diff --git a/src/mono/mono/mini/calls.c b/src/mono/mono/mini/calls.c
index a7d66f5ecf34ef..3689d8063897fd 100644
--- a/src/mono/mono/mini/calls.c
+++ b/src/mono/mono/mini/calls.c
@@ -111,7 +111,10 @@ ret_type_to_call_opcode (MonoCompile *cfg, MonoType *type, int calli, int virt)
 	case MONO_TYPE_U8:
 		return calli? OP_LCALL_REG: virt? OP_LCALL_MEMBASE: OP_LCALL;
 	case MONO_TYPE_R4:
-		return calli? OP_RCALL_REG: virt? OP_RCALL_MEMBASE: OP_RCALL;
+		if (cfg->r4fp)
+			return calli? OP_RCALL_REG: virt? OP_RCALL_MEMBASE: OP_RCALL;
+		else
+			return calli? OP_FCALL_REG: virt? OP_FCALL_MEMBASE: OP_FCALL;
 	case MONO_TYPE_R8:
 		return calli? OP_FCALL_REG: virt? OP_FCALL_MEMBASE: OP_FCALL;
 	case MONO_TYPE_VALUETYPE:
diff --git a/src/mono/mono/mini/intrinsics.c b/src/mono/mono/mini/intrinsics.c
index f19efd3c80ea76..e003caf4d24f3a 100644
--- a/src/mono/mono/mini/intrinsics.c
+++ b/src/mono/mono/mini/intrinsics.c
@@ -103,7 +103,7 @@ llvm_emit_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSign
 	MonoInst *ins = NULL;
 	int opcode = 0;
 	// Convert Math and MathF methods into LLVM intrinsics, e.g. MathF.Sin -> @llvm.sin.f32
-	if (in_corlib && !strcmp (m_class_get_name (cmethod->klass), "MathF")) {
+	if (in_corlib && !strcmp (m_class_get_name (cmethod->klass), "MathF") && cfg->r4fp) {
 		// (float)
 		if (fsig->param_count == 1 && fsig->params [0]->type == MONO_TYPE_R4) {
 			if (!strcmp (cmethod->name, "Ceiling")) {
@@ -1524,7 +1524,7 @@ mini_emit_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSign
 #endif
 				break;
 			case MONO_TYPE_R4:
-				ins->type = GINT_TO_UINT8 (STACK_R4);
+				ins->type = GINT_TO_UINT8 (cfg->r4_stack_type);
 				break;
 			case MONO_TYPE_R8:
 				ins->type = STACK_R8;
@@ -1662,7 +1662,7 @@ mini_emit_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSign
 #endif
 					break;
 				case MONO_TYPE_R4:
-					ins->type = GINT_TO_UINT8 (STACK_R4);
+					ins->type = GINT_TO_UINT8 (cfg->r4_stack_type);
 					break;
 				case MONO_TYPE_R8:
 					ins->type = STACK_R8;
diff --git a/src/mono/mono/mini/method-to-ir.c b/src/mono/mono/mini/method-to-ir.c
index f39e35642268b0..6578497e745008 100644
--- a/src/mono/mono/mini/method-to-ir.c
+++ b/src/mono/mono/mini/method-to-ir.c
@@ -337,7 +337,7 @@ mono_type_to_regmove (MonoCompile *cfg, MonoType *type)
 		return OP_LMOVE;
 #endif
 	case MONO_TYPE_R4:
-		return OP_RMOVE;
+		return cfg->r4fp ? OP_RMOVE : OP_FMOVE;
 	case MONO_TYPE_R8:
 		return OP_FMOVE;
 	case MONO_TYPE_VALUETYPE:
@@ -483,8 +483,9 @@ add_widen_op (MonoCompile *cfg, MonoInst *ins, MonoInst **arg1_ref, MonoInst **a
 	MonoInst *arg1 = *arg1_ref;
 	MonoInst *arg2 = *arg2_ref;
 
-	if ((arg1->type == STACK_R4 && arg2->type == STACK_R8) ||
-		(arg1->type == STACK_R8 && arg2->type == STACK_R4)) {
+	if (cfg->r4fp &&
+		((arg1->type == STACK_R4 && arg2->type == STACK_R8) ||
+		 (arg1->type == STACK_R8 && arg2->type == STACK_R4))) {
 		MonoInst *conv;
 
