llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT--> @llvm/pr-subscribers-clangir Author: Vishruth Thimmaiah (vishruth-thimmaiah) <details> <summary>Changes</summary> Adds support for the `__builtin_ia32_vpcom` and `__builtin_ia32_vpcomu` X86 builtins. Part of #<!-- -->167765 --- Full diff: https://github.com/llvm/llvm-project/pull/170362.diff 2 Files Affected: - (modified) clang/lib/CIR/CodeGen/CIRGenBuiltinX86.cpp (+62) - (added) clang/test/CIR/CodeGenBuiltins/X86/xop-builtins.c (+217) ``````````diff diff --git a/clang/lib/CIR/CodeGen/CIRGenBuiltinX86.cpp b/clang/lib/CIR/CodeGen/CIRGenBuiltinX86.cpp index 224a182ed17d1..7f8cd2768a10f 100644 --- a/clang/lib/CIR/CodeGen/CIRGenBuiltinX86.cpp +++ b/clang/lib/CIR/CodeGen/CIRGenBuiltinX86.cpp @@ -168,6 +168,62 @@ static mlir::Value emitVecInsert(CIRGenBuilderTy &builder, mlir::Location loc, return cir::VecInsertOp::create(builder, loc, vec, value, indexVal); } +static mlir::Value emitX86vpcom(CIRGenBuilderTy &builder, mlir::Location loc, + llvm::SmallVector<mlir::Value> ops, + bool isSigned) { + mlir::Value op0 = ops[0]; + mlir::Value op1 = ops[1]; + + cir::VectorType ty = cast<cir::VectorType>(op0.getType()); + mlir::Type elementTy = ty.getElementType(); + + uint64_t imm = + ops[2].getDefiningOp<cir::ConstantOp>().getIntValue().getZExtValue() & + 0x7; + + cir::CmpOpKind pred; + switch (imm) { + case 0x0: + pred = cir::CmpOpKind::lt; + break; + case 0x1: + pred = cir::CmpOpKind::le; + break; + case 0x2: + pred = cir::CmpOpKind::gt; + break; + case 0x3: + pred = cir::CmpOpKind::ge; + break; + case 0x4: + pred = cir::CmpOpKind::eq; + break; + case 0x5: + pred = cir::CmpOpKind::ne; + break; + case 0x6: + return builder.getNullValue(ty, loc); // FALSE + case 0x7: { + llvm::APInt allOnes = + llvm::APInt::getAllOnes(cast<cir::IntType>(elementTy).getWidth()); + return cir::VecSplatOp::create( + builder, loc, ty, + builder.getConstAPInt(loc, elementTy, allOnes)); // TRUE + } + default: + llvm_unreachable("Unexpected XOP vpcom/vpcomu predicate"); + } + + if (!isSigned) { + elementTy = builder.getUIntNTy(cast<cir::IntType>(elementTy).getWidth()); + ty = cir::VectorType::get(elementTy, ty.getSize()); + op0 = builder.createBitcast(op0, ty); + op1 = builder.createBitcast(op1, ty); + } + + return builder.createVecCompare(loc, pred, op0, op1); +} + mlir::Value CIRGenFunction::emitX86BuiltinExpr(unsigned builtinID, const CallExpr *expr) { if (builtinID == Builtin::BI__builtin_cpu_is) { @@ -900,14 +956,20 @@ mlir::Value CIRGenFunction::emitX86BuiltinExpr(unsigned builtinID, case X86::BI__builtin_ia32_ucmpq128_mask: case X86::BI__builtin_ia32_ucmpq256_mask: case X86::BI__builtin_ia32_ucmpq512_mask: + cgm.errorNYI(expr->getSourceRange(), + std::string("unimplemented X86 builtin call: ") + + getContext().BuiltinInfo.getName(builtinID)); + return {}; case