R600: Use legacy (0 * anything = 0) MUL instructions for pow intrinsics

Fixes wrong lighting in some corner cases with r600g and radeonsi, e.g.
manifested by failure of two piglit/glean tests and intermittent black
patches in many apps.

Tested on SI and RS880.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=62012 [radeonsi]
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=58150 [r600g]

NOTE: This is a candidate for the Mesa stable branch.

Reviewed-by: Christian König <christian.koenig@amd.com>
llvm-svn: 177730
This commit is contained in:
Michel Danzer 2013-03-22 14:09:10 +00:00
parent 4172375713
commit a2e28156b4
4 changed files with 5 additions and 15 deletions

View File

@ -58,7 +58,6 @@ R600TargetLowering::R600TargetLowering(TargetMachine &TM) :
setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i1, Custom);
setOperationAction(ISD::FPOW, MVT::f32, Custom);
setOperationAction(ISD::ROTL, MVT::i32, Custom);
@ -316,7 +315,6 @@ SDValue R600TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const
case ISD::SELECT: return LowerSELECT(Op, DAG);
case ISD::STORE: return LowerSTORE(Op, DAG);
case ISD::LOAD: return LowerLOAD(Op, DAG);
case ISD::FPOW: return LowerFPOW(Op, DAG);
case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
case ISD::INTRINSIC_VOID: {
SDValue Chain = Op.getOperand(0);
@ -918,15 +916,6 @@ SDValue R600TargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const
return DAG.getMergeValues(Ops, 2, DL);
}
SDValue R600TargetLowering::LowerFPOW(SDValue Op,
SelectionDAG &DAG) const {
DebugLoc DL = Op.getDebugLoc();
EVT VT = Op.getValueType();
SDValue LogBase = DAG.getNode(ISD::FLOG2, DL, VT, Op.getOperand(0));
SDValue MulLogBase = DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), LogBase);
return DAG.getNode(ISD::FEXP2, DL, VT, MulLogBase);
}
/// XXX Only kernel functions are supported, so we can assume for now that
/// every function is a kernel function, but in the future we should use
/// separate calling conventions for kernel and non-kernel functions.

View File

@ -59,7 +59,6 @@ private:
SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerFPTOUINT(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerFPOW(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerFrameIndex(SDValue Op, SelectionDAG &DAG) const;

View File

@ -1141,6 +1141,7 @@ let Predicates = [isR600] in {
def RECIP_UINT_r600 : RECIP_UINT_Common <0x78>;
defm DIV_r600 : DIV_Common<RECIP_IEEE_r600>;
def : POW_Common <LOG_IEEE_r600, EXP_IEEE_r600, MUL, R600_Reg32>;
def TGSI_LIT_Z_r600 : TGSI_LIT_Z_Common<MUL_LIT_r600, LOG_CLAMPED_r600, EXP_IEEE_r600>;
def : Pat<(fsqrt R600_Reg32:$src),
@ -1212,6 +1213,7 @@ def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
def SIN_eg : SIN_Common<0x8D>;
def COS_eg : COS_Common<0x8E>;
def : POW_Common <LOG_IEEE_eg, EXP_IEEE_eg, MUL, R600_Reg32>;
def : SIN_PAT <SIN_eg>;
def : COS_PAT <COS_eg>;
def : Pat<(fsqrt R600_Reg32:$src),
@ -1540,13 +1542,14 @@ def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>;
def EXP_IEEE_cm : EXP_IEEE_Common<0x81>;
def LOG_IEEE_ : LOG_IEEE_Common<0x83>;
def LOG_IEEE_cm : LOG_IEEE_Common<0x83>;
def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>;
def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>;
def SIN_cm : SIN_Common<0x8D>;
def COS_cm : COS_Common<0x8E>;
} // End isVector = 1
def : POW_Common <LOG_IEEE_cm, EXP_IEEE_cm, MUL, R600_Reg32>;
def : SIN_PAT <SIN_cm>;
def : COS_PAT <COS_cm>;

View File

@ -1442,8 +1442,7 @@ def : Pat <
/********** ================== **********/
/* llvm.AMDGPU.pow */
/* XXX: We are using IEEE MUL, not the 0 * anything = 0 MUL, is this correct? */
def : POW_Common <V_LOG_F32_e32, V_EXP_F32_e32, V_MUL_F32_e32, VReg_32>;
def : POW_Common <V_LOG_F32_e32, V_EXP_F32_e32, V_MUL_LEGACY_F32_e32, VReg_32>;
def : Pat <
(int_AMDGPU_div VSrc_32:$src0, VSrc_32:$src1),