[Mesa-dev] [PATCH 4/7] radeon/llvm: use llvm fabs intrinsic

[Vincent]

Le mardi 09 octobre 2012 à 10:49 -0400, Tom Stellard a écrit :
> On Mon, Oct 08, 2012 at 04:47:10PM +0200, Vincent Lejeune wrote:
> > ---
> >  src/gallium/drivers/radeon/AMDGPUISelLowering.cpp   | 3 +--
> >  src/gallium/drivers/radeon/AMDILIntrinsics.td       | 1 -
> >  src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c | 6 +++---
> >  3 files changed, 4 insertions(+), 6 deletions(-)
> > 
> > diff --git a/src/gallium/drivers/radeon/AMDGPUISelLowering.cpp b/src/gallium/drivers/radeon/AMDGPUISelLowering.cpp
> > index aee625d..30dd8f3 100644
> > --- a/src/gallium/drivers/radeon/AMDGPUISelLowering.cpp
> > +++ b/src/gallium/drivers/radeon/AMDGPUISelLowering.cpp
> > @@ -36,6 +36,7 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
> >    setOperationAction(ISD::FEXP2,  MVT::f32, Legal);
> >    setOperationAction(ISD::FPOW,   MVT::f32, Legal);
> >    setOperationAction(ISD::FLOG2,  MVT::f32, Legal);
> > +  setOperationAction(ISD::FABS,   MVT::f32, Legal);
> >    setOperationAction(ISD::FRINT,  MVT::f32, Legal);
> >  
> >    setOperationAction(ISD::UDIV, MVT::i32, Expand);
> > @@ -110,8 +111,6 @@ SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
> >        return LowerIntrinsicIABS(Op, DAG);
> >      case AMDGPUIntrinsic::AMDIL_exp:
> >        return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
> > -    case AMDGPUIntrinsic::AMDIL_fabs:
> > -      return DAG.getNode(ISD::FABS, DL, VT, Op.getOperand(1));
> >      case AMDGPUIntrinsic::AMDGPU_lrp:
> >        return LowerIntrinsicLRP(Op, DAG);
> >      case AMDGPUIntrinsic::AMDIL_fraction:
> > diff --git a/src/gallium/drivers/radeon/AMDILIntrinsics.td b/src/gallium/drivers/radeon/AMDILIntrinsics.td
> > index 4de5767..213c8bb 100644
> > --- a/src/gallium/drivers/radeon/AMDILIntrinsics.td
> > +++ b/src/gallium/drivers/radeon/AMDILIntrinsics.td
> > @@ -66,7 +66,6 @@ let TargetPrefix = "AMDIL", isTarget = 1 in {
> >  }
> >  
> >  let TargetPrefix = "AMDIL", isTarget = 1 in {
> > -  def int_AMDIL_fabs : GCCBuiltin<"__amdil_fabs">, UnaryIntFloat;
> >    def int_AMDIL_abs : GCCBuiltin<"__amdil_abs">, UnaryIntInt;
> >  
> >    def int_AMDIL_bit_extract_i32 : GCCBuiltin<"__amdil_ibit_extract">,
> > diff --git a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
> > index cc690c0..a8327ac 100644
> > --- a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
> > +++ b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
> > @@ -538,7 +538,7 @@ static void emit_prepare_cube_coords(
> >  		coords[i] = LLVMBuildExtractElement(builder, v, idx, "");
> >  	}
> >  
> > -	coords[2] = build_intrinsic(builder, "llvm.AMDIL.fabs.",
> > +	coords[2] = build_intrinsic(builder, "fabs",
> >  			type, &coords[2], 1, LLVMReadNoneAttribute);
> >  	coords[2] = build_intrinsic(builder, "llvm.AMDGPU.rcp",
> >  			type, &coords[2], 1, LLVMReadNoneAttribute);
> > @@ -1122,8 +1122,8 @@ void radeon_llvm_context_init(struct radeon_llvm_context * ctx)
> >  
> >  
> >  
> > -	bld_base->op_actions[TGSI_OPCODE_ABS].emit = build_tgsi_intrinsic_nomem;
> > -	bld_base->op_actions[TGSI_OPCODE_ABS].intr_name = "llvm.AMDIL.fabs.";
> > +	bld_base->op_actions[TGSI_OPCODE_ABS].emit = build_tgsi_intrinsic_readonly;
> > +	bld_base->op_actions[TGSI_OPCODE_ABS].intr_name = "fabs";
> 
> I've noticed this on a few of the patches.  Why fabs for the intrinsic
> name instead of llvm.fabs?


I didn't find "llvm.fabs" intrinsics ID in neither llvm 3.0 nor llvm
3.1, in the include/llvm/Intrinsics.{h,gen} files ; and radeon loader
always lower call to llvm.fabs/llvm.fabs.f32 to a function call.
It seems the intrinsic has been superseded by the "fabs" LibFunc symbol
in llvm. At the instruction selection stage, the fabs function call can
be lowered to the FABS Node if the target support it, otherwise it uses
a function call. In fact I noticed that this behavior is shared by a lot
of function from the libm library (floor, trunc, rint, ...) : there is
no llvm intrinsic, but the ISD node is generated when a call to the libm
function is found. (in
lib/Codegen/SelectionDAG/SelectionDAGBuilder.cpp).
Some libm function have still an llvm intrinsic like cos, sin, exp, log
and sqrt. IMHO this allow target support several implementations of
these functions, a native one and one that uses libm code.

Regards,
Vincent


> 
> -Tom
> 
> >  	bld_base->op_actions[TGSI_OPCODE_ARL].emit = build_tgsi_intrinsic_nomem;
> >  	bld_base->op_actions[TGSI_OPCODE_ARL].intr_name = "llvm.AMDGPU.arl";
> >  	bld_base->op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
> > -- 
> > 1.7.11.4
> > 
> > _______________________________________________
> > mesa-dev mailing list
> > mesa-dev at lists.freedesktop.org
> > http://lists.freedesktop.org/mailman/listinfo/mesa-dev