From 6f635b5488172a5410e46661638de6f4994bd9dd Mon Sep 17 00:00:00 2001 From: Jyotsna Verma Date: Tue, 5 Feb 2013 16:42:24 +0000 Subject: [PATCH] Hexagon: Add V4 compare instructions. Enable relationship mapping for the existing instructions. llvm-svn: 174389 --- llvm/lib/Target/Hexagon/HexagonInstrInfoV4.td | 159 +++++++++++++++++++++++--- llvm/test/CodeGen/Hexagon/cmp-to-genreg.ll | 34 ++++++ llvm/test/CodeGen/Hexagon/cmp-to-predreg.ll | 43 +++++++ 3 files changed, 220 insertions(+), 16 deletions(-) create mode 100644 llvm/test/CodeGen/Hexagon/cmp-to-genreg.ll create mode 100644 llvm/test/CodeGen/Hexagon/cmp-to-predreg.ll diff --git a/llvm/lib/Target/Hexagon/HexagonInstrInfoV4.td b/llvm/lib/Target/Hexagon/HexagonInstrInfoV4.td index 08225e4..f7b6a9d 100644 --- a/llvm/lib/Target/Hexagon/HexagonInstrInfoV4.td +++ b/llvm/lib/Target/Hexagon/HexagonInstrInfoV4.td @@ -251,6 +251,54 @@ def TFR_FI_immext_V4 : ALU32_ri<(outs IntRegs:$dst), []>, Requires<[HasV4T]>; +// Rd=cmp.eq(Rs,#s8) +let validSubTargets = HasV4SubT, isExtendable = 1, opExtendable = 2, +isExtentSigned = 1, opExtentBits = 8 in +def V4_A4_rcmpeqi : ALU32_ri<(outs IntRegs:$Rd), + (ins IntRegs:$Rs, s8Ext:$s8), + "$Rd = cmp.eq($Rs, #$s8)", + [(set (i32 IntRegs:$Rd), + (i32 (zext (i1 (seteq (i32 IntRegs:$Rs), + s8ExtPred:$s8)))))]>, + Requires<[HasV4T]>; + +// Preserve the TSTBIT generation +def : Pat <(i32 (zext (i1 (setne (i32 (and (i32 (shl 1, (i32 IntRegs:$src2))), + (i32 IntRegs:$src1))), 0)))), + (i32 (MUX_ii (i1 (TSTBIT_rr (i32 IntRegs:$src1), (i32 IntRegs:$src2))), + 1, 0))>; + +// Interfered with tstbit generation, above pattern preserves, see : tstbit.ll +// Rd=cmp.ne(Rs,#s8) +let validSubTargets = HasV4SubT, isExtendable = 1, opExtendable = 2, +isExtentSigned = 1, opExtentBits = 8 in +def V4_A4_rcmpneqi : ALU32_ri<(outs IntRegs:$Rd), + (ins IntRegs:$Rs, s8Ext:$s8), + "$Rd = !cmp.eq($Rs, #$s8)", + [(set (i32 IntRegs:$Rd), + (i32 (zext (i1 (setne (i32 IntRegs:$Rs), + s8ExtPred:$s8)))))]>, + Requires<[HasV4T]>; + +// Rd=cmp.eq(Rs,Rt) +let validSubTargets = HasV4SubT in +def V4_A4_rcmpeq : ALU32_ri<(outs IntRegs:$Rd), + (ins IntRegs:$Rs, IntRegs:$Rt), + "$Rd = cmp.eq($Rs, $Rt)", + [(set (i32 IntRegs:$Rd), + (i32 (zext (i1 (seteq (i32 IntRegs:$Rs), + IntRegs:$Rt)))))]>, + Requires<[HasV4T]>; + +// Rd=cmp.ne(Rs,Rt) +let validSubTargets = HasV4SubT in +def V4_A4_rcmpneq : ALU32_ri<(outs IntRegs:$Rd), + (ins IntRegs:$Rs, IntRegs:$Rt), + "$Rd = !cmp.eq($Rs, $Rt)", + [(set (i32 IntRegs:$Rd), + (i32 (zext (i1 (setne (i32 IntRegs:$Rs), + IntRegs:$Rt)))))]>, + Requires<[HasV4T]>; //===----------------------------------------------------------------------===// // ALU32 - @@ -3656,7 +3704,61 @@ def MEMb_ORr_MEM_V4 : MEMInst_V4<(outs), // incorrect code for negative numbers. // Pd=cmpb.eq(Rs,#u8) -let isCompare = 1 in +// p=!cmp.eq(r1,r2) +let isCompare = 1, validSubTargets = HasV4SubT in +def CMPnotEQ_rr : ALU32_rr<(outs PredRegs:$dst), + (ins IntRegs:$src1, IntRegs:$src2), + "$dst = !cmp.eq($src1, $src2)", + [(set (i1 PredRegs:$dst), + (setne (i32 IntRegs:$src1), (i32 IntRegs:$src2)))]>, + Requires<[HasV4T]>; + +// p=!cmp.eq(r1,#s10) +let isCompare = 1, validSubTargets = HasV4SubT in +def CMPnotEQ_ri : ALU32_ri<(outs PredRegs:$dst), + (ins IntRegs:$src1, s10Ext:$src2), + "$dst = !cmp.eq($src1, #$src2)", + [(set (i1 PredRegs:$dst), + (setne (i32 IntRegs:$src1), s10ImmPred:$src2))]>, + Requires<[HasV4T]>; + +// p=!cmp.gt(r1,r2) +let isCompare = 1, validSubTargets = HasV4SubT in +def CMPnotGT_rr : ALU32_rr<(outs PredRegs:$dst), + (ins IntRegs:$src1, IntRegs:$src2), + "$dst = !cmp.gt($src1, $src2)", + [(set (i1 PredRegs:$dst), + (not (setgt (i32 IntRegs:$src1), (i32 IntRegs:$src2))))]>, + Requires<[HasV4T]>; + +// p=!cmp.gt(r1,#s10) +let isCompare = 1, validSubTargets = HasV4SubT in +def CMPnotGT_ri : ALU32_ri<(outs PredRegs:$dst), + (ins IntRegs:$src1, s10Ext:$src2), + "$dst = !cmp.gt($src1, #$src2)", + [(set (i1 PredRegs:$dst), + (not (setgt (i32 IntRegs:$src1), s10ImmPred:$src2)))]>, + Requires<[HasV4T]>; + +// p=!cmp.gtu(r1,r2) +let isCompare = 1, validSubTargets = HasV4SubT in +def CMPnotGTU_rr : ALU32_rr<(outs PredRegs:$dst), + (ins IntRegs:$src1, IntRegs:$src2), + "$dst = !cmp.gtu($src1, $src2)", + [(set (i1 PredRegs:$dst), + (not (setugt (i32 IntRegs:$src1), (i32 IntRegs:$src2))))]>, + Requires<[HasV4T]>; + +// p=!cmp.gtu(r1,#u9) +let isCompare = 1, validSubTargets = HasV4SubT in +def CMPnotGTU_ri : ALU32_ri<(outs PredRegs:$dst), + (ins IntRegs:$src1, u9Ext:$src2), + "$dst = !cmp.gtu($src1, #$src2)", + [(set (i1 PredRegs:$dst), + (not (setugt (i32 IntRegs:$src1), u9ImmPred:$src2)))]>, + Requires<[HasV4T]>; + +let isCompare = 1, validSubTargets = HasV4SubT in def CMPbEQri_V4 : MInst<(outs PredRegs:$dst), (ins IntRegs:$src1, u8Imm:$src2), "$dst = cmpb.eq($src1, #$src2)", @@ -3664,6 +3766,12 @@ def CMPbEQri_V4 : MInst<(outs PredRegs:$dst), (seteq (and (i32 IntRegs:$src1), 255), u8ImmPred:$src2))]>, Requires<[HasV4T]>; +def : Pat <(brcond (i1 (setne (and (i32 IntRegs:$src1), 255), u8ImmPred:$src2)), + bb:$offset), + (JMP_cNot (CMPbEQri_V4 (i32 IntRegs:$src1), u8ImmPred:$src2), + bb:$offset)>, + Requires<[HasV4T]>; + // Pd=cmpb.eq(Rs,Rt) let isCompare = 1, validSubTargets = HasV4SubT in def CMPbEQrr_ubub_V4 : MInst<(outs PredRegs:$dst), @@ -3705,20 +3813,21 @@ def CMPbGTUri_V4 : MInst<(outs PredRegs:$dst), Requires<[HasV4T]>, ImmRegRel; // Pd=cmpb.gtu(Rs,Rt) -let isCompare = 1 in +let isCompare = 1, validSubTargets = HasV4SubT, CextOpcode = "CMPbGTU", +InputType = "reg" in def CMPbGTUrr_V4 : MInst<(outs PredRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2), "$dst = cmpb.gtu($src1, $src2)", [(set (i1 PredRegs:$dst), (setugt (and (i32 IntRegs:$src1), 255), (and (i32 IntRegs:$src2), 255)))]>, - Requires<[HasV4T]>; + Requires<[HasV4T]>, ImmRegRel; // Following instruction is not being extended as it results into the incorrect // code for negative numbers. // Signed half compare(.eq) ri. // Pd=cmph.eq(Rs,#s8) -let isCompare = 1 in +let isCompare = 1, validSubTargets = HasV4SubT in def CMPhEQri_V4 : MInst<(outs PredRegs:$dst), (ins IntRegs:$src1, s8Imm:$src2), "$dst = cmph.eq($src1, #$src2)", @@ -3732,7 +3841,7 @@ def CMPhEQri_V4 : MInst<(outs PredRegs:$dst), // r0=and(r0,#0xffff) // p0=cmp.eq(r0,#0) // Pd=cmph.eq(Rs,Rt) -let isCompare = 1 in +let isCompare = 1, validSubTargets = HasV4SubT in def CMPhEQrr_xor_V4 : MInst<(outs PredRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2), "$dst = cmph.eq($src1, $src2)", @@ -3747,7 +3856,7 @@ def CMPhEQrr_xor_V4 : MInst<(outs PredRegs:$dst), // r1=asl(r1,16) // p0=cmp.eq(r0,r1) // Pd=cmph.eq(Rs,Rt) -let isCompare = 1 in +let isCompare = 1, validSubTargets = HasV4SubT in def CMPhEQrr_shl_V4 : MInst<(outs PredRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2), "$dst = cmph.eq($src1, $src2)", @@ -3761,19 +3870,20 @@ used in the cmph.gt instruction. // Signed half compare(.gt) ri. // Pd=cmph.gt(Rs,#s8) -let isCompare = 1 in +let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 8, +isCompare = 1, validSubTargets = HasV4SubT in def CMPhGTri_V4 : MInst<(outs PredRegs:$dst), - (ins IntRegs:$src1, s8Imm:$src2), + (ins IntRegs:$src1, s8Ext:$src2), "$dst = cmph.gt($src1, #$src2)", [(set (i1 PredRegs:$dst), (setgt (shl (i32 IntRegs:$src1), (i32 16)), - s8ImmPred:$src2))]>, + s8ExtPred:$src2))]>, Requires<[HasV4T]>; */ // Signed half compare(.gt) rr. // Pd=cmph.gt(Rs,Rt) -let isCompare = 1 in +let isCompare = 1, validSubTargets = HasV4SubT in def CMPhGTrr_shl_V4 : MInst<(outs PredRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2), "$dst = cmph.gt($src1, $src2)", @@ -3784,24 +3894,41 @@ def CMPhGTrr_shl_V4 : MInst<(outs PredRegs:$dst), // Unsigned half compare rr (.gtu). // Pd=cmph.gtu(Rs,Rt) -let isCompare = 1 in +let isCompare = 1, validSubTargets = HasV4SubT, CextOpcode = "CMPhGTU", +InputType = "reg" in def CMPhGTUrr_V4 : MInst<(outs PredRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2), "$dst = cmph.gtu($src1, $src2)", [(set (i1 PredRegs:$dst), (setugt (and (i32 IntRegs:$src1), 65535), (and (i32 IntRegs:$src2), 65535)))]>, - Requires<[HasV4T]>; + Requires<[HasV4T]>, ImmRegRel; // Unsigned half compare ri (.gtu). // Pd=cmph.gtu(Rs,#u7) -let isCompare = 1 in +let isExtendable = 1, opExtendable = 2, isExtentSigned = 0, opExtentBits = 7, +isCompare = 1, validSubTargets = HasV4SubT, CextOpcode = "CMPhGTU", +InputType = "imm" in def CMPhGTUri_V4 : MInst<(outs PredRegs:$dst), - (ins IntRegs:$src1, u7Imm:$src2), + (ins IntRegs:$src1, u7Ext:$src2), "$dst = cmph.gtu($src1, #$src2)", [(set (i1 PredRegs:$dst), (setugt (and (i32 IntRegs:$src1), 65535), - u7ImmPred:$src2))]>, - Requires<[HasV4T]>; + u7ExtPred:$src2))]>, + Requires<[HasV4T]>, ImmRegRel; + +let validSubTargets = HasV4SubT in +def NTSTBIT_rr : SInst<(outs PredRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2), + "$dst = !tstbit($src1, $src2)", + [(set (i1 PredRegs:$dst), + (seteq (and (shl 1, (i32 IntRegs:$src2)), (i32 IntRegs:$src1)), 0))]>, + Requires<[HasV4T]>; + +let