//===---------------------------------------------------------------------===//
-Make the PPC branch selector target independant
+We should recognized various "overflow detection" idioms and translate them into
+llvm.uadd.with.overflow and similar intrinsics. Here is a multiply idiom:
+
+unsigned int mul(unsigned int a,unsigned int b) {
+ if ((unsigned long long)a*b>0xffffffff)
+ exit(0);
+ return a*b;
+}
+
+The legalization code for mul-with-overflow needs to be made more robust before
+this can be implemented though.
//===---------------------------------------------------------------------===//
//===---------------------------------------------------------------------===//
-Solve this DAG isel folding deficiency:
-
-int X, Y;
-
-void fn1(void)
-{
- X = X | (Y << 3);
-}
-
-compiles to
-
-fn1:
- movl Y, %eax
- shll $3, %eax
- orl X, %eax
- movl %eax, X
- ret
-
-The problem is the store's chain operand is not the load X but rather
-a TokenFactor of the load X and load Y, which prevents the folding.
-
-There are two ways to fix this:
-
-1. The dag combiner can start using alias analysis to realize that y/x
- don't alias, making the store to X not dependent on the load from Y.
-2. The generated isel could be made smarter in the case it can't
- disambiguate the pointers.
-
-Number 1 is the preferred solution.
-
-This has been "fixed" by a TableGen hack. But that is a short term workaround
-which will be removed once the proper fix is made.
-
-//===---------------------------------------------------------------------===//
-
On targets with expensive 64-bit multiply, we could LSR this:
for (i = ...; ++i) {
//===---------------------------------------------------------------------===//
-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.
//===---------------------------------------------------------------------===//
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.
//===---------------------------------------------------------------------===//
//===---------------------------------------------------------------------===//
-Turn this into a single byte store with no load (the other 3 bytes are
-unmodified):
-
-define void @test(i32* %P) {
- %tmp = load i32* %P
- %tmp14 = or i32 %tmp, 3305111552
- %tmp15 = and i32 %tmp14, 3321888767
- store i32 %tmp15, i32* %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
-
-//===---------------------------------------------------------------------===//
-
quantum_sigma_x in 462.libquantum contains the following loop:
for(i=0; i<reg->size; i++)
//===---------------------------------------------------------------------===//
-This should be optimized to one 'and' and one 'or', from PR4216:
-
-define i32 @test_bitfield(i32 %bf.prev.low) nounwind ssp {
-entry:
- %bf.prev.lo.cleared10 = or i32 %bf.prev.low, 32962 ; <i32> [#uses=1]
- %0 = and i32 %bf.prev.low, -65536 ; <i32> [#uses=1]
- %1 = and i32 %bf.prev.lo.cleared10, 40186 ; <i32> [#uses=1]
- %2 = or i32 %1, %0 ; <i32> [#uses=1]
- ret i32 %2
-}
-
-//===---------------------------------------------------------------------===//
-
This isn't recognized as bswap by instcombine (yes, it really is bswap):
unsigned long reverse(unsigned v) {
//===---------------------------------------------------------------------===//
+[LOOP RECOGNITION]
+
These idioms should be recognized as popcount (see PR1488):
unsigned countbits_slow(unsigned v) {
return count;
}
-This is a form of idiom recognition for loops, the same thing that could be
-useful for recognizing memset/memcpy.
+This sort of thing should be added to the loop idiom pass.
//===---------------------------------------------------------------------===//
//===---------------------------------------------------------------------===//
-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.
//===---------------------------------------------------------------------===//
extern THotKey m_HotKey;
THotKey GetHotKey () { return m_HotKey; }
-into (-O3 -fno-exceptions -static -fomit-frame-pointer):
-
-__Z9GetHotKeyv:
- pushl %esi
- movl 8(%esp), %eax
- movb _m_HotKey+3, %cl
- movb _m_HotKey+4, %dl
- movb _m_HotKey+2, %ch
- movw _m_HotKey, %si
- movw %si, (%eax)
- movb %ch, 2(%eax)
- movb %cl, 3(%eax)
- movb %dl, 4(%eax)
- popl %esi
- ret $4
-
-GCC produces:
-
-__Z9GetHotKeyv:
- movl _m_HotKey, %edx
- movl 4(%esp), %eax
- movl %edx, (%eax)
- movzwl _m_HotKey+4, %edx
- movw %dx, 4(%eax)
- ret $4
-
-The LLVM IR contains the needed alignment info, so we should be able to
-merge the loads and stores into 4-byte loads:
-
- %struct.THotKey = type { i16, i8, i8, i8 }
-define void @_Z9GetHotKeyv(%struct.THotKey* sret %agg.result) nounwind {
-...
- %tmp2 = load i16* getelementptr (@m_HotKey, i32 0, i32 0), align 8
- %tmp5 = load i8* getelementptr (@m_HotKey, i32 0, i32 1), align 2
- %tmp8 = load i8* getelementptr (@m_HotKey, i32 0, i32 2), align 1
- %tmp11 = load i8* getelementptr (@m_HotKey, i32 0, i32 3), align 2
-
-Alternatively, we should use a small amount of base-offset alias analysis
-to make it so the scheduler doesn't need to hold all the loads in regs at
-once.
