diff --git a/llvm/lib/CodeGen/ReplaceWithVeclib.cpp b/llvm/lib/CodeGen/ReplaceWithVeclib.cpp index 893aa4a91828d..56025aa5c45fb 100644 --- a/llvm/lib/CodeGen/ReplaceWithVeclib.cpp +++ b/llvm/lib/CodeGen/ReplaceWithVeclib.cpp @@ -6,9 +6,9 @@ // //===----------------------------------------------------------------------===// // -// Replaces calls to LLVM vector intrinsics (i.e., calls to LLVM intrinsics -// with vector operands) with matching calls to functions from a vector -// library (e.g., libmvec, SVML) according to TargetLibraryInfo. +// Replaces LLVM IR instructions with vector operands (i.e., the frem +// instruction or calls to LLVM intrinsics) with matching calls to functions +// from a vector library (e.g libmvec, SVML) using TargetLibraryInfo interface. // //===----------------------------------------------------------------------===// @@ -69,88 +69,98 @@ Function *getTLIFunction(Module *M, FunctionType *VectorFTy, return TLIFunc; } -/// Replace the call to the vector intrinsic ( \p CalltoReplace ) with a call to -/// the corresponding function from the vector library ( \p TLIVecFunc ). -static void replaceWithTLIFunction(CallInst &CalltoReplace, VFInfo &Info, +/// Replace the instruction \p I with a call to the corresponding function from +/// the vector library (\p TLIVecFunc). +static void replaceWithTLIFunction(Instruction &I, VFInfo &Info, Function *TLIVecFunc) { - IRBuilder<> IRBuilder(&CalltoReplace); - SmallVector Args(CalltoReplace.args()); + IRBuilder<> IRBuilder(&I); + auto *CI = dyn_cast(&I); + SmallVector Args(CI ? CI->args() : I.operands()); if (auto OptMaskpos = Info.getParamIndexForOptionalMask()) { - auto *MaskTy = VectorType::get(Type::getInt1Ty(CalltoReplace.getContext()), - Info.Shape.VF); + auto *MaskTy = + VectorType::get(Type::getInt1Ty(I.getContext()), Info.Shape.VF); Args.insert(Args.begin() + OptMaskpos.value(), Constant::getAllOnesValue(MaskTy)); } - // Preserve the operand bundles. + // If it is a call instruction, preserve the operand bundles. SmallVector OpBundles; - CalltoReplace.getOperandBundlesAsDefs(OpBundles); - CallInst *Replacement = IRBuilder.CreateCall(TLIVecFunc, Args, OpBundles); - CalltoReplace.replaceAllUsesWith(Replacement); + if (CI) + CI->getOperandBundlesAsDefs(OpBundles); + + auto *Replacement = IRBuilder.CreateCall(TLIVecFunc, Args, OpBundles); + I.replaceAllUsesWith(Replacement); // Preserve fast math flags for FP math. if (isa(Replacement)) - Replacement->copyFastMathFlags(&CalltoReplace); + Replacement->copyFastMathFlags(&I); } -/// Returns true when successfully replaced \p CallToReplace with a suitable -/// function taking vector arguments, based on available mappings in the \p TLI. -/// Currently only works when \p CallToReplace is a call to vectorized -/// intrinsic. +/// Returns true when successfully replaced \p I with a suitable function taking +/// vector arguments, based on available mappings in the \p TLI. Currently only +/// works when \p I is a call to vectorized intrinsic