Conanap created this revision.
Conanap added reviewers: power-llvm-team, PowerPC, nemanjai, saghir, hfinkel.
Conanap added projects: LLVM, clang, PowerPC.
Implemented following functions for Load VSX Vector Sign extend nad zero extend
instructions:
vector signed __int128 vec_xl_sext (signed long long, signed char *);
vector signed __int128 vec_xl_sext (signed long long, signed short *);
vector signed __int128 vec_xl_sext (signed long long, signed int *);
vector signed __int128 vec_xl_sext (signed long long, signed long long *);
vector unsigned __int128 vec_xl_zext (signed long long, unsigned char *);
vector unsigned __int128 vec_xl_zext (signed long long, unsigned short *);
vector unsigned __int128 vec_xl_zext (signed long long, unsigned int *);
vector unsigned __int128 vec_xl_zext (signed long long, unsigned long long *);
Repository:
rG LLVM Github Monorepo
https://reviews.llvm.org/D82502
Files:
clang/lib/Headers/altivec.h
clang/test/CodeGen/builtins-ppc-p10vector.c
llvm/lib/Target/PowerPC/PPCISelLowering.cpp
llvm/lib/Target/PowerPC/PPCISelLowering.h
llvm/lib/Target/PowerPC/PPCInstrPrefix.td
llvm/test/CodeGen/PowerPC/p10-vsx-builtins.ll
llvm/test/MC/Disassembler/PowerPC/p10insts.txt
llvm/test/MC/PowerPC/p10.s
Index: llvm/test/MC/PowerPC/p10.s
===================================================================
--- llvm/test/MC/PowerPC/p10.s
+++ llvm/test/MC/PowerPC/p10.s
@@ -33,3 +33,15 @@
# CHECK-BE: vclrrb 1, 4, 3 # encoding: [0x10,0x24,0x19,0xcd]
# CHECK-LE: vclrrb 1, 4, 3 # encoding: [0xcd,0x19,0x24,0x10]
vclrrb 1, 4, 3
+# CHECK-BE: lxvrbx 32, 1, 2 # encoding: [0x7c,0x01,0x10,0x1b]
+# CHECK-LE: lxvrbx 32, 1, 2 # encoding: [0x1b,0x10,0x01,0x7c]
+ lxvrbx 32, 1, 2
+# CHECK-BE: lxvrhx 33, 1, 2 # encoding: [0x7c,0x21,0x10,0x5b]
+# CHECK-LE: lxvrhx 33, 1, 2 # encoding: [0x5b,0x10,0x21,0x7c]
+ lxvrhx 33, 1, 2
+# CHECK-BE: lxvrdx 34, 1, 2 # encoding: [0x7c,0x41,0x10,0xdb]
+# CHECK-LE: lxvrdx 34, 1, 2 # encoding: [0xdb,0x10,0x41,0x7c]
+ lxvrdx 34, 1, 2
+# CHECK-BE: lxvrwx 35, 1, 2 # encoding: [0x7c,0x61,0x10,0x9b]
+# CHECK-LE: lxvrwx 35, 1, 2 # encoding: [0x9b,0x10,0x61,0x7c]
+ lxvrwx 35, 1, 2
Index: llvm/test/MC/Disassembler/PowerPC/p10insts.txt
===================================================================
--- llvm/test/MC/Disassembler/PowerPC/p10insts.txt
+++ llvm/test/MC/Disassembler/PowerPC/p10insts.txt
@@ -30,3 +30,15 @@
# CHECK: vclrrb 1, 4, 3
0x10 0x24 0x19 0xcd
+
+# CHECK: lxvrbx 32, 1, 2
+0x7c 0x01 0x10 0x1b
+
+# CHECK: lxvrhx 33, 1, 2
+0x7c 0x21 0x10 0x5b
+
+# CHECK: lxvrdx 34, 1, 2
+0x7c 0x41 0x10 0xdb
+
+# CHECK: lxvrwx 35, 1, 2
+0x7c 0x61 0x10 0x9b
Index: llvm/test/CodeGen/PowerPC/p10-vsx-builtins.ll
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/PowerPC/p10-vsx-builtins.ll
@@ -0,0 +1,74 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -verify-machineinstrs -mtriple=powerpc64le-unknown-linux-gnu \