 		/* Mixing r4/r8 is allowed by the spec */
@@ -826,7 +827,7 @@ mini_type_to_eval_stack_type (MonoCompile *cfg, MonoType *type, MonoInst *inst)
 		inst->type = STACK_I8;
 		return;
 	case MONO_TYPE_R4:
-		inst->type = GINT_TO_UINT8 (STACK_R4);
+		inst->type = GINT_TO_UINT8 (cfg->r4_stack_type);
 		break;
 	case MONO_TYPE_R8:
 		inst->type = STACK_R8;
@@ -1160,7 +1161,7 @@ type_from_op (MonoCompile *cfg, MonoInst *ins, MonoInst *src1, MonoInst *src2)
 		ins->opcode += ovf2ops_op_map [src1->type];
 		break;
 	case MONO_CEE_CONV_R4:
-		ins->type = GINT_TO_UINT8 (STACK_R4);
+		ins->type = GINT_TO_UINT8 (cfg->r4_stack_type);
 		ins->opcode += unops_op_map [src1->type];
 		break;
 	case MONO_CEE_CONV_R8:
@@ -1218,7 +1219,7 @@ type_from_op (MonoCompile *cfg, MonoInst *ins, MonoInst *src1, MonoInst *src2)
 		ins->type = STACK_I8;
 		break;
 	case OP_LOADR4_MEMBASE:
-		ins->type = GINT_TO_UINT8 (STACK_R4);
+		ins->type = GINT_TO_UINT8 (cfg->r4_stack_type);
 		break;
 	case OP_LOADR8_MEMBASE:
 		ins->type = STACK_R8;
@@ -1433,7 +1434,7 @@ mini_type_to_stack_type (MonoCompile *cfg, MonoType *t)
 	case MONO_TYPE_U8:
 		return STACK_I8;
 	case MONO_TYPE_R4:
-		return (MonoStackType)STACK_R4;
+		return (MonoStackType)cfg->r4_stack_type;
 	case MONO_TYPE_R8:
 		return STACK_R8;
 	case MONO_TYPE_VALUETYPE:
@@ -1916,7 +1917,7 @@ target_type_is_incompatible (MonoCompile *cfg, MonoType *target, MonoInst *arg)
 				return 1;
 		return 0;
 	case MONO_TYPE_R4:
-		if (arg->type != STACK_R4)
+		if (arg->type != cfg->r4_stack_type)
 			return 1;
 		return 0;
 	case MONO_TYPE_R8:
@@ -1979,6 +1980,8 @@ target_type_is_incompatible (MonoCompile *cfg, MonoType *target, MonoInst *arg)
 static MonoInst*
 convert_value (MonoCompile *cfg, MonoType *type, MonoInst *ins)
 {
+	if (!cfg->r4fp)
+		return ins;
 	type = mini_get_underlying_type (type);
 	switch (type->type) {
 	case MONO_TYPE_R4:
@@ -2071,7 +2074,7 @@ check_call_signature (MonoCompile *cfg, MonoMethodSignature *sig, MonoInst **arg
 				return TRUE;
 			continue;
 		case MONO_TYPE_R4:
-			if (args [i]->type != STACK_R4)
+			if (args [i]->type != cfg->r4_stack_type)
 				return TRUE;
 			continue;
 		case MONO_TYPE_R8:
@@ -4636,7 +4639,7 @@ mini_emit_init_rvar (MonoCompile *cfg, int dreg, MonoType *rtype)
 		MONO_EMIT_NEW_ICONST (cfg, dreg, 0);
 	} else if (t == MONO_TYPE_I8 || t == MONO_TYPE_U8) {
 		MONO_EMIT_NEW_I8CONST (cfg, dreg, 0);
-	} else if (t == MONO_TYPE_R4) {
+	} else if (cfg->r4fp && t == MONO_TYPE_R4) {
 		MONO_INST_NEW (cfg, ins, OP_R4CONST);
 		ins->type = STACK_R4;
 		ins->inst_p0 = (void*)&r4_0;
@@ -4672,7 +4675,7 @@ emit_dummy_init_rvar (MonoCompile *cfg, int dreg, MonoType *rtype)
 		MONO_EMIT_NEW_DUMMY_INIT (cfg, dreg, OP_DUMMY_ICONST);
 	} else if (t == MONO_TYPE_I8 || t == MONO_TYPE_U8) {
 		MONO_EMIT_NEW_DUMMY_INIT (cfg, dreg, OP_DUMMY_I8CONST);
-	} else if (t == MONO_TYPE_R4) {
+	} else if (cfg->r4fp && t == MONO_TYPE_R4) {
 		MONO_EMIT_NEW_DUMMY_INIT (cfg, dreg, OP_DUMMY_R4CONST);
 	} else if (t == MONO_TYPE_R4 || t == MONO_TYPE_R8) {
 		MONO_EMIT_NEW_DUMMY_INIT (cfg, dreg, OP_DUMMY_R8CONST);
@@ -7278,10 +7281,10 @@ mono_method_to_ir (MonoCompile *cfg, MonoMethod *method, MonoBasicBlock *start_b
 
