R600: Add support for 24-bit MUL instructions

Reviewed-by: Vincent Lejeune <vljn at ovi.com>
llvm-svn: 186922

lib/Target/R600/AMDGPUISelDAGToDAG.cpp

@@ -58,6 +58,9 @@ private:
   bool SelectADDRParam(SDValue Addr, SDValue& R1, SDValue& R2);
   bool SelectADDR(SDValue N, SDValue &R1, SDValue &R2);
   bool SelectADDR64(SDValue N, SDValue &R1, SDValue &R2);
+  SDValue SimplifyI24(SDValue &Op);
+  bool SelectI24(SDValue Addr, SDValue &Op);
+  bool SelectU24(SDValue Addr, SDValue &Op);
 
   static bool checkType(const Value *ptr, unsigned int addrspace);
 
@@ -674,7 +677,9 @@ const char *AMDGPUDAGToDAGISel::getPassName() const {
 #endif
 #undef DEBUGTMP
 
-///==== AMDGPU Functions ====///
+//===----------------------------------------------------------------------===//
+// Complex Patterns
+//===----------------------------------------------------------------------===//
 
 bool AMDGPUDAGToDAGISel::SelectGlobalValueConstantOffset(SDValue Addr,
     SDValue& IntPtr) {
@@ -741,6 +746,49 @@ bool AMDGPUDAGToDAGISel::SelectADDRIndirect(SDValue Addr, SDValue &Base,
   return true;
 }
 
+SDValue AMDGPUDAGToDAGISel::SimplifyI24(SDValue &Op) {
+  APInt Demanded = APInt(32, 0x00FFFFFF);
+  APInt KnownZero, KnownOne;
+  TargetLowering::TargetLoweringOpt TLO(*CurDAG, true, true);
+  const TargetLowering *TLI = getTargetLowering();
+  if (TLI->SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO)) {
+    CurDAG->ReplaceAllUsesWith(Op, TLO.New);
+    CurDAG->RepositionNode(Op.getNode(), TLO.New.getNode());
+    return SimplifyI24(TLO.New);
+  } else {
+    return  Op;
+  }
+}
+
+bool AMDGPUDAGToDAGISel::SelectI24(SDValue Op, SDValue &I24) {
+
+  assert(Op.getValueType() == MVT::i32);
+
+  if (CurDAG->ComputeNumSignBits(Op) == 9) {
+    I24 = SimplifyI24(Op);
+    return true;
+  }
+  return false;
+}
+
+bool AMDGPUDAGToDAGISel::SelectU24(SDValue Op, SDValue &U24) {
+  APInt KnownZero;
+  APInt KnownOne;
+  CurDAG->ComputeMaskedBits(Op, KnownZero, KnownOne);
+
+  assert (Op.getValueType() == MVT::i32);
+
+  // ANY_EXTEND and EXTLOAD operations can only be done on types smaller than
+  // i32.  These smaller types are legal to use with the i24 instructions.
+  if ((KnownZero & APInt(KnownZero.getBitWidth(), 0xFF000000)) == 0xFF000000 ||
+       Op.getOpcode() == ISD::ANY_EXTEND ||
+       ISD::isEXTLoad(Op.getNode())) {
+    U24 = SimplifyI24(Op);
+    return true;
+  }
+  return false;
+}
+
 void AMDGPUDAGToDAGISel::PostprocessISelDAG() {
 
   if (Subtarget.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS) {

lib/Target/R600/AMDGPUInstructions.td

@@ -173,6 +173,9 @@ def FP_ONE : PatLeaf <
   [{return N->isExactlyValue(1.0);}]
 >;
 
+def U24 : ComplexPattern<i32, 1, "SelectU24", [], []>;
+def I24 : ComplexPattern<i32, 1, "SelectI24", [], []>;
+
 let isCodeGenOnly = 1, isPseudo = 1 in {
 
 let usesCustomInserter = 1  in {
@@ -366,6 +369,16 @@ class ROTRPattern <Instruction BIT_ALIGN> : Pat <
   (BIT_ALIGN $src0, $src0, $src1)
 >;
 
