Break type inference for log with unrelated module

[simonster]

julia> import Primes

julia> f(x) = log(x);

julia> @code_llvm f(1.0)

define double @julia_f_70658(double) #0 {
top:
  %ptls_i8 = call i8* asm "movq %fs:0, $0;\0Aaddq $$-2672, $0", "=r,~{dirflag},~{fpsr},~{flags}"() #2
  %ptls = bitcast i8* %ptls_i8 to %jl_value_t***
  %1 = alloca [4 x %jl_value_t*], align 8
  %.sub = getelementptr inbounds [4 x %jl_value_t*], [4 x %jl_value_t*]* %1, i64 0, i64 0
  %2 = getelementptr [4 x %jl_value_t*], [4 x %jl_value_t*]* %1, i64 0, i64 2
  %3 = bitcast %jl_value_t** %2 to i8*
  call void @llvm.memset.p0i8.i32(i8* %3, i8 0, i32 16, i32 8, i1 false)
  %4 = bitcast [4 x %jl_value_t*]* %1 to i64*
  store i64 4, i64* %4, align 8
  %5 = getelementptr [4 x %jl_value_t*], [4 x %jl_value_t*]* %1, i64 0, i64 1
  %6 = bitcast i8* %ptls_i8 to i64*
  %7 = load i64, i64* %6, align 8
  %8 = bitcast %jl_value_t** %5 to i64*
  store i64 %7, i64* %8, align 8
  store %jl_value_t** %.sub, %jl_value_t*** %ptls, align 8
  %9 = getelementptr [4 x %jl_value_t*], [4 x %jl_value_t*]* %1, i64 0, i64 3
  store %jl_value_t* inttoptr (i64 140530114606896 to %jl_value_t*), %jl_value_t** %2, align 8
  %10 = call %jl_value_t* @jl_gc_pool_alloc(i8* %ptls_i8, i32 1432, i32 16)
  %11 = getelementptr inbounds %jl_value_t, %jl_value_t* %10, i64 -1, i32 0
  store %jl_value_t* inttoptr (i64 140530057424720 to %jl_value_t*), %jl_value_t** %11, align 8
  %12 = bitcast %jl_value_t* %10 to double*
  store double %0, double* %12, align 8
  store %jl_value_t* %10, %jl_value_t** %9, align 8
  %13 = call %jl_value_t* @jl_invoke(%jl_value_t* inttoptr (i64 140530176032384 to %jl_value_t*), %jl_value_t** %2, i32 2)
  %14 = bitcast %jl_value_t* %13 to double*
  %15 = load double, double* %14, align 16
  %16 = load i64, i64* %8, align 8
  store i64 %16, i64* %6, align 8
  ret double %15
}

julia> Pkg.installed("Primes")
v"0.1.2"

julia> versioninfo()
Julia Version 0.5.1-pre+55
Commit 8d4ef37* (2017-02-13 09:11 UTC)
Platform Info:
  System: Linux (x86_64-linux-gnu)
  CPU: Intel(R) Core(TM) i7-3930K CPU @ 3.20GHz
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Sandybridge)
  LAPACK: libopenblas64_
  LIBM: libopenlibm
  LLVM: libLLVM-3.7.1 (ORCJIT, sandybridge)

Compare to:

julia> f(x) = log(x);

julia> @code_llvm f(1.0)

define double @julia_f_70640(double) #0 {
top:
  %1 = call double @julia_log_70641(double %0) #0
  ret double %1
}

This does not happen if Julia is started with --compilecache=no. Maybe related to #18465? But @code_warntype f(1.0) looks fine regardless of whether Primes is loaded.

[simonster]

Also, for fun:

julia> import Primes

julia> f(x) = log(x);

julia> @code_llvm f(1.0)

define double @julia_f_70651(double) #0 {
top:
  %ptls_i8 = call i8* asm "movq %fs:0, $0;\0Aaddq $$-2672, $0", "=r,~{dirflag},~{fpsr},~{flags}"() #2
  %ptls = bitcast i8* %ptls_i8 to %jl_value_t***
  %1 = alloca [4 x %jl_value_t*], align 8
  %.sub = getelementptr inbounds [4 x %jl_value_t*], [4 x %jl_value_t*]* %1, i64 0, i64 0
  %2 = getelementptr [4 x %jl_value_t*], [4 x %jl_value_t*]* %1, i64 0, i64 2
  %3 = bitcast %jl_value_t** %2 to i8*
  call void @llvm.memset.p0i8.i32(i8* %3, i8 0, i32 16, i32 8, i1 false)
  %4 = bitcast [4 x %jl_value_t*]* %1 to i64*
  store i64 4, i64* %4, align 8
  %5 = getelementptr [4 x %jl_value_t*], [4 x %jl_value_t*]* %1, i64 0, i64 1
  %6 = bitcast i8* %ptls_i8 to i64*
  %7 = load i64, i64* %6, align 8
  %8 = bitcast %jl_value_t** %5 to i64*
  store i64 %7, i64* %8, align 8
  store %jl_value_t** %.sub, %jl_value_t*** %ptls, align 8
  %9 = getelementptr [4 x %jl_value_t*], [4 x %jl_value_t*]* %1, i64 0, i64 3
  store %jl_value_t* inttoptr (i64 140028128294704 to %jl_value_t*), %jl_value_t** %2, align 8
  %10 = call %jl_value_t* @jl_gc_pool_alloc(i8* %ptls_i8, i32 1432, i32 16)
  %11 = getelementptr inbounds %jl_value_t, %jl_value_t* %10, i64 -1, i32 0
  store %jl_value_t* inttoptr (i64 140028071112528 to %jl_value_t*), %jl_value_t** %11, align 8
  %12 = bitcast %jl_value_t* %10 to double*
  store double %0, double* %12, align 8
  store %jl_value_t* %10, %jl_value_t** %9, align 8
  %13 = call %jl_value_t* @jl_invoke(%jl_value_t* inttoptr (i64 140028189719904 to %jl_value_t*), %jl_value_t** %2, i32 2)
  %14 = bitcast %jl_value_t* %13 to double*
  %15 = load double, double* %14, align 16
  %16 = load i64, i64* %8, align 8
  store i64 %16, i64* %6, align 8
  ret double %15
}

julia> f(1.0)
0.0

julia> @code_llvm f(1.0)

define double @julia_f_70651(double) #0 {
top:
  %1 = call double @julia_log_70659(double %0) #0
  ret double %1
}

(and it's not just code_llvm that is showing the suboptimal calling convention; the code itself is allocating after importing Primes)

[simonster]

Also, FWIW, this does not happen on master.

[yuyichao]

Dup of #18465