[mono][llvm] Improve LLVM 9 compatibility. (#34182)

Improve LLVM 9 compatibility.

Several intrinsics were removed in LLVM 7 and above; replace them with
IR sequences that generate the desired instructions.

Fix some type mismatch problems with the generated IR (in
Sse2.StoreAlignedNonTemporal, Vector128.CreateScalarUnsafe, and
Sse2.MaskMove); these were caught by running the coreclr hardware
intrinsic JIT tests.

Support the two-argument forms of Sse.SqrtScalar, Sse.ReciprocalScalar,
and Sse.ReciprocalSqrtScalar.

llvm::linkCoreCLRGC(), in llvm/CodeGen/GCs.h, was removed and replaced with
llvm::linkAllBuiltinGCs(), in llvm/CodeGen/BuiltinGCs.h.

In mono_llvm_dump_value and mono_llvm_dump_module, flush the c library's
stdout stream before using (and flushing) LLVM's own buffered output
functions.

src/mono/mono/mini/llvm-intrinsics.h

@@ -88,17 +88,25 @@ INTRINS(SSE_PSRL_Q, x86_sse2_psrl_q)
 INTRINS(SSE_PSLL_W, x86_sse2_psll_w)
 INTRINS(SSE_PSLL_D, x86_sse2_psll_d)
 INTRINS(SSE_PSLL_Q, x86_sse2_psll_q)
+#if LLVM_API_VERSION < 700
+// These intrinsics were removed in LLVM 7 (bcaab53d479e7005ee69e06321bbb493f9b7f5e6).
 INTRINS(SSE_SQRT_PS, x86_sse_sqrt_ps)
+INTRINS(SSE_SQRT_SS, x86_sse_sqrt_ss)
+INTRINS(SSE_SQRT_PD, x86_sse2_sqrt_pd)
+INTRINS(SSE_SQRT_SD, x86_sse2_sqrt_sd)
+INTRINS(SSE_PMULUDQ, x86_sse2_pmulu_dq)
+#else
+INTRINS_OVR(SSE_SQRT_PD, sqrt)
+INTRINS_OVR(SSE_SQRT_PS, sqrt)
+INTRINS_OVR(SSE_SQRT_SD, sqrt)
+INTRINS_OVR(SSE_SQRT_SS, sqrt)
+#endif
 INTRINS(SSE_RCP_PS, x86_sse_rcp_ps)
 INTRINS(SSE_RSQRT_PS, x86_sse_rsqrt_ps)
-INTRINS(SSE_SQRT_SS, x86_sse_sqrt_ss)
 INTRINS(SSE_RCP_SS, x86_sse_rcp_ss)
 INTRINS(SSE_RSQRT_SS, x86_sse_rsqrt_ss)
-INTRINS(SSE_SQRT_PD, x86_sse2_sqrt_pd)
-INTRINS(SSE_SQRT_SD, x86_sse2_sqrt_sd)
 INTRINS(SSE_CVTTPD2DQ, x86_sse2_cvttpd2dq)
 INTRINS(SSE_CVTTPS2DQ, x86_sse2_cvttps2dq)
-INTRINS(SSE_CVTDQ2PS, x86_sse2_cvtdq2ps)
 INTRINS(SSE_CVTPD2DQ, x86_sse2_cvtpd2dq)
 INTRINS(SSE_CVTPS2DQ, x86_sse2_cvtps2dq)
 INTRINS(SSE_CVTPD2PS, x86_sse2_cvtpd2ps)
@@ -110,10 +118,6 @@ INTRINS(SSE_CVTSD2SI, x86_sse2_cvtsd2si)
 INTRINS(SSE_CVTTSD2SI, x86_sse2_cvttsd2si)
 INTRINS(SSE_CVTSD2SI64, x86_sse2_cvtsd2si64)
 INTRINS(SSE_CVTTSD2SI64, x86_sse2_cvttsd2si64)
-INTRINS(SSE_CVTSI2SS, x86_sse_cvtsi2ss)
-INTRINS(SSE_CVTSI2SS64, x86_sse_cvtsi642ss)
-INTRINS(SSE_CVTSI2SD, x86_sse2_cvtsi2sd)
-INTRINS(SSE_CVTSI2SD64, x86_sse2_cvtsi642sd)
 INTRINS(SSE_CVTSD2SS, x86_sse2_cvtsd2ss)
 INTRINS(SSE_CMPPD, x86_sse2_cmp_pd)
 INTRINS(SSE_CMPPS, x86_sse_cmp_ps)
@@ -161,19 +165,10 @@ INTRINS(SSE_MINSD, x86_sse2_min_sd)
 INTRINS(SSE_HADDPD, x86_sse3_hadd_pd)
 INTRINS(SSE_HSUBPD, x86_sse3_hsub_pd)
 INTRINS(SSE_ADDSUBPD, x86_sse3_addsub_pd)
-INTRINS(SSE_PADDSW, x86_sse2_padds_w)
-INTRINS(SSE_PSUBSW, x86_sse2_psubs_w)
-INTRINS(SSE_PADDUSW, x86_sse2_paddus_w)
-INTRINS(SSE_PSUBUSW, x86_sse2_psubus_w)
 INTRINS(SSE_PMULHW, x86_sse2_pmulh_w)
 INTRINS(SSE_PMULHU, x86_sse2_pmulhu_w)
-INTRINS(SSE_PMULUDQ, x86_sse2_pmulu_dq)
 INTRINS(SSE_PMULHUW, x86_sse2_pmulhu_w)
 INTRINS(SSE_PMADDWD, x86_sse2_pmadd_wd)
-INTRINS(SSE_PADDSB, x86_sse2_padds_b)
-INTRINS(SSE_PSUBSB, x86_sse2_psubs_b)
-INTRINS(SSE_PADDUSB, x86_sse2_paddus_b)
-INTRINS(SSE_PSUBUSB, x86_sse2_psubus_b)
 INTRINS(SSE_PSADBW, x86_sse2_psad_bw)
 INTRINS(SSE_PAUSE, x86_sse2_pause)
 INTRINS(SSE_MASKMOVDQU, x86_sse2_maskmov_dqu)
@@ -208,11 +203,21 @@ INTRINS_OVR(SSE_SADD_SATI8, sadd_sat)
 INTRINS_OVR(SSE_UADD_SATI8, uadd_sat)
 INTRINS_OVR(SSE_SADD_SATI16, sadd_sat)
 INTRINS_OVR(SSE_UADD_SATI16, uadd_sat)
+
+INTRINS_OVR(SSE_SSUB_SATI8, ssub_sat)
+INTRINS_OVR(SSE_USUB_SATI8, usub_sat)
+INTRINS_OVR(SSE_SSUB_SATI16, ssub_sat)
+INTRINS_OVR(SSE_USUB_SATI16, usub_sat)
 #else
 INTRINS(SSE_SADD_SATI8, x86_sse2_padds_b)
 INTRINS(SSE_UADD_SATI8, x86_sse2_paddus_b)
 INTRINS(SSE_SADD_SATI16, x86_sse2_padds_w)
 INTRINS(SSE_UADD_SATI16, x86_sse2_paddus_w)
+
+INTRINS(SSE_SSUB_SATI8, x86_sse2_psubs_b)
+INTRINS(SSE_USUB_SATI8, x86_sse2_psubus_b)
+INTRINS(SSE_SSUB_SATI16, x86_sse2_psubs_w)
+INTRINS(SSE_USUB_SATI16, x86_sse2_psubus_w)
 #endif
 #endif
 #if defined(TARGET_WASM) && LLVM_API_VERSION >= 800