or the frem instruction. static bool replaceWithCallToVeclib(const TargetLibraryInfo &TLI, - CallInst &CallToReplace) { - if (!CallToReplace.getCalledFunction()) - return false; + Instruction &I) { + // At the moment VFABI assumes the return type is always widened unless it is + // a void type. + auto *VTy = dyn_cast(I.getType()); + ElementCount EC(VTy ? VTy->getElementCount() : ElementCount::getFixed(0)); - auto IntrinsicID = CallToReplace.getCalledFunction()->getIntrinsicID(); - // Replacement is only performed for intrinsic functions. - if (IntrinsicID == Intrinsic::not_intrinsic) - return false; - - // Compute arguments types of the corresponding scalar call. Additionally - // checks if in the vector call, all vector operands have the same EC. - ElementCount VF = ElementCount::getFixed(0); - SmallVector ScalarArgTypes; - for (auto Arg : enumerate(CallToReplace.args())) { - auto *ArgTy = Arg.value()->getType(); - if (isVectorIntrinsicWithScalarOpAtArg(IntrinsicID, Arg.index())) { - ScalarArgTypes.push_back(ArgTy); - } else if (auto *VectorArgTy = dyn_cast(ArgTy)) { - ScalarArgTypes.push_back(ArgTy->getScalarType()); - // Disallow vector arguments with different VFs. When processing the first - // vector argument, store it's VF, and for the rest ensure that they match - // it. - if (VF.isZero()) - VF = VectorArgTy->getElementCount(); - else if (VF != VectorArgTy->getElementCount()) + // Compute the argument types of the corresponding scalar call and the scalar + // function name. For calls, it additionally finds the function to replace + // and checks that all vector operands match the previously found EC. + SmallVector ScalarArgTypes; + std::string ScalarName; + Function *FuncToReplace = nullptr; + if (auto *CI = dyn_cast(&I)) { + FuncToReplace = CI->getCalledFunction(); + Intrinsic::ID IID = FuncToReplace->getIntrinsicID(); + assert(IID != Intrinsic::not_intrinsic && "Not an intrinsic"); + for (auto Arg : enumerate(CI->args())) { + auto *ArgTy = Arg.value()->getType(); + if (isVectorIntrinsicWithScalarOpAtArg(IID, Arg.index())) { + ScalarArgTypes.push_back(ArgTy); + } else if (auto *VectorArgTy = dyn_cast(ArgTy)) { + ScalarArgTypes.push_back(VectorArgTy->getElementType()); + // When return type is void, set EC to the first vector argument, and + // disallow vector arguments with different ECs. + if (EC.isZero()) + EC = VectorArgTy->getElementCount(); + else if (EC != VectorArgTy->getElementCount()) + return false; + } else + // Exit when it is supposed to be a vector argument but it isn't. return false; - } else - // Exit when it is supposed to be a vector argument but it isn't. + } + // Try to reconstruct the name for the scalar version of the instruction, + // using scalar argument types. + ScalarName = Intrinsic::isOverloaded(IID) + ? Intrinsic::getName(IID, ScalarArgTypes, I.getModule()) + : Intrinsic::getName(IID).str(); + } else { + assert(VTy && "Return type must be a vector"); + auto *ScalarTy = VTy->getScalarType(); + LibFunc Func; + if (!TLI.getLibFunc(I.getOpcode(), ScalarTy, Func)) return false; + ScalarName = TLI.getName(Func); + ScalarArgTypes = {ScalarTy, ScalarTy}; } - // Try to reconstruct the name for the scalar version of this intrinsic using - // the intrinsic ID and the argument types converted to scalar above. - std::string ScalarName = - (Intrinsic::isOverloaded(IntrinsicID) - ? Intrinsic::getName(IntrinsicID, ScalarArgTypes, - CallToReplace.getModule()) - : Intrinsic::getName(IntrinsicID).str()); - // Try to find the mapping for the scalar version of this intrinsic and the // exact vector width of the call operands in the TargetLibraryInfo. First, // check with a non-masked variant, and if that fails try with a masked one. const VecDesc *VD = - TLI.getVectorMappingInfo(ScalarName, VF, /*Masked*/ false); - if (!VD && !(VD = TLI.getVectorMappingInfo(ScalarName, VF, /*Masked*/ true))) + TLI.getVectorMappingInfo(ScalarName, EC, /*Masked*/ false); + if (!VD && !(VD = TLI.getVectorMappingInfo(ScalarName, EC, /*Masked*/ true))) return false; LLVM_DEBUG(dbgs() << DEBUG_TYPE << ": Found TLI mapping from: `" << ScalarName - << "` and vector width " << VF << " to: `" + << "` and vector width " << EC << " to: `" << VD->getVectorFnName() << "`.\n"); // Replace the call to the intrinsic with a call to the vector library // function. - Type *ScalarRetTy = CallToReplace.getType()->getScalarType(); + Type *ScalarRetTy = I.getType()->getScalarType(); FunctionType *ScalarFTy = FunctionType::get(ScalarRetTy, ScalarArgTypes, /*isVarArg*/ false); const std::string MangledName = VD->getVectorFunctionABIVariantString(); @@ -162,27 +172,37 @@ static bool replaceWithCallToVeclib(const TargetLibraryInfo &TLI, if (!VectorFTy) return false; - Function *FuncToReplace = CallToReplace.getCalledFunction(); - Function *TLIFunc = getTLIFunction(CallToReplace.getModule(), VectorFTy, + Function *TLIFunc = getTLIFunction(I.getModule(), VectorFTy, VD->getVectorFnName(), FuncToReplace); - replaceWithTLIFunction(CallToReplace, *OptInfo, TLIFunc); - - LLVM_DEBUG(dbgs() << DEBUG_TYPE << ": Replaced call to `" - << FuncToReplace->getName() << "` with call to `" - << TLIFunc->getName() << "`.\n"); + replaceWithTLIFunction(I, *OptInfo, TLIFunc); + LLVM_DEBUG(dbgs() << DEBUG_TYPE << ": Replaced call to `" << ScalarName + << "` with call to `" << TLIFunc->getName() << "`.\n"); ++NumCallsReplaced; return true; } +/// Supported instruction \p I must be a vectorized frem or a call to an +/// intrinsic that returns either void or a vector. +static bool isSupportedInstruction(Instruction *I) { + Type *Ty = I->getType(); + if (auto *CI = dyn_cast(I)) + return (Ty->isVectorTy() || Ty->isVoidTy()) && CI->getCalledFunction() && + CI->getCalledFunction()->getIntrinsicID() != + Intrinsic::not_intrinsic; + if (I->getOpcode() == Instruction::FRem && Ty->isVectorTy()) + return true; + return false; +} + static bool runImpl(const TargetLibraryInfo &TLI, Function &F) { bool Changed = false; - SmallVector ReplacedCalls; + SmallVector