+; RUN: -mcpu=pwr10 -ppc-asm-full-reg-names -ppc-vsr-nums-as-vr < %s | \
+; RUN: FileCheck %s
+
+; ZEXT TEST CASES
+
+; i8
+; CHECK: lxvrbx
+; Function Attrs: norecurse nounwind readonly
+define dso_local <1 x i128> @vec_xl_zext(i64 %__offset, i8* nocapture readonly %__pointer) local_unnamed_addr #0 {
+; CHECK-LABEL: vec_xl_zext:
+; CHECK: # %bb.0: # %entry
+; CHECK-NEXT: lxvrbx v2, r4, r3
+; CHECK-NEXT: blr
+entry:
+ %add.ptr = getelementptr inbounds i8, i8* %__pointer, i64 %__offset
+ %0 = load i8, i8* %add.ptr, align 1
+ %conv = zext i8 %0 to i128
+ %splat.splatinsert = insertelement <1 x i128> undef, i128 %conv, i32 0
+ ret <1 x i128> %splat.splatinsert
+}
+
+; i16
+; CHECK: lxvrhx
+; Function Attrs: norecurse nounwind readonly
+define dso_local <1 x i128> @vec_xl_zext_short(i64 %__offset, i16* nocapture readonly %__pointer) local_unnamed_addr #0 {
+; CHECK-LABEL: vec_xl_zext_short:
+; CHECK: # %bb.0: # %entry
+; CHECK-NEXT: sldi r3, r3, 1
+; CHECK-NEXT: lxvrhx v2, r4, r3
+; CHECK-NEXT: blr
+entry:
+ %add.ptr = getelementptr inbounds i16, i16* %__pointer, i64 %__offset
+ %0 = load i16, i16* %add.ptr, align 2
+ %conv = zext i16 %0 to i128
+ %splat.splatinsert = insertelement <1 x i128> undef, i128 %conv, i32 0
+ ret <1 x i128> %splat.splatinsert
+}
+
+; i32
+; CHECK: lxvrwx
+; Function Attrs: norecurse nounwind readonly
+define dso_local <1 x i128> @vec_xl_zext_word(i64 %__offset, i32* nocapture readonly %__pointer) local_unnamed_addr #0 {
+; CHECK-LABEL: vec_xl_zext_word:
+; CHECK: # %bb.0: # %entry
+; CHECK-NEXT: sldi r3, r3, 2
+; CHECK-NEXT: lxvrwx v2, r4, r3
+; CHECK-NEXT: blr
+entry:
+ %add.ptr = getelementptr inbounds i32, i32* %__pointer, i64 %__offset
+ %0 = load i32, i32* %add.ptr, align 4
+ %conv = zext i32 %0 to i128
+ %splat.splatinsert = insertelement <1 x i128> undef, i128 %conv, i32 0
+ ret <1 x i128> %splat.splatinsert
+}
+
+; i64
+; CHECK: lxvrdx
+; Function Attrs: norecurse nounwind readonly
+define dso_local <1 x i128> @vec_xl_zext_dw(i64 %__offset, i64* nocapture readonly %__pointer) local_unnamed_addr #0 {
+; CHECK-LABEL: vec_xl_zext_dw:
+; CHECK: # %bb.0: # %entry
+; CHECK-NEXT: sldi r3, r3, 3
+; CHECK-NEXT: lxvrdx v2, r4, r3
+; CHECK-NEXT: blr
+entry:
+ %add.ptr = getelementptr inbounds i64, i64* %__pointer, i64 %__offset
+ %0 = load i64, i64* %add.ptr, align 8
+ %conv = zext i64 %0 to i128
+ %splat.splatinsert = insertelement <1 x i128> undef, i128 %conv, i32 0
+ ret <1 x i128> %splat.splatinsert
+}
+
Index: llvm/lib/Target/PowerPC/PPCInstrPrefix.td
===================================================================
--- llvm/lib/Target/PowerPC/PPCInstrPrefix.td
+++ llvm/lib/Target/PowerPC/PPCInstrPrefix.td
@@ -209,6 +209,21 @@
def PrefixInstrs : Predicate<"PPCSubTarget->hasPrefixInstrs()">;
def IsISA3_1 : Predicate<"PPCSubTarget->isISA3_1()">;
+let mayLoad = 1, mayStore = 0, Predicates = [IsISA3_1] in {
+ // The XFormMemOp flag is set on the instruction format.