+// 24-bit arithmetic patterns
+def umul24 : PatFrag <(ops node:$x, node:$y), (mul node:$x, node:$y)>;
+
+/*
+class UMUL24Pattern <Instruction UMUL24> : Pat <
+  (mul U24:$x, U24:$y),
+  (UMUL24 $x, $y)
+>;
+*/
+
 include "R600Instructions.td"
 
 include "SIInstrInfo.td"

lib/Target/R600/R600Instructions.td

@@ -1473,6 +1473,9 @@ let Predicates = [isEGorCayman] in {
   def CNDGE_eg : CNDGE_Common<0x1B>;
   def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
   def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
+  def MUL_UINT24_eg : R600_2OP <0xB5, "MUL_UINT24",
+    [(set i32:$dst, (mul U24:$src0, U24:$src1))], VecALU
+  >;
   def DOT4_eg : DOT4_Common<0xBE>;
   defm CUBE_eg : CUBE_Common<0xC0>;
 
@@ -1703,6 +1706,10 @@ defm R600_ : RegisterLoadStore <R600_Reg32, FRAMEri, ADDRIndirect>;
 
 let Predicates = [isCayman] in {
 
+def MUL_INT24_cm : R600_2OP <0x5B, "MUL_INT24",
+  [(set i32:$dst, (mul I24:$src0, I24:$src1))], VecALU
+>;
+
 let isVector = 1 in {
 
 def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;

lib/Target/R600/SIInstructions.td

@@ -866,14 +866,16 @@ defm V_MUL_F32 : VOP2_32 <0x00000008, "V_MUL_F32",
   [(set f32:$dst, (fmul f32:$src0, f32:$src1))]
 >;
 
-} // End isCommutable = 1
 
-//defm V_MUL_I32_I24 : VOP2_32 <0x00000009, "V_MUL_I32_I24", []>;
+defm V_MUL_I32_I24 : VOP2_32 <0x00000009, "V_MUL_I32_I24",
+  [(set i32:$dst, (mul I24:$src0, I24:$src1))]
+>;
 //defm V_MUL_HI_I32_I24 : VOP2_32 <0x0000000a, "V_MUL_HI_I32_I24", []>;
-//defm V_MUL_U32_U24 : VOP2_32 <0x0000000b, "V_MUL_U32_U24", []>;
+defm V_MUL_U32_U24 : VOP2_32 <0x0000000b, "V_MUL_U32_U24",
+  [(set i32:$dst, (mul U24:$src0, U24:$src1))]
+>;
 //defm V_MUL_HI_U32_U24 : VOP2_32 <0x0000000c, "V_MUL_HI_U32_U24", []>;
 
-let isCommutable = 1 in {
 
 defm V_MIN_LEGACY_F32 : VOP2_32 <0x0000000d, "V_MIN_LEGACY_F32",
   [(set f32:$dst, (AMDGPUfmin f32:$src0, f32:$src1))]

test/CodeGen/R600/mul_int24.ll

@@ -0,0 +1,19 @@
+; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s --check-prefix=EG-CHECK
+; RUN: llc < %s -march=r600 -mcpu=cayman | FileCheck %s --check-prefix=CM-CHECK
+; RUN: llc < %s -march=r600 -mcpu=SI | FileCheck %s --check-prefix=SI-CHECK
+
+; EG-CHECK: @i32_mul24
+; Signed 24-bit multiply is not supported on pre-Cayman GPUs.
+; EG-CHECK: MULLO_INT
+; CM-CHECK: MUL_INT24 {{[ *]*}}T{{[0-9].[XYZW]}}, KC0[2].Z, KC0[2].W
+; SI-CHECK: V_MUL_I32_I24
+define void @i32_mul24(i32 addrspace(1)* %out, i32 %a, i32 %b) {
+entry:
+  %0 = shl i32 %a, 8
+  %a_24 = ashr i32 %0, 8
+  %1 = shl i32 %b, 8
+  %b_24 = ashr i32 %1, 8
+  %2 = mul i32 %a_24, %b_24
+  store i32 %2, i32 addrspace(1)* %out
+  ret void
+}

