From: Richard Sandiford Date: Fri, 23 Aug 2013 11:18:53 +0000 (+0000) Subject: [SystemZ] Prefer LHI;ST... over LAY;MV... X-Git-Url: http://demsky.eecs.uci.edu/git/?a=commitdiff_plain;h=65ddcfa8c1c05aeecd9d4fb062bb121e376aaceb;p=oota-llvm.git [SystemZ] Prefer LHI;ST... over LAY;MV... If we had a store of an integer to memory, and the integer and store size were suitable for a form of MV..., we used MV... no matter what. We could then have sequences like: lay %r2, 0(%r3,%r4) mvi 0(%r2), 4 In these cases it seems better to force the constant into a register and use a normal store: lhi %r2, 4 stc %r2, 0(%r3, %r4) since %r2 is more likely to be hoisted and is easier to rematerialize. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@189098 91177308-0d34-0410-b5e6-96231b3b80d8 --- diff --git a/lib/Target/SystemZ/SystemZISelDAGToDAG.cpp b/lib/Target/SystemZ/SystemZISelDAGToDAG.cpp index d9794b1d643..e0d543791b0 100644 --- a/lib/Target/SystemZ/SystemZISelDAGToDAG.cpp +++ b/lib/Target/SystemZ/SystemZISelDAGToDAG.cpp @@ -159,6 +159,12 @@ class SystemZDAGToDAGISel : public SelectionDAGISel { bool selectBDAddr(SystemZAddressingMode::DispRange DR, SDValue Addr, SDValue &Base, SDValue &Disp); + // Try to match Addr as a FormBDX address with displacement type DR. + // Return true on success and if the result had no index. Store the + // base and displacement in Base and Disp respectively. + bool selectMVIAddr(SystemZAddressingMode::DispRange DR, SDValue Addr, + SDValue &Base, SDValue &Disp); + // Try to match Addr as a FormBDX* address of form Form with // displacement type DR. Return true on success, storing the base, // displacement and index in Base, Disp and Index respectively. @@ -189,6 +195,14 @@ class SystemZDAGToDAGISel : public SelectionDAGISel { return selectBDAddr(SystemZAddressingMode::Disp20Pair, Addr, Base, Disp); } + // MVI matching routines used by SystemZOperands.td. + bool selectMVIAddr12Pair(SDValue Addr, SDValue &Base, SDValue &Disp) { + return selectMVIAddr(SystemZAddressingMode::Disp12Pair, Addr, Base, Disp); + } + bool selectMVIAddr20Pair(SDValue Addr, SDValue &Base, SDValue &Disp) { + return selectMVIAddr(SystemZAddressingMode::Disp20Pair, Addr, Base, Disp); + } + // BDX matching routines used by SystemZOperands.td. bool selectBDXAddr12Only(SDValue Addr, SDValue &Base, SDValue &Disp, SDValue &Index) { @@ -575,6 +589,17 @@ bool SystemZDAGToDAGISel::selectBDAddr(SystemZAddressingMode::DispRange DR, return true; } +bool SystemZDAGToDAGISel::selectMVIAddr(SystemZAddressingMode::DispRange DR, + SDValue Addr, SDValue &Base, + SDValue &Disp) { + SystemZAddressingMode AM(SystemZAddressingMode::FormBDXNormal, DR); + if (!selectAddress(Addr, AM) || AM.Index.getNode()) + return false; + + getAddressOperands(AM, Addr.getValueType(), Base, Disp); + return true; +} + bool SystemZDAGToDAGISel::selectBDXAddr(SystemZAddressingMode::AddrForm Form, SystemZAddressingMode::DispRange DR, SDValue Addr, SDValue &Base, diff --git a/lib/Target/SystemZ/SystemZInstrFormats.td b/lib/Target/SystemZ/SystemZInstrFormats.td index 39b76397054..f53833210ec 100644 --- a/lib/Target/SystemZ/SystemZInstrFormats.td +++ b/lib/Target/SystemZ/SystemZInstrFormats.td @@ -627,27 +627,33 @@ class StoreMultipleRSY opcode, RegisterOperand cls> let mayStore = 1; } +// StoreSI* instructions are used to store an integer to