+into (-m64 -O3 -fno-exceptions -static -fomit-frame-pointer):
+
+__Z9GetHotKeyv: ## @_Z9GetHotKeyv
+ movq _m_HotKey@GOTPCREL(%rip), %rax
+ movzwl (%rax), %ecx
+ movzbl 2(%rax), %edx
+ shlq $16, %rdx
+ orq %rcx, %rdx
+ movzbl 3(%rax), %ecx
+ shlq $24, %rcx
+ orq %rdx, %rcx
+ movzbl 4(%rax), %eax
+ shlq $32, %rax
+ orq %rcx, %rax
+ ret
//===---------------------------------------------------------------------===//
Consider:
int test() {
- long long input[8] = {1,1,1,1,1,1,1,1};
+ long long input[8] = {1,0,1,0,1,0,1,0};
foo(input);
}
-We currently compile this into a memcpy from a global array since the
-initializer is fairly large and not memset'able. This is good, but the memcpy
-gets lowered to load/stores in the code generator. This is also ok, except
-that the codegen lowering for memcpy doesn't handle the case when the source
-is a constant global. This gives us atrocious code like this:
+Clang compiles this into:
- call "L1$pb"
-"L1$pb":
- popl %eax
- movl _C.0.1444-"L1$pb"+32(%eax), %ecx
- movl %ecx, 40(%esp)
- movl _C.0.1444-"L1$pb"+20(%eax), %ecx
- movl %ecx, 28(%esp)
- movl _C.0.1444-"L1$pb"+36(%eax), %ecx
- movl %ecx, 44(%esp)
- movl _C.0.1444-"L1$pb"+44(%eax), %ecx
- movl %ecx, 52(%esp)
- movl _C.0.1444-"L1$pb"+40(%eax), %ecx
- movl %ecx, 48(%esp)
- movl _C.0.1444-"L1$pb"+12(%eax), %ecx
- movl %ecx, 20(%esp)
- movl _C.0.1444-"L1$pb"+4(%eax), %ecx
-...
+ call void @llvm.memset.p0i8.i64(i8* %tmp, i8 0, i64 64, i32 16, i1 false)
+ %0 = getelementptr [8 x i64]* %input, i64 0, i64 0
+ store i64 1, i64* %0, align 16
+ %1 = getelementptr [8 x i64]* %input, i64 0, i64 2
+ store i64 1, i64* %1, align 16
+ %2 = getelementptr [8 x i64]* %input, i64 0, i64 4
+ store i64 1, i64* %2, align 16
+ %3 = getelementptr [8 x i64]* %input, i64 0, i64 6
+ store i64 1, i64* %3, align 16
-instead of:
- movl $1, 16(%esp)
- movl $0, 20(%esp)
- movl $1, 24(%esp)
- movl $0, 28(%esp)
- movl $1, 32(%esp)
- movl $0, 36(%esp)
- ...
+Which gets codegen'd into:
+
+ pxor %xmm0, %xmm0
+ movaps %xmm0, -16(%rbp)
+ movaps %xmm0, -32(%rbp)
+ movaps %xmm0, -48(%rbp)
+ movaps %xmm0, -64(%rbp)
+ movq $1, -64(%rbp)
+ movq $1, -48(%rbp)
+ movq $1, -32(%rbp)
+ movq $1, -16(%rbp)
+
+It would be better to have 4 movq's of 0 instead of the movaps's.
//===---------------------------------------------------------------------===//
matching code in dag combine doesn't look through truncates aggressively
enough. Here are some testcases reduces from GCC PR17886:
-unsigned long long f(unsigned long long x, int y) {
- return (x << y) | (x >> 64-y);
-}
-unsigned f2(unsigned x, int y){
- return (x << y) | (x >> 32-y);
-}
-unsigned long long f3(unsigned long long x){
- int y = 9;
- return (x << y) | (x >> 64-y);
-}
-unsigned f4(unsigned x){
- int y = 10;
- return (x << y) | (x >> 32-y);
-}
unsigned long long f5(unsigned long long x, unsigned long long y) {
return (x << 8) | ((y >> 48) & 0xffull);
}
}
}
-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.
-
//===---------------------------------------------------------------------===//
-We do a number of simplifications in simplify libcalls to strength reduce
-standard library functions, but we don't currently merge them together. For
-example, it is useful to merge memcpy(a,b,strlen(b)) -> strcpy. This can only
-be done safely if "b" isn't modified between the strlen and memcpy of course.
+This (and similar related idioms):
-//===---------------------------------------------------------------------===//
+unsigned int foo(unsigned char i) {
+ return i | (i<<8) | (i<<16) | (i<<24);
+}
-Reassociate should turn things like:
+compiles into:
-int factorial(int X) {
- return X*X*X*X*X*X*X*X;
+define i32 @foo(i8 zeroext %i) nounwind readnone ssp noredzone {
+entry:
+ %conv = zext i8 %i to i32
+ %shl = shl i32 %conv, 8
+ %shl5 = shl i32 %conv, 16
+ %shl9 = shl i32 %conv, 24
+ %or = or i32 %shl9, %conv
+ %or6 = or i32 %or, %shl5
+ %or10 = or i32 %or6, %shl
+ ret i32 %or10
}
-into llvm.powi calls, allowing the code generator to produce balanced
-multiplication trees.
