From d89493a9617fe30c987b0d5903c98ddcd3319613 Mon Sep 17 00:00:00 2001
From: Zoltan Varga <vargaz@gmail.com>
Date: Wed, 25 Oct 2023 14:12:02 -0700
Subject: [PATCH] Implement gsharedvt-in support.

---
 src/mono/mono/mini/mini-amd64-gsharedvt.c  | 34 +++++++++++++---------
 src/mono/mono/mini/mini-amd64-gsharedvt.h  |  1 +
 src/mono/mono/mini/tramp-amd64-gsharedvt.c |  7 +++--
 3 files changed, 26 insertions(+), 16 deletions(-)

diff --git a/src/mono/mono/mini/mini-amd64-gsharedvt.c b/src/mono/mono/mini/mini-amd64-gsharedvt.c
index 0aa2f1559006b3..d839cf858276f6 100644
--- a/src/mono/mono/mini/mini-amd64-gsharedvt.c
+++ b/src/mono/mono/mini/mini-amd64-gsharedvt.c
@@ -163,6 +163,7 @@ get_arg_slots (ArgInfo *ainfo, int **out_slots, gboolean is_source_argument)
 		break;
 	case ArgInDoubleSSEReg:
 	case ArgInFloatSSEReg:
+	case ArgSIMDInSSEReg:
 		nsrc = 1;
 		src = g_malloc (nsrc * sizeof (int));
 		src [0] = map_freg (sreg);
@@ -224,20 +225,21 @@ static void
 handle_marshal_when_dst_gsharedvt (ArgInfo *src_info, int *arg_marshal)
 {
 	switch (src_info->storage) {
-		case ArgInIReg:
-		case ArgInDoubleSSEReg:
-		case ArgInFloatSSEReg:
-		case ArgValuetypeInReg:
-		case ArgOnStack:
-			*arg_marshal = GSHAREDVT_ARG_BYVAL_TO_BYREF;
-			break;
-		case ArgValuetypeAddrInIReg:
-		case ArgValuetypeAddrOnStack:
-			*arg_marshal = GSHAREDVT_ARG_NONE;
-			break;
-		default:
-			NOT_IMPLEMENTED; // See above
-			break;
+	case ArgInIReg:
+	case ArgInDoubleSSEReg:
+	case ArgInFloatSSEReg:
+	case ArgSIMDInSSEReg:
+	case ArgValuetypeInReg:
+	case ArgOnStack:
+		*arg_marshal = GSHAREDVT_ARG_BYVAL_TO_BYREF;
+		break;
+	case ArgValuetypeAddrInIReg:
+	case ArgValuetypeAddrOnStack:
+		*arg_marshal = GSHAREDVT_ARG_NONE;
+		break;
+	default:
+		NOT_IMPLEMENTED; // See above
+		break;
 	}
 }
 
@@ -331,6 +333,7 @@ mono_arch_get_gsharedvt_call_info (MonoMemoryManager *mem_manager, gpointer addr
 		case ArgInIReg:
 		case ArgInDoubleSSEReg:
 		case ArgInFloatSSEReg:
+		case ArgSIMDInSSEReg:
 		case ArgValuetypeInReg:
 		case ArgOnStack:
 			nsrc = get_arg_slots (src_info, &src, TRUE);
@@ -500,6 +503,9 @@ mono_arch_get_gsharedvt_call_info (MonoMemoryManager *mem_manager, gpointer addr
 		case ArgInFloatSSEReg:
 			info->ret_marshal = GSHAREDVT_RET_R8;
 			break;
+		case ArgSIMDInSSEReg:
+			info->ret_marshal = GSHAREDVT_RET_SIMD;
+			break;
 		case ArgValuetypeAddrInIReg:
 			break;
 		default:
diff --git a/src/mono/mono/mini/mini-amd64-gsharedvt.h b/src/mono/mono/mini/mini-amd64-gsharedvt.h
index 539dff9ff51e16..fea5d385b7d968 100644
--- a/src/mono/mono/mini/mini-amd64-gsharedvt.h
+++ b/src/mono/mono/mini/mini-amd64-gsharedvt.h
@@ -37,6 +37,7 @@ typedef enum {
 	GSHAREDVT_RET_I8,      // 8 byte integer
 	GSHAREDVT_RET_IREGS_1, // Load in first return register
 	GSHAREDVT_RET_R8,     // Double
+	GSHAREDVT_RET_SIMD,     // SIMD
 	GSHAREDVT_RET_NUM,
 } GSharedVtRetMarshal;
 
diff --git a/src/mono/mono/mini/tramp-amd64-gsharedvt.c b/src/mono/mono/mini/tramp-amd64-gsharedvt.c
index 43b575b9f3d7dd..57860c794829d5 100644
--- a/src/mono/mono/mini/tramp-amd64-gsharedvt.c
+++ b/src/mono/mono/mini/tramp-amd64-gsharedvt.c
@@ -224,7 +224,7 @@ mono_arch_get_gsharedvt_trampoline (MonoTrampInfo **info, gboolean aot)
 
 	/*callconv in regs */
 	caller_reg_area_offset = offset;
-	reg_area_size = ALIGN_TO ((n_arg_regs + n_arg_fregs) * 8, MONO_ARCH_FRAME_ALIGNMENT);
+	reg_area_size = ALIGN_TO ((n_arg_regs * 8) + (n_arg_fregs * 16), MONO_ARCH_FRAME_ALIGNMENT);
 	offset += reg_area_size;
 
 	framesize = offset;
@@ -266,7 +266,7 @@ mono_arch_get_gsharedvt_trampoline (MonoTrampInfo **info, gboolean aot)
 		amd64_mov_membase_reg (code, AMD64_RSP, caller_reg_area_offset + i * 8, param_regs [i], sizeof (target_mgreg_t));
 
 	for (i = 0; i < n_arg_fregs; ++i)
-		amd64_sse_movsd_membase_reg (code, AMD64_RSP, caller_reg_area_offset + (i + n_arg_regs) * 8, i);
+		amd64_sse_movups_membase_reg (code, AMD64_RSP, caller_reg_area_offset + (n_arg_regs * 8) + (i * 16), i);
 
 	/* TODO Allocate stack area used to pass arguments to the method */
 
@@ -411,6 +411,9 @@ mono_arch_get_gsharedvt_trampoline (MonoTrampInfo **info, gboolean aot)
 		case GSHAREDVT_RET_R8:
 			amd64_sse_movsd_reg_membase (code, AMD64_XMM0, AMD64_R11, 0);
 			break;
+		case GSHAREDVT_RET_SIMD:
+			amd64_sse_movups_reg_membase (code, AMD64_XMM0, AMD64_R11, 0);
+			break;
 		default:
 			x86_breakpoint (code); /* can't handle specific case */
 		}