Carlo Alberto Ferraris
2012-Feb-25 11:17 UTC
[LLVMdev] Missed optimization on array initialization
Prompted by a SO post (http://stackoverflow.com/questions/9441882/compiler-instruction-reordering-optimizations-in-c-and-what-inhibits-them/9442363) I checked and found that LLVM yields the same (seemingly) suboptimal code as MSVC. Consider the following, simplified, C snippet: extern void bar(int*); void foo(int a) { int ar[100] = {a}; if (a) return; bar(ar); } Ideally, the array initialization should be sank after the return, but in Clang/LLVM 3.0 this doesn't happen: ; ModuleID = '/tmp/webcompile/_11079_0.bc' target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128" target triple = "x86_64-unknown-linux-gnu" define void @_Z3fooi(i32 %a) uwtable { %ar =alloca [100 xi32],align 16 %1 =bitcast [100 xi32]* %arto i8* call void @llvm.memset.p0i8.i64(i8* %1,i8 0,i64 400,i32 16,i1 false) %2 =getelementptr inbounds [100 xi32]* %ar,i64 0,i64 0 store i32 %a,i32* %2,align 16, !tbaa !0 %3 =icmp eq i32 %a, 0 br i1 %3,label %4,label %5 ;<label>:4 ; preds = %0 call void @_Z3barPi(i32* %2) br label %5 ;<label>:5 ; preds = %4, %0 ret void } declare void @llvm.memset.p0i8.i64(i8*nocapture,i8,i64,i32,i1)nounwind declare void @_Z3barPi(i32*) !0 = metadata !{metadata !"int", metadata !1} !1 = metadata !{metadata !"omnipotent char", metadata !2} !2 = metadata !{metadata !"Simple C/C++ TBAA",null} and this gets emitted as (for x64, but x86 is similar): # BB#0: pushq %rbx .Ltmp3: .cfi_def_cfa_offset 16 subq $400, %rsp # imm = 0x190 .Ltmp4: .cfi_def_cfa_offset 416 .Ltmp5: .cfi_offset %rbx, -16 movl %edi, %ebx leaq (%rsp), %rdi xorl %esi, %esi movl $400, %edx # imm = 0x190 callq memset movl %ebx, (%rsp) testl %ebx, %ebx jne .LBB0_2 # BB#1: leaq (%rsp), %rdi callq _Z3barPi .LBB0_2: addq $400, %rsp # imm = 0x190 popq %rbx ret I don't have ToT at hand, so I don't know if this is still the case. Any idea why this might be happening? -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20120225/4969db25/attachment.html> -------------- next part -------------- A non-text attachment was scrubbed... Name: cafxx.vcf Type: text/x-vcard Size: 230 bytes Desc: not available URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20120225/4969db25/attachment.vcf>
Hi Carlo, for what it's worth, gcc-4.7 doesn't get this either. Ciao, Duncan. On 25/02/12 12:17, Carlo Alberto Ferraris wrote:> Prompted by a SO post > (http://stackoverflow.com/questions/9441882/compiler-instruction-reordering-optimizations-in-c-and-what-inhibits-them/9442363) > I checked and found that LLVM yields the same (seemingly) suboptimal code as MSVC. > Consider the following, simplified, C snippet: > > extern void bar(int*); > > void foo(int a) > { > int ar[100] = {a}; > if (a) > return; > bar(ar); > } > > Ideally, the array initialization should be sank after the return, but in > Clang/LLVM 3.0 this doesn't happen: > > ; ModuleID ='/tmp/webcompile/_11079_0.bc' > target datalayout ="e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128" > target triple ="x86_64-unknown-linux-gnu" > > define void @_Z3fooi(i32 %a) uwtable { > %ar =alloca [100 xi32],align 16 > %1 =bitcast [100 xi32]* %arto i8* > call void @llvm.memset.p0i8.i64(i8* %1,i8 0,i64 400,i32 16,i1 false) > %2 =getelementptr inbounds [100 xi32]* %ar,i64 0,i64 0 > store i32 %a,i32* %2,align 16, !tbaa !0 > %3 =icmp eq i32 %a, 0 > br i1 %3,label %4,label %5 > > ;<label>:4 ; preds = %0 > call void @_Z3barPi(i32* %2) > br label %5 > > ;<label>:5 ; preds = %4, %0 > ret void > } > > declare void @llvm.memset.p0i8.i64(i8*nocapture,i8,i64,i32,i1)nounwind > > declare void @_Z3barPi(i32*) > > !0 = metadata !{metadata !"int", metadata !1} > !1 = metadata !{metadata !"omnipotent char", metadata !2} > !2 = metadata !{metadata !"Simple C/C++ TBAA",null} > > and this gets emitted as (for x64, but x86 is similar): > > # BB#0: > pushq %rbx > .Ltmp3: > .cfi_def_cfa_offset 16 > subq $400, %rsp # imm = 0x190 > .Ltmp4: > .cfi_def_cfa_offset 416 > .Ltmp5: > .cfi_offset %rbx, -16 > movl %edi, %ebx > leaq (%rsp), %rdi > xorl %esi, %esi > movl $400, %edx # imm = 0x190 > callq memset > movl %ebx, (%rsp) > testl %ebx, %ebx > jne .LBB0_2 > # BB#1: > leaq (%rsp), %rdi > callq _Z3barPi > .LBB0_2: > addq $400, %rsp # imm = 0x190 > popq %rbx > ret > > I don't have ToT at hand, so I don't know if this is still the case. Any idea > why this might be happening? > > > > _______________________________________________ > LLVM Developers mailing list > LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu > http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
