[PATCH 1/1] R600: Use optimized 24bit path in udivrem
Jan Vesely
jan.vesely at rutgers.edu
Tue Jul 29 15:40:21 PDT 2014
v2: drop enum keyword
use correct extension mode
don;t bother computing the sign in unsinged case
Signed-off-by: Jan Vesely <jan.vesely at rutgers.edu>
---
lib/Target/R600/AMDGPUISelLowering.cpp | 55 +++++---
lib/Target/R600/AMDGPUISelLowering.h | 2 +-
test/CodeGen/R600/udivrem24.ll | 244 +++++++++++++++++++++++++++++++++
3 files changed, 283 insertions(+), 18 deletions(-)
create mode 100644 test/CodeGen/R600/udivrem24.ll
diff --git a/lib/Target/R600/AMDGPUISelLowering.cpp b/lib/Target/R600/AMDGPUISelLowering.cpp
index 3996093..70d101a 100644
--- a/lib/Target/R600/AMDGPUISelLowering.cpp
+++ b/lib/Target/R600/AMDGPUISelLowering.cpp
@@ -1381,7 +1381,7 @@ SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
// This is a shortcut for integer division because we have fast i32<->f32
// conversions, and fast f32 reciprocal instructions. The fractional part of a
// float is enough to accurately represent up to a 24-bit integer.
-SDValue AMDGPUTargetLowering::LowerSDIVREM24(SDValue Op, SelectionDAG &DAG) const {
+SDValue AMDGPUTargetLowering::LowerDIVREM24(SDValue Op, SelectionDAG &DAG, bool sign) const {
SDLoc DL(Op);
EVT VT = Op.getValueType();
SDValue LHS = Op.getOperand(0);
@@ -1389,6 +1389,9 @@ SDValue AMDGPUTargetLowering::LowerSDIVREM24(SDValue Op, SelectionDAG &DAG) cons
MVT IntVT = MVT::i32;
MVT FltVT = MVT::f32;
+ ISD::NodeType ToFp = sign ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
+ ISD::NodeType ToInt = sign ? ISD::FP_TO_SINT : ISD::FP_TO_UINT;
+
if (VT.isVector()) {
unsigned NElts = VT.getVectorNumElements();
IntVT = MVT::getVectorVT(MVT::i32, NElts);
@@ -1397,29 +1400,35 @@ SDValue AMDGPUTargetLowering::LowerSDIVREM24(SDValue Op, SelectionDAG &DAG) cons
unsigned BitSize = VT.getScalarType().getSizeInBits();
- // char|short jq = ia ^ ib;
- SDValue jq = DAG.getNode(ISD::XOR, DL, VT, LHS, RHS);
+ SDValue jq = DAG.getConstant(1, IntVT);
+
+ if (sign) {
+ // char|short jq = ia ^ ib;
+ jq = DAG.getNode(ISD::XOR, DL, VT, LHS, RHS);
- // jq = jq >> (bitsize - 2)
- jq = DAG.getNode(ISD::SRA, DL, VT, jq, DAG.getConstant(BitSize - 2, VT));
+ // jq = jq >> (bitsize - 2)
+ jq = DAG.getNode(ISD::SRA, DL, VT, jq, DAG.getConstant(BitSize - 2, VT));
- // jq = jq | 0x1
- jq = DAG.getNode(ISD::OR, DL, VT, jq, DAG.getConstant(1, VT));
+ // jq = jq | 0x1
+ jq = DAG.getNode(ISD::OR, DL, VT, jq, DAG.getConstant(1, VT));
- // jq = (int)jq
- jq = DAG.getSExtOrTrunc(jq, DL, IntVT);
+ // jq = (int)jq
+ jq = DAG.getSExtOrTrunc(jq, DL, IntVT);
+ }
// int ia = (int)LHS;
- SDValue ia = DAG.getSExtOrTrunc(LHS, DL, IntVT);
+ SDValue ia = sign ?
