RFC: Remove sqrt_llvm intrinsic

This is more of a "Do we want to move in this direction RFC". As mentioned in #43786,
we currently have three implementations of these intrinsics:

1. The code generated by LLVM for the intrinsic
2. The code LLVM uses for constant folding the intrinsic
3. Our own runtime intrinsic used by the interpreter

This basically removes the third one, which will be required if we want to do
something about #26434 because we just forward these to libm. Of course we'll
still have to do something to teach LLVM how to constant fold these in a manner
compatible with what will actually end up running, but that's a separate issue.

Author: Keno

base/compiler/optimize.jl

@@ -618,7 +618,6 @@ function is_pure_intrinsic_infer(f::IntrinsicFunction)
              f === Intrinsics.pointerset || # this one is never effect-free
              f === Intrinsics.llvmcall ||   # this one is never effect-free
              f === Intrinsics.arraylen ||   # this one is volatile
-             f === Intrinsics.sqrt_llvm ||  # this one may differ at runtime (by a few ulps)
              f === Intrinsics.sqrt_llvm_fast ||  # this one may differ at runtime (by a few ulps)
              f === Intrinsics.have_fma ||  # this one depends on the runtime environment
              f === Intrinsics.cglobal)  # cglobal lookup answer changes at runtime

base/compiler/tfuncs.jl

@@ -167,7 +167,6 @@ add_tfunc(ceil_llvm, 1, 1, math_tfunc, 10)
 add_tfunc(floor_llvm, 1, 1, math_tfunc, 10)
 add_tfunc(trunc_llvm, 1, 1, math_tfunc, 10)
 add_tfunc(rint_llvm, 1, 1, math_tfunc, 10)
-add_tfunc(sqrt_llvm, 1, 1, math_tfunc, 20)
 add_tfunc(sqrt_llvm_fast, 1, 1, math_tfunc, 20)
     ## same-type comparisons ##
 cmp_tfunc(@nospecialize(x), @nospecialize(y)) = Bool

base/math.jl

@@ -25,8 +25,6 @@ using .Base: sign_mask, exponent_mask, exponent_one,
             significand_bits, exponent_bits, exponent_bias,
             exponent_max, exponent_raw_max
 
-using Core.Intrinsics: sqrt_llvm
-
 using .Base: IEEEFloat
 
 @noinline function throw_complex_domainerror(f::Symbol, x)
@@ -586,6 +584,8 @@ Stacktrace:
 """
 log1p(x)
 
+sqrt_llvm(x::Float32) = ccall("llvm.sqrt.f32", llvmcall, Float32, (Float32,), x)
+sqrt_llvm(x::Float64) = ccall("llvm.sqrt.f64", llvmcall, Float64, (Float64,), x)
 @inline function sqrt(x::Union{Float32,Float64})
     x < zero(x) && throw_complex_domainerror(:sqrt, x)
     sqrt_llvm(x)

src/intrinsics.cpp

@@ -65,7 +65,6 @@ static void jl_init_intrinsic_functions_codegen(void)
     float_func[floor_llvm] = true;
     float_func[trunc_llvm] = true;
     float_func[rint_llvm] = true;
-    float_func[sqrt_llvm] = true;
     float_func[sqrt_llvm_fast] = true;
 }
 
@@ -1485,10 +1484,6 @@ static Value *emit_untyped_intrinsic(jl_codectx_t &ctx, intrinsic f, Value **arg
         FunctionCallee rintintr = Intrinsic::getDeclaration(jl_Module, Intrinsic::rint, makeArrayRef(t));
         return ctx.builder.CreateCall(rintintr, x);
     }
-    case sqrt_llvm: {
-        FunctionCallee sqrtintr = Intrinsic::getDeclaration(jl_Module, Intrinsic::sqrt, makeArrayRef(t));
-        return ctx.builder.CreateCall(sqrtintr, x);
-    }
     case sqrt_llvm_fast: {
         FunctionCallee sqrtintr = Intrinsic::getDeclaration(jl_Module, Intrinsic::sqrt, makeArrayRef(t));
         return math_builder(ctx, true)().CreateCall(sqrtintr, x);

src/intrinsics.h

@@ -86,7 +86,6 @@
     ADD_I(floor_llvm, 1) \
     ADD_I(trunc_llvm, 1) \
     ADD_I(rint_llvm, 1) \
-    ADD_I(sqrt_llvm, 1) \
     ADD_I(sqrt_llvm_fast, 1) \
     /*  pointer access */ \
     ADD_I(pointerref, 3) \

src/julia_internal.h

@@ -1233,7 +1233,6 @@ JL_DLLEXPORT jl_value_t *jl_ceil_llvm(jl_value_t *a);
 JL_DLLEXPORT jl_value_t *jl_floor_llvm(jl_value_t *a);
 JL_DLLEXPORT jl_value_t *jl_trunc_llvm(jl_value_t *a);
 JL_DLLEXPORT jl_value_t *jl_rint_llvm(jl_value_t *a);
-JL_DLLEXPORT jl_value_t *jl_sqrt_llvm(jl_value_t *a);
 JL_DLLEXPORT jl_value_t *jl_sqrt_llvm_fast(jl_value_t *a);
 JL_DLLEXPORT jl_value_t *jl_abs_float(jl_value_t *a);
 JL_DLLEXPORT jl_value_t *jl_copysign_float(jl_value_t *a, jl_value_t *b);

src/runtime_intrinsics.c

@@ -1445,7 +1445,6 @@ un_fintrinsic(ceil_float,ceil_llvm)
 un_fintrinsic(floor_float,floor_llvm)
 un_fintrinsic(trunc_float,trunc_llvm)
 un_fintrinsic(rint_float,rint_llvm)
-un_fintrinsic(sqrt_float,sqrt_llvm)
 un_fintrinsic(sqrt_float,sqrt_llvm_fast)
 
 JL_DLLEXPORT jl_value_t *jl_arraylen(jl_value_t *a)