[PATCH] Fix a performance problem in gep(gep ...) merging

[Wei Mi]

I collected the stat data for spec2000. Overall, reaching the limit of
MaxLookupSearchDepth is not quite often. After we disable gep merging,
the times when the limit being reached only increase a little for two
files in spec2000. The stat data is attached.

Without the patch -- with gep merging:
reach_limit = 27, total = 81402, ratio = 0.03%, file = sv.c

With the patch -- without gep merging:
reach_limit = 28, total = 81486, ratio = 0.03%, file = sv.c
reach_limit = 8, total = 1107, ratio = 0.72%, file = tietze.c

I also tested the performance for spec2000 and internal benchmarks.
There were no performance differences.

I added the stat in -stat result. I will send it out in a separate patch.

Wei.


On Tue, Apr 21, 2015 at 4:05 PM, Wei Mi <wmi at google.com> wrote:
> REPOSITORY
>   rL LLVM
>
> http://reviews.llvm.org/D8911
>
> Files:
>   llvm/trunk/lib/Transforms/InstCombine/InstructionCombining.cpp
>   llvm/trunk/test/Transforms/InstCombine/descale-zero.ll
>   llvm/trunk/test/Transforms/InstCombine/getelementptr.ll
>
> Index: llvm/trunk/test/Transforms/InstCombine/descale-zero.ll
> ===================================================================
> --- llvm/trunk/test/Transforms/InstCombine/descale-zero.ll
> +++ llvm/trunk/test/Transforms/InstCombine/descale-zero.ll
> @@ -1,20 +0,0 @@
> -; RUN: opt < %s -instcombine -S | FileCheck %s
> -
> -target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
> -target triple = "x86_64-apple-macosx10.10.0"
> -
> -define internal i8* @descale_zero() {
> -entry:
> -; CHECK: load i8*, i8** inttoptr (i64 48 to i8**), align 16
> -; CHECK-NEXT: ret i8*
> -  %i16_ptr = load i16*, i16** inttoptr (i64 48 to i16**), align 16
> -  %num = load i64, i64* inttoptr (i64 64 to i64*), align 64
> -  %num_times_2 = shl i64 %num, 1
> -  %num_times_2_plus_4 = add i64 %num_times_2, 4
> -  %i8_ptr = bitcast i16* %i16_ptr to i8*
> -  %i8_ptr_num_times_2_plus_4 = getelementptr i8, i8* %i8_ptr, i64 %num_times_2_plus_4
> -  %num_times_neg2 = mul i64 %num, -2
> -  %num_times_neg2_minus_4 = add i64 %num_times_neg2, -4
> -  %addr = getelementptr i8, i8* %i8_ptr_num_times_2_plus_4, i64 %num_times_neg2_minus_4
> -  ret i8* %addr
> -}
> Index: llvm/trunk/test/Transforms/InstCombine/getelementptr.ll
> ===================================================================
> --- llvm/trunk/test/Transforms/InstCombine/getelementptr.ll
> +++ llvm/trunk/test/Transforms/InstCombine/getelementptr.ll
> @@ -104,8 +104,8 @@
>          %B = getelementptr i32, i32* %A, i64 %D
>          ret i32* %B
>  ; CHECK-LABEL: @test7(
> -; CHECK: %A.sum = add i64 %C, %D
> -; CHECK: getelementptr i32, i32* %I, i64 %A.sum
> +; CHECK: %A = getelementptr i32, i32* %I, i64 %C
> +; CHECK: %B = getelementptr i32, i32* %A, i64 %D
>  }
>
>  define i8* @test8([10 x i32]* %X) {
> Index: llvm/trunk/lib/Transforms/InstCombine/InstructionCombining.cpp
> ===================================================================
> --- llvm/trunk/lib/Transforms/InstCombine/InstructionCombining.cpp
> +++ llvm/trunk/lib/Transforms/InstCombine/InstructionCombining.cpp
> @@ -1467,6 +1467,11 @@
>          // normalized.
>          if (SO1->getType() != GO1->getType())
>            return nullptr;
> +        // Only do the combine when GO1 and SO1 are both constants. Only in
> +        // this case, we are sure the cost after the merge is never more than
> +        // that before the merge.
> +        if (!isa<Constant>(GO1) || !isa<Constant>(SO1))
> +          return nullptr;
>          Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
>        }
>
> EMAIL PREFERENCES
>   http://reviews.llvm.org/settings/panel/emailpreferences/
-------------- next part --------------
A non-text attachment was scrubbed...
Name: disable.gepmerging.stat
Type: application/octet-stream
Size: 2274 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150511/097c052c/attachment.obj>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: gepmerging.stat
Type: application/octet-stream
Size: 2150 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150511/097c052c/attachment-0001.obj>