1; RUN: opt < %s -instcombine -S | grep "align 32" | count 1 2 3; It's tempting to have an instcombine in which the src pointer of a 4; memcpy is aligned up to the alignment of the destination, however 5; there are pitfalls. If the src is an alloca, aligning it beyond what 6; the target's stack pointer is aligned at will require dynamic 7; stack realignment, which can require functions that don't otherwise 8; need a frame pointer to need one. 9; 10; Abstaining from this transform is not the only way to approach this 11; issue. Some late phase could be smart enough to reduce alloca 12; alignments when they are greater than they need to be. Or, codegen 13; could do dynamic alignment for just the one alloca, and leave the 14; main stack pointer at its standard alignment. 15 16@dst = global [1024 x i8] zeroinitializer, align 32 17 18define void @foo() nounwind { 19entry: 20 %src = alloca [1024 x i8], align 1 21 %src1 = getelementptr [1024 x i8], [1024 x i8]* %src, i32 0, i32 0 22 call void @llvm.memcpy.p0i8.p0i8.i32(i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @dst, i32 0, i32 0), i8* %src1, i32 1024, i32 1, i1 false) 23 call void @frob(i8* %src1) nounwind 24 ret void 25} 26 27declare void @frob(i8*) 28 29declare void @llvm.memcpy.p0i8.p0i8.i32(i8* nocapture, i8* nocapture, i32, i32, i1) nounwind 30