From: Rhys Weatherley <rweather@southern-storm.com.au>
Date: Fri, 11 Jun 2004 03:48:00 +0000 (+0000)
Subject: Optimize the x86 function epilog when we are certain that the
X-Git-Tag: r.0.0.4~17
X-Git-Url: https://git.unchartedbackwaters.co.uk/w/?a=commitdiff_plain;h=4b1a2d8f4c86b27e2adc9768e9ed352d2ce6551e;p=francis%2Flibjit.git

Optimize the x86 function epilog when we are certain that the
stack height doesn't change between entry and exit (i.e. the
function is a leaf and there are no alloca's).
---

diff --git a/ChangeLog b/ChangeLog
index 964f220..201d0c2 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -12,6 +12,11 @@
 	jit/jit-rules-interp.c, jit/jit-rules-x86.c, jit/jit-rules.h:
 	implement global register allocation for parameters in stack slots.
 
+	* jit/jit-rules-x86.c, jit/jit-rules-x86.sel, jit/jit-rules.h:
+	optimize the x86 function epilog when we are certain that the
+	stack height doesn't change between entry and exit (i.e. the
+	function is a leaf and there are no alloca's).
+
 2004-06-10  Rhys Weatherley  <rweather@southern-storm.com.au>
 
 	* jit/jit-apply-arm.c, jit/jit-gen-arm.c, jit/jit-gen-arm.h,
diff --git a/jit/jit-rules-x86.c b/jit/jit-rules-x86.c
index bbe9a60..a5f210d 100644
--- a/jit/jit-rules-x86.c
+++ b/jit/jit-rules-x86.c
@@ -819,20 +819,37 @@ void _jit_gen_epilog(jit_gencode_t gen, jit_function_t func)
 	}
 
 	/* Restore the callee save registers that we used */
-	offset = -(func->builder->frame_size);
-	for(reg = 0; reg <= 7; ++reg)
+	if(gen->stack_changed)
 	{
-		if(jit_reg_is_used(gen->touched, reg) &&
-		   (_jit_reg_info[reg].flags & JIT_REG_CALL_USED) == 0)
+		offset = -(func->builder->frame_size);
+		for(reg = 0; reg <= 7; ++reg)
 		{
-			offset -= sizeof(void *);
-			x86_mov_reg_membase(inst, _jit_reg_info[reg].cpu_reg,
-								X86_EBP, offset, sizeof(void *));
+			if(jit_reg_is_used(gen->touched, reg) &&
+			   (_jit_reg_info[reg].flags & JIT_REG_CALL_USED) == 0)
+			{
+				offset -= sizeof(void *);
+				x86_mov_reg_membase(inst, _jit_reg_info[reg].cpu_reg,
+									X86_EBP, offset, sizeof(void *));
+			}
+		}
+	}
+	else
+	{
+		for(reg = 7; reg >= 0; --reg)
+		{
+			if(jit_reg_is_used(gen->touched, reg) &&
+			   (_jit_reg_info[reg].flags & JIT_REG_CALL_USED) == 0)
+			{
+				x86_pop_reg(inst, _jit_reg_info[reg].cpu_reg);
+			}
 		}
 	}
 
 	/* Pop the stack frame and restore the saved copy of ebp */
-	x86_mov_reg_reg(inst, X86_ESP, X86_EBP, sizeof(void *));
+	if(gen->stack_changed || func->builder->frame_size > 0)
+	{
+		x86_mov_reg_reg(inst, X86_ESP, X86_EBP, sizeof(void *));
+	}
 	x86_pop_reg(inst, X86_EBP);
 
 	/* Return from the current function */
diff --git a/jit/jit-rules-x86.sel b/jit/jit-rules-x86.sel
index 3da10d6..8a5adc8 100644
--- a/jit/jit-rules-x86.sel
+++ b/jit/jit-rules-x86.sel
@@ -1708,12 +1708,15 @@ JIT_OP_RETURN_REG: manual
 JIT_OP_PUSH_INT: unary_note
 	[imm] -> {
 		x86_push_imm(inst, $1);
+		gen->stack_changed = 1;
 	}
 	[local] -> {
 		x86_push_membase(inst, X86_EBP, $1);
+		gen->stack_changed = 1;
 	}
 	[reg] -> {
 		x86_push_reg(inst, $1);
+		gen->stack_changed = 1;
 	}
 