+ DAG.getSExtOrTrunc(LHS, DL, IntVT) : DAG.getZExtOrTrunc(LHS, DL, IntVT);
// int ib, (int)RHS;
- SDValue ib = DAG.getSExtOrTrunc(RHS, DL, IntVT);
+ SDValue ib = sign ?
+ DAG.getSExtOrTrunc(RHS, DL, IntVT) : DAG.getZExtOrTrunc(RHS, DL, IntVT);
// float fa = (float)ia;
- SDValue fa = DAG.getNode(ISD::SINT_TO_FP, DL, FltVT, ia);
+ SDValue fa = DAG.getNode(ToFp, DL, FltVT, ia);
// float fb = (float)ib;
- SDValue fb = DAG.getNode(ISD::SINT_TO_FP, DL, FltVT, ib);
+ SDValue fb = DAG.getNode(ToFp, DL, FltVT, ib);
// float fq = native_divide(fa, fb);
SDValue fq = DAG.getNode(ISD::FMUL, DL, FltVT,
@@ -1436,7 +1445,7 @@ SDValue AMDGPUTargetLowering::LowerSDIVREM24(SDValue Op, SelectionDAG &DAG) cons
DAG.getNode(ISD::FMUL, DL, FltVT, fqneg, fb), fa);
// int iq = (int)fq;
- SDValue iq = DAG.getNode(ISD::FP_TO_SINT, DL, IntVT, fq);
+ SDValue iq = DAG.getNode(ToInt, DL, IntVT, fq);
// fr = fabs(fr);
fr = DAG.getNode(ISD::FABS, DL, FltVT, fr);
@@ -1452,11 +1461,13 @@ SDValue AMDGPUTargetLowering::LowerSDIVREM24(SDValue Op, SelectionDAG &DAG) cons
// jq = (cv ? jq : 0);
jq = DAG.getNode(ISD::SELECT, DL, VT, cv, jq, DAG.getConstant(0, VT));
- // dst = iq + jq;
- iq = DAG.getSExtOrTrunc(iq, DL, VT);
+ // dst = trunc/extend to legal type
+ iq = sign ? DAG.getSExtOrTrunc(iq, DL, VT) : DAG.getZExtOrTrunc(iq, DL, VT);
+ // dst = iq + jq;
SDValue Div = DAG.getNode(ISD::ADD, DL, VT, iq, jq);
+ // Rem needs compensation, it's easier to recompute it
SDValue Rem = DAG.getNode(ISD::MUL, DL, VT, Div, RHS);
Rem = DAG.getNode(ISD::SUB, DL, VT, LHS, Rem);
@@ -1475,6 +1486,16 @@ SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
SDValue Num = Op.getOperand(0);
SDValue Den = Op.getOperand(1);
+ if (VT == MVT::i32) {
+ if (DAG.MaskedValueIsZero(Op.getOperand(0), APInt(32, 0xff << 24)) &&
+ DAG.MaskedValueIsZero(Op.getOperand(1), APInt(32, 0xff << 24))) {
+ // TODO: We technically could do this for i64, but shouldn't that just be
+ // handled by something generally reducing 64-bit division on 32-bit
+ // values to 32-bit?
+ return LowerDIVREM24(Op, DAG, false);
+ }
+ }
+
// RCP = URECIP(Den) = 2^32 / Den + e
// e is rounding error.
SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
@@ -1585,7 +1606,7 @@ SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
// TODO: We technically could do this for i64, but shouldn't that just be
// handled by something generally reducing 64-bit division on 32-bit
// values to 32-bit?