-
-//===---------------------------------------------------------------------===//
+it would be better as:
-We generate a horrible libcall for llvm.powi. For example, we compile:
+unsigned int bar(unsigned char i) {
+ unsigned int j=i | (i << 8);
+ return j | (j<<16);
+}
-#include <cmath>
-double f(double a) { return std::pow(a, 4); }
+aka:
-into:
+define i32 @bar(i8 zeroext %i) nounwind readnone ssp noredzone {
+entry:
+ %conv = zext i8 %i to i32
+ %shl = shl i32 %conv, 8
+ %or = or i32 %shl, %conv
+ %shl5 = shl i32 %or, 16
+ %or6 = or i32 %shl5, %or
+ ret i32 %or6
+}
-__Z1fd:
- subl $12, %esp
- movsd 16(%esp), %xmm0
- movsd %xmm0, (%esp)
- movl $4, 8(%esp)
- call L___powidf2$stub
- addl $12, %esp
- ret
+or even i*0x01010101, depending on the speed of the multiplier. The best way to
+handle this is to canonicalize it to a multiply in IR and have codegen handle
+lowering multiplies to shifts on cpus where shifts are faster.
-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 do a number of simplifications in simplify libcalls to strength reduce
+standard library functions, but we don't currently merge them together. For
+example, it is useful to merge memcpy(a,b,strlen(b)) -> strcpy. This can only
+be done safely if "b" isn't modified between the strlen and memcpy of course.
//===---------------------------------------------------------------------===//
//===---------------------------------------------------------------------===//
-We miss some instcombines for stuff like this:
-void bar (void);
-void foo (unsigned int a) {
- /* This one is equivalent to a >= (3 << 2). */
- if ((a >> 2) >= 3)
- bar ();
-}
-
-A few other related ones are in GCC PR14753.
-
-//===---------------------------------------------------------------------===//
-
Divisibility by constant can be simplified (according to GCC PR12849) from
being a mulhi to being a mul lo (cheaper). Testcase:
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".
//===---------------------------------------------------------------------===//
//===---------------------------------------------------------------------===//
-Instcombine will merge comparisons like (x >= 10) && (x < 20) by producing (x -
-10) u< 10, but only when the comparisons have matching sign.
-
-This could be converted with a similiar technique. (PR1941)
-
-define i1 @test(i8 %x) {
- %A = icmp uge i8 %x, 5
- %B = icmp slt i8 %x, 20
- %C = and i1 %A, %B
- ret i1 %C
-}
-
-//===---------------------------------------------------------------------===//
-
These functions perform the same computation, but produce different assembly.
define i8 @select(i8 %x) readnone nounwind {
//===---------------------------------------------------------------------===//
-From GCC Bug 15241:
-unsigned int
-foo (unsigned int a, unsigned int b)
-{
- if (a <= 7 && b <= 7)
- baz ();
-}
-Should combine to "(a|b) <= 7". Currently not optimized with "clang
--emit-llvm-bc | opt -std-compile-opts".
-
-//===---------------------------------------------------------------------===//
-
-From GCC Bug 3756:
-int
-pn (int n)
-{
- return (n >= 0 ? 1 : -1);
-}
-Should combine to (n >> 31) | 1. Currently not optimized with "clang
--emit-llvm-bc | opt -std-compile-opts | llc".
-
-//===---------------------------------------------------------------------===//
-
-From GCC Bug 28685:
-int test(int a, int b)
-{
- int lt = a < b;
- int eq = a == b;
-
- return (lt || eq);
-}
-Should combine to "a <= b". Currently not optimized with "clang
--emit-llvm-bc | opt -std-compile-opts | llc".
-
-//===---------------------------------------------------------------------===//
-
-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".
-
-//===---------------------------------------------------------------------===//
-
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;}
//===---------------------------------------------------------------------===//
-int a(unsigned char* b) {return *b > 99;}
-There's an unnecessary zext in the generated code with "clang
--emit-llvm-bc | opt -std-compile-opts".
-
-//===---------------------------------------------------------------------===//
-
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 a) {return ((a | 1) & 3) | (a & -4);}
-Should combine to "a | 1". 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 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".
//===---------------------------------------------------------------------===//
+[STORE SINKING]
+
Store sinking: This code:
void f (int n, int *cond, int *res) {
It would be better to do the mul once to reduce codesize above the if.
This is GCC PR38204.
+
+//===---------------------------------------------------------------------===//
+This simple function from 179.art:
+
+int winner, numf2s;
+struct { double y; int reset; } *Y;
+
+void find_match() {
+ int i;
+ winner = 0;
+ for (i=0;i<numf2s;i++)
+ if (Y[i].y > Y[winner].y)
+ winner =i;
+}
+
+Compiles into (with clang TBAA):
+
+for.body: ; preds = %for.inc, %bb.nph
+ %indvar = phi i64 [ 0, %bb.nph ], [ %indvar.next, %for.inc ]
+ %i.01718 = phi i32 [ 0, %bb.nph ], [ %i.01719, %for.inc ]
+ %tmp4 = getelementptr inbounds %struct.anon* %tmp3, i64 %indvar, i32 0
+ %tmp5 = load double* %tmp4, align 8, !tbaa !4
+ %idxprom7 = sext i32 %i.01718 to i64
+ %tmp10 = getelementptr inbounds %struct.anon* %tmp3, i64 %idxprom7, i32 0
+ %tmp11 = load double* %tmp10, align 8, !tbaa !4
+ %cmp12 = fcmp ogt double %tmp5, %tmp11
+ br i1 %cmp12, label %if.then, label %for.inc
+
+if.then: ; preds = %for.body
+ %i.017 = trunc i64 %indvar to i32
+ br label %for.inc
+
+for.inc: ; preds = %for.body, %if.then
+ %i.01719 = phi i32 [ %i.01718, %for.body ], [ %i.017, %if.then ]
+ %indvar.next = add i64 %indvar, 1
+ %exitcond = icmp eq i64 %indvar.next, %tmp22
+ br i1 %exitcond, label %for.cond.for.end_crit_edge, label %for.body
+
+
+It is good that we hoisted the reloads of numf2's, and Y out of the loop and
+sunk the store to winner out.
