[llvm] r226881 - R600: Try to use lower types for 64bit division if possible
Jan Vesely
jan.vesely at rutgers.edu
Thu Jan 22 15:42:44 PST 2015
Author: jvesely
Date: Thu Jan 22 17:42:43 2015
New Revision: 226881
URL: http://llvm.org/viewvc/llvm-project?rev=226881&view=rev
Log:
R600: Try to use lower types for 64bit division if possible
v2: add and enable tests for SI
Signed-off-by: Jan Vesely <jan.vesely at rutgers.edu>
Reviewed-by: Matt Arsenault <Matthew.Arsenault at amd.com>
Added:
llvm/trunk/test/CodeGen/R600/sdivrem64.ll
Modified:
llvm/trunk/lib/Target/R600/AMDGPUISelLowering.cpp
llvm/trunk/lib/Target/R600/AMDGPUISelLowering.h
llvm/trunk/test/CodeGen/R600/udivrem64.ll
Modified: llvm/trunk/lib/Target/R600/AMDGPUISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/R600/AMDGPUISelLowering.cpp?rev=226881&r1=226880&r2=226881&view=diff
==============================================================================
--- llvm/trunk/lib/Target/R600/AMDGPUISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/R600/AMDGPUISelLowering.cpp Thu Jan 22 17:42:43 2015
@@ -1661,6 +1661,20 @@ void AMDGPUTargetLowering::LowerUDIVREM6
SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
+ if (VT == MVT::i64 &&
+ DAG.MaskedValueIsZero(RHS, APInt::getHighBitsSet(64, 32)) &&
+ DAG.MaskedValueIsZero(LHS, APInt::getHighBitsSet(64, 32))) {
+
+ SDValue Res = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
+ LHS_Lo, RHS_Lo);
+
+ SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, Res.getValue(0), zero);
+ SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, Res.getValue(1), zero);
+ Results.push_back(DIV);
+ Results.push_back(REM);
+ return;
+ }
+
// Get Speculative values
SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
@@ -1722,8 +1736,8 @@ SDValue AMDGPUTargetLowering::LowerUDIVR
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))) {
+ if (DAG.MaskedValueIsZero(Num, APInt::getHighBitsSet(32, 8)) &&
+ DAG.MaskedValueIsZero(Den, APInt::getHighBitsSet(32, 8))) {
// 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?
@@ -1835,19 +1849,31 @@ SDValue AMDGPUTargetLowering::LowerSDIVR
SDValue LHS = Op.getOperand(0);
SDValue RHS = Op.getOperand(1);
- if (VT == MVT::i32) {
- if (DAG.ComputeNumSignBits(Op.getOperand(0)) > 8 &&
- DAG.ComputeNumSignBits(Op.getOperand(1)) > 8) {
- // 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, true);
- }
- }
-
SDValue Zero = DAG.getConstant(0, VT);
SDValue NegOne = DAG.getConstant(-1, VT);
+ if (VT == MVT::i32 &&
+ DAG.ComputeNumSignBits(LHS) > 8 &&
+ DAG.ComputeNumSignBits(RHS) > 8) {
+ return LowerDIVREM24(Op, DAG, true);
+ }
+ if (VT == MVT::i64 &&
+ DAG.ComputeNumSignBits(LHS) > 32 &&
+ DAG.ComputeNumSignBits(RHS) > 32) {
+ EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
+
+ //HiLo split
+ SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, Zero);
+ SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, Zero);
+ SDValue DIVREM = DAG.getNode(ISD::SDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
+ LHS_Lo, RHS_Lo);
+ SDValue Res[2] = {
+ DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(0)),
+ DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(1))
+ };
+ return DAG.getMergeValues(Res, DL);
+ }
+
SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
Modified: llvm/trunk/lib/Target/R600/AMDGPUISelLowering.h
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/R600/AMDGPUISelLowering.h?rev=226881&r1=226880&r2=226881&view=diff
==============================================================================
--- llvm/trunk/lib/Target/R600/AMDGPUISelLowering.h (original)
+++ llvm/trunk/lib/Target/R600/AMDGPUISelLowering.h Thu Jan 22 17:42:43 2015
@@ -43,7 +43,6 @@ private:
/// \brief Split a vector store into multiple scalar stores.
/// \returns The resulting chain.