- return LowerSDIVREM24(Op, DAG);
+ return LowerDIVREM24(Op, DAG, true);
}
}
diff --git a/lib/Target/R600/AMDGPUISelLowering.h b/lib/Target/R600/AMDGPUISelLowering.h
index 574e9b2..fe576a3 100644
--- a/lib/Target/R600/AMDGPUISelLowering.h
+++ b/lib/Target/R600/AMDGPUISelLowering.h
@@ -82,7 +82,7 @@ protected:
SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerSDIVREM(SDValue Op, SelectionDAG &DAG) const;
- SDValue LowerSDIVREM24(SDValue Op, SelectionDAG &DAG) const;
+ SDValue LowerDIVREM24(SDValue Op, SelectionDAG &DAG, bool sign) const;
bool isHWTrueValue(SDValue Op) const;
bool isHWFalseValue(SDValue Op) const;
diff --git a/test/CodeGen/R600/udivrem24.ll b/test/CodeGen/R600/udivrem24.ll
new file mode 100644
index 0000000..219c662
--- /dev/null
+++ b/test/CodeGen/R600/udivrem24.ll
@@ -0,0 +1,244 @@
+; RUN: llc -march=r600 -mcpu=SI < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
+; RUN: llc -march=r600 -mcpu=redwood < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s
+
+; FUNC-LABEL: @udiv24_i8
+; SI: V_CVT_F32_UBYTE
+; SI: V_CVT_F32_UBYTE
+; SI: V_RCP_F32
+; SI: V_CVT_U32_F32
+
+; EG: UINT_TO_FLT
+; EG-DAG: UINT_TO_FLT
+; EG-DAG: RECIP_IEEE
+; EG: FLT_TO_UINT
+define void @udiv24_i8(i8 addrspace(1)* %out, i8 addrspace(1)* %in) {
+ %den_ptr = getelementptr i8 addrspace(1)* %in, i8 1
+ %num = load i8 addrspace(1) * %in
+ %den = load i8 addrspace(1) * %den_ptr
+ %result = udiv i8 %num, %den
+ store i8 %result, i8 addrspace(1)* %out
+ ret void
+}
+
+; FUNC-LABEL: @udiv24_i16
+; SI: V_CVT_F32_U32
+; SI: V_CVT_F32_U32
+; SI: V_RCP_F32
+; SI: V_CVT_U32_F32
+
+; EG: UINT_TO_FLT
+; EG-DAG: UINT_TO_FLT
+; EG-DAG: RECIP_IEEE
+; EG: FLT_TO_UINT
+define void @udiv24_i16(i16 addrspace(1)* %out, i16 addrspace(1)* %in) {
+ %den_ptr = getelementptr i16 addrspace(1)* %in, i16 1
+ %num = load i16 addrspace(1) * %in, align 2
+ %den = load i16 addrspace(1) * %den_ptr, align 2
+ %result = udiv i16 %num, %den
+ store i16 %result, i16 addrspace(1)* %out, align 2
+ ret void
+}
+
+; FUNC-LABEL: @udiv24_i32
+; SI: V_CVT_F32_U32
+; SI-DAG: V_CVT_F32_U32
+; SI-DAG: V_RCP_F32
+; SI: V_CVT_U32_F32
+
+; EG: UINT_TO_FLT
+; EG-DAG: UINT_TO_FLT
+; EG-DAG: RECIP_IEEE
+; EG: FLT_TO_UINT
+define void @udiv24_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
+ %den_ptr = getelementptr i32 addrspace(1)* %in, i32 1
+ %num = load i32 addrspace(1) * %in, align 4
+ %den = load i32 addrspace(1) * %den_ptr, align 4
+ %num.i24.0 = shl i32 %num, 8
+ %den.i24.0 = shl i32 %den, 8
+ %num.i24 = lshr i32 %num.i24.0, 8
+ %den.i24 = lshr i32 %den.i24.0, 8
+ %result = udiv i32 %num.i24, %den.i24
+ store i32 %result, i32 addrspace(1)* %out, align 4
+ ret void
+}
+
+; FUNC-LABEL: @udiv25_i32
+; RCP_IFLAG is for URECIP in the full 32b alg
+; SI: V_RCP_IFLAG
