[loop-idiom] GVN fails to remove loads after loop-idiom recognition

[atrick]

Bugzilla Link	10872
Resolution	FIXED
Resolved on	Sep 06, 2011 18:34
Version	trunk
OS	All
Attachments	-O3 optimized, -O3 -unroll-scev, test bitcode
CC	@asl,@sunfishcode

Extended Description

Test case: SingleSource/Benchmarks/Stanford test.simple.Puzzle on A9 is 14% slower with -unroll-scev.

Running with -O3 optimizes to O3.ll (0.54s)

-unroll-scev produces O3-unroll-scev.ll (0.64s)

These runs were using r138990. -unroll-scev will soon be default, but -disable-unroll-scev will be available.

llc -mcpu=cortex-a9 -relocation-model=pic -disable-fp-elim -disable-non-leaf-fp-elim O3.ll

-unroll-scev exposes more opportunities for memset_pattern, resulting in:

call void @​memset_pattern16(i8* bitcast (i32* getelementptr inbounds ([13 x [512 x i32]]* @​p, i32 0, i32 6, i32 0) to i8*), i8* bitcast ([4 x i32]* @.memset_pattern3 to i8*), i32 12) nounwind

This is fine, but then GVN fails to remove the subsequent loads:

%tmp2.i = load i32* getelementptr inbounds ([13 x i32]* @​piecemax, i32 0, i32 0), align 4, !tbaa !​3

for.end.i: ; preds = %for.inc.i8, %if.then
%tmp14.i = load i32* getelementptr inbounds ([13 x i32]* @​class, i32 0, i32 0), align 4, !tbaa !​3

Removing %tmp2.i exposes lots of constant folding, but it is the removal of %tmp14.i that speeds up the benchmark.

Also rdar://10065079

[llvmbot]

Fixed in r139204.