Canonicalizing load -> store to use integer type confuses LICM (II) #22849
Description
| Bugzilla Link | 22475 |
| Version | trunk |
| OS | Windows NT |
| Reporter | LLVM Bugzilla Contributor |
| CC | @chandlerc,@hfinkel |
Extended Description
This is somewhat similar to, but distinct from, #22834 .
For this code:
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-pc-windows-msvc"
@in = internal unnamed_addr global i32* null, align 8
@out = internal unnamed_addr global i32* null, align 8
define i64 @bar(i32 %N) {
entry:
br label %do.body
do.body: ; preds = %l2, %entry
%i.0 = phi i32 [ 0, %entry ], [ %inc, %l2 ]
%total = phi i64 [ 0, %entry ], [ %next, %l2 ]
%c = icmp eq i32 %N, 6
br i1 %c, label %l1, label %do.body.l2_crit_edge
do.body.l2_crit_edge: ; preds = %do.body
%inval.pre = load i32** @in, align 8
br label %l2
l1: ; preds = %do.body
%0 = load i32** @in, align 8
store i32* %0, i32** @out, align 8
br label %l2
l2: ; preds = %do.body.l2_crit_edge, %l1
%inval = phi i32* [ %inval.pre, %do.body.l2_crit_edge ], [ %0, %l1 ]
store i32* %inval , i32** @out, align 8
%int = ptrtoint i32* %inval to i64
%next = add i64 %total, %int
%inc = add nsw i32 %i.0, 1
%cmp = icmp slt i32 %inc, %N
br i1 %cmp, label %do.body, label %do.end
do.end: ; preds = %l2
ret i64 %total
}
LICM is able to merge the two stores and sink them into the exit block.
However, if the load/store pair above was canonicalized (before the PHI was created by GVN), and l1 is replaced with:
l1: ; preds = %do.body
%v1 = load i64* bitcast (i32** @in to i64*), align 8
store i64 %v1, i64* bitcast (i32** @out to i64*), align 8
%0 = inttoptr i64 %v1 to i32*
br label %l2
LICM fails to sink the store.
The problem is that LICM::PromoteAliasSet() requires all stores in the alias set to have the same type - while here we have two stores with different types. Relaxing this condition to all stores having the same size doesn't work on its own, because LoopPromoter also relies on the "same type" assumption.