8]* @pgn_event, i32 0, i32 %1906 +%1908 = call i8* @strcpy(i8* %1907, i8* %1905) nounwind align 1 +%1909 = call i32 @strlen(i8* getelementptr ([32 x i8]* @pgn_event, i32 0,i32 0)) +%1910 = getelementptr [32 x i8]* @pgn_event, i32 0, i32 %1909 + +The last strlen is computable as 1908-@pgn_event, which means 1910=1908. + +//===---------------------------------------------------------------------===// + +186.crafty has this interesting pattern with the "out.4543" variable: + +call void @llvm.memcpy.i32( + i8* getelementptr ([10 x i8]* @out.4543, i32 0, i32 0), + i8* getelementptr ([7 x i8]* @"\01LC28700", i32 0, i32 0), i32 7, i32 1) +%101 = call@printf(i8* ... @out.4543, i32 0, i32 0)) nounwind + +It is basically doing: + + memcpy(globalarray, "string"); + printf(..., globalarray); + +Anyway, by knowing that printf just reads the memory and forward substituting +the string directly into the printf, this eliminates reads from globalarray. +Since this pattern occurs frequently in crafty (due to the "DisplayTime" and +other similar functions) there are many stores to "out". Once all the printfs +stop using "out", all that is left is the memcpy's into it. This should allow +globalopt to remove the "stored only" global. + +//===---------------------------------------------------------------------===// + +This code: + +define inreg i32 @foo(i8* inreg %p) nounwind { + %tmp0 = load i8* %p + %tmp1 = ashr i8 %tmp0, 5 + %tmp2 = sext i8 %tmp1 to i32 + ret i32 %tmp2 +} + +could be dagcombine'd to a sign-extending load with a shift. +For example, on x86 this currently gets this: + + movb (%eax), %al + sarb $5, %al + movsbl %al, %eax + +while it could get this: + + movsbl (%eax), %eax + sarl $5, %eax + +//===---------------------------------------------------------------------===// + +GCC PR31029: + +int test(int x) { return 1-x == x; } // --> return false +int test2(int x) { return 2-x == x; } // --> return x == 1 ? + +Always foldable for odd constants, what is the rule for even? + +//===---------------------------------------------------------------------===// + +PR 3381: GEP to field of size 0 inside a struct could be turned into GEP +for next field in struct (which is at same address). + +For example: store of float into { {{}}, float } could be turned into a store to +the float directly. + +//===---------------------------------------------------------------------===// + +#include +double foo(double a) { return sin(a); } + +This compiles into this on x86-64 Linux: +foo: + subq $8, %rsp + call sin + addq $8, %rsp + ret +vs: + +foo: + jmp sin + +//===---------------------------------------------------------------------===// + +The arg promotion pass should make use of nocapture to make its alias analysis +stuff much more precise. + +//===---------------------------------------------------------------------===// + +The following functions should be optimized to use a select instead of a +branch (from gcc PR40072): + +char char_int(int m) {if(m>7) return 0; return m;} +int int_char(char m) {if(m>7) return 0; return m;} + +//===---------------------------------------------------------------------===// + +int func(int a, int b) { if (a & 0x80) b |= 0x80; else b &= ~0x80; return b; } + +Generates this: + +define i32 @func(i32 %a, i32 %b) nounwind readnone ssp { +entry: + %0 = and i32 %a, 128 ; [#uses=1] + %1 = icmp eq i32 %0, 0 ; [#uses=1] + %2 = or i32 %b, 128 ; [#uses=1] + %3 = and i32 %b, -129 ; [#uses=1] + %b_addr.0 = select i1 %1, i32 %3, i32 %2 ; [#uses=1] + ret i32 %b_addr.0 +} + +However, it's functionally equivalent to: + + b = (b & ~0x80) | (a & 0x80); + +Which generates this: + +define i32 @func(i32 %a, i32 %b) nounwind readnone ssp { +entry: + %0 = and i32 %b, -129 ; [#uses=1] + %1 = and i32 %a, 128 ; [#uses=1] + %2 = or i32 %0, %1 ; [#uses=1] + ret i32 %2 +} + +This can be generalized for other forms: + + b = (b & ~0x80) | (a & 0x40) << 1; + +//===---------------------------------------------------------------------===// + +These two functions produce different code. They shouldn't: + +#include + +uint8_t p1(uint8_t b, uint8_t a) { + b = (b & ~0xc0) | (a & 0xc0); + return (b); +} + +uint8_t p2(uint8_t b, uint8_t a) { + b = (b & ~0x40) | (a & 0x40); + b = (b & ~0x80) | (a & 0x80); + return (b); +} + +define zeroext i8 @p1(i8 zeroext %b, i8 zeroext %a) nounwind readnone ssp { +entry: + %0 = and i8 %b, 63 ; [#uses=1] + %1 = and i8 %a, -64 ; [#uses=1] + %2 = or i8 %1, %0 ; [#uses=1] + ret i8 %2 +} + +define zeroext i8 @p2(i8 zeroext %b, i8 zeroext %a) nounwind readnone ssp { +entry: + %0 = and i8 %b, 63 ; [#uses=1] + %.masked = and i8 %a, 64 ; [#uses=1] + %1 = and i8 %a, -128 ; [#uses=1] + %2 = or i8 %1, %0 ; [#uses=1] + %3 = or i8 %2, %.masked ; [#uses=1] + ret i8 %3 +} + +//===---------------------------------------------------------------------===// + +IPSCCP does not currently propagate argument dependent constants through +functions where it does not not all of the callers. This includes functions +with normal external linkage as well as templates, C99 inline functions etc. +Specifically, it does nothing to: + +define i32 @test(i32 %x, i32 %y, i32 %z) nounwind { +entry: + %0 = add nsw i32 %y, %z + %1 = mul i32 %0, %x + %2 = mul i32 %y, %z + %3 = add nsw i32 %1, %2 + ret i32 %3 +} + +define i32 @test2() nounwind { +entry: + %0 = call i32 @test(i32 1, i32 2, i32 4) nounwind + ret i32 %0 +} + +It would be interesting extend IPSCCP to be able to handle simple cases like +this, where all of the arguments to a call are constant. Because IPSCCP runs +before inlining, trivial templates and inline functions are not yet inlined. +The results for a function + set of constant arguments should be memoized in a +map. + +//===---------------------------------------------------------------------===// + +The libcall constant folding stuff should be moved out of SimplifyLibcalls into +libanalysis' constantfolding logic. This would allow IPSCCP to be able to +handle simple things like this: + +static int foo(const char *X) { return strlen(X); } +int bar() { return foo("abcd"); } + +//===---------------------------------------------------------------------===// + +InstCombine should use SimplifyDemandedBits to remove the or instruction: + +define i1 @test(i8 %x, i8 %y) { + %A = or i8 %x, 1 + %B = icmp ugt i8 %A, 3 + ret i1 %B +} + +Currently instcombine calls SimplifyDemandedBits with either all bits or just +the sign bit, if the comparison is obviously a sign test. In this case, we only +need all but the bottom two bits from %A, and if we gave that mask to SDB it +would delete the or instruction for us. + +//===---------------------------------------------------------------------===// + +functionattrs doesn't know much about memcpy/memset. This function should be +marked readnone rather than readonly, since it only twiddles local memory, but +functionattrs doesn't handle memset/memcpy/memmove aggressively: + +struct X { int *p; int *q; }; +int foo() { + int i = 0, j = 1; + struct X x, y; + int **p; + y.p = &i; + x.q = &j; + p = __builtin_memcpy (&x, &y, sizeof (int *)); + return **p; +} + +//===---------------------------------------------------------------------===// + +Missed instcombine transformation: +define i1 @a(i32 %x) nounwind readnone { +entry: + %cmp = icmp eq i32 %x, 30 + %sub = add i32 %x, -30 + %cmp2 = icmp ugt i32 %sub, 9 + %or = or i1 %cmp, %cmp2 + ret i1 %or +} +This should be optimized to a single compare. Testcase derived from gcc. + +//===---------------------------------------------------------------------===// + +Missed instcombine transformation: +void b(); +void a(int x) { if (((1<47)&(b<58); } + +The sgt and slt should be combined into a single comparison. Testcase derived +from gcc. + +//===---------------------------------------------------------------------===// + +Missed instcombine transformation: +define i32 @a(i32 %x) nounwind readnone { +entry: + %rem = srem i32 %x, 32 + %shl = shl i32 1, %rem + ret i32 %shl +} + +The srem can be transformed to an and because if x is negative, the shift is +undefined. Testcase derived from gcc. + +//===---------------------------------------------------------------------===// + +Missed instcombine/dagcombine transformation: +define i32 @a(i32 %x, i32 %y) nounwind readnone { +entry: + %mul = mul i32 %y, -8 + %sub = sub i32 %x, %mul + ret i32 %sub +} + +Should compile to something like x+y*8, but currently compiles to an +inefficient result. Testcase derived from gcc. + +//===---------------------------------------------------------------------===// + +Missed instcombine/dagcombine transformation: +define void @lshift_lt(i8 zeroext %a) nounwind { +entry: + %conv = zext i8 %a to i32 + %shl = shl i32 %conv, 3 + %cmp = icmp ult i32 %shl, 33 + br i1 %cmp, label %if.then, label %if.end + +if.then: + tail call void @bar() nounwind + ret void + +if.end: + ret void +} +declare void @bar() nounwind + +The shift should be eliminated. Testcase derived from gcc. + +//===---------------------------------------------------------------------===// + +These compile into different code, one gets recognized as a switch and the +other doesn't due to phase ordering issues (PR6212): + +int test1(int mainType, int subType) { + if (mainType == 7) + subType = 4; + else if (mainType == 9) + subType = 6; + else if (mainType == 11) + subType = 9; + return subType; +} + +int test2(int mainType, int subType) { + if (mainType == 7) + subType = 4; + if (mainType == 9) + subType = 6; + if (mainType == 11) + subType = 9; + return