memory, but the +// addresses are more restricted than for normal stores. If we are in the +// situation of having to force either the address into a register or the +// constant into a register, it's usually better to do the latter. +// We therefore match the address in the same way as a normal store and +// only use the StoreSI* instruction if the matched address is suitable. class StoreSI opcode, SDPatternOperator operator, - Immediate imm, AddressingMode mode = bdaddr12only> - : InstSI + : InstSI { + [(operator imm:$I2, mviaddr12pair:$BD1)]> { let mayStore = 1; } class StoreSIY opcode, SDPatternOperator operator, - Immediate imm, AddressingMode mode = bdaddr20only> - : InstSIY + : InstSIY { + [(operator imm:$I2, mviaddr20pair:$BD1)]> { let mayStore = 1; } class StoreSIL opcode, SDPatternOperator operator, Immediate imm> - : InstSIL { + [(operator imm:$I2, mviaddr12pair:$BD1)]> { let mayStore = 1; } @@ -655,9 +661,9 @@ multiclass StoreSIPair siOpcode, bits<16> siyOpcode, SDPatternOperator operator, Immediate imm> { let DispKey = mnemonic in { let DispSize = "12" in - def "" : StoreSI; + def "" : StoreSI; let DispSize = "20" in - def Y : StoreSIY; + def Y : StoreSIY; } } diff --git a/lib/Target/SystemZ/SystemZOperands.td b/lib/Target/SystemZ/SystemZOperands.td index 9d794392280..eb96dba0f2d 100644 --- a/lib/Target/SystemZ/SystemZOperands.td +++ b/lib/Target/SystemZ/SystemZOperands.td @@ -435,6 +435,7 @@ def BDLAddr64Disp12Len8 : AddressAsmOperand<"BDLAddr", "64", "12", "Len8">; // is one of: // shift : base + displacement (32-bit) // bdaddr : base + displacement +// mviaddr : like bdaddr, but reject cases with a natural index // bdxaddr : base + displacement + index // laaddr : like bdxaddr, but used for Load Address operations // dynalloc : base + displacement + index + ADJDYNALLOC @@ -460,6 +461,8 @@ def bdaddr12only : BDMode <"BDAddr", "64", "12", "Only">; def bdaddr12pair : BDMode <"BDAddr", "64", "12", "Pair">; def bdaddr20only : BDMode <"BDAddr", "64", "20", "Only">; def bdaddr20pair : BDMode <"BDAddr", "64", "20", "Pair">; +def mviaddr12pair : BDMode <"MVIAddr", "64", "12", "Pair">; +def mviaddr20pair : BDMode <"MVIAddr", "64", "20", "Pair">; def bdxaddr12only : BDXMode<"BDXAddr", "64", "12", "Only">; def bdxaddr12pair : BDXMode<"BDXAddr", "64", "12", "Pair">; def bdxaddr20only : BDXMode<"BDXAddr", "64", "20", "Only">; diff --git a/test/CodeGen/SystemZ/alloca-02.ll b/test/CodeGen/SystemZ/alloca-02.ll index b6ed7f7741d..b5787b10235 100644 --- a/test/CodeGen/SystemZ/alloca-02.ll +++ b/test/CodeGen/SystemZ/alloca-02.ll @@ -21,18 +21,21 @@ define i64 @f1(i64 %length, i64 %index) { ; ; CHECK-C-LABEL: f1: ; CHECK-C: lgr %r15, [[ADDR:%r[1-5]]] -; CHECK-C: la [[TMP:%r[1-5]]], 160(%r3,[[ADDR]]) -; CHECK-C: mvi 0([[TMP]]), 2 +; CHECK-C-DAG: la %r2, 160([[ADDR]]) +; CHECK-C-DAG: lhi [[TMP:%r[0-5]]], 2 +; CHECK-C: stc [[TMP]], 0({{%r3,%r2|%r2,%r3}}) ; ; CHECK-D-LABEL: f1: ; CHECK-D: lgr %r15, [[ADDR:%r[1-5]]] -; CHECK-D: la [[TMP:%r[1-5]]], 160(%r3,[[ADDR]]) -; CHECK-D: mvi 4095([[TMP]]), 3 +; CHECK-D-DAG: la %r2, 160([[ADDR]]) +; CHECK-D-DAG: lhi [[TMP:%r[0-5]]], 3 +; CHECK-D: stc [[TMP]], 4095({{%r3,%r2|%r2,%r3}}) ; ; CHECK-E-LABEL: f1: ; CHECK-E: lgr %r15, [[ADDR:%r[1-5]]] -; CHECK-E: la [[TMP:%r[1-5]]], 160(%r3,[[ADDR]]) -; CHECK-E: mviy 4096([[TMP]]), 4 +; CHECK-E-DAG: la %r2, 