//===---------------------------------------------------------------------===//
-We should recognize idioms for add-with-carry and turn it into the appropriate
-intrinsics. This example:
+We should recognized various "overflow detection" idioms and translate them into
+llvm.uadd.with.overflow and similar intrinsics. Here is a multiply idiom:
-unsigned add32carry(unsigned sum, unsigned x) {
- unsigned z = sum + x;
- if (sum + x < x)
- z++;
- return z;
+unsigned int mul(unsigned int a,unsigned int b) {
+ if ((unsigned long long)a*b>0xffffffff)
+ exit(0);
+ return a*b;
}
-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 ; <i32> [#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.
-
-//===---------------------------------------------------------------------===//
-
-With the recent changes to make the implicit def/use set explicit in
-machineinstrs, we should change the target descriptions for 'call' instructions
-so that the .td files don't list all the call-clobbered registers as implicit
-defs. Instead, these should be added by the code generator (e.g. on the dag).
-
-This has a number of uses:
-
-1. PPC32/64 and X86 32/64 can avoid having multiple copies of call instructions
- for their different impdef sets.
-2. Targets with multiple calling convs (e.g. x86) which have different clobber
- sets don't need copies of call instructions.
-3. 'Interprocedural register allocation' can be done to reduce the clobber sets
- of calls.
-
-//===---------------------------------------------------------------------===//
-
-Make the PPC branch selector target independant
+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) {
//===---------------------------------------------------------------------===//
-For vector types, TargetData.cpp::getTypeInfo() returns alignment that is equal
+For vector types, DataLayout.cpp::getTypeInfo() returns alignment that is equal
to the type size. It works but can be overly conservative as the alignment of
specific vector types are target dependent.
return v ^ (t >> 8);
}
-Neither is this (very standard idiom):
+//===---------------------------------------------------------------------===//
+
+[LOOP DELETION]
-int f(int n)
-{
- return (((n) << 24) | (((n) & 0xff00) << 8)
- | (((n) >> 8) & 0xff00) | ((n) >> 24));
+We don't delete this output free loop, because trip count analysis doesn't
+realize that it is finite (if it were infinite, it would be undefined). Not
+having this blocks Loop Idiom from matching strlen and friends.
+
+void foo(char *C) {
+ int x = 0;
+ while (*C)
+ ++x,++C;
}
//===---------------------------------------------------------------------===//
c += v & 1;
return c;
}
-unsigned countbits_fast(unsigned v){
- unsigned c;
- for (c = 0; v; c++)
- v &= v - 1; // clear the least significant bit set
- return c;
-}
-BITBOARD = unsigned long long
-int PopCnt(register BITBOARD a) {
- register int c=0;
- while(a) {
- c++;
- a &= a - 1;
- }
- return c;
-}
unsigned int popcount(unsigned int input) {
unsigned int count = 0;
for (unsigned int i = 0; i < 4 * 8; i++)
return count;
}
-This is a form of idiom recognition for loops, the same thing that could be
-useful for recognizing memset/memcpy.
+This should be recognized as CLZ: rdar://8459039
+
+unsigned clz_a(unsigned a) {
+ int i;
+ for (i=0;i<32;i++)
+ if (a & (1<<(31-i)))
+ return i;
+ return 32;
+}
+
+This sort of thing should be added to the loop idiom pass.
//===---------------------------------------------------------------------===//
//===---------------------------------------------------------------------===//
-[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
//===---------------------------------------------------------------------===//
-LSR should know what GPR types a target has from TargetData. This code:
-
-volatile short X, Y; // globals
-
-void foo(int N) {
- int i;
- for (i = 0; i < N; i++) { X = i; Y = i*4; }
-}
-
-produces two near identical IV's (after promotion) on PPC/ARM:
-
-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.
-
-//===---------------------------------------------------------------------===//
-
Tail call elim should be more aggressive, checking to see if the call is
followed by an uncond branch to an exit block.
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
+
+Which gets codegen'd into:
-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)
- ...
+ 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.
+//===---------------------------------------------------------------------===//
+
+This (and similar related idioms):
+
+unsigned int foo(unsigned char i) {
+ return i | (i<<8) | (i<<16) | (i<<24);
+}
+
+compiles into:
+
+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
+}
+
+it would be better as:
+
+unsigned int bar(unsigned char i) {
+ unsigned int j=i | (i << 8);
+ return j | (j<<16);
+}
+
+aka:
+
+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
+}
+
+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.