+; SI-NOT: V_RCP_F32
+
+; EG-NOT: UINT_TO_FLT
+; EG-NOT: RECIP_IEEE
+define void @udiv25_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
+ %den_ptr = getelementptr i32 addrspace(1)* %in, i32 1
+ %num = load i32 addrspace(1) * %in, align 4
+ %den = load i32 addrspace(1) * %den_ptr, align 4
+ %num.i24.0 = shl i32 %num, 7
+ %den.i24.0 = shl i32 %den, 7
+ %num.i24 = lshr i32 %num.i24.0, 7
+ %den.i24 = lshr i32 %den.i24.0, 7
+ %result = udiv i32 %num.i24, %den.i24
+ store i32 %result, i32 addrspace(1)* %out, align 4
+ ret void
+}
+
+; FUNC-LABEL: @test_no_udiv24_i32_1
+; RCP_IFLAG is for URECIP in the full 32b alg
+; SI: V_RCP_IFLAG
+; SI-NOT: V_RCP_F32
+
+; EG-NOT: UINT_TO_FLT
+; EG-NOT: RECIP_IEEE
+define void @test_no_udiv24_i32_1(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
+ %den_ptr = getelementptr i32 addrspace(1)* %in, i32 1
+ %num = load i32 addrspace(1) * %in, align 4
+ %den = load i32 addrspace(1) * %den_ptr, align 4
+ %num.i24.0 = shl i32 %num, 8
+ %den.i24.0 = shl i32 %den, 7
+ %num.i24 = lshr i32 %num.i24.0, 8
+ %den.i24 = lshr i32 %den.i24.0, 7
+ %result = udiv i32 %num.i24, %den.i24
+ store i32 %result, i32 addrspace(1)* %out, align 4
+ ret void
+}
+
+; FUNC-LABEL: @test_no_udiv24_i32_2
+; RCP_IFLAG is for URECIP in the full 32b alg
+; SI: V_RCP_IFLAG
+; SI-NOT: V_RCP_F32
+
+; EG-NOT: UINT_TO_FLT
+; EG-NOT: RECIP_IEEE
+define void @test_no_udiv24_i32_2(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
+ %den_ptr = getelementptr i32 addrspace(1)* %in, i32 1
+ %num = load i32 addrspace(1) * %in, align 4
+ %den = load i32 addrspace(1) * %den_ptr, align 4
+ %num.i24.0 = shl i32 %num, 7
+ %den.i24.0 = shl i32 %den, 8
+ %num.i24 = lshr i32 %num.i24.0, 7
+ %den.i24 = lshr i32 %den.i24.0, 8
+ %result = udiv i32 %num.i24, %den.i24
+ store i32 %result, i32 addrspace(1)* %out, align 4
+ ret void
+}
+
+; FUNC-LABEL: @urem24_i8
+; SI: V_CVT_F32_UBYTE
+; SI: V_CVT_F32_UBYTE
+; SI: V_RCP_F32
+; SI: V_CVT_U32_F32
+
+; EG: UINT_TO_FLT
+; EG-DAG: UINT_TO_FLT
+; EG-DAG: RECIP_IEEE
+; EG: FLT_TO_UINT
+define void @urem24_i8(i8 addrspace(1)* %out, i8 addrspace(1)* %in) {
+ %den_ptr = getelementptr i8 addrspace(1)* %in, i8 1
+ %num = load i8 addrspace(1) * %in
+ %den = load i8 addrspace(1) * %den_ptr
+ %result = urem i8 %num, %den
+ store i8 %result, i8 addrspace(1)* %out
+ ret void
+}
+
+; FUNC-LABEL: @urem24_i16
+; SI: V_CVT_F32_U32
+; SI: V_CVT_F32_U32
+; SI: V_RCP_F32
+; SI: V_CVT_U32_F32
+
+; EG: UINT_TO_FLT
+; EG-DAG: UINT_TO_FLT
+; EG-DAG: RECIP_IEEE
+; EG: FLT_TO_UINT
+define void @urem24_i16(i16 addrspace(1)* %out, i16 addrspace(1)* %in) {
+ %den_ptr = getelementptr i16 addrspace(1)* %in, i16 1
+ %num = load i16 addrspace(1) * %in, align 2
+ %den = load i16 addrspace(1) * %den_ptr, align 2