 				dreg = alloc_freg (cfg);
 				EMIT_NEW_LOAD_MEMBASE (cfg, ins, OP_LOADR4_MEMBASE, dreg, cons->dreg, 0);
-				ins->type = GINT_TO_UINT8 (STACK_R4);
+				ins->type = GINT_TO_UINT8 (cfg->r4_stack_type);
 			} else {
 				MONO_INST_NEW (cfg, ins, OP_R4CONST);
-				ins->type = GINT_TO_UINT8 (STACK_R4);
+				ins->type = GINT_TO_UINT8 (cfg->r4_stack_type);
 				ins->dreg = alloc_dreg (cfg, STACK_R8);
 				ins->inst_p0 = f;
 				MONO_ADD_INS (cfg->cbb, ins);
diff --git a/src/mono/mono/mini/mini-amd64.h b/src/mono/mono/mini/mini-amd64.h
index 46e2828f868ed0..80c570e58f5bea 100644
--- a/src/mono/mono/mini/mini-amd64.h
+++ b/src/mono/mono/mini/mini-amd64.h
@@ -453,6 +453,7 @@ typedef struct {
 #define MONO_ARCH_HAVE_SDB_TRAMPOLINES 1
 #define MONO_ARCH_HAVE_OP_GENERIC_CLASS_INIT 1
 #define MONO_ARCH_HAVE_GENERAL_RGCTX_LAZY_FETCH_TRAMPOLINE 1
+#define MONO_ARCH_FLOAT32_SUPPORTED 1
 #define MONO_ARCH_LLVM_TARGET_LAYOUT "e-i64:64-i128:128-n8:16:32:64-S128"
 
 #define MONO_ARCH_HAVE_INTERP_PINVOKE_TRAMP
diff --git a/src/mono/mono/mini/mini-arm.h b/src/mono/mono/mini/mini-arm.h
index 13b94cd8c8293c..73bcc8bb8bb4b8 100644
--- a/src/mono/mono/mini/mini-arm.h
+++ b/src/mono/mono/mini/mini-arm.h
@@ -363,6 +363,7 @@ typedef struct MonoCompileArch {
 #define MONO_ARCH_HAVE_OBJC_GET_SELECTOR 1
 #define MONO_ARCH_HAVE_SDB_TRAMPOLINES 1
 #define MONO_ARCH_HAVE_OP_GENERIC_CLASS_INIT 1
+#define MONO_ARCH_FLOAT32_SUPPORTED 1
 #define MONO_ARCH_LLVM_TARGET_LAYOUT "e-p:32:32-n32-S64"
 
 #define MONO_ARCH_HAVE_INTERP_ENTRY_TRAMPOLINE 1
diff --git a/src/mono/mono/mini/mini-arm64.h b/src/mono/mono/mini/mini-arm64.h
index 3b9218b32c0b1a..17195886b76725 100644
--- a/src/mono/mono/mini/mini-arm64.h
+++ b/src/mono/mono/mini/mini-arm64.h
@@ -177,6 +177,7 @@ typedef struct {
 #define MONO_ARCH_HAVE_OP_GENERIC_CLASS_INIT 1
 #define MONO_ARCH_HAVE_OPCODE_NEEDS_EMULATION 1
 #define MONO_ARCH_HAVE_DECOMPOSE_LONG_OPTS 1
+#define MONO_ARCH_FLOAT32_SUPPORTED 1
 #define MONO_ARCH_HAVE_INTERP_PINVOKE_TRAMP 1
 #define MONO_ARCH_HAVE_INIT_MRGCTX 1
 #define MONO_ARCH_LLVM_TARGET_LAYOUT "e-i64:64-i128:128-n32:64-S128"
diff --git a/src/mono/mono/mini/mini-s390x.h b/src/mono/mono/mini/mini-s390x.h
index 8dafec4aefb560..b9deb1a2683d63 100644
--- a/src/mono/mono/mini/mini-s390x.h
+++ b/src/mono/mono/mini/mini-s390x.h
@@ -83,6 +83,7 @@ struct SeqPointInfo {
 #define MONO_ARCH_HAVE_SDB_TRAMPOLINES			1
 #define MONO_ARCH_HAVE_SETUP_RESUME_FROM_SIGNAL_HANDLER_CTX	1
 #define MONO_ARCH_HAVE_UNWIND_BACKTRACE 		1
+#define MONO_ARCH_FLOAT32_SUPPORTED			1
 
