Save and resume full 128-bits of xmm argument registers in trampoline…

…. Add definition of MonoContextSimdReg on amd64 for Windows platform and use for size/offset calculations.

mono/arch/amd64/amd64-codegen.h

@@ -105,6 +105,10 @@ typedef enum
 #define AMD64_ARGUMENT_REGS ((1<<AMD64_RDX) | (1<<AMD64_RCX) | (1<<AMD64_R8) | (1<<AMD64_R9))
 #define AMD64_IS_ARGUMENT_REG(reg) (AMD64_ARGUMENT_REGS & (1 << (reg)))
 
+/* xmm0-xmm3 for standard calling convention, additionally xmm4-xmm5 for __vectorcall (not currently used) */
+#define AMD64_ARGUMENT_XREGS ((1<<AMD64_XMM0) | (1<<AMD64_XMM1) | (1<<AMD64_XMM2) | (1<<AMD64_XMM3) | (1<<AMD64_XMM4) | (1<<AMD64_XMM5))
+#define AMD64_IS_ARGUMENT_XREG(reg) (AMD64_ARGUMENT_XREGS & (1 << (reg)))
+
 #define AMD64_CALLEE_SAVED_REGS ((1<<AMD64_RDI) | (1<<AMD64_RSI) | (1<<AMD64_RBX) | (1<<AMD64_R12) | (1<<AMD64_R13) | (1<<AMD64_R14) | (1<<AMD64_R15) | (1<<AMD64_RBP))
 #define AMD64_IS_CALLEE_SAVED_REG(reg) (AMD64_CALLEE_SAVED_REGS & (1 << (reg)))
 #else
@@ -114,6 +118,9 @@ typedef enum
 #define AMD64_ARGUMENT_REGS ((1<<AMD64_RDI) | (1<<AMD64_RSI) | (1<<AMD64_RDX) | (1<<AMD64_RCX) | (1<<AMD64_R8) | (1<<AMD64_R9))
 #define AMD64_IS_ARGUMENT_REG(reg) (AMD64_ARGUMENT_REGS & (1 << (reg)))
 
+#define AMD64_ARGUMENT_XREGS ((1<<AMD64_XMM0) | (1<<AMD64_XMM1) | (1<<AMD64_XMM2) | (1<<AMD64_XMM3) | (1<<AMD64_XMM4) | (1<<AMD64_XMM5) | (1<<AMD64_XMM6) | (1<<AMD64_XMM7))
+#define AMD64_IS_ARGUMENT_XREG(reg) (AMD64_ARGUMENT_XREGS & (1 << (reg)))
+
 #define AMD64_CALLEE_SAVED_REGS ((1<<AMD64_RBX) | (1<<AMD64_R12) | (1<<AMD64_R13) | (1<<AMD64_R14) | (1<<AMD64_R15) | (1<<AMD64_RBP))
 #define AMD64_IS_CALLEE_SAVED_REG(reg) (AMD64_CALLEE_SAVED_REGS & (1 << (reg)))
 #endif
@@ -537,6 +544,17 @@ typedef union {
 		amd64_codegen_post(inst); \
 	} while (0)
 
+#define amd64_movdqu_reg_membase(inst,reg,basereg,disp)	\
+	do {	\
+		amd64_codegen_pre(inst); \
+		x86_prefix((inst), 0xf3); \
+		amd64_emit_rex(inst, 0, (reg), 0, (basereg)); \
+		*(inst)++ = (unsigned char)0x0f;	\
+		*(inst)++ = (unsigned char)0x6f;	\
+		x86_membase_emit ((inst), (reg) & 0x7, (basereg) & 0x7, (disp));	\
+		amd64_codegen_post(inst); \
+	} while (0)
+
 #define amd64_movsd_reg_membase(inst,reg,basereg,disp)	\
 	do {	\
 		amd64_codegen_pre(inst); \
@@ -559,6 +577,17 @@ typedef union {
 		amd64_codegen_post(inst); \
 	} while (0)
 
+#define amd64_movdqu_membase_reg(inst,basereg,disp,reg)	\
+	do {	\
+		amd64_codegen_pre(inst); \
+		x86_prefix((inst), 0xf3); \
+		amd64_emit_rex(inst, 0, (reg), 0, (basereg)); \
+		*(inst)++ = (unsigned char)0x0f;	\
+		*(inst)++ = (unsigned char)0x7f;	\
+		x86_membase_emit ((inst), (reg) & 0x7, (basereg) & 0x7, (disp));	\
+		amd64_codegen_post(inst); \
+	} while (0)
+
 #define amd64_movsd_membase_reg(inst,basereg,disp,reg)	\
 	do {	\
 		amd64_codegen_pre(inst); \

mono/mini/tramp-amd64.c

@@ -360,8 +360,9 @@ mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInf
 		/* cfa = rbp + cfa_offset */
 		mono_add_unwind_op_offset (unwind_ops, code, buf, i, - cfa_offset + saved_regs_offset + (i * sizeof (mgreg_t)));
 	}
-	for (i = 0; i < 8; ++i)
-		amd64_movsd_membase_reg (code, AMD64_RBP, saved_fpregs_offset + (i * sizeof(mgreg_t)), i);
+	for (i = 0; i < AMD64_XMM_NREG; ++i)
+		if (AMD64_IS_ARGUMENT_XREG (i))
+			amd64_movdqu_membase_reg (code, AMD64_RBP, saved_fpregs_offset + (i * sizeof(MonoContextSimdReg)), i);
 
 	/* Check that the stack is aligned */
 	amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, sizeof (mgreg_t));
@@ -543,8 +544,9 @@ mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInf
 	for (i = 0; i < AMD64_NREG; ++i)
 		if (AMD64_IS_ARGUMENT_REG (i) || i == AMD64_R10 || i == AMD64_RAX)
 			amd64_mov_reg_membase (code, i, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), sizeof(mgreg_t));
-	for (i = 0; i < 8; ++i)
-		amd64_movsd_reg_membase (code, i, AMD64_RBP, saved_fpregs_offset + (i * sizeof(mgreg_t)));
+	for (i = 0; i < AMD64_XMM_NREG; ++i)
+		if (AMD64_IS_ARGUMENT_XREG (i))
+			amd64_movdqu_reg_membase (code, i, AMD64_RBP, saved_fpregs_offset + (i * sizeof(MonoContextSimdReg)));
 
 	/* Restore stack */
 #if TARGET_WIN32

mono/utils/mono-context.h

@@ -31,6 +31,9 @@ typedef struct __darwin_xmm_reg MonoContextSimdReg;
 typedef struct __darwin_xmm_reg MonoContextSimdReg;
 #elif defined(__linux__) && defined(__GLIBC__)
 typedef struct _libc_xmmreg MonoContextSimdReg;
+#elif defined(HOST_WIN32)
+#include <emmintrin.h>
+typedef __m128d MonoContextSimdReg;
 #endif
 #elif defined(TARGET_ARM64)
 typedef __uint128_t MonoContextSimdReg;
@@ -247,7 +250,7 @@ typedef struct {
 
 typedef struct {
 	mgreg_t gregs [AMD64_NREG];
-#if defined(__APPLE__) || (defined(__linux__) && defined(__GLIBC__))
+#if defined(__APPLE__) || (defined(__linux__) && defined(__GLIBC__)) || defined(HOST_WIN32)
 	MonoContextSimdReg fregs [AMD64_XMM_NREG];
 #else
 	double fregs [AMD64_XMM_NREG];