31 lines · plain
1; RUN: opt < %s -passes=instcombine -S | grep "align 32" | count 22 3; It's tempting to have an instcombine in which the src pointer of a4; memcpy is aligned up to the alignment of the destination, however5; there are pitfalls. If the src is an alloca, aligning it beyond what6; the target's stack pointer is aligned at will require dynamic7; stack realignment, which can require functions that don't otherwise8; need a frame pointer to need one.9;10; Abstaining from this transform is not the only way to approach this11; issue. Some late phase could be smart enough to reduce alloca12; alignments when they are greater than they need to be. Or, codegen13; could do dynamic alignment for just the one alloca, and leave the14; main stack pointer at its standard alignment.15;16 17 18@dst = global [1024 x i8] zeroinitializer, align 3219 20define void @foo() nounwind {21entry:22 %src = alloca [1024 x i8], align 6423 call void @llvm.memcpy.p0.p0.i32(ptr align 32 @dst, ptr align 32 %src, i32 1024, i1 false)24 call void @frob(ptr %src) nounwind25 ret void26}27 28declare void @frob(ptr)29 30declare void @llvm.memcpy.p0.p0.i32(ptr nocapture, ptr nocapture, i32, i1) nounwind31