 #define S390_STACK_ALIGNMENT		 8
 #define S390_FIRST_ARG_REG 		s390_r2
diff --git a/src/mono/mono/mini/mini-wasm.h b/src/mono/mono/mini/mini-wasm.h
index f8234087ea8c05..61e81e8a81fbdb 100644
--- a/src/mono/mono/mini/mini-wasm.h
+++ b/src/mono/mono/mini/mini-wasm.h
@@ -29,6 +29,7 @@
 #define MONO_ARCH_EMULATE_FCONV_TO_U4 1
 #define MONO_ARCH_NO_EMULATE_LONG_SHIFT_OPS 1
 #define MONO_ARCH_NO_EMULATE_LONG_MUL_OPTS 1
+#define MONO_ARCH_FLOAT32_SUPPORTED 1
 
 //mini-codegen stubs - this doesn't do anything
 #define MONO_ARCH_CALLEE_REGS (1 << 0)
diff --git a/src/mono/mono/mini/mini-x86.h b/src/mono/mono/mini/mini-x86.h
index 7235aa7781f47e..ed839af690732e 100644
--- a/src/mono/mono/mini/mini-x86.h
+++ b/src/mono/mono/mini/mini-x86.h
@@ -229,6 +229,7 @@ typedef struct {
 #define MONO_ARCH_HAVE_OP_TAILCALL_REG 1
 #define MONO_ARCH_HAVE_SDB_TRAMPOLINES 1
 #define MONO_ARCH_LLVM_TARGET_LAYOUT "e-p:32:32-n32-S128"
+#define MONO_ARCH_FLOAT32_SUPPORTED 1
 #define MONO_ARCH_NEED_SIMD_BANK 1
 #define MONO_ARCH_USE_SHARED_FP_SIMD_BANK 1
 
diff --git a/src/mono/mono/mini/mini.c b/src/mono/mono/mini/mini.c
index 9ceb79caa97dc6..0c68db6c01c9ef 100644
--- a/src/mono/mono/mini/mini.c
+++ b/src/mono/mono/mini/mini.c
@@ -3133,6 +3133,14 @@ mini_method_compile (MonoMethod *method, guint32 opts, JitFlags flags, int parts
 	try_llvm = mono_use_llvm || llvm;
 #endif
 
+#ifdef MONO_ARCH_FLOAT32_SUPPORTED
+	/* Force float32 mode on platforms where its supported */
+	opts |= MONO_OPT_FLOAT32;
+#else
+	opts &= ~MONO_OPT_FLOAT32;
+	g_assert (!llvm);
+#endif
+
  restart_compile:
 	if (method_is_gshared) {
 		method_to_compile = method;
@@ -3209,6 +3217,8 @@ mini_method_compile (MonoMethod *method, guint32 opts, JitFlags flags, int parts
 
 	if (!is_simd_supported (cfg))
 		cfg->opt &= ~MONO_OPT_SIMD;
+	cfg->r4fp = (cfg->opt & MONO_OPT_FLOAT32) ? 1 : 0;
+	cfg->r4_stack_type = cfg->r4fp ? STACK_R4 : STACK_R8;
 
 	if (cfg->gen_seq_points)
 		cfg->seq_points = g_ptr_array_new ();
diff --git a/src/mono/mono/mini/mini.h b/src/mono/mono/mini/mini.h
index 32d057bd18fb83..a21e3c331af3e0 100644
--- a/src/mono/mono/mini/mini.h
+++ b/src/mono/mono/mini/mini.h
@@ -1501,6 +1501,7 @@ typedef struct {
 	guint            gshared : 1;
 	guint            gsharedvt : 1;
 	guint            gsharedvt_min : 1;
+	guint            r4fp : 1;
 	guint            llvm_only : 1;
 	guint            interp : 1;
 	guint            use_current_cpu : 1;
@@ -1512,6 +1513,7 @@ typedef struct {
 	guint            deopt : 1;
 	guint            prefer_instances : 1;
 	guint8           uses_simd_intrinsics;
+	int              r4_stack_type;
 	gpointer         debug_info;
 	guint32          lmf_offset;
 	guint16          *intvars;