//===---------------------------------------------------------------------===//
//===---------------------------------------------------------------------===//
-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:
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".
//===---------------------------------------------------------------------===//
-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".
//===---------------------------------------------------------------------===//
+int g(int x) { return (x - 10) < 0; }
+Should combine to "x <= 9" (the sub has nsw). Currently not
+optimized with "clang -emit-llvm-bc | opt -std-compile-opts".
+
+//===---------------------------------------------------------------------===//
+
+int g(int x) { return (x + 10) < 0; }
+Should combine to "x < -10" (the add has nsw). Currently not
+optimized with "clang -emit-llvm-bc | opt -std-compile-opts".
+
+//===---------------------------------------------------------------------===//
+
+int f(int i, int j) { return i < j + 1; }
+int g(int i, int j) { return j > i - 1; }
+Should combine to "i <= j" (the add/sub has nsw). Currently not
+optimized with "clang -emit-llvm-bc | opt -std-compile-opts".
+
+//===---------------------------------------------------------------------===//
+
+unsigned f(unsigned x) { return ((x & 7) + 1) & 15; }
+The & 15 part should be optimized away, it doesn't change the result. 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
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]
actually a conditional increment: loadpre18.c loadpre19.c
+//===---------------------------------------------------------------------===//
+
+[LOAD PRE / STORE SINKING / SPEC HACK]
+
+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.
//===---------------------------------------------------------------------===//
//===---------------------------------------------------------------------===//
-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 |
//===---------------------------------------------------------------------===//
+simplifylibcalls should turn these snprintf idioms into memcpy (GCC PR47917)
+
+char buf1[6], buf2[6], buf3[4], buf4[4];
+int i;
+
+int foo (void) {
+ int ret = snprintf (buf1, sizeof buf1, "abcde");
+ ret += snprintf (buf2, sizeof buf2, "abcdef") * 16;
+ ret += snprintf (buf3, sizeof buf3, "%s", i++ < 6 ? "abc" : "def") * 256;
+ ret += snprintf (buf4, sizeof buf4, "%s", i++ > 10 ? "abcde" : "defgh")*4096;
+ return ret;
+}
+
+//===---------------------------------------------------------------------===//
+
"gas" uses this idiom:
else if (strchr ("+-/*%|&^:[]()~", *intel_parser.op_string))
..
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 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(
//===---------------------------------------------------------------------===//
-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:
return **p;
}
+This can be seen at:
+$ clang t.c -S -o - -mkernel -O0 -emit-llvm | opt -functionattrs -S
+
+
//===---------------------------------------------------------------------===//
Missed instcombine transformation:
//===---------------------------------------------------------------------===//
-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
- 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 [
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
+ ja LBB2_2
+ cmpl $4, %edi
+ jne LBB2_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
+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
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)
-Looks related to PR3100
-
-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 ();
-}
//===---------------------------------------------------------------------===//
//===---------------------------------------------------------------------===//
+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 the addition only occurs in 64-bit mode, and appears to be at
+least partially caused by Scalar Evolution not keeping its cache updated: it
+returns the "wrong" result immediately after indvars runs, but figures out the
+expected result if it is run from scratch on IR resulting from running indvars.
+
+//===---------------------------------------------------------------------===//
+
+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.
+
+//===---------------------------------------------------------------------===//
+
+We don't fold (icmp (add) (add)) unless the two adds only have a single use.
+There are a lot of cases that we're refusing to fold in (e.g.) 256.bzip2, for
+example:
+
+ %indvar.next90 = add i64 %indvar89, 1 ;; Has 2 uses
+ %tmp96 = add i64 %tmp95, 1 ;; Has 1 use
+ %exitcond97 = icmp eq i64 %indvar.next90, %tmp96
+
+We don't fold this because we don't want to introduce an overlapped live range
+of the ivar. However if we can make this more aggressive without causing
+performance issues in two ways:
+
+1. If *either* the LHS or RHS has a single use, we can definitely do the
+ transformation. In the overlapping liverange case we're trading one register
+ use for one fewer operation, which is a reasonable trade. Before doing this
+ we should verify that the llc output actually shrinks for some benchmarks.
+2. If both ops have multiple uses, we can still fold it if the operations are
+ both sinkable to *after* the icmp (e.g. in a subsequent block) which doesn't
+ increase register pressure.
+
+There are a ton of icmp's we aren't simplifying because of the reg pressure
+concern. Care is warranted here though because many of these are induction
+variables and other cases that matter a lot to performance, like the above.