X86::BI__builtin_ia32_vpcomb: case X86::BI__builtin_ia32_vpcomw: case X86::BI__builtin_ia32_vpcomd: case X86::BI__builtin_ia32_vpcomq: + return emitX86vpcom(builder, getLoc(expr->getExprLoc()), ops, true); case X86::BI__builtin_ia32_vpcomub: case X86::BI__builtin_ia32_vpcomuw: case X86::BI__builtin_ia32_vpcomud: case X86::BI__builtin_ia32_vpcomuq: + return emitX86vpcom(builder, getLoc(expr->getExprLoc()), ops, false); case X86::BI__builtin_ia32_kortestcqi: case X86::BI__builtin_ia32_kortestchi: case X86::BI__builtin_ia32_kortestcsi: diff --git a/clang/test/CIR/CodeGenBuiltins/X86/xop-builtins.c b/clang/test/CIR/CodeGenBuiltins/X86/xop-builtins.c new file mode 100644 index 0000000000000..10ef735fbbdb7 --- /dev/null +++ b/clang/test/CIR/CodeGenBuiltins/X86/xop-builtins.c @@ -0,0 +1,217 @@ +// RUN: %clang_cc1 -x c -flax-vector-conversions=none -ffreestanding %s -triple=x86_64-unknown-linux -target-feature +xop -fclangir -emit-cir -o %t.cir -Wall -Werror +// RUN: FileCheck --check-prefix=CIR --input-file=%t.cir %s +// RUN: %clang_cc1 -x c -flax-vector-conversions=none -ffreestanding %s -triple=x86_64-unknown-linux -target-feature +xop -fclangir -emit-llvm -o %t.ll -Wall -Werror +// RUN: FileCheck --check-prefix=LLVM --input-file=%t.ll %s + +// RUN: %clang_cc1 -x c++ -flax-vector-conversions=none -ffreestanding %s -triple=x86_64-unknown-linux -target-feature +xop -fclangir -emit-cir -o %t.cir -Wall -Werror +// RUN: FileCheck --check-prefix=CIR --input-file=%t.cir %s +// RUN: %clang_cc1 -x c++ -flax-vector-conversions=none -ffreestanding %s -triple=x86_64-unknown-linux -target-feature +xop -fclangir -emit-llvm -o %t.ll -Wall -Werror +// RUN: FileCheck --check-prefix=LLVM --input-file=%t.ll %s + +// RUN: %clang_cc1 -x c -flax-vector-conversions=none -ffreestanding %s -triple=x86_64-unknown-linux -target-feature +xop -emit-llvm -o - -Wall -Werror | FileCheck %s -check-prefix=OGCG +// RUN: %clang_cc1 -x c++ -flax-vector-conversions=none -ffreestanding %s -triple=x86_64-unknown-linux -target-feature +xop -emit-llvm -o - -Wall -Werror | FileCheck %s -check-prefix=OGCG + +// This test mimics clang/test/CodeGen/X86/xop-builtins.c, which eventually +// CIR shall be able to support fully. + +#include <x86intrin.h> + +__m128i test_mm_com_epu8(__m128i a, __m128i b) { + // CIR-LABEL: test_mm_com_epu8 + // CIR: %[[A:.*]] = cir.cast bitcast %{{.*}} : !cir.vector<16 x !s8i> -> !cir.vector<16 x !u8i> + // CIR: %[[B:.*]] = cir.cast bitcast %{{.*}} : !cir.vector<16 x !s8i> -> !cir.vector<16 x !u8i> + // CIR: %[[CMP:.*]] = cir.vec.cmp(lt, %[[A]], %[[B]]) : !cir.vector<16 x !u8i>, !cir.vector<16 x !s8i> + // CIR: %[[RES:.*]] = cir.cast bitcast %[[CMP]] : !cir.vector<16 x !s8i> -> !cir.vector<2 x !s64i> + + // LLVM-LABEL: test_mm_com_epu8 + // LLVM: %[[CMP:.*]] = icmp ult <16 x i8> %{{.*}}, %{{.*}} + // LLVM: %[[RES:.