- SDValue LowerUDIVREM(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerFREM(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerFCEIL(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const;
@@ -90,6 +89,7 @@ protected:
SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerSDIVREM(SDValue Op, SelectionDAG &DAG) const;
+ SDValue LowerUDIVREM(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerDIVREM24(SDValue Op, SelectionDAG &DAG, bool sign) const;
void LowerUDIVREM64(SDValue Op, SelectionDAG &DAG,
SmallVectorImpl<SDValue> &Results) const;
Added: llvm/trunk/test/CodeGen/R600/sdivrem64.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/R600/sdivrem64.ll?rev=226881&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/R600/sdivrem64.ll (added)
+++ llvm/trunk/test/CodeGen/R600/sdivrem64.ll Thu Jan 22 17:42:43 2015
@@ -0,0 +1,218 @@
+;RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %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: {{^}}test_sdiv:
+;EG: RECIP_UINT
+;EG: LSHL {{.*}}, 1,
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI: v_bfe_u32
+;SI-NOT: v_mad_f32
+;SI-NOT: v_lshr_64
+;SI: s_endpgm
+define void @test_sdiv(i64 addrspace(1)* %out, i64 %x, i64 %y) {
+ %result = sdiv i64 %x, %y
+ store i64 %result, i64 addrspace(1)* %out
+ ret void
+}
+
+;FUNC-LABEL: {{^}}test_srem:
+;EG: RECIP_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: BFE_UINT
+;EG: AND_INT {{.*}}, 1,
+
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI-NOT: v_mad_f32
+;SI-NOT: v_lshr_64
+;SI: s_endpgm
+define void @test_srem(i64 addrspace(1)* %out, i64 %x, i64 %y) {
+ %result = urem i64 %x, %y
+ store i64 %result, i64 addrspace(1)* %out
+ ret void
+}
+
+;FUNC-LABEL: {{^}}test_sdiv3264:
+;EG: RECIP_UINT
+;EG-NOT: BFE_UINT
+
+;SI-NOT: s_bfe_u32
+;SI-NOT: v_mad_f32
+;SI-NOT: v_lshr_64
+;SI: s_endpgm
+define void @test_sdiv3264(i64 addrspace(1)* %out, i64 %x, i64 %y) {
+ %1 = ashr i64 %x, 33
+ %2 = ashr i64 %y, 33
+ %result = sdiv i64 %1, %2
+ store i64 %result, i64 addrspace(1)* %out
+ ret void
+}
+
+;FUNC-LABEL: {{^}}test_srem3264:
+;EG: RECIP_UINT
+;EG-NOT: BFE_UINT
+
+;SI-NOT: s_bfe_u32
+;SI-NOT: v_mad_f32
+;SI-NOT: v_lshr_64
+;SI: s_endpgm
+define void @test_srem3264(i64 addrspace(1)* %out, i64 %x, i64 %y) {
+ %1 = ashr i64 %x, 33
+ %2 = ashr i64 %y, 33
+ %result = srem i64 %1, %2
+ store i64 %result, i64 addrspace(1)* %out
+ ret void
+}
+
+;FUNC-LABEL: {{^}}test_sdiv2464:
+;EG: INT_TO_FLT
+;EG: INT_TO_FLT
+;EG: FLT_TO_INT
+;EG-NOT: RECIP_UINT
+;EG-NOT: BFE_UINT
+
+;SI-NOT: s_bfe_u32
+;SI: v_mad_f32
+;SI-NOT: v_lshr_64
+;SI: s_endpgm
+define void @test_sdiv2464(i64 addrspace(1)* %out, i64 %x, i64 %y) {
+ %1 = ashr i64 %x, 40
+ %2 = ashr i64 %y, 40
+ %result = sdiv i64 %1, %2
+ store i64 %result, i64 addrspace(1)* %out
+ ret void
+}
+
+;FUNC-LABEL: {{^}}test_srem2464:
+;EG: INT_TO_FLT
+;EG: INT_TO_FLT
+;EG: FLT_TO_INT
+;EG-NOT: RECIP_UINT
+;EG-NOT: BFE_UINT
+
+;SI-NOT: s_bfe_u32
+;SI: v_mad_f32
+;SI-NOT: v_lshr_64
+;SI: s_endpgm