validSubTargets = HasV4SubT in +def NTSTBIT_ri : SInst<(outs PredRegs:$dst), (ins IntRegs:$src1, u5Imm:$src2), + "$dst = !tstbit($src1, $src2)", + [(set (i1 PredRegs:$dst), + (seteq (and (shl 1, u5ImmPred:$src2), (i32 IntRegs:$src1)), 0))]>, + Requires<[HasV4T]>; //===----------------------------------------------------------------------===// // XTYPE/PRED - diff --git a/llvm/test/CodeGen/Hexagon/cmp-to-genreg.ll b/llvm/test/CodeGen/Hexagon/cmp-to-genreg.ll new file mode 100644 index 0000000..97cf51c --- /dev/null +++ b/llvm/test/CodeGen/Hexagon/cmp-to-genreg.ll @@ -0,0 +1,34 @@ +; RUN: llc -march=hexagon -mcpu=hexagonv4 < %s | FileCheck %s +; Check that we generate compare to general register. + +define i32 @compare1(i32 %a) nounwind { +; CHECK: r{{[0-9]+}}{{ *}}={{ *}}cmp.eq(r{{[0-9]+}},{{ *}}#120) +entry: + %cmp = icmp eq i32 %a, 120 + %conv = zext i1 %cmp to i32 + ret i32 %conv +} + +define i32 @compare2(i32 %a) nounwind readnone { +; CHECK: r{{[0-9]+}}{{ *}}={{ *}}!cmp.eq(r{{[0-9]+}},{{ *}}#120) +entry: + %cmp = icmp ne i32 %a, 120 + %conv = zext i1 %cmp to i32 + ret i32 %conv +} + +define i32 @compare3(i32 %a, i32 %b) nounwind readnone { +; CHECK: r{{[0-9]+}}{{ *}}={{ *}}cmp.eq(r{{[0-9]+}},{{ *}}r{{[0-9]+}}) +entry: + %cmp = icmp eq i32 %a, %b + %conv = zext i1 %cmp to i32 + ret i32 %conv +} + +define i32 @compare4(i32 %a, i32 %b) nounwind readnone { +; CHECK: r{{[0-9]+}}{{ *}}={{ *}}!cmp.eq(r{{[0-9]+}},{{ *}}r{{[0-9]+}}) +entry: + %cmp = icmp ne i32 %a, %b + %conv = zext i1 %cmp to i32 + ret i32 %conv +} diff --git a/llvm/test/CodeGen/Hexagon/cmp-to-predreg.ll b/llvm/test/CodeGen/Hexagon/cmp-to-predreg.ll new file mode 100644 index 0000000..d430b90 --- /dev/null +++ b/llvm/test/CodeGen/Hexagon/cmp-to-predreg.ll @@ -0,0 +1,43 @@ +; RUN: llc -march=hexagon -mcpu=hexagonv4 < %s | FileCheck %s +; Check that we generate compare to predicate register. + +define i32 @compare1(i32 %a, i32 %b) nounwind { +; CHECK: p{{[0-3]}}{{ *}}={{ *}}!cmp.eq(r{{[0-9]+}},{{ *}}r{{[0-9]+}}) +entry: + %cmp = icmp ne i32 %a, %b + %add = add nsw i32 %a, %b + %sub = sub nsw i32 %a, %b + %add.sub = select i1 %cmp, i32 %add, i32 %sub + ret i32 %add.sub +} + +define i32 @compare2(i32 %a) nounwind { +; CHECK: p{{[0-3]}}{{ *}}={{ *}}!cmp.eq(r{{[0-9]+}},{{ *}}#10) +entry: + %cmp = icmp ne i32 %a, 10 + %add = add nsw i32 %a, 10 + %sub = sub nsw i32 %a, 10 + %add.sub = select i1 %cmp, i32 %add, i32 %sub + ret i32 %add.sub +} + +define i32 @compare3(i32 %a, i32 %b) nounwind { +; CHECK: p{{[0-3]}}{{ *}}={{ *}}cmp.gt(r{{[0-9]+}},{{ *}}r{{[0-9]+}}) +entry: + %cmp = icmp sgt i32 %a, %b + %sub = sub nsw i32 %a, %b + %add = add nsw i32 %a, %b + %sub.add = select i1 %cmp, i32 %sub, i32 %add + ret i32 %sub.add +} + +define i32 @compare4(i32 %a) nounwind { +; CHECK: p{{[0-3]}}{{ *}}={{ *}}cmp.gt(r{{[0-9]+}},{{ *}}#10) +entry: + %cmp = icmp sgt i32 %a, 10 + %sub = sub nsw i32 %a, 10 + %add = add nsw i32 %a, 10 + %sub.add = select i1 %cmp, i32 %sub, i32 %add + ret i32 %sub.add +} + -- 2.7.4