[PATCH] R600/SI: Add a pattern for i32 ctpop inside a branch

[Tom Stellard]

On Wed, Jun 18, 2014 at 10:28:47AM -0700, Matt Arsenault wrote:
> 
> On Jun 18, 2014, at 9:25 AM, Tom Stellard <thomas.stellard at amd.com> wrote:
> 
> > ---
> > lib/Target/R600/SIInstructions.td |  5 +++++
> > test/CodeGen/R600/ctpop.ll        | 30 ++++++++++++++++++++++++++++++
> > 2 files changed, 35 insertions(+)
> > 
> > diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
> > index 7dde8c3..2121a1e 100644
> > --- a/lib/Target/R600/SIInstructions.td
> > +++ b/lib/Target/R600/SIInstructions.td
> > @@ -2714,6 +2714,11 @@ def : Pat <
> >> ;
> > 
> > def : Pat <
> > +   (i32 (ctpop i32:$popcnt)),
> > +   (V_BCNT_U32_B32_e32 $popcnt, (V_MOV_B32_e32 0))
> > +>;
> > +
> > +def : Pat <
> >   (i64 (ctpop i64:$src)),
> >   (INSERT_SUBREG (INSERT_SUBREG (i64 (IMPLICIT_DEF)),
> >     (S_BCNT1_I32_B64 $src), sub0),
> > diff --git a/test/CodeGen/R600/ctpop.ll b/test/CodeGen/R600/ctpop.ll
> > index e4d11e0..15be8e1 100644
> > --- a/test/CodeGen/R600/ctpop.ll
> > +++ b/test/CodeGen/R600/ctpop.ll
> > @@ -252,3 +252,33 @@ define void @v_ctpop_i32_add_vvar_inv(i32 addrspace(1)* noalias %out, i32 addrsp
> >   store i32 %add, i32 addrspace(1)* %out, align 4
> >   ret void
> > }
> > +
> > +; FIXME: We currently disallow SALU instructions in all branches,
> > +; but there are some cases when the should be allowed.
> > +
> > +; FUNC-LABEL: @ctpop_i32_in_br
> > +; SI: BUFFER_LOAD_DWORD [[VAL:v[0-9]+]],
> > +; SI: V_BCNT_U32_B32_e64 [[RESULT:v[0-9]+]], [[VAL]], 0
> > +; SI: BUFFER_STORE_DWORD [[RESULT]],
> > +; SI: S_ENDPGM
> > +; EG: BCNT_INT
> > +define void @ctpop_i32_in_br(i32 addrspace(1)* %out, i32 addrspace(1)* %in, i32 %cond) {
> > +entry:
> > +  %0 = icmp eq i32 %cond, 0
> > +  br i1 %0, label %if, label %else
> > +
> > +if:
> > +  %1 = load i32 addrspace(1)* %in
> > +  %2 = call i32 @llvm.ctpop.i32(i32 %1)
> > +  br label %endif
> > +
> > +else:
> > +  %3 = getelementptr i32 addrspace(1)* %in, i32 1
> > +  %4 = load i32 addrspace(1)* %3
> > +  br label %endif
> > +
> > +endif:
> > +  %5 = phi i32 [%2, %if], [%4, %else]
> > +  store i32 %5, i32 addrspace(1)* %out
> > +  ret void
> > +}
> > -- 
> > 1.8.1.5
> 
> Why not also i64, unless this is supposed to be a very temporary hack

Here is an Updated patch with an i64 pattern.

-Tom

> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-R600-SI-Add-patterns-for-ctpop-inside-a-branch.patch
Type: text/x-diff
Size: 5162 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140619/ae71adda/attachment.patch>