[2019-08] [arm/arm64] fix fconv/rconv issues with LLVM (#16532)

* [ci] add make -C mono/mini llvmaotcheck to FullAOT LLVM lanes

* [arm64] use opcode emulation for fconv/rconv

LLVM eagerly optimizes constants such as NaN in a way that does not align with .NET behaviour. In the future we can change it so that we only do opcode emulation when LLVM is used.

Fixes #16411

* [arm] account only 4 bytes on stack for single precision arguments

Fixes `test_0_arm64_small_stack_args` in gshared.cs when
* `test_0_arm64_small_stack_args` is compiled with LLVM
* `Foo3.Floats` is compiled with Mini

* [arm] use opcode emulation for fconv/rconv

* [llvmaotcheck] require sse3 being available for regression tests

* [amd64] use opcode emulation for fconv/rconv

* fix cast and signature for mono_rconv_u4

mono/mini/Makefile.am.in

@@ -881,11 +881,16 @@ gctest: mono gc-test.exe
 	MONO_DEBUG_OPTIONS=clear-nursery-at-gc $(MINI_RUNTIME) --regression gc-test.exe
 
 LLVM_AOT_RUNTIME_OPTS=$(if $(LLVM),--llvm,)
+if AMD64
+LLVM_AOT_COMPILER_OPTS=$(if $(LLVM),"=llvmllc=-mcpu=generic -mattr=+sse3",)
+else
+LLVM_AOT_COMPILER_OPTS=
+endif
 GSHAREDVT_RUNTIME_OPTS=$(if $(GSHAREDVT),-O=gsharedvt,)
 
 aotcheck: mono $(regtests)
 	rm -rf *.exe.so *.exe.dylib *.exe.dylib.dSYM
-	$(MINI_RUNTIME) $(LLVM_AOT_RUNTIME_OPTS) --aot $(regtests) || exit 1
+	$(MINI_RUNTIME) $(LLVM_AOT_RUNTIME_OPTS) --aot$(LLVM_AOT_COMPILER_OPTS) $(regtests) || exit 1
 	for i in $(regtests); do $(RUNTIME_AOTCHECK) --regression $$i || exit 1; done
 	rm -rf *.exe.so *.exe.dylib *.exe.dylib.dSYM
 

mono/mini/jit-icalls.c

@@ -925,12 +925,12 @@ mono_fconv_u4_2 (double v)
 	return mono_fconv_u4 (v);
 }
 
-gint32
+guint32
 mono_rconv_u4 (float v)
 {
 	if (mono_isinf (v) || mono_isnan (v))
 		return 0;
-	return (gint32)v;
+	return (guint32) v;
 }
 #endif
 

mono/mini/jit-icalls.h

@@ -80,7 +80,7 @@ G_EXTERN_C gint64 mono_fconv_i8 (double v);
 G_EXTERN_C guint32 mono_fconv_u4 (double v);
 G_EXTERN_C guint32 mono_fconv_u4_2 (double v);
 
-G_EXTERN_C gint32 mono_rconv_u4 (float v);
+G_EXTERN_C guint32 mono_rconv_u4 (float v);
 
 G_EXTERN_C gint64 mono_fconv_ovf_i8 (double v);
 

mono/mini/mini-amd64.h

@@ -414,10 +414,8 @@ typedef struct {
 
 #define MONO_ARCH_EMULATE_CONV_R8_UN 1
 #define MONO_ARCH_EMULATE_FCONV_TO_U8 1
-// The Windows x64 FullAOT+LLVM fails to pass the basic-float tests without this.
-#ifdef TARGET_WIN32
+// x64 FullAOT+LLVM fails to pass the basic-float tests without this.
 #define MONO_ARCH_EMULATE_FCONV_TO_U4 1
-#endif
 #define MONO_ARCH_EMULATE_FREM 1
 #define MONO_ARCH_HAVE_IS_INT_OVERFLOW 1
 #define MONO_ARCH_HAVE_INVALIDATE_METHOD 1

mono/mini/mini-arm.c

@@ -1214,7 +1214,7 @@ add_float (guint *fpr, guint *stack_size, ArgInfo *ainfo, gboolean is_double, gi
 		ainfo->reg = ARMREG_SP;
 		ainfo->storage = RegTypeBase;
 
-		*stack_size += 8;
+		*stack_size += is_double ? 8 : 4;
 	}
 }
 

mono/mini/mini-arm.h

@@ -310,6 +310,7 @@ typedef struct MonoCompileArch {
 	int thunks_size;
 } MonoCompileArch;
 
+#define MONO_ARCH_EMULATE_FCONV_TO_U4 1
 #define MONO_ARCH_EMULATE_FCONV_TO_I8 1
 #define MONO_ARCH_EMULATE_FCONV_TO_U8 1
 #define MONO_ARCH_EMULATE_LCONV_TO_R8 1

mono/mini/mini-arm64.h

@@ -112,13 +112,14 @@ typedef struct {
 	int thunks_size;
 } MonoCompileArch;
 
+#define MONO_ARCH_EMULATE_FCONV_TO_U4 1
+#define MONO_ARCH_EMULATE_FCONV_TO_U8 1
 #ifdef MONO_ARCH_ILP32
 /* For the watch (starting with series 4), a new ABI is introduced: arm64_32.
  * We can still use the older AOT compiler to produce bitcode, because it's
  * "offset compatible". However, since it is targeting arm7k, it makes certain
  * assumptions that we need to align here. */
 #define MONO_ARCH_EMULATE_FCONV_TO_I8 1
-#define MONO_ARCH_EMULATE_FCONV_TO_U8 1
 #define MONO_ARCH_EMULATE_LCONV_TO_R8 1
 #define MONO_ARCH_EMULATE_LCONV_TO_R4 1
 #define MONO_ARCH_EMULATE_LCONV_TO_R8_UN 1

scripts/ci/run-test-testing_aot_full.sh

@@ -10,6 +10,11 @@ then
 ${TESTCMD} --label=mini --timeout=25m make -j ${CI_CPU_COUNT} -w -C mono/mini -k llvmonlycheck
 else
 ${TESTCMD} --label=mini --timeout=25m make -j ${CI_CPU_COUNT} -w -C mono/mini -k fullaotcheck
+if [[ ${CI_TAGS} == *'_llvm'* ]]; then
+	${TESTCMD} --label=mini-aotcheck --timeout=25m make -j ${CI_CPU_COUNT} -w -C mono/mini -k llvmaotcheck
+	# FIXME: https://github.com/mono/mono/issues/15999
+	# ${TESTCMD} --label=mini-aotcheck --timeout=25m make -j ${CI_CPU_COUNT} -w -C mono/mini -k llvmfullaotcheck
+fi
 fi
 
 ${TESTCMD} --label=runtime --timeout=160m make -w -C mono/tests -k test-wrench V=1