ReplacedCalls; for (auto &I : instructions(F)) { - if (auto *CI = dyn_cast(&I)) { - if (replaceWithCallToVeclib(TLI, *CI)) { - ReplacedCalls.push_back(CI); - Changed = true; - } + if (!isSupportedInstruction(&I)) + continue; + if (replaceWithCallToVeclib(TLI, I)) { + ReplacedCalls.push_back(&I); + Changed = true; } } // Erase the calls to the intrinsics that have been replaced diff --git a/llvm/test/CodeGen/AArch64/replace-intrinsics-with-veclib-armpl.ll b/llvm/test/CodeGen/AArch64/replace-with-veclib-armpl.ll similarity index 91% rename from llvm/test/CodeGen/AArch64/replace-intrinsics-with-veclib-armpl.ll rename to llvm/test/CodeGen/AArch64/replace-with-veclib-armpl.ll index d41870ec6e791..4480a90a2728d 100644 --- a/llvm/test/CodeGen/AArch64/replace-intrinsics-with-veclib-armpl.ll +++ b/llvm/test/CodeGen/AArch64/replace-with-veclib-armpl.ll @@ -15,7 +15,7 @@ declare @llvm.cos.nxv2f64() declare @llvm.cos.nxv4f32() ;. -; CHECK: @llvm.compiler.used = appending global [32 x ptr] [ptr @armpl_vcosq_f64, ptr @armpl_vcosq_f32, ptr @armpl_svcos_f64_x, ptr @armpl_svcos_f32_x, ptr @armpl_vsinq_f64, ptr @armpl_vsinq_f32, ptr @armpl_svsin_f64_x, ptr @armpl_svsin_f32_x, ptr @armpl_vexpq_f64, ptr @armpl_vexpq_f32, ptr @armpl_svexp_f64_x, ptr @armpl_svexp_f32_x, ptr @armpl_vexp2q_f64, ptr @armpl_vexp2q_f32, ptr @armpl_svexp2_f64_x, ptr @armpl_svexp2_f32_x, ptr @armpl_vexp10q_f64, ptr @armpl_vexp10q_f32, ptr @armpl_svexp10_f64_x, ptr @armpl_svexp10_f32_x, ptr @armpl_vlogq_f64, ptr @armpl_vlogq_f32, ptr @armpl_svlog_f64_x, ptr @armpl_svlog_f32_x, ptr @armpl_vlog2q_f64, ptr @armpl_vlog2q_f32, ptr @armpl_svlog2_f64_x, ptr @armpl_svlog2_f32_x, ptr @armpl_vlog10q_f64, ptr @armpl_vlog10q_f32, ptr @armpl_svlog10_f64_x, ptr @armpl_svlog10_f32_x], section "llvm.metadata" +; CHECK: @llvm.compiler.used = appending global [36 x ptr] [ptr @armpl_vcosq_f64, ptr @armpl_vcosq_f32, ptr @armpl_svcos_f64_x, ptr @armpl_svcos_f32_x, ptr @armpl_vsinq_f64, ptr @armpl_vsinq_f32, ptr @armpl_svsin_f64_x, ptr @armpl_svsin_f32_x, ptr @armpl_vexpq_f64, ptr @armpl_vexpq_f32, ptr @armpl_svexp_f64_x, ptr @armpl_svexp_f32_x, ptr @armpl_vexp2q_f64, ptr @armpl_vexp2q_f32, ptr @armpl_svexp2_f64_x, ptr @armpl_svexp2_f32_x, ptr @armpl_vexp10q_f64, ptr @armpl_vexp10q_f32, ptr @armpl_svexp10_f64_x, ptr @armpl_svexp10_f32_x, ptr @armpl_vlogq_f64, ptr @armpl_vlogq_f32, ptr @armpl_svlog_f64_x, ptr @armpl_svlog_f32_x, ptr @armpl_vlog2q_f64, ptr @armpl_vlog2q_f32, ptr @armpl_svlog2_f64_x, ptr @armpl_svlog2_f32_x, ptr @armpl_vlog10q_f64, ptr @armpl_vlog10q_f32, ptr @armpl_svlog10_f64_x, ptr @armpl_svlog10_f32_x, ptr @armpl_vfmodq_f64, ptr @armpl_vfmodq_f32, ptr @armpl_svfmod_f64_x, ptr @armpl_svfmod_f32_x], section "llvm.metadata" ;. define <2 x double> @llvm_cos_f64(<2 x double> %in) { ; CHECK-LABEL: define <2 x double> @llvm_cos_f64 @@ -424,6 +424,46 @@ define @llvm_pow_vscale_f32( %in, %1 } +define <2 x double> @frem_f64(<2 x double> %in) { +; CHECK-LABEL: define <2 x double> @frem_f64 +; CHECK-SAME: (<2 x double> [[IN:%.