Dead store elimination kills non-dead store as part of memcpy

[jsonn]

Bugzilla Link	11455
Resolution	FIXED
Resolved on	Dec 02, 2011 18:04
Version	trunk
OS	Linux
Attachments	Test case
CC	@efriedma-quic,@cooperp

Extended Description

Changeset 144239 introduced more aggressive DSE. This created a codegen bug shown in the attached input. With -O2 for AMD64, tf->tf_rdi is no longer copied.

[jsonn]

assigned to @cooperp

[llvmbot]

Testcase:

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"

%struct.trapframe = type { i64, i64 }

define void @​cpu_lwp_fork(%struct.trapframe* %md_regs, i64 %pcb_rsp0) nounwind uwtable {
entry:
%tmp = inttoptr i64 %pcb_rsp0 to %struct.trapframe*
%add.ptr = getelementptr inbounds %struct.trapframe* %tmp, i64 -1
%tmp1 = bitcast %struct.trapframe* %add.ptr to i8*
%tmp2 = bitcast %struct.trapframe* %md_regs to i8*
call void @​llvm.memcpy.p0i8.p0i8.i64(i8* %tmp1, i8* %tmp2, i64 16, i32 1, i1 false)
%tf_trapno = getelementptr inbounds %struct.trapframe* %tmp, i64 -1, i32 1
store i64 3, i64* %tf_trapno, align 8
ret void
}

declare void @​llvm.memcpy.p0i8.p0i8.i64(i8* nocapture, i8* nocapture, i64, i32, i1) nounwind

opt -dse changes the "i64 16" to "i64 8" in the size of the memcpy.

[efriedma-quic]

opt -dse changes the "i64 16" to "i64 8" in the size of the memcpy.

I don't follow how that transformation is incorrect; the store overlaps with memcpy.

[jsonn]

What surprises me is that clang -O2 -S -emit-llvm output for the original test case does not use memcpy intrinsic.

[efriedma-quic]

What surprises me is that clang -O2 -S -emit-llvm output for the original test
case does not use memcpy intrinsic.

instcombine will transform an 8-byte memcpy into a load+store.

[efriedma-quic]

Real testcase:

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-apple-macosx10.7.2"
%struct.trapframe = type { i64, i64, i64 }
declare void @​llvm.memcpy.p0i8.p0i8.i64(i8* nocapture, i8* nocapture, i64, i32, i1) nounwind
define void @​cpu_lwp_fork(%struct.trapframe* %md_regs, i64 %pcb_rsp0) nounwind uwtable noinline ssp {
entry:
%0 = inttoptr i64 %pcb_rsp0 to %struct.trapframe*
%add.ptr = getelementptr inbounds %struct.trapframe* %0, i64 -1
%1 = bitcast %struct.trapframe* %add.ptr to i8*
%2 = bitcast %struct.trapframe* %md_regs to i8*
call void @​llvm.memcpy.p0i8.p0i8.i64(i8* %1, i8* %2, i64 24, i32 1, i1 false)
%tf_trapno = getelementptr inbounds %struct.trapframe* %0, i64 -1, i32 1
store i64 3, i64* %tf_trapno, align 8
ret void
}

Reproduces with opt -dse. Looks like isOverwrite in DeadStoreElimination.cpp is doing some incorrect math.

[cooperp]

Thanks Eli. I'll take a look.

[cooperp]

Fixed in r145731