[llvm] r208479 - R600/SI: Prettier display of input modifiers

[Tom Stellard]

On Tue, May 13, 2014 at 03:48:12PM -0700, Matt Arsenault wrote:
> On 05/13/2014 03:45 PM, Tom Stellard wrote:
> > On Mon, May 12, 2014 at 05:06:28PM -0700, Matt Arsenault wrote:
> >> On May 10, 2014, at 12:18 PM, Vincent Lejeune <vljn at ovi.com> wrote:
> >>
> >>> class VOP3_32 <bits<9> op, string opName, list<dag> pattern> : VOP3 <
> >>>    op, (outs VReg_32:$dst),
> >>> -  (ins VSrc_32:$src0, VSrc_32:$src1, VSrc_32:$src2,
> >>> -   InstFlag:$abs, InstFlag:$clamp, InstFlag:$omod, InstFlag:$neg),
> >>> -  opName#" $dst, $src0, $src1, $src2, $abs, $clamp, $omod, $neg", pattern
> >>> +  (ins InputMods: $src0_modifiers, VSrc_32:$src0, InputMods:$src1_modifiers,
> >>> +   VSrc_32:$src1, InputMods:$src2_modifiers, VSrc_32:$src2,
> >>> +   InstFlag:$clamp, InstFlag:$omod),
> >>> +  opName#" $dst, $src0_modifiers, $src1, $src2, $clamp, $omod", pattern
> >>>> , VOP <opName>;
> >> This seems to be breaking BFE instructions, since the bitwidth always ends up as 0. Is $src1_modifiers supposed to be included in the asm string instead of src1?
> >>
> >> e.g.
> >>
> >> declare i32 @llvm.AMDGPU.bfe.i32(i32, i32, i32) nounwind readnone
> >>
> >> define void @foo(i32 addrspace(1)* %out, i32 addrspace(1)* %ptr) nounwind {
> >>    %load = load i32 addrspace(1)* %ptr, align 4
> >>    %bfe0 = call i32 @llvm.AMDGPU.bfe.i32(i32 %load, i32 0, i32 8) nounwind readnone
> >>    store i32 %bfe0, i32 addrspace(1)* %out, align 4
> >>    ret void
> >> }
> >>
> >> Now produces
> >>
> >> V_BFE_I32 v0, v0, 0, 0, 0, 0
> > What is the expected output?
> >
> > -Tom
> 
> It should be V_BFE_I32 destreg, srcreg, offset, width, which in this 
> case should be V_BFE_I32 v0, v0, 0, 8

Could this be a bug in the convertToVALU code?

-Tom