*]]) { +; CHECK-NEXT: [[TMP1:%.*]] = call <2 x double> @armpl_vfmodq_f64(<2 x double> [[IN]], <2 x double> [[IN]]) +; CHECK-NEXT: ret <2 x double> [[TMP1]] +; + %1= frem <2 x double> %in, %in + ret <2 x double> %1 +} + +define <4 x float> @frem_f32(<4 x float> %in) { +; CHECK-LABEL: define <4 x float> @frem_f32 +; CHECK-SAME: (<4 x float> [[IN:%.*]]) { +; CHECK-NEXT: [[TMP1:%.*]] = call <4 x float> @armpl_vfmodq_f32(<4 x float> [[IN]], <4 x float> [[IN]]) +; CHECK-NEXT: ret <4 x float> [[TMP1]] +; + %1= frem <4 x float> %in, %in + ret <4 x float> %1 +} + +define @frem_vscale_f64( %in) #0 { +; CHECK-LABEL: define @frem_vscale_f64 +; CHECK-SAME: ( [[IN:%.*]]) #[[ATTR1]] { +; CHECK-NEXT: [[TMP1:%.*]] = call @armpl_svfmod_f64_x( [[IN]], [[IN]], shufflevector ( insertelement ( poison, i1 true, i64 0), poison, zeroinitializer)) +; CHECK-NEXT: ret [[TMP1]] +; + %1= frem %in, %in + ret %1 +} + +define @frem_vscale_f32( %in) #0 { +; CHECK-LABEL: define @frem_vscale_f32 +; CHECK-SAME: ( [[IN:%.*]]) #[[ATTR1]] { +; CHECK-NEXT: [[TMP1:%.*]] = call @armpl_svfmod_f32_x( [[IN]], [[IN]], shufflevector ( insertelement ( poison, i1 true, i64 0), poison, zeroinitializer)) +; CHECK-NEXT: ret [[TMP1]] +; + %1= frem %in, %in + ret %1 +} + attributes #0 = { "target-features"="+sve" } ;. ; CHECK: attributes #[[ATTR0:[0-9]+]] = { nocallback nofree nosync nounwind speculatable willreturn memory(none) } diff --git a/llvm/test/CodeGen/AArch64/replace-intrinsics-with-veclib-sleef-scalable.ll b/llvm/test/CodeGen/AArch64/replace-with-veclib-sleef-scalable.ll similarity index 95% rename from llvm/test/CodeGen/AArch64/replace-intrinsics-with-veclib-sleef-scalable.ll rename to llvm/test/CodeGen/AArch64/replace-with-veclib-sleef-scalable.ll index c2ff6014bc694..590dd9effac0e 100644 --- a/llvm/test/CodeGen/AArch64/replace-intrinsics-with-veclib-sleef-scalable.ll +++ b/llvm/test/CodeGen/AArch64/replace-with-veclib-sleef-scalable.ll @@ -4,7 +4,7 @@ target triple = "aarch64-unknown-linux-gnu" ;. -; CHECK: @llvm.compiler.used = appending global [16 x ptr] [ptr @_ZGVsMxv_cos, ptr @_ZGVsMxv_cosf, ptr @_ZGVsMxv_exp, ptr @_ZGVsMxv_expf, ptr @_ZGVsMxv_exp2, ptr @_ZGVsMxv_exp2f, ptr @_ZGVsMxv_exp10, ptr @_ZGVsMxv_exp10f, ptr @_ZGVsMxv_log, ptr @_ZGVsMxv_logf, ptr @_ZGVsMxv_log10, ptr @_ZGVsMxv_log10f, ptr @_ZGVsMxv_log2, ptr @_ZGVsMxv_log2f, ptr @_ZGVsMxv_sin, ptr @_ZGVsMxv_sinf], section "llvm.metadata" +; CHECK: @llvm.compiler.used = appending global [18 x ptr] [ptr @_ZGVsMxv_cos, ptr @_ZGVsMxv_cosf, ptr @_ZGVsMxv_exp, ptr @_ZGVsMxv_expf, ptr @_ZGVsMxv_exp2, ptr @_ZGVsMxv_exp2f, ptr @_ZGVsMxv_exp10, ptr @_ZGVsMxv_exp10f, ptr @_ZGVsMxv_log, ptr @_ZGVsMxv_logf, ptr @_ZGVsMxv_log10, ptr @_ZGVsMxv_log10f, ptr @_ZGVsMxv_log2, ptr @_ZGVsMxv_log2f, ptr @_ZGVsMxv_sin, ptr @_ZGVsMxv_sinf, ptr @_ZGVsMxvv_fmod, ptr @_ZGVsMxvv_fmodf], section "llvm.metadata" ;. define @llvm_ceil_vscale_f64( %in) { ; CHECK-LABEL: @llvm_ceil_vscale_f64( @@ -384,6 +384,24 @@ define @llvm_trunc_vscale_f32( %in) { ret %1 } +define @frem_f64( %in) { +; CHECK-LABEL: @frem_f64( +; CHECK-NEXT: [[TMP1:%.