*]] = sext <16 x i1> %[[CMP]] to <16 x i8> + // LLVM: %{{.*}} = bitcast <16 x i8> %[[RES]] to <2 x i64> + + // OGCG-LABEL: test_mm_com_epu8 + // OGCG: %[[CMP:.*]] = icmp ult <16 x i8> %{{.*}}, %{{.*}} + // OGCG: %[[RES:.*]] = sext <16 x i1> %[[CMP]] to <16 x i8> + // OGCG: %{{.*}} = bitcast <16 x i8> %[[RES]] to <2 x i64> + return _mm_com_epu8(a, b, 0); +} + +__m128i test_mm_com_epu16(__m128i a, __m128i b) { + // CIR-LABEL: test_mm_com_epu16 + // CIR: %[[VAL1:.*]] = cir.cast bitcast %{{.*}} : !cir.vector<8 x !s16i> -> !cir.vector<8 x !u16i> + // CIR: %[[VAL2:.*]] = cir.cast bitcast %{{.*}} : !cir.vector<8 x !s16i> -> !cir.vector<8 x !u16i> + // CIR: %[[CMP:.*]] = cir.vec.cmp(lt, %[[VAL1]], %[[VAL2]]) : !cir.vector<8 x !u16i>, !cir.vector<8 x !s16i> + // CIR: %[[RES:.*]] = cir.cast bitcast %[[CMP]] : !cir.vector<8 x !s16i> -> !cir.vector<2 x !s64i> + + // LLVM-LABEL: test_mm_com_epu16 + // LLVM: %[[CMP:.*]] = icmp ult <8 x i16> %{{.*}}, %{{.*}} + // LLVM: %[[RES:.*]] = sext <8 x i1> %[[CMP]] to <8 x i16> + // LLVM: %{{.*}} = bitcast <8 x i16> %[[RES]] to <2 x i64> + + // OGCG-LABEL: test_mm_com_epu16 + // OGCG: %[[CMP:.*]] = icmp ult <8 x i16> %{{.*}}, %{{.*}} + // OGCG: %[[RES:.*]] = sext <8 x i1> %[[CMP]] to <8 x i16> + // OGCG: %{{.*}} = bitcast <8 x i16> %[[RES]] to <2 x i64> + return _mm_com_epu16(a, b, 0); +} + +__m128i test_mm_com_epu32(__m128i a, __m128i b) { + // CIR-LABEL: test_mm_com_epu32 + // CIR: %[[VAL1:.*]] = cir.cast bitcast %{{.*}} : !cir.vector<4 x !s32i> -> !cir.vector<4 x !u32i> + // CIR: %[[VAL2:.*]] = cir.cast bitcast %{{.*}} : !cir.vector<4 x !s32i> -> !cir.vector<4 x !u32i> + // CIR: %[[CMP:.*]] = cir.vec.cmp(lt, %[[VAL1]], %[[VAL2]]) : !cir.vector<4 x !u32i>, !cir.vector<4 x !s32i> + // CIR: %[[RES:.*]] = cir.cast bitcast %[[CMP]] : !cir.vector<4 x !s32i> -> !cir.vector<2 x !s64i> + + // LLVM-LABEL: test_mm_com_epu32 + // LLVM: %[[CMP:.*]] = icmp ult <4 x i32> %{{.*}}, %{{.*}} + // LLVM: %[[RES:.*]] = sext <4 x i1> %[[CMP]] to <4 x i32> + // LLVM: %{{.*}} = bitcast <4 x i32> %[[RES]] to <2 x i64> + + // OGCG-LABEL: test_mm_com_epu32 + // OGCG: %[[CMP:.*]] = icmp ult <4 x i32> %{{.*}}, %{{.*}} + // OGCG: %[[RES:.*]] = sext <4 x i1> %[[CMP]] to <4 x i32> + // OGCG: %{{.*}} = bitcast <4 x i32> %[[RES]] to <2 x i64> + return _mm_com_epu32(a, b, 0); +} + +__m128i test_mm_com_epu64(__m128i a, __m128i b) { + // CIR-LABEL: test_mm_com_epu64 + // CIR: %[[VAL1:.*]] = cir.cast bitcast %{{.*}} : !cir.vector<2 x !s64i> -> !cir.vector<2 x !u64i> + // CIR: %[[VAL2:.*]] = cir.cast bitcast %{{.*}} : !cir.vector<2 x !s64i> -> !cir.vector<2 x !u64i> + // CIR: %[[CMP:.*]] = cir.vec.cmp(lt, %[[VAL1]], %[[VAL2]]) : !cir.vector<2 x !u64i>, !cir.vector<2 x !s64i> + + // LLVM-LABEL: test_mm_com_epu64 + // LLVM: %[[CMP:.*]] = icmp ult <2 x i64> %{{.*}}, %{{.*}} + // LLVM: %[[RES:.