subType; +} + +//===---------------------------------------------------------------------===// + +The following test case (from PR6576): + +define i32 @mul(i32 %a, i32 %b) nounwind readnone { +entry: + %cond1 = icmp eq i32 %b, 0 ; [#uses=1] + br i1 %cond1, label %exit, label %bb.nph +bb.nph: ; preds = %entry + %tmp = mul i32 %b, %a ; [#uses=1] + ret i32 %tmp +exit: ; preds = %entry + ret i32 0 +} + +could be reduced to: + +define i32 @mul(i32 %a, i32 %b) nounwind readnone { +entry: + %tmp = mul i32 %b, %a + ret i32 %tmp +} + +//===---------------------------------------------------------------------===// + +We should use DSE + llvm.lifetime.end to delete dead vtable pointer updates. +See GCC PR34949 + +Another interesting case is that something related could be used for variables +that go const after their ctor has finished. In these cases, globalopt (which +can statically run the constructor) could mark the global const (so it gets put +in the readonly section). A testcase would be: + +#include +using namespace std; +const complex should_be_in_rodata (42,-42); +complex should_be_in_data (42,-42); +complex should_be_in_bss; + +Where we currently evaluate the ctors but the globals don't become const because +the optimizer doesn't know they "become const" after the ctor is done. See +GCC PR4131 for more examples. + +//===---------------------------------------------------------------------===// + +In this code: + +long foo(long x) { + return x > 1 ? x : 1; +} + +LLVM emits a comparison with 1 instead of 0. 0 would be equivalent +and cheaper on most targets. + +LLVM prefers comparisons with zero over non-zero in general, but in this +case it choses instead to keep the max operation obvious. + +//===---------------------------------------------------------------------===// + +Take the following testcase on x86-64 (similar testcases exist for all targets +with addc/adde): + +define void @a(i64* nocapture %s, i64* nocapture %t, i64 %a, i64 %b, +i64 %c) nounwind { +entry: + %0 = zext i64 %a to i128 ; [#uses=1] + %1 = zext i64 %b to i128 ; [#uses=1] + %2 = add i128 %1, %0 ; [#uses=2] + %3 = zext i64 %c to i128 ; [#uses=1] + %4 = shl i128 %3, 64 ; [#uses=1] + %5 = add i128 %4, %2 ; [#uses=1] + %6 = lshr i128 %5, 64 ; [#uses=1] + %7 = trunc i128 %6 to i64 ; [#uses=1] + store i64 %7, i64* %s, align 8 + %8 = trunc i128 %2 to i64 ; [#uses=1] + store i64 %8, i64* %t, align 8 + ret void +} + +Generated code: + addq %rcx, %rdx + movl $0, %eax + adcq $0, %rax + addq %r8, %rax + movq %rax, (%rdi) + movq %rdx, (%rsi) + ret + +Expected code: + addq %rcx, %rdx + adcq $0, %r8 + movq %r8, (%rdi) + movq %rdx, (%rsi) + ret + +The generated SelectionDAG has an ADD of an ADDE, where both operands of the +ADDE are zero. Replacing one of the operands of the ADDE with the other operand +of the ADD, and replacing the ADD with the ADDE, should give the desired result. + +(That said, we are doing a lot better than gcc on this testcase. :) ) + +//===---------------------------------------------------------------------===// + +Switch lowering generates less than ideal code for the following switch: +define void @a(i32 %x) nounwind { +entry: + switch i32 %x, label %if.end [ + i32 0, label %if.then + i32 1, label %if.then + i32 2, label %if.then + i32 3, label %if.then + i32 5, label %if.then + ] +if.then: + tail call void @foo() nounwind + ret void +if.end: + ret void +} +declare void @foo() + +Generated code on x86-64 (other platforms give similar results): +a: + cmpl $5, %edi + ja .LBB0_2 + movl %edi, %eax + movl $47, %ecx + btq %rax, %rcx + jb .LBB0_3 +.LBB0_2: + ret +.LBB0_3: + jmp foo # TAILCALL + +The movl+movl+btq+jb could be simplified to a cmpl+jne. + +Or, if we wanted to be really clever, we could simplify the whole thing to +something like the following, which eliminates a branch: + xorl $1, %edi + cmpl $4, %edi + ja .LBB0_2 + ret +.LBB0_2: + jmp foo # TAILCALL +//===---------------------------------------------------------------------===// +Given a branch where the two target blocks are identical ("ret i32 %b" in +both), simplifycfg will simplify them away. But not so for a switch statement: + +define i32 @f(i32 %a, i32 %b) nounwind readnone { +entry: + switch i32 %a, label %bb3 [ + i32 4, label %bb + i32 6, label %bb + ] + +bb: ; preds = %entry, %entry + ret i32 %b + +bb3: ; preds = %entry + ret i32 %b +} +//===---------------------------------------------------------------------===// + +clang -O3 fails to devirtualize this virtual inheritance case: (GCC PR45875) + +struct c1 {}; +struct c10 : c1{ + virtual void foo (); +}; +struct c11 : c10, c1{ + virtual void f6 (); +}; +struct c28 : virtual c11{ + void f6 (); +}; +void check_c28 () { + c28 obj; + c11 *ptr = &obj; + ptr->f6 (); +} + +//===---------------------------------------------------------------------===//