+Here's a blob of code that you can drop into the bottom of visitICmp to see some
+missed cases:
+
+ { Value *A, *B, *C, *D;
+ if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
+ match(Op1, m_Add(m_Value(C), m_Value(D))) &&
+ (A == C || A == D || B == C || B == D)) {
+ errs() << "OP0 = " << *Op0 << " U=" << Op0->getNumUses() << "\n";
+ errs() << "OP1 = " << *Op1 << " U=" << Op1->getNumUses() << "\n";
+ errs() << "CMP = " << I << "\n\n";
+ }
+ }
+
+//===---------------------------------------------------------------------===//
+
+define i1 @test1(i32 %x) nounwind {
+ %and = and i32 %x, 3
+ %cmp = icmp ult i32 %and, 2
+ ret i1 %cmp
+}
+
+Can be folded to (x & 2) == 0.
+
+define i1 @test2(i32 %x) nounwind {
+ %and = and i32 %x, 3
+ %cmp = icmp ugt i32 %and, 1
+ ret i1 %cmp
+}
+
+Can be folded to (x & 2) != 0.
+
+SimplifyDemandedBits shrinks the "and" constant to 2 but instcombine misses the
+icmp transform.
+
+//===---------------------------------------------------------------------===//
+
+This code:
+
+typedef struct {
+int f1:1;
+int f2:1;
+int f3:1;
+int f4:29;
+} t1;
+
+typedef struct {
+int f1:1;
+int f2:1;
+int f3:30;
+} t2;
+
+t1 s1;
+t2 s2;
+
+void func1(void)
+{
+s1.f1 = s2.f1;
+s1.f2 = s2.f2;
+}
+
+Compiles into this IR (on x86-64 at least):
+
+%struct.t1 = type { i8, [3 x i8] }
+@s2 = global %struct.t1 zeroinitializer, align 4
+@s1 = global %struct.t1 zeroinitializer, align 4
+define void @func1() nounwind ssp noredzone {
+entry:
+ %0 = load i32* bitcast (%struct.t1* @s2 to i32*), align 4
+ %bf.val.sext5 = and i32 %0, 1
+ %1 = load i32* bitcast (%struct.t1* @s1 to i32*), align 4
+ %2 = and i32 %1, -4
+ %3 = or i32 %2, %bf.val.sext5
+ %bf.val.sext26 = and i32 %0, 2
+ %4 = or i32 %3, %bf.val.sext26
+ store i32 %4, i32* bitcast (%struct.t1* @s1 to i32*), align 4
+ ret void
+}
+
+The two or/and's should be merged into one each.
+
+//===---------------------------------------------------------------------===//
+
+Machine level code hoisting can be useful in some cases. For example, PR9408
+is about:
+
+typedef union {
+ void (*f1)(int);
+ void (*f2)(long);
+} funcs;
+
+void foo(funcs f, int which) {
+ int a = 5;
+ if (which) {
+ f.f1(a);
+ } else {
+ f.f2(a);
+ }
+}
+
+which we compile to:
+
+foo: # @foo
+# BB#0: # %entry
+ pushq %rbp
+ movq %rsp, %rbp
+ testl %esi, %esi
+ movq %rdi, %rax
+ je .LBB0_2
+# BB#1: # %if.then
+ movl $5, %edi
+ callq *%rax
+ popq %rbp
+ ret
+.LBB0_2: # %if.else
+ movl $5, %edi
+ callq *%rax
+ popq %rbp
+ ret
+
+Note that bb1 and bb2 are the same. This doesn't happen at the IR level
+because one call is passing an i32 and the other is passing an i64.
+
+//===---------------------------------------------------------------------===//
+
+I see this sort of pattern in 176.gcc in a few places (e.g. the start of
+store_bit_field). The rem should be replaced with a multiply and subtract:
+
+ %3 = sdiv i32 %A, %B
+ %4 = srem i32 %A, %B
+
+Similarly for udiv/urem. Note that this shouldn't be done on X86 or ARM,
+which can do this in a single operation (instruction or libcall). It is
+probably best to do this in the code generator.
+
+//===---------------------------------------------------------------------===//
+
+unsigned foo(unsigned x, unsigned y) { return (x & y) == 0 || x == 0; }
+should fold to (x & y) == 0.
+
+//===---------------------------------------------------------------------===//
+
+unsigned foo(unsigned x, unsigned y) { return x > y && x != 0; }
+should fold to x > y.
+
+//===---------------------------------------------------------------------===//