*]] = call @_ZGVsMxvv_fmod( [[IN:%.*]], [[IN]], shufflevector ( insertelement ( poison, i1 true, i64 0), poison, zeroinitializer)) +; CHECK-NEXT: ret [[TMP1]] +; + %1= frem %in, %in + ret %1 +} + +define @frem_f32( %in) { +; CHECK-LABEL: @frem_f32( +; CHECK-NEXT: [[TMP1:%.*]] = call @_ZGVsMxvv_fmodf( [[IN:%.*]], [[IN]], shufflevector ( insertelement ( poison, i1 true, i64 0), poison, zeroinitializer)) +; CHECK-NEXT: ret [[TMP1]] +; + %1= frem %in, %in + ret %1 +} + declare @llvm.ceil.nxv2f64() declare @llvm.ceil.nxv4f32() declare @llvm.copysign.nxv2f64(, ) diff --git a/llvm/test/CodeGen/AArch64/replace-intrinsics-with-veclib-sleef.ll b/llvm/test/CodeGen/AArch64/replace-with-veclib-sleef.ll similarity index 95% rename from llvm/test/CodeGen/AArch64/replace-intrinsics-with-veclib-sleef.ll rename to llvm/test/CodeGen/AArch64/replace-with-veclib-sleef.ll index be247de368056..865a46009b205 100644 --- a/llvm/test/CodeGen/AArch64/replace-intrinsics-with-veclib-sleef.ll +++ b/llvm/test/CodeGen/AArch64/replace-with-veclib-sleef.ll @@ -4,7 +4,7 @@ target triple = "aarch64-unknown-linux-gnu" ;. -; CHECK: @llvm.compiler.used = appending global [16 x ptr] [ptr @_ZGVnN2v_cos, ptr @_ZGVnN4v_cosf, ptr @_ZGVnN2v_exp, ptr @_ZGVnN4v_expf, ptr @_ZGVnN2v_exp2, ptr @_ZGVnN4v_exp2f, ptr @_ZGVnN2v_exp10, ptr @_ZGVnN4v_exp10f, ptr @_ZGVnN2v_log, ptr @_ZGVnN4v_logf, ptr @_ZGVnN2v_log10, ptr @_ZGVnN4v_log10f, ptr @_ZGVnN2v_log2, ptr @_ZGVnN4v_log2f, ptr @_ZGVnN2v_sin, ptr @_ZGVnN4v_sinf], section "llvm.metadata" +; CHECK: @llvm.compiler.used = appending global [18 x ptr] [ptr @_ZGVnN2v_cos, ptr @_ZGVnN4v_cosf, ptr @_ZGVnN2v_exp, ptr @_ZGVnN4v_expf, ptr @_ZGVnN2v_exp2, ptr @_ZGVnN4v_exp2f, ptr @_ZGVnN2v_exp10, ptr @_ZGVnN4v_exp10f, ptr @_ZGVnN2v_log, ptr @_ZGVnN4v_logf, ptr @_ZGVnN2v_log10, ptr @_ZGVnN4v_log10f, ptr @_ZGVnN2v_log2, ptr @_ZGVnN4v_log2f, ptr @_ZGVnN2v_sin, ptr @_ZGVnN4v_sinf, ptr @_ZGVnN2vv_fmod, ptr @_ZGVnN4vv_fmodf], section "llvm.metadata" ;. define <2 x double> @llvm_ceil_f64(<2 x double> %in) { ; CHECK-LABEL: @llvm_ceil_f64( @@ -384,6 +384,24 @@ define <4 x float> @llvm_trunc_f32(<4 x float> %in) { ret <4 x float> %1 } +define <2 x double> @frem_f64(<2 x double> %in) { +; CHECK-LABEL: @frem_f64( +; CHECK-NEXT: [[TMP1:%.*]] = call <2 x double> @_ZGVnN2vv_fmod(<2 x double> [[IN:%.*]], <2 x double> [[IN]]) +; CHECK-NEXT: ret <2 x double> [[TMP1]] +; + %1= frem <2 x double> %in, %in + ret <2 x double> %1 +} + +define <4 x float> @frem_f32(<4 x float> %in) { +; CHECK-LABEL: @frem_f32( +; CHECK-NEXT: [[TMP1:%.*]] = call <4 x float> @_ZGVnN4vv_fmodf(<4 x float> [[IN:%.*]], <4 x float> [[IN]]) +; CHECK-NEXT: ret <4 x float> [[TMP1]] +; + %1= frem <4 x float> %in, %in + ret <4 x float> %1 +} + declare <2 x double> @llvm.ceil.v2f64(<2 x double>) declare <4 x float> @llvm.ceil.v4f32(<4 x float>) declare <2 x double> @llvm.copysign.v2f64(<2 x double>, <2 x double>)