test/CodeGen/R600/mul_uint24.ll

@@ -0,0 +1,65 @@
+; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s --check-prefix=EG-CHECK
+; RUN: llc < %s -march=r600 -mcpu=cayman | FileCheck %s --check-prefix=EG-CHECK
+; RUN: llc < %s -march=r600 -mcpu=SI | FileCheck %s --check-prefix=SI-CHECK
+
+; EG-CHECK: @u32_mul24
+; EG-CHECK: MUL_UINT24 {{[* ]*}}T{{[0-9]\.[XYZW]}}, KC0[2].Z, KC0[2].W
+; SI-CHECK: @u32_mul24
+; SI-CHECK: V_MUL_U32_U24
+
+define void @u32_mul24(i32 addrspace(1)* %out, i32 %a, i32 %b) {
+entry:
+  %0 = shl i32 %a, 8
+  %a_24 = lshr i32 %0, 8
+  %1 = shl i32 %b, 8
+  %b_24 = lshr i32 %1, 8
+  %2 = mul i32 %a_24, %b_24
+  store i32 %2, i32 addrspace(1)* %out
+  ret void
+}
+
+; EG-CHECK: @i16_mul24
+; EG-CHECK-DAG: VTX_READ_16 [[A:T[0-9]\.X]], T{{[0-9]}}.X, 40
+; EG-CHECK-DAG: VTX_READ_16 [[B:T[0-9]\.X]], T{{[0-9]}}.X, 44
+; The order of A and B does not matter.
+; EG-CHECK: MUL_UINT24 {{[* ]*}}T{{[0-9]}}.[[MUL_CHAN:[XYZW]]], [[A]], [[B]]
+; The result must be sign-extended
+; EG-CHECK: LSHL {{[* ]*}}T{{[0-9]}}.[[LSHL_CHAN:[XYZW]]], PV.[[MUL_CHAN]], literal.x
+; EG-CHECK: 16
+; EG-CHECK: ASHR {{[* ]*}}T{{[0-9]\.[XYZW]}}, PV.[[LSHL_CHAN]], literal.x
+; EG-CHECK: 16
+; SI-CHECK: @i16_mul24
+; SI-CHECK: V_MUL_U32_U24_e{{(32|64)}} [[MUL:VGPR[0-9]]], {{[SV]GPR[0-9], [SV]GPR[0-9]}}
+; SI-CHECK: V_LSHLREV_B32_e32 [[LSHL:VGPR[0-9]]], 16, [[MUL]]
+; SI-CHECK: V_ASHRREV_I32_e32 VGPR{{[0-9]}}, 16, [[LSHL]]
+
+define void @i16_mul24(i32 addrspace(1)* %out, i16 %a, i16 %b) {
+entry:
+  %0 = mul i16 %a, %b
+  %1 = sext i16 %0 to i32
+  store i32 %1, i32 addrspace(1)* %out
+  ret void
+}
+
+; EG-CHECK: @i8_mul24
+; EG-CHECK-DAG: VTX_READ_8 [[A:T[0-9]\.X]], T{{[0-9]}}.X, 40
+; EG-CHECK-DAG: VTX_READ_8 [[B:T[0-9]\.X]], T{{[0-9]}}.X, 44
+; The order of A and B does not matter.
+; EG-CHECK: MUL_UINT24 {{[* ]*}}T{{[0-9]}}.[[MUL_CHAN:[XYZW]]], [[A]], [[B]]
+; The result must be sign-extended
+; EG-CHECK: LSHL {{[* ]*}}T{{[0-9]}}.[[LSHL_CHAN:[XYZW]]], PV.[[MUL_CHAN]], literal.x
+; EG-CHECK: 24
+; EG-CHECK: ASHR {{[* ]*}}T{{[0-9]\.[XYZW]}}, PV.[[LSHL_CHAN]], literal.x
+; EG-CHECK: 24
+; SI-CHECK: @i8_mul24
+; SI-CHECK: V_MUL_U32_U24_e{{(32|64)}} [[MUL:VGPR[0-9]]], {{[SV]GPR[0-9], [SV]GPR[0-9]}}
+; SI-CHECK: V_LSHLREV_B32_e32 [[LSHL:VGPR[0-9]]], 24, [[MUL]]
+; SI-CHECK: V_ASHRREV_I32_e32 VGPR{{[0-9]}}, 24, [[LSHL]]
+
+define void @i8_mul24(i32 addrspace(1)* %out, i8 %a, i8 %b) {
+entry:
+  %0 = mul i8 %a, %b
+  %1 = sext i8 %0 to i32
+  store i32 %1, i32 addrspace(1)* %out
+  ret void
+}