Make have_fma consistent between interpreter and compiled (#52206)

Currently the interpreter always returns false. Which isn't very good.
Make it follow whatever the JIT will do.

Author: gbaraldi

src/jl_exported_funcs.inc

@@ -200,6 +200,7 @@
     XX(jl_get_binding_wr) \
     XX(jl_get_cpu_name) \
     XX(jl_get_cpu_features) \
+    XX(jl_cpu_has_fma) \
     XX(jl_get_current_task) \
     XX(jl_get_default_sysimg_path) \
     XX(jl_get_excstack) \

src/llvm-cpufeatures.cpp

@@ -59,15 +59,15 @@ static bool have_fma(Function &intr, Function &caller, const Triple &TT) JL_NOTS
     StringRef FS =
         FSAttr.isValid() ? FSAttr.getValueAsString() : jl_ExecutionEngine->getTargetFeatureString();
 
-    SmallVector<StringRef, 6> Features;
+    SmallVector<StringRef, 128> Features;
     FS.split(Features, ',');
     for (StringRef Feature : Features)
     if (TT.isARM()) {
       if (Feature == "+vfp4")
         return typ == "f32" || typ == "f64";
       else if (Feature == "+vfp4sp")
         return typ == "f32";
-    } else {
+    } else if (TT.isX86()) {
       if (Feature == "+fma" || Feature == "+fma4")
         return typ == "f32" || typ == "f64";
     }

src/processor.h

@@ -226,6 +226,8 @@ JL_DLLEXPORT jl_value_t *jl_get_cpu_name(void);
 // Return the features of the host CPU as a julia string.
 JL_DLLEXPORT jl_value_t *jl_get_cpu_features(void);
 // Dump the name and feature set of the host CPU
+JL_DLLEXPORT jl_value_t *jl_cpu_has_fma(int bits);
+// Check if the CPU has native FMA instructions;
 // For debugging only
 JL_DLLEXPORT void jl_dump_host_cpu(void);
 JL_DLLEXPORT jl_value_t* jl_check_pkgimage_clones(char* data);

src/processor_arm.cpp

@@ -1808,6 +1808,22 @@ JL_DLLEXPORT jl_value_t *jl_get_cpu_features(void)
     return jl_cstr_to_string(jl_get_cpu_features_llvm().c_str());
 }
 
+JL_DLLEXPORT jl_value_t *jl_cpu_has_fma(int bits)
+{
+#ifdef _CPU_AARCH64_
+    return jl_true;
+#else
+    TargetData<feature_sz> target = jit_targets.front();
+    FeatureList<feature_sz> features = target.en.features;
+    if (bits == 32 && test_nbit(features, Feature::vfp4sp))
+        return jl_true;
+    else if ((bits == 64 || bits == 32) && test_nbit(features, Feature::vfp4))
+        return jl_true;
+    else
+        return jl_false;
+#endif
+}
+
 jl_image_t jl_init_processor_sysimg(void *hdl)
 {
     if (!jit_targets.empty())

src/processor_fallback.cpp

@@ -172,6 +172,11 @@ JL_DLLEXPORT jl_value_t *jl_get_cpu_features(void)
     return jl_cstr_to_string(jl_get_cpu_features_llvm().c_str());
 }
 
+JL_DLLEXPORT jl_value_t *jl_cpu_has_fma(int bits)
+{
+    return jl_false; // Match behaviour of have_fma in src/llvm-cpufeatures.cpp (assume false)
+}
+
 JL_DLLEXPORT void jl_dump_host_cpu(void)
 {
     jl_safe_printf("CPU: %s\n", host_cpu_name().c_str());

src/processor_x86.cpp

@@ -4,6 +4,7 @@
 
 // CPUID
 
+#include "julia.h"
 extern "C" JL_DLLEXPORT void jl_cpuid(int32_t CPUInfo[4], int32_t InfoType)
 {
     asm volatile (
@@ -1062,6 +1063,16 @@ JL_DLLEXPORT jl_value_t *jl_get_cpu_features(void)
     return jl_cstr_to_string(jl_get_cpu_features_llvm().c_str());
 }
 
+JL_DLLEXPORT jl_value_t *jl_cpu_has_fma(int bits)
+{
+    TargetData<feature_sz> target = jit_targets.front();
+    FeatureList<feature_sz> features = target.en.features;
+    if ((bits == 32 || bits == 64) && (test_nbit(features, Feature::fma) || test_nbit(features, Feature::fma4)))
+        return jl_true;
+    else
+        return jl_false;
+}
+
 jl_image_t jl_init_processor_sysimg(void *hdl)
 {
     if (!jit_targets.empty())

src/runtime_intrinsics.c

@@ -1671,10 +1671,15 @@ 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_value_t *jl_cpu_has_fma(int bits);
 
 JL_DLLEXPORT jl_value_t *jl_have_fma(jl_value_t *typ)
 {
-    JL_TYPECHK(have_fma, datatype, typ);
-    // TODO: run-time feature check?
-    return jl_false;
+    JL_TYPECHK(have_fma, datatype, typ); // TODO what about float16/bfloat16?
+    if (typ == (jl_value_t*)jl_float32_type)
+        return jl_cpu_has_fma(32);
+    else if (typ == (jl_value_t*)jl_float64_type)
+        return jl_cpu_has_fma(64);
+    else
+        return jl_false;
 }

test/llvmpasses/cpu-features.ll

@@ -3,6 +3,8 @@
 ; RUN: opt -enable-new-pm=1 --opaque-pointers=0 --load-pass-plugin=libjulia-codegen%shlibext -passes='CPUFeatures,simplifycfg' -S %s | FileCheck %s
 
 ; RUN: opt -enable-new-pm=1 --opaque-pointers=1 --load-pass-plugin=libjulia-codegen%shlibext -passes='CPUFeatures,simplifycfg' -S %s | FileCheck %s
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128-ni:10:11:12:13"
+target triple = "x86_64-linux-gnu"
 
 declare i1 @julia.cpu.have_fma.f64()
 declare double @with_fma(double %0, double %1, double %2)

test/sysinfo.jl

@@ -12,6 +12,8 @@ Base.Sys.loadavg()
 
 @test length(ccall(:jl_get_cpu_name, String, ())) != 0
 @test length(ccall(:jl_get_cpu_features, String, ())) >= 0
+foo_fma() = Core.Intrinsics.have_fma(Int64)
+@test ccall(:jl_cpu_has_fma, Bool, (Cint,), 64) == foo_fma()
 
 if Sys.isunix()
     mktempdir() do tempdir