+ %result = urem i16 %num, %den
+ store i16 %result, i16 addrspace(1)* %out, align 2
+ ret void
+}
+
+; FUNC-LABEL: @urem24_i32
+; SI: V_CVT_F32_U32
+; SI: V_CVT_F32_U32
+; SI: V_RCP_F32
+; SI: V_CVT_U32_F32
+
+; EG: UINT_TO_FLT
+; EG-DAG: UINT_TO_FLT
+; EG-DAG: RECIP_IEEE
+; EG: FLT_TO_UINT
+define void @urem24_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
+ %den_ptr = getelementptr i32 addrspace(1)* %in, i32 1
+ %num = load i32 addrspace(1) * %in, align 4
+ %den = load i32 addrspace(1) * %den_ptr, align 4
+ %num.i24.0 = shl i32 %num, 8
+ %den.i24.0 = shl i32 %den, 8
+ %num.i24 = lshr i32 %num.i24.0, 8
+ %den.i24 = lshr i32 %den.i24.0, 8
+ %result = urem i32 %num.i24, %den.i24
+ store i32 %result, i32 addrspace(1)* %out, align 4
+ ret void
+}
+
+; FUNC-LABEL: @urem25_i32
+; RCP_IFLAG is for URECIP in the full 32b alg
+; SI: V_RCP_IFLAG
+; SI-NOT: V_RCP_F32
+
+; EG-NOT: UINT_TO_FLT
+; EG-NOT: RECIP_IEEE
+define void @urem25_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
+ %den_ptr = getelementptr i32 addrspace(1)* %in, i32 1
+ %num = load i32 addrspace(1) * %in, align 4
+ %den = load i32 addrspace(1) * %den_ptr, align 4
+ %num.i24.0 = shl i32 %num, 7
+ %den.i24.0 = shl i32 %den, 7
+ %num.i24 = lshr i32 %num.i24.0, 7
+ %den.i24 = lshr i32 %den.i24.0, 7
+ %result = urem i32 %num.i24, %den.i24
+ store i32 %result, i32 addrspace(1)* %out, align 4
+ ret void
+}
+
+; FUNC-LABEL: @test_no_urem24_i32_1
+; RCP_IFLAG is for URECIP in the full 32b alg
+; SI: V_RCP_IFLAG
+; SI-NOT: V_RCP_F32
+
+; EG-NOT: UINT_TO_FLT
+; EG-NOT: RECIP_IEEE
+define void @test_no_urem24_i32_1(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
+ %den_ptr = getelementptr i32 addrspace(1)* %in, i32 1
+ %num = load i32 addrspace(1) * %in, align 4
+ %den = load i32 addrspace(1) * %den_ptr, align 4
+ %num.i24.0 = shl i32 %num, 8
+ %den.i24.0 = shl i32 %den, 7
+ %num.i24 = lshr i32 %num.i24.0, 8
+ %den.i24 = lshr i32 %den.i24.0, 7
+ %result = urem i32 %num.i24, %den.i24
+ store i32 %result, i32 addrspace(1)* %out, align 4
+ ret void
+}
+
+; FUNC-LABEL: @test_no_urem24_i32_2
+; RCP_IFLAG is for URECIP in the full 32b alg
+; SI: V_RCP_IFLAG
+; SI-NOT: V_RCP_F32
+
+; EG-NOT: UINT_TO_FLT
+; EG-NOT: RECIP_IEEE
+define void @test_no_urem24_i32_2(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
+ %den_ptr = getelementptr i32 addrspace(1)* %in, i32 1
+ %num = load i32 addrspace(1) * %in, align 4
+ %den = load i32 addrspace(1) * %den_ptr, align 4
+ %num.i24.0 = shl i32 %num, 7
+ %den.i24.0 = shl i32 %den, 8
+ %num.i24 = lshr i32 %num.i24.0, 7
+ %den.i24 = lshr i32 %den.i24.0, 8
+ %result = urem i32 %num.i24, %den.i24
+ store i32 %result, i32 addrspace(1)* %out, align 4
+ ret void
+}
--
1.9.3
More information about the llvm-commits
mailing list