+
+However, this is awful on several levels: the conditional truncate in the loop
+(-indvars at fault? why can't we completely promote the IV to i64?).
+
+Beyond that, we have a partially redundant load in the loop: if "winner" (aka
+%i.01718) isn't updated, we reload Y[winner].y the next time through the loop.
+Similarly, the addressing that feeds it (including the sext) is redundant. In
+the end we get this generated assembly:
+
+LBB0_2: ## %for.body
+ ## =>This Inner Loop Header: Depth=1
+ movsd (%rdi), %xmm0
+ movslq %edx, %r8
+ shlq $4, %r8
+ ucomisd (%rcx,%r8), %xmm0
+ jbe LBB0_4
+ movl %esi, %edx
+LBB0_4: ## %for.inc
+ addq $16, %rdi
+ incq %rsi
+ cmpq %rsi, %rax
+ jne LBB0_2
+
+All things considered this isn't too bad, but we shouldn't need the movslq or
+the shlq instruction, or the load folded into ucomisd every time through the
+loop.
+
+On an x86-specific topic, if the loop can't be restructure, the movl should be a
+cmov.
+
//===---------------------------------------------------------------------===//
+[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.
//===---------------------------------------------------------------------===//
-[LOAD PRE with NON-AVAILABLE ADDRESS]
+[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
%11 is partially redundant, an in BB2 it should have the value %8.
-GCC PR33344 is a similar case.
+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. There are many pre testcases as ssa-pre-*.c
+GCC testsuite, ones we don't get yet are (checked through loadpre25):
-//===---------------------------------------------------------------------===//
+[CRIT EDGE BREAKING]
+loadpre3.c predcom-4.c
-There are some interesting cases in testsuite/gcc.dg/tree-ssa/pred-comm* in the
-GCC testsuite. For example, predcom-1.c is:
-
- for (i = 2; i < 1000; i++)
- fib[i] = (fib[i-1] + fib[i - 2]) & 0xffff;
-
-which compiles into:
-
-bb1: ; preds = %bb1, %bb1.thread
- %indvar = phi i32 [ 0, %bb1.thread ], [ %0, %bb1 ]
- %i.0.reg2mem.0 = add i32 %indvar, 2
- %0 = add i32 %indvar, 1 ; <i32> [#uses=3]
- %1 = getelementptr [1000 x i32]* @fib, i32 0, i32 %0
- %2 = load i32* %1, align 4 ; <i32> [#uses=1]
- %3 = getelementptr [1000 x i32]* @fib, i32 0, i32 %indvar
- %4 = load i32* %3, align 4 ; <i32> [#uses=1]
- %5 = add i32 %4, %2 ; <i32> [#uses=1]
- %6 = and i32 %5, 65535 ; <i32> [#uses=1]
- %7 = getelementptr [1000 x i32]* @fib, i32 0, i32 %i.0.reg2mem.0
- store i32 %6, i32* %7, align 4
- %exitcond = icmp eq i32 %0, 998 ; <i1> [#uses=1]
- br i1 %exitcond, label %return, label %bb1
-
-This is basically:
- LOAD fib[i+1]
- LOAD fib[i]
- STORE fib[i+2]
-
-instead of handling this as a loop or other xform, all we'd need to do is teach
-load PRE to phi translate the %0 add (i+1) into the predecessor as (i'+1+1) =
-(i'+2) (where i' is the previous iteration of i). This would find the store
-which feeds it.
-
-predcom-2.c is apparently the same as predcom-1.c
-predcom-3.c is very similar but needs loads feeding each other instead of
-store->load.
-predcom-4.c seems the same as the rest.
+[PRE OF READONLY CALL]
+loadpre5.c
+
+[TURN SELECT INTO BRANCH]
+loadpre14.c loadpre15.c
+actually a conditional increment: loadpre18.c loadpre19.c
//===---------------------------------------------------------------------===//
-Other simple load PRE cases:
-http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35287 [LPRE crit edge splitting]
+[LOAD PRE / STORE SINKING / SPEC HACK]
-http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34677 (licm does this, LPRE crit edge)
- llvm-gcc t2.c -S -o - -O0 -emit-llvm | llvm-as | opt -mem2reg -simplifycfg -gvn | llvm-dis
+This is a chunk of code from 456.hmmer:
+
+int f(int M, int *mc, int *mpp, int *tpmm, int *ip, int *tpim, int *dpp,
+ int *tpdm, int xmb, int *bp, int *ms) {
+ int k, sc;
+ for (k = 1; k <= M; k++) {
+ mc[k] = mpp[k-1] + tpmm[k-1];
+ if ((sc = ip[k-1] + tpim[k-1]) > mc[k]) mc[k] = sc;
+ if ((sc = dpp[k-1] + tpdm[k-1]) > mc[k]) mc[k] = sc;
+ if ((sc = xmb + bp[k]) > mc[k]) mc[k] = sc;
+ mc[k] += ms[k];
+ }
+}
+
+It is very profitable for this benchmark to turn the conditional stores to mc[k]
+into a conditional move (select instr in IR) and allow the final store to do the
+store. See GCC PR27313 for more details. Note that this is valid to xform even
+with the new C++ memory model, since mc[k] is previously loaded and later
+stored.
//===---------------------------------------------------------------------===//
-Type based alias analysis:
-http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14705
+[SCALAR PRE]
+There are many PRE testcases in testsuite/gcc.dg/tree-ssa/ssa-pre-*.c in the
+GCC testsuite.