 JIT_OP_PUSH_LONG: spill_before
@@ -1732,20 +1735,24 @@ JIT_OP_PUSH_LONG: spill_before
 			x86_push_membase(inst, X86_EBP, offset + 4);
 			x86_push_membase(inst, X86_EBP, offset);
 		}
+		gen->stack_changed = 1;
 	}
 
 JIT_OP_PUSH_FLOAT32: unary_note, stack
 	[imm] -> {
 		jit_int *ptr = (jit_int *)($1);
 		x86_push_imm(inst, ptr[0]);
+		gen->stack_changed = 1;
 	}
 	[local] -> {
 		x86_push_membase(inst, X86_EBP, $1);
+		gen->stack_changed = 1;
 	}
 	[reg] -> {
 		x86_alu_reg_imm(inst, X86_SUB, X86_ESP, sizeof(jit_float32));
 		x86_fst_membase(inst, X86_ESP, 0, 0, 1);
 		_jit_regs_free_reg(gen, reg, 1);
+		gen->stack_changed = 1;
 	}
 
 JIT_OP_PUSH_FLOAT64: unary_note, stack
@@ -1753,15 +1760,18 @@ JIT_OP_PUSH_FLOAT64: unary_note, stack
 		jit_int *ptr = (jit_int *)($1);
 		x86_push_imm(inst, ptr[1]);
 		x86_push_imm(inst, ptr[0]);
+		gen->stack_changed = 1;
 	}
 	[local] -> {
 		x86_push_membase(inst, X86_EBP, $1 + 4);
 		x86_push_membase(inst, X86_EBP, $1);
+		gen->stack_changed = 1;
 	}
 	[reg] -> {
 		x86_alu_reg_imm(inst, X86_SUB, X86_ESP, sizeof(jit_float64));
 		x86_fst_membase(inst, X86_ESP, 0, 1, 1);
 		_jit_regs_free_reg(gen, reg, 1);
+		gen->stack_changed = 1;
 	}
 
 JIT_OP_PUSH_NFLOAT: unary_note, stack
@@ -1773,6 +1783,7 @@ JIT_OP_PUSH_NFLOAT: unary_note, stack
 		}
 		x86_push_imm(inst, ptr[1]);
 		x86_push_imm(inst, ptr[0]);
+		gen->stack_changed = 1;
 	}
 	[local] -> {
 		if(sizeof(jit_nfloat) != sizeof(jit_float64))
@@ -1781,6 +1792,7 @@ JIT_OP_PUSH_NFLOAT: unary_note, stack
 		}
 		x86_push_membase(inst, X86_EBP, $1 + 4);
 		x86_push_membase(inst, X86_EBP, $1);
+		gen->stack_changed = 1;
 	}
 	[freg] -> {
 		if(sizeof(jit_nfloat) != sizeof(jit_float64))
@@ -1794,17 +1806,20 @@ JIT_OP_PUSH_NFLOAT: unary_note, stack
 			x86_fst_membase(inst, X86_ESP, 0, 1, 1);
 		}
 		_jit_regs_free_reg(gen, reg, 1);
+		gen->stack_changed = 1;
 	}
 
 JIT_OP_PUSH_STRUCT: unary_note
 	[reg] -> {
 		/* TODO */
 		TODO();
+		gen->stack_changed = 1;
 	}
 
 JIT_OP_POP_STACK:
 	[] -> {
 		x86_alu_reg_imm(inst, X86_ADD, X86_ESP, insn->value1->address);
+		gen->stack_changed = 1;
 	}
 
 JIT_OP_FLUSH_SMALL_STRUCT:
@@ -2451,4 +2466,5 @@ JIT_OP_ALLOCA: unary
 		x86_alu_reg_imm(inst, X86_AND, $1, ~15);
 		x86_alu_reg_reg(inst, X86_SUB, X86_ESP, $1);
 		x86_mov_reg_reg(inst, $1, X86_ESP, 4);
+		gen->stack_changed = 1;
 	}
diff --git a/jit/jit-rules.h b/jit/jit-rules.h
index 3a70175..fad1600 100644
--- a/jit/jit-rules.h
+++ b/jit/jit-rules.h
@@ -147,6 +147,7 @@ struct jit_gencode
 	jit_extra_gen_state;			/* CPU-specific extra information */
 #endif
 	void			   *epilog_fixup; /* Fixup list for function epilogs */
+	int					stack_changed; /* Stack top changed since entry */
 };
 
 /*