*]] = sext <2 x i1> %[[CMP]] to <2 x i64> + + // OGCG-LABEL: test_mm_com_epu64 + // OGCG: %[[CMP:.*]] = icmp ult <2 x i64> %{{.*}}, %{{.*}} + // OGCG: %[[RES:.*]] = sext <2 x i1> %[[CMP]] to <2 x i64> + return _mm_com_epu64(a, b, 0); +} + +__m128i test_mm_com_epi8(__m128i a, __m128i b) { + // CIR-LABEL: test_mm_com_epi8 + // CIR: %[[CMP:.*]] = cir.vec.cmp(lt, %{{.*}}, %{{.*}}) : !cir.vector<16 x !s8i>, !cir.vector<16 x !s8i> + // CIR: %[[RES:.*]] = cir.cast bitcast %[[CMP]] : !cir.vector<16 x !s8i> -> !cir.vector<2 x !s64i> + + // LLVM-LABEL: test_mm_com_epi8 + // LLVM: %[[CMP:.*]] = icmp slt <16 x i8> %{{.*}}, %{{.*}} + // LLVM: %[[RES:.*]] = sext <16 x i1> %[[CMP]] to <16 x i8> + // LLVM: %{{.*}} = bitcast <16 x i8> %[[RES]] to <2 x i64> + + // OGCG-LABEL: test_mm_com_epi8 + // OGCG: %[[CMP:.*]] = icmp slt <16 x i8> %{{.*}}, %{{.*}} + // OGCG: %[[RES:.*]] = sext <16 x i1> %[[CMP]] to <16 x i8> + // OGCG: %{{.*}} = bitcast <16 x i8> %[[RES]] to <2 x i64> + return _mm_com_epi8(a, b, 0); +} + +__m128i test_mm_com_epi16(__m128i a, __m128i b) { + // CIR-LABEL: test_mm_com_epi16 + // CIR: %[[CMP:.*]] = cir.vec.cmp(lt, %{{.*}}, %{{.*}}) : !cir.vector<8 x !s16i>, !cir.vector<8 x !s16i> + // CIR: %[[RES:.*]] = cir.cast bitcast %[[CMP]] : !cir.vector<8 x !s16i> -> !cir.vector<2 x !s64i> + + // LLVM-LABEL: test_mm_com_epi16 + // LLVM: %[[CMP:.*]] = icmp slt <8 x i16> %{{.*}}, %{{.*}} + // LLVM: %[[RES:.*]] = sext <8 x i1> %[[CMP]] to <8 x i16> + // LLVM: %{{.*}} = bitcast <8 x i16> %[[RES]] to <2 x i64> + + // OGCG-LABEL: test_mm_com_epi16 + // OGCG: %[[CMP:.*]] = icmp slt <8 x i16> %{{.*}}, %{{.*}} + // OGCG: %[[RES:.*]] = sext <8 x i1> %[[CMP]] to <8 x i16> + // OGCG: %{{.*}} = bitcast <8 x i16> %[[RES]] to <2 x i64> + return _mm_com_epi16(a, b, 0); +} + +__m128i test_mm_com_epi32(__m128i a, __m128i b) { + // CIR-LABEL: test_mm_com_epi32 + // CIR: %[[CMP:.*]] = cir.vec.cmp(lt, %{{.*}}, %{{.*}}) : !cir.vector<4 x !s32i>, !cir.vector<4 x !s32i> + // CIR: %[[RES:.*]] = cir.cast bitcast %[[CMP]] : !cir.vector<4 x !s32i> -> !cir.vector<2 x !s64i> + + // LLVM-LABEL: test_mm_com_epi32 + // LLVM: %[[CMP:.*]] = icmp slt <4 x i32> %{{.*}}, %{{.*}} + // LLVM: %[[RES:.*]] = sext <4 x i1> %[[CMP]] to <4 x i32> + // LLVM: %{{.*}} = bitcast <4 x i32> %[[RES]] to <2 x i64> + + // OGCG-LABEL: test_mm_com_epi32 + // OGCG: %[[CMP:.*]] = icmp slt <4 x i32> %{{.*}}, %{{.*}} + // OGCG: %[[RES:.*]] = sext <4 x i1> %[[CMP]] to <4 x i32> + // OGCG: %{{.*}} = bitcast <4 x i32> %[[RES]] to <2 x i64> + return _mm_com_epi32(a, b, 0); +} + +__m128i test_mm_com_epi64(__m128i a, __m128i b) { + // CIR-LABEL: test_mm_com_epi64 + // CIR: %[[CMP:.*]] = cir.vec.cmp(lt, %{{.*}}, %{{.*}}) : !cir.vector<2 x !s64i>, !cir.vector<2 x !s64i> + + // LLVM-LABEL: test_mm_com_epi64 + // LLVM: %[[CMP:.*]] = icmp slt <2 x i64> %{{.*}}, %{{.*}} + // LLVM: %[[RES:.