//===---------------------------------------------------------------------===//
-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
+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];
-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;
+__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
+
//===---------------------------------------------------------------------===//
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) GEP PHI TRAN, 2) CRIT EDGE 3) MULTIPLE DIFFERENT
+we miss it because we need 1) CRIT EDGE 2) MULTIPLE DIFFERENT
VALS PRODUCED BY ONE BLOCK OVER DIFFERENT PATHS
//===---------------------------------------------------------------------===//
simplifylibcalls should do several optimizations for strspn/strcspn:
-strcspn(x, "") -> strlen(x)
-strcspn("", x) -> 0
-strspn("", x) -> 0
-strspn(x, "") -> strlen(x)
-strspn(x, "a") -> strchr(x, 'a')-x
-
strcspn(x, "a") -> inlined loop for up to 3 letters (similarly for strspn):
size_t __strcspn_c3 (__const char *__s, int __reject1, int __reject2,
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:
+The memcpy+strlen strlen can be replaced with:
%3074 = call i32 @strlen([5 x i8]* @"\01LC42") nounwind readonly
//===---------------------------------------------------------------------===//
-186.crafty contains this interesting pattern:
-
-%77 = call i8* @strstr(i8* getelementptr ([6 x i8]* @"\01LC5", i32 0, i32 0),
- i8* %30)
-%phitmp648 = icmp eq i8* %77, getelementptr ([6 x i8]* @"\01LC5", i32 0, i32 0)
-br i1 %phitmp648, label %bb70, label %bb76
-
-bb70: ; preds = %OptionMatch.exit91, %bb69
- %78 = call i32 @strlen(i8* %30) nounwind readonly align 1 ; <i32> [#uses=1]
-
-This is basically:
- cststr = "abcdef";
- if (strstr(cststr, P) == cststr) {
- x = strlen(P);
- ...
-
-The strstr call would be significantly cheaper written as:
-
-cststr = "abcdef";
-if (memcmp(P, str, strlen(P)))
- x = strlen(P);
-
-This is memcmp+strlen instead of strstr. This also makes the strlen fully
-redundant.
-
-//===---------------------------------------------------------------------===//
-
-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(
//===---------------------------------------------------------------------===//
-#include <math.h>
-double foo(double a) { return sin(a); }
-
-This compiles into this on x86-64 Linux:
-foo:
- subq $8, %rsp
- call sin
- addq $8, %rsp
- ret
-vs:
-
-foo:
- jmp sin
-
-//===---------------------------------------------------------------------===//
-
The arg promotion pass should make use of nocapture to make its alias analysis
stuff much more precise.
//===---------------------------------------------------------------------===//
-InstCombine should use SimplifyDemandedBits to remove the or instruction:
+functionattrs doesn't know much about memcpy/memset. This function should be
+marked readnone rather than readonly, since it only twiddles local memory, but
+functionattrs doesn't handle memset/memcpy/memmove aggressively:
+
+struct X { int *p; int *q; };
+int foo() {
+ int i = 0, j = 1;
+ struct X x, y;
+ int **p;
+ y.p = &i;
+ x.q = &j;
+ p = __builtin_memcpy (&x, &y, sizeof (int *));
+ return **p;
+}
+
+This can be seen at:
+$ clang t.c -S -o - -mkernel -O0 -emit-llvm | opt -functionattrs -S
+
+
+//===---------------------------------------------------------------------===//
+
+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 or reassociate transformation:
+int a(int a, int b) { return (a==12)&(b>47)&(b<58); }
+
+The sgt and slt should be combined into a single comparison. Testcase derived
+from gcc.
+
+//===---------------------------------------------------------------------===//
+
+Missed instcombine transformation:
+
+ %382 = srem i32 %tmp14.i, 64 ; [#uses=1]
+ %383 = zext i32 %382 to i64 ; [#uses=1]
+ %384 = shl i64 %381, %383 ; [#uses=1]
+ %385 = icmp slt i32 %tmp14.i, 64 ; [#uses=1]
+
+The srem can be transformed to an and because if %tmp14.i is negative, the
+shift is undefined. Testcase derived from 403.gcc.
+
+//===---------------------------------------------------------------------===//
+
+This is a range comparison on a divided result (from 403.gcc):
+
+ %1337 = sdiv i32 %1336, 8 ; [#uses=1]
+ %.off.i208 = add i32 %1336, 7 ; [#uses=1]
+ %1338 = icmp ult i32 %.off.i208, 15 ; [#uses=1]
+
+We already catch this (removing the sdiv) if there isn't an add, we should
+handle the 'add' as well. This is a common idiom with it's builtin_alloca code.
+C testcase:
+
+int a(int x) { return (unsigned)(x/16+7) < 15; }
+
+Another similar case involves truncations on 64-bit targets:
+
+ %361 = sdiv i64 %.046, 8 ; [#uses=1]
+ %362 = trunc i64 %361 to i32 ; [#uses=2]
+...
+ %367 = icmp eq i32 %362, 0 ; [#uses=1]
+
+//===---------------------------------------------------------------------===//
+
+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
-define i1 @test(i8 %x, i8 %y) {
- %A = or i8 %x, 1
- %B = icmp ugt i8 %A, 3
- ret i1 %B
+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.
-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.