On Feb 25, 2012, at 3:17 AM, Carlo Alberto Ferraris wrote:> Prompted by a SO post (http://stackoverflow.com/questions/9441882/compiler-instruction-reordering-optimizations-in-c-and-what-inhibits-them/9442363) I checked and found that LLVM yields the same (seemingly) suboptimal code as MSVC. > Consider the following, simplified, C snippet:> extern void bar(int*); > > void foo(int a) > { > int ar[100] = {a}; > if (a) > return; > bar(ar); > } > > Ideally, the array initialization should be sank after the return, but in Clang/LLVM 3.0 this doesn't happen:This is a straight-forward form of code motion we don't implement, which would be built on partially dead store analysis. Our dead store analysis in general isn't very powerful, and cannot see across blocks. It turns out that it is pretty expensive and doesn't often lead to big performance wins. That said, it is certainly an area that should be improved. I'll note that the original example from SO is more complex. Instead of a single store, it is a whole loop that initializes the array. Handling this case requires moving the entire loop, which requires fairly heroic compiler analysis. The saving grace is that that case is equivalent to a memcpy, so we may be able to handle *that* someday.> %ar = alloca [100 x i32], align 16 > %1 = bitcast [100 x i32]* %ar to i8* > call void @llvm.memset.p0i8.i64(i8* %1, i8 0, i64 400, i32 16, i1 false) > %2 = getelementptr inbounds [100 x i32]* %ar, i64 0, i64 0 > store i32 %a, i32* %2, align 16, !tbaa !0I'm surprised that we're not shortening the memset to skip setting the dead element. That *is* something that we should be able to handle. Pete, didn't you implement this a while ago? -Chris -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20120225/975f3c38/attachment.html>
On Feb 25, 2012, at 10:32 AM, Chris Lattner <clattner at apple.com> wrote:> > On Feb 25, 2012, at 3:17 AM, Carlo Alberto Ferraris wrote: > >> Prompted by a SO post (http://stackoverflow.com/questions/9441882/compiler-instruction-reordering-optimizations-in-c-and-what-inhibits-them/9442363) I checked and found that LLVM yields the same (seemingly) suboptimal code as MSVC. >> Consider the following, simplified, C snippet: > >> extern void bar(int*); >> >> void foo(int a) >> { >> int ar[100] = {a}; >> if (a) >> return; >> bar(ar); >> } >> >> Ideally, the array initialization should be sank after the return, but in Clang/LLVM 3.0 this doesn't happen: > > This is a straight-forward form of code motion we don't implement, which would be built on partially dead store analysis. Our dead store analysis in general isn't very powerful, and cannot see across blocks. It turns out that it is pretty expensive and doesn't often lead to big performance wins. That said, it is certainly an area that should be improved. > > I'll note that the original example from SO is more complex. Instead of a single store, it is a whole loop that initializes the array. Handling this case requires moving the entire loop, which requires fairly heroic compiler analysis. The saving grace is that that case is equivalent to a memcpy, so we may be able to handle *that* someday. > >> %ar = alloca [100 x i32], align 16 >> %1 = bitcast [100 x i32]* %ar to i8* >> call void @llvm.memset.p0i8.i64(i8* %1, i8 0, i64 400, i32 16, i1 false) >> %2 = getelementptr inbounds [100 x i32]* %ar, i64 0, i64 0 >> store i32 %a, i32* %2, align 16, !tbaa !0 > > I'm surprised that we're not shortening the memset to skip setting the dead element. That *is* something that we should be able to handle. Pete, didn't you implement this a while ago?Yeah. I think my implementation only trimmed stores to the end of the memset but this is the start. I'll take a look at improving that. Will probably only want to shorten the start of the memset when it's not going to shorten it to a horribly unaligned start position but that's ok here. Pete> > -Chris > > _______________________________________________ > LLVM Developers mailing list > LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu > http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20120225/90e8d62b/attachment.html>