+ def LXVRBX : X_XT6_RA5_RB5<31, 13, "lxvrbx", vsrc, []>;
+ def LXVRHX : X_XT6_RA5_RB5<31, 45, "lxvrhx", vsrc, []>;
+ def LXVRWX : X_XT6_RA5_RB5<31, 77, "lxvrwx", vsrc, []>;
+ def LXVRDX : X_XT6_RA5_RB5<31, 109, "lxvrdx", vsrc, []>;
+}
+
+def SDT_PPCLXVRZX : SDTypeProfile<1, 2, [
+ SDTCisVT<0, v1i128>, SDTCisPtrTy<1>, SDTCisPtrTy<2>
+]>;
+
+def PPClxvrzx : SDNode<"PPCISD::LXVRZX", SDT_PPCLXVRZX,
+ [SDNPHasChain, SDNPMayLoad]>;
+
let Predicates = [PrefixInstrs] in {
let Interpretation64Bit = 1, isCodeGenOnly = 1 in {
defm PADDI8 :
@@ -512,7 +527,7 @@
(PPCmatpcreladdr pcreladdr:$dst), 8),
(PSTXSDpc (COPY_TO_REGCLASS (XSCVQPUDZ f128:$src), VFRC), $dst, 0)>;
- // If the PPCmatpcreladdr node is not caught by any other pattern it should be
+ // If the PPCmatpcreladdr node is not caught by any other pattern it should be
// caught here and turned into a paddi instruction to materialize the address.
def : Pat<(PPCmatpcreladdr pcreladdr:$addr), (PADDI8pc 0, $addr)>;
}
@@ -564,4 +579,16 @@
(v4i32 (COPY_TO_REGCLASS (XXGENPCVWM $VRB, imm:$IMM), VRRC))>;
def : Pat<(v2i64 (int_ppc_vsx_xxgenpcvdm v2i64:$VRB, imm:$IMM)),
(v2i64 (COPY_TO_REGCLASS (XXGENPCVDM $VRB, imm:$IMM), VRRC))>;
+
+ // Special isntructions for PPC10 load vsx vector with zero extend
+ // Utilize the appropriate Load VSX Vector Rightmost instruction depending
+ // on the width of PPClxvrzx.
+ def : Pat <(v1i128 (PPClxvrzx xoaddr:$src, 8)),
+ (v1i128 (COPY_TO_REGCLASS (LXVRBX xoaddr:$src), VRRC))>;
+ def : Pat <(v1i128 (PPClxvrzx xoaddr:$src, 16)),
+ (v1i128 (COPY_TO_REGCLASS (LXVRHX xoaddr:$src), VRRC))>;
+ def : Pat <(v1i128 (PPClxvrzx xoaddr:$src, 32)),
+ (v1i128 (COPY_TO_REGCLASS (LXVRWX xoaddr:$src), VRRC))>;
+ def : Pat <(v1i128 (PPClxvrzx xoaddr:$src, 64)),
+ (v1i128 (COPY_TO_REGCLASS (LXVRDX xoaddr:$src), VRRC))>;
}
Index: llvm/lib/Target/PowerPC/PPCISelLowering.h
===================================================================
--- llvm/lib/Target/PowerPC/PPCISelLowering.h
+++ llvm/lib/Target/PowerPC/PPCISelLowering.h
@@ -519,7 +519,13 @@
/// GPRC = TOC_ENTRY GA, TOC
/// Loads the entry for GA from the TOC, where the TOC base is given by
/// the last operand.
- TOC_ENTRY
+ TOC_ENTRY,
+
+ /// LXVRZX - Load VSX Vector Rightmost and Zero Extend
+ /// This node represents v1i128 BUILD_VECTOR of a zero extending load
+ /// instruction from <byte, halfword, word, or doubleword> to i128.
+ /// Allows utilization of the Load VSX Vector Rightmost Instructions.