src/mono/mono/mini/llvm-jit.cpp

@@ -33,7 +33,12 @@
 #include "llvm/ExecutionEngine/Orc/RTDyldObjectLinkingLayer.h"
 #include "llvm/ExecutionEngine/JITSymbol.h"
 #include "llvm/Transforms/Scalar.h"
+
+#if LLVM_API_VERSION >= 800
+#include "llvm/CodeGen/BuiltinGCs.h"
+#else
 #include "llvm/CodeGen/GCs.h"
+#endif
 
 #include <cstdlib>
 
@@ -50,6 +55,17 @@ mono_llvm_set_unhandled_exception_handler (void)
 {
 }
 
+// noop function that merely ensures that certain symbols are not eliminated
+// from the resulting binary.
+static void
+link_gc () {
+#if LLVM_API_VERSION >= 800
+	llvm::linkAllBuiltinGCs();
+#else
+	llvm::linkCoreCLRGC(); // Mono uses built-in "coreclr" GCStrategy
+#endif
+}
+
 template <typename T>
 static std::vector<T> singletonSet(T t) {
   std::vector<T> Vec;
@@ -304,7 +320,6 @@ class MonoLLVMJIT {
 		initializeInstCombine(registry);
 		initializeTarget(registry);
 		initializeLoopIdiomRecognizeLegacyPassPass(registry);
-		linkCoreCLRGC(); // Mono uses built-in "coreclr" GCStrategy
 
 		// FIXME: find optimal mono specific order of passes
 		// see https://llvm.org/docs/Frontend/PerformanceTips.html#pass-ordering
@@ -445,6 +460,8 @@ mono_llvm_jit_init ()
 {
 	if (jit != nullptr) return;
 
+	link_gc ();
+
 	mono_native_tls_alloc (&current_cfg_tls_id, NULL);
 
 	InitializeNativeTarget ();

src/mono/mono/mini/mini-llvm-cpp.cpp

@@ -51,16 +51,18 @@ void
 mono_llvm_dump_value (LLVMValueRef value)
 {
 	/* Same as LLVMDumpValue (), but print to stdout */
-	outs () << (*unwrap<Value> (value)) << "\n";
 	fflush (stdout);
+	outs () << (*unwrap<Value> (value)) << "\n";
+	outs ().flush ();
 }
 
 void
 mono_llvm_dump_module (LLVMModuleRef module)
 {
 	/* Same as LLVMDumpModule (), but print to stdout */
-	outs () << (*unwrap (module));
 	fflush (stdout);
+	outs () << (*unwrap (module));
+	outs ().flush ();
 }
 
 /* Missing overload for building an alloca with an alignment */