160([[ADDR]]) +; CHECK-E-DAG: lhi [[TMP:%r[0-5]]], 4 +; CHECK-E: stcy [[TMP]], 4096({{%r3,%r2|%r2,%r3}}) %a = alloca i8, i64 %length store volatile i8 0, i8 *%a %b = getelementptr i8 *%a, i64 4095 diff --git a/test/CodeGen/SystemZ/frame-13.ll b/test/CodeGen/SystemZ/frame-13.ll index 1d383542694..60bff508d93 100644 --- a/test/CodeGen/SystemZ/frame-13.ll +++ b/test/CodeGen/SystemZ/frame-13.ll @@ -182,17 +182,16 @@ define void @f8() { } ; Check a case where the original displacement is out of range. The backend -; should force an LAY from the outset. We don't yet do any kind of anchor -; optimization, so there should be no offset on the MVHI itself. +; should force STY to be used instead. define void @f9() { ; CHECK-NOFP-LABEL: f9: -; CHECK-NOFP: lay %r1, 12296(%r15) -; CHECK-NOFP: mvhi 0(%r1), 42 +; CHECK-NOFP: lhi [[TMP:%r[0-5]]], 42 +; CHECK-NOFP: sty [[TMP]], 12296(%r15) ; CHECK-NOFP: br %r14 ; ; CHECK-FP-LABEL: f9: -; CHECK-FP: lay %r1, 12296(%r11) -; CHECK-FP: mvhi 0(%r1), 42 +; CHECK-FP: lhi [[TMP:%r[0-5]]], 42 +; CHECK-FP: sty [[TMP]], 12296(%r11) ; CHECK-FP: br %r14 %region1 = alloca [2006 x i32], align 8 %region2 = alloca [2006 x i32], align 8 diff --git a/test/CodeGen/SystemZ/int-const-03.ll b/test/CodeGen/SystemZ/int-const-03.ll index 78db9637e4e..af1cef2c138 100644 --- a/test/CodeGen/SystemZ/int-const-03.ll +++ b/test/CodeGen/SystemZ/int-const-03.ll @@ -139,11 +139,11 @@ define void @f14(i8 *%src) { ret void } -; Check that MVI does not allow an index +; Check that MVI does not allow an index. We prefer STC in that case. define void @f15(i64 %src, i64 %index) { ; CHECK-LABEL: f15: -; CHECK: agr %r2, %r3 -; CHECK: mvi 4095(%r2), 42 +; CHECK: lhi [[TMP:%r[0-5]]], 42 +; CHECK: stc [[TMP]], 4095({{%r2,%r3|%r3,%r2}} ; CHECK: br %r14 %add1 = add i64 %src, %index %add2 = add i64 %add1, 4095 @@ -152,11 +152,11 @@ define void @f15(i64 %src, i64 %index) { ret void } -; Check that MVIY does not allow an index +; Check that MVIY does not allow an index. We prefer STCY in that case. define void @f16(i64 %src, i64 %index) { ; CHECK-LABEL: f16: -; CHECK: agr %r2, %r3 -; CHECK: mviy 4096(%r2), 42 +; CHECK: lhi [[TMP:%r[0-5]]], 42 +; CHECK: stcy [[TMP]], 4096({{%r2,%r3|%r3,%r2}} ; CHECK: br %r14 %add1 = add i64 %src, %index %add2 = add i64 %add1, 4096 diff --git a/test/CodeGen/SystemZ/int-const-04.ll b/test/CodeGen/SystemZ/int-const-04.ll index c109faab20d..aced50b5601 100644 --- a/test/CodeGen/SystemZ/int-const-04.ll +++ b/test/CodeGen/SystemZ/int-const-04.ll @@ -75,34 +75,34 @@ define void @f8(i16 *%a) { ret void } -; Check the next halfword up, which needs separate address logic. -; Other sequences besides this one would be OK. +; Check the next halfword up, which is out of range. We prefer STHY +; in that case. define void @f9(i16 *%a) { ; CHECK-LABEL: f9: -; CHECK: aghi %r2, 4096 -; CHECK: mvhhi 0(%r2), 42 +; CHECK: lhi [[TMP:%r[0-5]]], 42 +; CHECK: sthy [[TMP]], 4096(%r2) ; CHECK: br %r14 %ptr = getelementptr i16 *%a, i64 2048 store i16 42, i16 *%ptr ret void } -; Check negative displacements, which also need separate address logic. +; Check negative displacements, for which we again prefer STHY. define void @f10(i16 *%a) { ; CHECK-LABEL: f10: -; CHECK: aghi %r2, -2 -; CHECK: mvhhi 0(%r2), 42 +; CHECK: lhi [[TMP:%r[0-5]]], 42 +; CHECK: sthy [[TMP]], -2(%r2) ; CHECK: br %r14 %ptr = getelementptr