Fix OP_FMAX an OP_FMIN on amd64 and lift the arch limitation (#65445)

dotnet · Feb 17, 2022 · 158f089 · 158f089
1 parent f8fb217
commit 158f089
Show file tree

Hide file tree

Showing 2 changed files with 28 additions and 33 deletions.
diff --git a/src/mono/mono/mini/mini-llvm.c b/src/mono/mono/mini/mini-llvm.c
@@ -7718,24 +7718,33 @@ process_bb (EmitContext *ctx, MonoBasicBlock *bb)
 				LLVMValueRef args [] = { l, r };
 #if defined(TARGET_X86) || defined(TARGET_AMD64)
 				LLVMTypeRef t = LLVMTypeOf (l);
+				LLVMTypeRef elem_t = LLVMGetElementType (t);
 				unsigned int elems = LLVMGetVectorSize (t);
-				gboolean is_r4 = ins->inst_c1 == MONO_TYPE_R4;
-				int iid = -1;
-				if (ins->inst_c0 == OP_FMAX) {
-					if (elems == 1)
-						iid = is_r4 ? INTRINS_SSE_MAXSS : INTRINS_SSE_MAXSD;
-					else
-						iid = is_r4 ? INTRINS_SSE_MAXPS : INTRINS_SSE_MAXPD;
+				unsigned int elem_bits = mono_llvm_get_prim_size_bits (elem_t);
+				unsigned int v_size = elems * elem_bits;
+				if (v_size == 128) {
+					gboolean is_r4 = ins->inst_c1 == MONO_TYPE_R4;
+					int iid = -1;
+					if (ins->inst_c0 == OP_FMAX) {
+						if (elems == 1)
+							iid = is_r4 ? INTRINS_SSE_MAXSS : INTRINS_SSE_MAXSD;
+						else
+							iid = is_r4 ? INTRINS_SSE_MAXPS : INTRINS_SSE_MAXPD;
+					} else {
+						if (elems == 1)
+							iid = is_r4 ? INTRINS_SSE_MINSS : INTRINS_SSE_MINSD;
+						else
+							iid = is_r4 ? INTRINS_SSE_MINPS : INTRINS_SSE_MINPD;
+					}
+					result = call_intrins (ctx, iid, args, dname);
 				} else {
-					if (elems == 1)
-						iid = is_r4 ? INTRINS_SSE_MINSS : INTRINS_SSE_MINSD;
-					else
-						iid = is_r4 ? INTRINS_SSE_MINPS : INTRINS_SSE_MINPD;
+					LLVMRealPredicate op = ins->inst_c0 == OP_FMAX ? LLVMRealUGE : LLVMRealULE;
+					LLVMValueRef cmp = LLVMBuildFCmp (builder, op, l, r, "");
+					result = LLVMBuildSelect (builder, cmp, l, r, "");
 				}
-				result = call_intrins (ctx, iid, args, dname);
+
 #elif defined(TARGET_ARM64)
-				int instc0_arm64 = ins->inst_c0 == OP_FMAX ? INTRINS_AARCH64_ADV_SIMD_FMAX : INTRINS_AARCH64_ADV_SIMD_FMIN;
-				IntrinsicId iid = (IntrinsicId) instc0_arm64;
+				IntrinsicId iid = ins->inst_c0 == OP_FMAX ? INTRINS_AARCH64_ADV_SIMD_FMAX : INTRINS_AARCH64_ADV_SIMD_FMIN;
 				llvm_ovr_tag_t ovr_tag = ovr_tag_from_mono_vector_class (ins->klass);
 				result = call_overloaded_intrins (ctx, iid, ovr_tag, args, "");
 #else
@@ -7762,18 +7771,17 @@ process_bb (EmitContext *ctx, MonoBasicBlock *bb)
 					LLVMValueRef cmp = LLVMBuildICmp (builder, op, l, r, "");
 					result = LLVMBuildSelect (builder, cmp, l, r, "");
 				} else {
-					int instc0_arm64;
+					IntrinsicId iid;
 					switch (ins->inst_c0) {
 					case OP_IMAX:
-						instc0_arm64 = is_unsigned ? INTRINS_AARCH64_ADV_SIMD_UMAX : INTRINS_AARCH64_ADV_SIMD_SMAX;
+						iid = is_unsigned ? INTRINS_AARCH64_ADV_SIMD_UMAX : INTRINS_AARCH64_ADV_SIMD_SMAX;
 						break;
 					case OP_IMIN:
-						instc0_arm64 = is_unsigned ? INTRINS_AARCH64_ADV_SIMD_UMIN : INTRINS_AARCH64_ADV_SIMD_SMIN;
+						iid = is_unsigned ? INTRINS_AARCH64_ADV_SIMD_UMIN : INTRINS_AARCH64_ADV_SIMD_SMIN;
 						break;
 					default:
 						g_assert_not_reached ();
 					}
-					IntrinsicId iid = (IntrinsicId) instc0_arm64;
 					LLVMValueRef args [] = { l, r };
 					llvm_ovr_tag_t ovr_tag = ovr_tag_from_mono_vector_class (ins->klass);
 					result = call_overloaded_intrins (ctx, iid, ovr_tag, args, "");

diff --git a/src/mono/mono/mini/simd-intrinsics.c b/src/mono/mono/mini/simd-intrinsics.c
@@ -3186,7 +3186,7 @@ mono_emit_simd_intrinsics (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSign
 	if (m_class_get_nested_in (cmethod->klass))
 		class_ns = m_class_get_name_space (m_class_get_nested_in (cmethod->klass));
 
-#if defined(TARGET_ARM64)
+#if defined(TARGET_ARM64) || defined(TARGET_AMD64)
 	if (!strcmp (class_ns, "System.Runtime.Intrinsics")) {
 		if (!strcmp (class_name, "Vector128") || !strcmp (class_name, "Vector64"))
 			return emit_sri_vector (cfg, cmethod, fsig, args);
@@ -3196,20 +3196,7 @@ mono_emit_simd_intrinsics (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSign
 		if (!strcmp (class_name, "Vector128`1") || !strcmp (class_name, "Vector64`1"))
 			return emit_vector64_vector128_t (cfg, cmethod, fsig, args);
 	}
-#endif // defined(TARGET_ARM64)
-
-// There isn't any SIMD intrinsitcs to replace with for Vector64 methods on non-arm architectures
-#if defined(TARGET_AMD64)
-	if (!strcmp (class_ns, "System.Runtime.Intrinsics")) {
-		if (!strcmp (class_name, "Vector128"))
-			return emit_sri_vector (cfg, cmethod, fsig, args);
-	}
-
-	if (!strcmp (class_ns, "System.Runtime.Intrinsics")) {
-		if (!strcmp (class_name, "Vector128`1"))
-			return emit_vector64_vector128_t (cfg, cmethod, fsig, args);
-	}
-#endif // defined(TARGET_AMD64)
+#endif // defined(TARGET_ARM64) || defined(TARGET_AMD64)
 
 	return emit_simd_intrinsics (class_ns, class_name, cfg, cmethod, fsig, args);
 }