+//===---------------------------------------------------------------------===//
+
+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 ; <i1> [#uses=1]
+ br i1 %cond1, label %exit, label %bb.nph
+bb.nph: ; preds = %entry
+ %tmp = mul i32 %b, %a ; <i32> [#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 <complex>
+using namespace std;
+const complex<char> should_be_in_rodata (42,-42);
+complex<char> should_be_in_data (42,-42);
+complex<char> 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 ; <i128> [#uses=1]
+ %1 = zext i64 %b to i128 ; <i128> [#uses=1]
+ %2 = add i128 %1, %0 ; <i128> [#uses=2]
+ %3 = zext i64 %c to i128 ; <i128> [#uses=1]
+ %4 = shl i128 %3, 64 ; <i128> [#uses=1]
+ %5 = add i128 %4, %2 ; <i128> [#uses=1]
+ %6 = lshr i128 %5, 64 ; <i128> [#uses=1]
+ %7 = trunc i128 %6 to i64 ; <i64> [#uses=1]
+ store i64 %7, i64* %s, align 8
+ %8 = trunc i128 %2 to i64 ; <i64> [#uses=1]
+ store i64 %8, i64* %t, align 8
+ ret void
+}
+
+Generated code:
+ addq %rcx, %rdx
+ sbbq %rax, %rax
+ subq %rax, %r8
+ movq %r8, (%rdi)
+ movq %rdx, (%rsi)
+ ret
+
+Expected code:
+ addq %rcx, %rdx
+ adcq $0, %r8
+ movq %r8, (%rdi)
+ movq %rdx, (%rsi)
+ ret
+
+//===---------------------------------------------------------------------===//
+
+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
+
+//===---------------------------------------------------------------------===//
+
+We compile this:
+
+int foo(int a) { return (a & (~15)) / 16; }
+
+Into:
+
+define i32 @foo(i32 %a) nounwind readnone ssp {
+entry:
+ %and = and i32 %a, -16
+ %div = sdiv i32 %and, 16
+ ret i32 %div
+}
+
+but this code (X & -A)/A is X >> log2(A) when A is a power of 2, so this case
+should be instcombined into just "a >> 4".
+
+We do get this at the codegen level, so something knows about it, but
+instcombine should catch it earlier:
+
+_foo: ## @foo
+## BB#0: ## %entry
+ movl %edi, %eax
+ sarl $4, %eax
+ ret
+
+//===---------------------------------------------------------------------===//
+
+This code (from GCC PR28685):
+
+int test(int a, int b) {
+ int lt = a < b;
+ int eq = a == b;
+ if (lt)
+ return 1;
+ return eq;
+}
+
+Is compiled to:
+
+define i32 @test(i32 %a, i32 %b) nounwind readnone ssp {
+entry:
+ %cmp = icmp slt i32 %a, %b
+ br i1 %cmp, label %return, label %if.end
+
+if.end: ; preds = %entry
+ %cmp5 = icmp eq i32 %a, %b
+ %conv6 = zext i1 %cmp5 to i32
+ ret i32 %conv6
+
+return: ; preds = %entry
+ ret i32 1
+}
+
+it could be:
+
+define i32 @test__(i32 %a, i32 %b) nounwind readnone ssp {
+entry:
+ %0 = icmp sle i32 %a, %b
+ %retval = zext i1 %0 to i32
+ ret i32 %retval
+}
+
+//===---------------------------------------------------------------------===//
+
+This code can be seen in viterbi:
+
+ %64 = call noalias i8* @malloc(i64 %62) nounwind
+...
+ %67 = call i64 @llvm.objectsize.i64(i8* %64, i1 false) nounwind
+ %68 = call i8* @__memset_chk(i8* %64, i32 0, i64 %62, i64 %67) nounwind
+
+llvm.objectsize.i64 should be taught about malloc/calloc, allowing it to
+fold to %62. This is a security win (overflows of malloc will get caught)
+and also a performance win by exposing more memsets to the optimizer.
+
+This occurs several times in viterbi.
+
+Note that this would change the semantics of @llvm.objectsize which by its
+current definition always folds to a constant. We also should make sure that
+we remove checking in code like
+
+ char *p = malloc(strlen(s)+1);
+ __strcpy_chk(p, s, __builtin_objectsize(p, 0));
+
+//===---------------------------------------------------------------------===//
+
+This code (from Benchmarks/Dhrystone/dry.c):
+
+define i32 @Func1(i32, i32) nounwind readnone optsize ssp {
+entry:
+ %sext = shl i32 %0, 24
+ %conv = ashr i32 %sext, 24
+ %sext6 = shl i32 %1, 24
+ %conv4 = ashr i32 %sext6, 24
+ %cmp = icmp eq i32 %conv, %conv4
+ %. = select i1 %cmp, i32 10000, i32 0
+ ret i32 %.
+}
+
+Should be simplified into something like:
+
+define i32 @Func1(i32, i32) nounwind readnone optsize ssp {
+entry:
+ %sext = shl i32 %0, 24
+ %conv = and i32 %sext, 0xFF000000
+ %sext6 = shl i32 %1, 24
+ %conv4 = and i32 %sext6, 0xFF000000
+ %cmp = icmp eq i32 %conv, %conv4
+ %. = select i1 %cmp, i32 10000, i32 0
+ ret i32 %.
+}
+
+and then to:
+
+define i32 @Func1(i32, i32) nounwind readnone optsize ssp {
+entry:
+ %conv = and i32 %0, 0xFF
+ %conv4 = and i32 %1, 0xFF
+ %cmp = icmp eq i32 %conv, %conv4
+ %. = select i1 %cmp, i32 10000, i32 0
+ ret i32 %.