+ LXVRZX
};
} // end namespace PPCISD
Index: llvm/lib/Target/PowerPC/PPCISelLowering.cpp
===================================================================
--- llvm/lib/Target/PowerPC/PPCISelLowering.cpp
+++ llvm/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -1573,6 +1573,7 @@
case PPCISD::MAT_PCREL_ADDR: return "PPCISD::MAT_PCREL_ADDR";
case PPCISD::LD_SPLAT: return "PPCISD::LD_SPLAT";
case PPCISD::FNMSUB: return "PPCISD::FNMSUB";
+ case PPCISD::LXVRZX: return "PPCISD::LXVRZX";
}
return nullptr;
}
@@ -13760,6 +13761,51 @@
return SDValue();
}
+static SDValue combineBVZEXTLOAD(SDNode *N, SelectionDAG &DAG) {
+ SDLoc dl(N);
+
+ // Look for the pattern of a load from a narrow width to i128, feeding
+ // into a BUILD_VECTOR of v1i128. Replace this sequence with a PPCISD node
+ // (LXVRZX). This node represents a zero extending load that will be matched
+ // to the Load VSX Vector Rightmost instructions.
+
+ // This combine is only eligible for a BUILD_VECTOR of v1i128.
+ // Other return types are not valid for the LXVRZX replacement.
+ if (N->getValueType(0) != MVT::v1i128) {
+ return SDValue();
+ }
+
+ SDValue Operand = N->getOperand(0);
+ // Proceed with the transformation if the operand to the BUILD_VECTOR
+ // is a load instruction.
+ if (Operand.getOpcode() != ISD::LOAD) {
+ return SDValue();
+ }
+
+ LoadSDNode *LD = dyn_cast<LoadSDNode>(Operand);
+ EVT MemoryType = LD->getMemoryVT();
+
+ // This transformation is only valid if the we are loading either a byte,
+ // halfword, word, or doubleword.
+ bool ValidLDType = MemoryType == MVT::i8 || MemoryType == MVT::i16 ||
+ MemoryType == MVT::i32 || MemoryType == MVT::i64;
+
+ // Ensure that the load from the narrow width is being zero extended to i128.
+ if ((LD->getValueType(0) != MVT::i128) ||
+ (LD->getExtensionType() != ISD::ZEXTLOAD) || (!ValidLDType))
+ return SDValue();
+
+ // The width of the narrow type becomes an operand of the LXVRZX node
+ // we are creating in order to pattern match to the appropriate instruction
+ // in the backend.
+ SDValue Width = DAG.getIntPtrConstant(MemoryType.getScalarSizeInBits(), dl);
+ SDValue LoadOps[] = {LD->getChain(), LD->getBasePtr(), Width};
+
+ return DAG.getMemIntrinsicNode(PPCISD::LXVRZX, dl,
+ DAG.getVTList(MVT::v1i128, MVT::Other),
+ LoadOps, MemoryType, LD->getMemOperand());
+}
+
SDValue PPCTargetLowering::DAGCombineBuildVector(SDNode *N,
DAGCombinerInfo &DCI) const {
assert(N->getOpcode() == ISD::BUILD_VECTOR &&
@@ -13797,6 +13843,14 @@
return Reduced;
}
+ // On Power10, the Load VSX Vector Rightmost instructions can be utilized
+ // if this is a BUILD_VECTOR of v1i128, and if the operand to the BUILD_VECTOR
+ // is a load from <valid narrow width> to i128.
+ if (Subtarget.isISA3_1()) {
+ SDValue BVOfZLoad = combineBVZEXTLOAD(N, DAG);
+ if (BVOfZLoad)
+ return BVOfZLoad;
+ }
if (N->getValueType(0) != MVT::v2f64)
return SDValue();
Index: clang/test/CodeGen/builtins-ppc-p10vector.c
===================================================================
--- clang/test/CodeGen/builtins-ppc-p10vector.c
+++ clang/test/CodeGen/builtins-ppc-p10vector.c
@@ -9,8 +9,16 @@
vector unsigned char vuca;
vector unsigned short vusa;
vector unsigned int vuia;
+signed char *ca;
+unsigned char *uca;
+
vector unsigned long long vulla, vullb;
-unsigned int uia;
+unsigned int uia, *uiap;
+signed int *ia;
+signed short *sia;
+unsigned short *usia;
+signed long long *llia, llib;
+unsigned long long *ullia;
vector unsigned long long test_vpdepd(void) {
