[memcpyopt] memset not created out of per-field store instructions

[llvmbot]

Bugzilla Link	9417
Version	trunk
OS	Linux
Attachments	testcase .ll
Reporter	LLVM Bugzilla Contributor
CC	@lattner

Extended Description

The attached file is a .ll with a simple loop doing 3 stores of 0 to each member in @​glbl. Loop idiom recognizer won't merge it because each store has a stride larger than the width of the store.

Who's responsible for merging:

A = gep foo, 0
B = gep foo, 1
store 0, A
store 0, B

into one store/memset instruction?

This is pulled from MultiSource/Benchmarks/Ptrdist/ks which has gcc/llc = 0.46 on x86-64 linux.

[lattner]

memcpyopt is responsible for doing this.

[lattner]

All of the time in PtrDist/ks is spent inside a pointer chasing loop:

for (modNode = nets[(*netNode).net];
     modNode != NULL;
     modNode = (*modNode).next) {
    if ((*modNode).module == module) {
	gain = gain + netCost;
    }
}

The issue here is code layout.

[lattner]

This is lame, but I'm not going to have time to look at it. If we're still that far behind GCC, it implies that there is something bigger going on. I don't think GCC is doing this level of optimization, maybe it's a frontend difference.

[arsenm]

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"
target triple = "x86_64-unknown-linux-gnu"

%s = type { i64, i64, i64 }

@glbl = unnamed_addr global [256 x %s] zeroinitializer, align 16

define void @test() {
entry:
  br label %loop

loop:                                             ; preds = %loop, %entry
  %indvar = phi i64 [ 0, %entry ], [ %indvar2, %loop ]
  %x = getelementptr [256 x %s], ptr @glbl, i64 0, i64 %indvar, i32 0
  %y = getelementptr [256 x %s], ptr @glbl, i64 0, i64 %indvar, i32 1
  %z = getelementptr [256 x %s], ptr @glbl, i64 0, i64 %indvar, i32 2
  store i64 0, ptr %x, align 8
  store i64 0, ptr %y, align 8
  store i64 0, ptr %z, align 8
  %indvar2 = add i64 %indvar, 1
  %exitcond = icmp eq i64 %indvar2, 256
  br i1 %exitcond, label %exit, label %loop

exit:                                             ; preds = %loop
  ret void
}

opt -p=memcpyopt does produce a memset. opt -O3 reduces the whole function to just the memset