+}
+//===---------------------------------------------------------------------===//
+
+clang -O3 currently compiles this code
+
+int g(unsigned int a) {
+ unsigned int c[100];
+ c[10] = a;
+ c[11] = a;
+ unsigned int b = c[10] + c[11];
+ if(b > a*2) a = 4;
+ else a = 8;
+ return a + 7;
+}
+
+into
+
+define i32 @g(i32 a) nounwind readnone {
+ %add = shl i32 %a, 1
+ %mul = shl i32 %a, 1
+ %cmp = icmp ugt i32 %add, %mul
+ %a.addr.0 = select i1 %cmp, i32 11, i32 15
+ ret i32 %a.addr.0
+}
+
+The icmp should fold to false. This CSE opportunity is only available
+after GVN and InstCombine have run.
+
+//===---------------------------------------------------------------------===//
+
+memcpyopt should turn this:
+
+define i8* @test10(i32 %x) {
+ %alloc = call noalias i8* @malloc(i32 %x) nounwind
+ call void @llvm.memset.p0i8.i32(i8* %alloc, i8 0, i32 %x, i32 1, i1 false)
+ ret i8* %alloc
+}
+
+into a call to calloc. We should make sure that we analyze calloc as
+aggressively as malloc though.
+
+//===---------------------------------------------------------------------===//
+
+clang -O3 doesn't optimize this:
+
+void f1(int* begin, int* end) {
+ std::fill(begin, end, 0);
+}
+
+into a memset. This is PR8942.
+
+//===---------------------------------------------------------------------===//
+
+clang -O3 -fno-exceptions currently compiles this code:
+
+void f(int N) {
+ std::vector<int> v(N);
+
+ extern void sink(void*); sink(&v);
+}
+
+into
+
+define void @_Z1fi(i32 %N) nounwind {
+entry:
+ %v2 = alloca [3 x i32*], align 8
+ %v2.sub = getelementptr inbounds [3 x i32*]* %v2, i64 0, i64 0
+ %tmpcast = bitcast [3 x i32*]* %v2 to %"class.std::vector"*
+ %conv = sext i32 %N to i64
+ store i32* null, i32** %v2.sub, align 8, !tbaa !0
+ %tmp3.i.i.i.i.i = getelementptr inbounds [3 x i32*]* %v2, i64 0, i64 1
+ store i32* null, i32** %tmp3.i.i.i.i.i, align 8, !tbaa !0
+ %tmp4.i.i.i.i.i = getelementptr inbounds [3 x i32*]* %v2, i64 0, i64 2
+ store i32* null, i32** %tmp4.i.i.i.i.i, align 8, !tbaa !0
+ %cmp.i.i.i.i = icmp eq i32 %N, 0
+ br i1 %cmp.i.i.i.i, label %_ZNSt12_Vector_baseIiSaIiEEC2EmRKS0_.exit.thread.i.i, label %cond.true.i.i.i.i
+
+_ZNSt12_Vector_baseIiSaIiEEC2EmRKS0_.exit.thread.i.i: ; preds = %entry
+ store i32* null, i32** %v2.sub, align 8, !tbaa !0
+ store i32* null, i32** %tmp3.i.i.i.i.i, align 8, !tbaa !0
+ %add.ptr.i5.i.i = getelementptr inbounds i32* null, i64 %conv
+ store i32* %add.ptr.i5.i.i, i32** %tmp4.i.i.i.i.i, align 8, !tbaa !0
+ br label %_ZNSt6vectorIiSaIiEEC1EmRKiRKS0_.exit
+
+cond.true.i.i.i.i: ; preds = %entry
+ %cmp.i.i.i.i.i = icmp slt i32 %N, 0
+ br i1 %cmp.i.i.i.i.i, label %if.then.i.i.i.i.i, label %_ZNSt12_Vector_baseIiSaIiEEC2EmRKS0_.exit.i.i
+
+if.then.i.i.i.i.i: ; preds = %cond.true.i.i.i.i
+ call void @_ZSt17__throw_bad_allocv() noreturn nounwind
+ unreachable
+
+_ZNSt12_Vector_baseIiSaIiEEC2EmRKS0_.exit.i.i: ; preds = %cond.true.i.i.i.i
+ %mul.i.i.i.i.i = shl i64 %conv, 2
+ %call3.i.i.i.i.i = call noalias i8* @_Znwm(i64 %mul.i.i.i.i.i) nounwind
+ %0 = bitcast i8* %call3.i.i.i.i.i to i32*
+ store i32* %0, i32** %v2.sub, align 8, !tbaa !0
+ store i32* %0, i32** %tmp3.i.i.i.i.i, align 8, !tbaa !0
+ %add.ptr.i.i.i = getelementptr inbounds i32* %0, i64 %conv
+ store i32* %add.ptr.i.i.i, i32** %tmp4.i.i.i.i.i, align 8, !tbaa !0
+ call void @llvm.memset.p0i8.i64(i8* %call3.i.i.i.i.i, i8 0, i64 %mul.i.i.i.i.i, i32 4, i1 false)
+ br label %_ZNSt6vectorIiSaIiEEC1EmRKiRKS0_.exit
+
+This is just the handling the construction of the vector. Most surprising here
+is the fact that all three null stores in %entry are dead (because we do no
+cross-block DSE).
+
+Also surprising is that %conv isn't simplified to 0 in %....exit.thread.i.i.
+This is a because the client of LazyValueInfo doesn't simplify all instruction
+operands, just selected ones.