+define void @test_srem2464(i64 addrspace(1)* %out, i64 %x, i64 %y) {
+ %1 = ashr i64 %x, 40
+ %2 = ashr i64 %y, 40
+ %result = srem i64 %1, %2
+ store i64 %result, i64 addrspace(1)* %out
+ ret void
+}
Modified: llvm/trunk/test/CodeGen/R600/udivrem64.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/R600/udivrem64.ll?rev=226881&r1=226880&r2=226881&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/R600/udivrem64.ll (original)
+++ llvm/trunk/test/CodeGen/R600/udivrem64.ll Thu Jan 22 17:42:43 2015
@@ -1,5 +1,5 @@
-;XUN: llc < %s -march=amdgcn -mcpu=SI -verify-machineinstrs| FileCheck --check-prefix=SI --check-prefix=FUNC %s
-;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck --check-prefix=EG --check-prefix=FUNC %s
+;RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %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: {{^}}test_udiv:
;EG: RECIP_UINT
@@ -34,6 +34,39 @@
;EG: BFE_UINT
;EG: BFE_UINT
;EG: BFE_UINT
+
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI-NOT: v_mad_f32
+;SI-NOT: v_lshr_64
;SI: s_endpgm
define void @test_udiv(i64 addrspace(1)* %out, i64 %x, i64 %y) {
%result = udiv i64 %x, %y
@@ -74,9 +107,110 @@ define void @test_udiv(i64 addrspace(1)*
;EG: BFE_UINT
;EG: BFE_UINT
;EG: AND_INT {{.*}}, 1,
+
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI: s_bfe_u32
+;SI-NOT: v_mad_f32
+;SI-NOT: v_lshr_64
;SI: s_endpgm
define void @test_urem(i64 addrspace(1)* %out, i64 %x, i64 %y) {
%result = urem i64 %x, %y
store i64 %result, i64 addrspace(1)* %out
ret void
}
+
+;FUNC-LABEL: {{^}}test_udiv3264:
+;EG: RECIP_UINT
+;EG-NOT: BFE_UINT
+
+;SI-NOT: s_bfe_u32
+;SI-NOT: v_mad_f32
+;SI-NOT: v_lshr_64
+;SI: s_endpgm
+define void @test_udiv3264(i64 addrspace(1)* %out, i64 %x, i64 %y) {
+ %1 = lshr i64 %x, 33
+ %2 = lshr i64 %y, 33
+ %result = udiv i64 %1, %2
+ store i64 %result, i64 addrspace(1)* %out
+ ret void
+}
+
+;FUNC-LABEL: {{^}}test_urem3264:
+;EG: RECIP_UINT
+;EG-NOT: BFE_UINT
+
+;SI-NOT: s_bfe_u32
+;SI-NOT: v_mad_f32
+;SI-NOT: v_lshr_64
+;SI: s_endpgm
+define void @test_urem3264(i64 addrspace(1)* %out, i64 %x, i64 %y) {
+ %1 = lshr i64 %x, 33
+ %2 = lshr i64 %y, 33
+ %result = urem i64 %1, %2
+ store i64 %result, i64 addrspace(1)* %out
+ ret void
+}
+
+;FUNC-LABEL: {{^}}test_udiv2464:
+;EG: UINT_TO_FLT
+;EG: UINT_TO_FLT
+;EG: FLT_TO_UINT
+;EG-NOT: RECIP_UINT
+;EG-NOT: BFE_UINT
+
+;SI-NOT: v_lshr_64
+;SI: v_mad_f32
+;SI: s_endpgm
+define void @test_udiv2464(i64 addrspace(1)* %out, i64 %x, i64 %y) {
+ %1 = lshr i64 %x, 40
+ %2 = lshr i64 %y, 40
+ %result = udiv i64 %1, %2
+ store i64 %result, i64 addrspace(1)* %out
+ ret void
+}
+
+;FUNC-LABEL: {{^}}test_urem2464:
+;EG: UINT_TO_FLT
+;EG: UINT_TO_FLT
+;EG: FLT_TO_UINT
+;EG-NOT: RECIP_UINT
+;EG-NOT: BFE_UINT
+
+;SI-NOT: v_lshr_64
+;SI: v_mad_f32
+;SI: s_endpgm
+define void @test_urem2464(i64 addrspace(1)* %out, i64 %x, i64 %y) {
+ %1 = lshr i64 %x, 40
+ %2 = lshr i64 %y, 40
+ %result = urem i64 %1, %2
+ store i64 %result, i64 addrspace(1)* %out
+ ret void
+}
More information about the llvm-commits
mailing list