*]] = sext <2 x i1> %[[CMP]] to <2 x i64> + + // OGCG-LABEL: test_mm_com_epi64 + // OGCG: %[[CMP:.*]] = icmp slt <2 x i64> %{{.*}}, %{{.*}} + // OGCG: %[[RES:.*]] = sext <2 x i1> %[[CMP]] to <2 x i64> + return _mm_com_epi64(a, b, 0); +} + +__m128i test_mm_com_epi32_false(__m128i a, __m128i b) { + // CIR-LABEL: test_mm_com_epi32_false + // CIR: %[[ZERO:.*]] = cir.const #cir.zero : !cir.vector<4 x !s32i> + // CIR: %{{.*}} = cir.cast bitcast %[[ZERO]] : !cir.vector<4 x !s32i> -> !cir.vector<2 x !s64i> + + // LLVM-LABEL: test_mm_com_epi32_false + // LLVM: store <2 x i64> zeroinitializer, ptr %[[A:.*]], align 16 + // LLVM: %[[ZERO:.*]] = load <2 x i64>, ptr %[[A]], align 16 + // LLVM: ret <2 x i64> %[[ZERO]] + + // OGCG-LABEL: test_mm_com_epi32_false + // OGCG: ret <2 x i64> zeroinitializer + return _mm_com_epi32(a, b, 6); +} + +__m128i test_mm_com_epu32_false(__m128i a, __m128i b) { + // CIR-LABEL: test_mm_com_epu32_false + // CIR: %[[ZERO:.*]] = cir.const #cir.zero : !cir.vector<4 x !s32i> + // CIR: %{{.*}} = cir.cast bitcast %[[ZERO]] : !cir.vector<4 x !s32i> -> !cir.vector<2 x !s64i> + + // LLVM-LABEL: test_mm_com_epu32_false + // LLVM: store <2 x i64> zeroinitializer, ptr %[[A:.*]], align 16 + // LLVM: %[[ZERO:.*]] = load <2 x i64>, ptr %[[A]], align 16 + // LLVM: ret <2 x i64> %[[ZERO]] + + // OGCG-LABEL: test_mm_com_epu32_false + // OGCG: ret <2 x i64> zeroinitializer + return _mm_com_epu32(a, b, 6); +} + +__m128i test_mm_com_epi32_true(__m128i a, __m128i b) { + // CIR-LABEL: test_mm_com_epi32_true + // CIR: %[[VAL:.*]] = cir.const #cir.int<-1> : !s32i + // CIR: %[[SPLAT:.*]] = cir.vec.splat %[[VAL]] : !s32i, !cir.vector<4 x !s32i> + // CIR: %{{.*}} = cir.cast bitcast %[[SPLAT]] : !cir.vector<4 x !s32i> -> !cir.vector<2 x !s64i> + + // LLVM-LABEL: test_mm_com_epi32_true + // LLVM: store <2 x i64> splat (i64 -1), ptr %[[VAL:.*]], align 16 + // LLVM: %[[SPLAT:.*]] = load <2 x i64>, ptr %[[VAL]], align 16 + // LLVM: ret <2 x i64> %[[SPLAT]] + + // OGCG-LABEL: test_mm_com_epi32_true + // OGCG: ret <2 x i64> splat (i64 -1) + return _mm_com_epi32(a, b, 7); +} + +__m128i test_mm_com_epu32_true(__m128i a, __m128i b) { + // CIR-LABEL: test_mm_com_epu32_true + // CIR: %[[VAL:.*]] = cir.const #cir.int<-1> : !s32i + // CIR: %[[SPLAT:.*]] = cir.vec.splat %[[VAL]] : !s32i, !cir.vector<4 x !s32i> + // CIR: %{{.*}} = cir.cast bitcast %[[SPLAT]] : !cir.vector<4 x !s32i> -> !cir.vector<2 x !s64i> + + // LLVM-LABEL: test_mm_com_epu32_true + // LLVM: store <2 x i64> splat (i64 -1), ptr %[[VAL:.*]], align 16 + // LLVM: %[[SPLAT:.*]] = load <2 x i64>, ptr %[[VAL]], align 16 + // LLVM: ret <2 x i64> %[[SPLAT]] + + // OGCG-LABEL: test_mm_com_epu32_true + // OGCG: ret <2 x i64> splat (i64 -1) + return _mm_com_epu32(a, b, 7); +} `````````` </details> https://github.com/llvm/llvm-project/pull/170362 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