+
+//===---------------------------------------------------------------------===//
+
+clang -O3 -fno-exceptions currently compiles this code:
+
+void f(char* a, int n) {
+ __builtin_memset(a, 0, n);
+ for (int i = 0; i < n; ++i)
+ a[i] = 0;
+}
+
+into:
+
+define void @_Z1fPci(i8* nocapture %a, i32 %n) nounwind {
+entry:
+ %conv = sext i32 %n to i64
+ tail call void @llvm.memset.p0i8.i64(i8* %a, i8 0, i64 %conv, i32 1, i1 false)
+ %cmp8 = icmp sgt i32 %n, 0
+ br i1 %cmp8, label %for.body.lr.ph, label %for.end
+
+for.body.lr.ph: ; preds = %entry
+ %tmp10 = add i32 %n, -1
+ %tmp11 = zext i32 %tmp10 to i64
+ %tmp12 = add i64 %tmp11, 1
+ call void @llvm.memset.p0i8.i64(i8* %a, i8 0, i64 %tmp12, i32 1, i1 false)
+ ret void
+
+for.end: ; preds = %entry
+ ret void
+}
+
+This shouldn't need the ((zext (%n - 1)) + 1) game, and it should ideally fold
+the two memset's together. The issue with %n seems to stem from poor handling
+of the original loop.
+
+To simplify this, we need SCEV to know that "n != 0" because of the dominating
+conditional. That would turn the second memset into a simple memset of 'n'.
+
+//===---------------------------------------------------------------------===//
+
+clang -O3 -fno-exceptions currently compiles this code:
+
+struct S {
+ unsigned short m1, m2;
+ unsigned char m3, m4;
+};
+
+void f(int N) {
+ std::vector<S> v(N);
+ extern void sink(void*); sink(&v);
+}
+
+into poor code for zero-initializing 'v' when N is >0. The problem is that
+S is only 6 bytes, but each element is 8 byte-aligned. We generate a loop and
+4 stores on each iteration. If the struct were 8 bytes, this gets turned into
+a memset.
+
+In order to handle this we have to:
+ A) Teach clang to generate metadata for memsets of structs that have holes in
+ them.
+ B) Teach clang to use such a memset for zero init of this struct (since it has
+ a hole), instead of doing elementwise zeroing.
+
+//===---------------------------------------------------------------------===//
+
+clang -O3 currently compiles this code:
+
+extern const int magic;
+double f() { return 0.0 * magic; }
+
+into
+
+@magic = external constant i32
+
+define double @_Z1fv() nounwind readnone {
+entry:
+ %tmp = load i32* @magic, align 4, !tbaa !0
+ %conv = sitofp i32 %tmp to double
+ %mul = fmul double %conv, 0.000000e+00
+ ret double %mul
+}
+
+We should be able to fold away this fmul to 0.0. More generally, fmul(x,0.0)
+can be folded to 0.0 if we can prove that the LHS is not -0.0, not a NaN, and
+not an INF. The CannotBeNegativeZero predicate in value tracking should be
+extended to support general "fpclassify" operations that can return
+yes/no/unknown for each of these predicates.
+
+In this predicate, we know that uitofp is trivially never NaN or -0.0, and
+we know that it isn't +/-Inf if the floating point type has enough exponent bits
+to represent the largest integer value as < inf.
+
+//===---------------------------------------------------------------------===//
+
+When optimizing a transformation that can change the sign of 0.0 (such as the
+0.0*val -> 0.0 transformation above), it might be provable that the sign of the
+expression doesn't matter. For example, by the above rules, we can't transform
+fmul(sitofp(x), 0.0) into 0.0, because x might be -1 and the result of the
+expression is defined to be -0.0.
+
+If we look at the uses of the fmul for example, we might be able to prove that
+all uses don't care about the sign of zero. For example, if we have:
+
+ fadd(fmul(sitofp(x), 0.0), 2.0)
+
+Since we know that x+2.0 doesn't care about the sign of any zeros in X, we can
+transform the fmul to 0.0, and then the fadd to 2.0.
+
+//===---------------------------------------------------------------------===//
+
+We should enhance memcpy/memcpy/memset to allow a metadata node on them
+indicating that some bytes of the transfer are undefined. This is useful for
+frontends like clang when lowering struct copies, when some elements of the
+struct are undefined. Consider something like this:
+
+struct x {
+ char a;
+ int b[4];
+};
+void foo(struct x*P);
+struct x testfunc() {
+ struct x V1, V2;
+ foo(&V1);
+ V2 = V1;
+
+ return V2;
+}
+
+We currently compile this to:
+$ clang t.c -S -o - -O0 -emit-llvm | opt -scalarrepl -S
+
+
+%struct.x = type { i8, [4 x i32] }
+
+define void @testfunc(%struct.x* sret %agg.result) nounwind ssp {
+entry:
+ %V1 = alloca %struct.x, align 4
+ call void @foo(%struct.x* %V1)
+ %tmp1 = bitcast %struct.x* %V1 to i8*
+ %0 = bitcast %struct.x* %V1 to i160*
+ %srcval1 = load i160* %0, align 4
+ %tmp2 = bitcast %struct.x* %agg.result to i8*
+ %1 = bitcast %struct.x* %agg.result to i160*
+ store i160 %srcval1, i160* %1, align 4
+ ret void
+}
+
+This happens because SRoA sees that the temp alloca has is being memcpy'd into
+and out of and it has holes and it has to be conservative. If we knew about the
+holes, then this could be much much better.
+
+Having information about these holes would also improve memcpy (etc) lowering at
+llc time when it gets inlined, because we can use smaller transfers. This also
+avoids partial register stalls in some important cases.
//===---------------------------------------------------------------------===//
projects / oota-llvm.git / blobdiff