// CHECK: @llvm.ppc.altivec.vpdepd(<2 x i64>
@@ -79,3 +87,59 @@
// CHECK-LE-NEXT: ret <16 x i8>
return vec_clrr(vuca, uia);
}
+
+vector signed __int128 test_vec_xl_sext_i8(void) {
+ // CHECK: load i8
+ // CHECK: sext i8
+ // CHECK: ret <1 x i128>
+ return vec_xl_sext(llib, ca);
+}
+
+vector signed __int128 test_vec_xl_sext_i16(void) {
+ // CHECK: load i16
+ // CHECK: sext i16
+ // CHECK: ret <1 x i128>
+ return vec_xl_sext(llib, sia);
+}
+
+vector signed __int128 test_vec_xl_sext_i32(void) {
+ // CHECK: load i32
+ // CHECK: sext i32
+ // CHECK: ret <1 x i128>
+ return vec_xl_sext(llib, ia);
+}
+
+vector signed __int128 test_vec_xl_sext_i64(void) {
+ // CHECK: load i64
+ // CHECK: sext i64
+ // CHECK: ret <1 x i128>
+ return vec_xl_sext(llib, llia);
+}
+
+vector unsigned __int128 test_vec_xl_zext_i8(void) {
+ // CHECK: load i8
+ // CHECK: zext i8
+ // CHECK: ret <1 x i128>
+ return vec_xl_zext(llib, uca);
+}
+
+vector unsigned __int128 test_vec_xl_zext_i16(void) {
+ // CHECK: load i16
+ // CHECK: zext i16
+ // CHECK: ret <1 x i128>
+ return vec_xl_zext(llib, usia);
+}
+
+vector unsigned __int128 test_vec_xl_zext_i32(void) {
+ // CHECK: load i32
+ // CHECK: zext i32
+ // CHECK: ret <1 x i128>
+ return vec_xl_zext(llib, uiap);
+}
+
+vector unsigned __int128 test_vec_xl_zext_i64(void) {
+ // CHECK: load i64
+ // CHECK: zext i64
+ // CHECK: ret <1 x i128>
+ return vec_xl_zext(llib, ullia);
+}
Index: clang/lib/Headers/altivec.h
===================================================================
--- clang/lib/Headers/altivec.h
+++ clang/lib/Headers/altivec.h
@@ -16509,6 +16509,52 @@
#define vec_xl_be vec_xl
#endif
+#if defined(__POWER10_VECTOR) && defined(__VSX)
+
+/* vect_xl_sext */
+static __inline__ vector signed __int128 __ATTRS_o_ai
+vec_xl_sext(signed long long __offset, signed char *__pointer) {
+ return (vector signed __int128)*(__pointer + __offset);
+}
+
+static __inline__ vector signed __int128 __ATTRS_o_ai
+vec_xl_sext(signed long long __offset, signed short *__pointer) {
+ return (vector signed __int128)*(__pointer + __offset);
+}
+
+static __inline__ vector signed __int128 __ATTRS_o_ai
+vec_xl_sext(signed long long __offset, signed int *__pointer) {
+ return (vector signed __int128)*(__pointer + __offset);
+}
+
+static __inline__ vector signed __int128 __ATTRS_o_ai
+vec_xl_sext(signed long long __offset, signed long long *__pointer) {
+ return (vector signed __int128)*(__pointer + __offset);
+}
+
+/* vec_xl_zext */
+static __inline__ vector unsigned __int128 __ATTRS_o_ai
+vec_xl_zext(signed long long __offset, unsigned char *__pointer) {
+ return (vector unsigned __int128)*(__pointer + __offset);
+}
+
+static __inline__ vector unsigned __int128 __ATTRS_o_ai
+vec_xl_zext(signed long long __offset, unsigned short *__pointer) {
+ return (vector unsigned __int128)*(__pointer + __offset);
+}
+
+static __inline__ vector unsigned __int128 __ATTRS_o_ai
+vec_xl_zext(signed long long __offset, unsigned int *__pointer) {
+ return (vector unsigned __int128)*(__pointer + __offset);
+}
+
+static __inline__ vector unsigned __int128 __ATTRS_o_ai
+vec_xl_zext(signed long long __offset, unsigned long long *__pointer) {
+ return (vector unsigned __int128)*(__pointer + __offset);
+}
+
+#endif
+
/* vec_xst */
static inline __ATTRS_o_ai void vec_xst(vector signed char __vec,
@@ -16830,6 +16876,7 @@
return __builtin_altivec_vclrrb((vector signed char)__a, __n);
#endif
}
+
#endif /* __POWER10_VECTOR__ */
#undef __ATTRS_o_ai
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