i16 *%a, i64 -1 store i16 42, i16 *%ptr ret void } -; Check that MVHHI does not allow an index +; Check that MVHHI does not allow an index. define void @f11(i64 %src, i64 %index) { ; CHECK-LABEL: f11: -; CHECK: agr %r2, %r3 -; CHECK: mvhhi 0(%r2), 42 +; CHECK: lhi [[TMP:%r[0-5]]], 42 +; CHECK: sth [[TMP]], 0({{%r2,%r3|%r3,%r2}}) ; CHECK: br %r14 %add = add i64 %src, %index %ptr = inttoptr i64 %add to i16 * diff --git a/test/CodeGen/SystemZ/int-const-05.ll b/test/CodeGen/SystemZ/int-const-05.ll index d0c85691493..98d6851c197 100644 --- a/test/CodeGen/SystemZ/int-const-05.ll +++ b/test/CodeGen/SystemZ/int-const-05.ll @@ -66,34 +66,33 @@ define void @f7(i32 *%a) { ret void } -; Check the next word up, which needs separate address logic. -; Other sequences besides this one would be OK. +; Check the next word up, which is out of range. We prefer STY in that case. define void @f8(i32 *%a) { ; CHECK-LABEL: f8: -; CHECK: aghi %r2, 4096 -; CHECK: mvhi 0(%r2), 42 +; CHECK: lhi [[TMP:%r[0-5]]], 42 +; CHECK: sty [[TMP]], 4096(%r2) ; CHECK: br %r14 %ptr = getelementptr i32 *%a, i64 1024 store i32 42, i32 *%ptr ret void } -; Check negative displacements, which also need separate address logic. +; Check negative displacements, for which we again prefer STY. define void @f9(i32 *%a) { ; CHECK-LABEL: f9: -; CHECK: aghi %r2, -4 -; CHECK: mvhi 0(%r2), 42 +; CHECK: lhi [[TMP:%r[0-5]]], 42 +; CHECK: sty [[TMP]], -4(%r2) ; CHECK: br %r14 %ptr = getelementptr i32 *%a, i64 -1 store i32 42, i32 *%ptr ret void } -; Check that MVHI does not allow an index +; Check that MVHI does not allow an index. define void @f10(i64 %src, i64 %index) { ; CHECK-LABEL: f10: -; CHECK: agr %r2, %r3 -; CHECK: mvhi 0(%r2), 42 +; CHECK: lhi [[TMP:%r[0-5]]], 42 +; CHECK: st [[TMP]], 0({{%r2,%r3|%r3,%r2}}) ; CHECK: br %r14 %add = add i64 %src, %index %ptr = inttoptr i64 %add to i32 * diff --git a/test/CodeGen/SystemZ/int-const-06.ll b/test/CodeGen/SystemZ/int-const-06.ll index 12a555c61e4..cf07c665dde 100644 --- a/test/CodeGen/SystemZ/int-const-06.ll +++ b/test/CodeGen/SystemZ/int-const-06.ll @@ -66,34 +66,34 @@ define void @f7(i64 *%a) { ret void } -; Check the next doubleword up, which needs separate address logic. -; Other sequences besides this one would be OK. +; Check the next doubleword up, which is out of range. We prefer STG +; in that case. define void @f8(i64 *%a) { ; CHECK-LABEL: f8: -; CHECK: aghi %r2, 4096 -; CHECK: mvghi 0(%r2), 42 +; CHECK: lghi [[TMP:%r[0-5]]], 42 +; CHECK: stg [[TMP]], 4096(%r2) ; CHECK: br %r14 %ptr = getelementptr i64 *%a, i64 512 store i64 42, i64 *%ptr ret void } -; Check negative displacements, which also need separate address logic. +; Check negative displacements, for which we again prefer STG. define void @f9(i64 *%a) { ; CHECK-LABEL: f9: -; CHECK: aghi %r2, -8 -; CHECK: mvghi 0(%r2), 42 +; CHECK: lghi [[TMP:%r[0-5]]], 42 +; CHECK: stg [[TMP]], -8(%r2) ; CHECK: br %r14 %ptr = getelementptr i64 *%a, i64 -1 store i64 42, i64 *%ptr ret void } -; Check that MVGHI does not allow an index +; Check that MVGHI does not allow an index. define void @f10(i64 %src, i64 %index) { ; CHECK-LABEL: f10: -; CHECK: agr %r2, %r3 -; CHECK: mvghi 0(%r2), 42 +; CHECK: lghi [[TMP:%r[0-5]]], 42 +; CHECK: stg [[TMP]], 0({{%r2,%r3|%r3,%r2}}) ; CHECK: br %r14 %add = add i64 %src, %index %ptr = inttoptr i64 %add to i64 *