# "attr" => "attitional attribute arguments for constructor"
# "init_attr" => "emit attribute initialization template"
# "rd_constructor" => "c source code which constructs an ir_node"
+# "latency" => "latency of this operation (can be float)"
# },
#
# ... # (all nodes you need to describe)
#
# outs: if a node defines more than one output, the names of the projections
# nodes having outs having automatically the mode mode_T
+# One can also annotate some flags for each out, additional to irn_flags.
+# They are separated from name with a colon ':', and concatenated by pipe '|'
+# Only I and S are available at the moment (same meaning as in irn_flags).
+# example: [ "frame:I", "stack:I|S", "M" ]
#
# comment: OPTIONAL comment for the node constructor
#
# return res
#
# NOTE: rd_constructor and args are only optional if and only if arity is 0,1,2 or 3
+#
+# latency: the latency of the operation, default is 1
+#
# register types:
# 0 - no special type
# 1 - caller save (register must be saved by the caller of a function)
# 2 - callee save (register must be saved by the called function)
# 4 - ignore (do not assign this register)
+# 8 - emitter can choose an arbitrary register of this class
+# 16 - the register is a virtual one
# NOTE: Last entry of each class is the largest Firm-Mode a register can hold
%reg_classes = (
"gp" => [
# { "name" => "r32", "type" => 1 },
{ "name" => "ebp", "type" => 2 },
{ "name" => "esp", "type" => 4 },
- { "name" => "gp_NOREG", "type" => 6 }, # we need a dummy register for NoReg nodes
- { "name" => "gp_UKNWN", "type" => 6 }, # we need a dummy register for Unknown nodes
+ { "name" => "gp_NOREG", "type" => 2 | 4 | 16 }, # we need a dummy register for NoReg nodes
+ { "name" => "gp_UKNWN", "type" => 2 | 4 | 8 | 16}, # we need a dummy register for Unknown nodes
{ "mode" => "mode_P" }
],
"xmm" => [
{ "name" => "xmm5", "type" => 1 },
{ "name" => "xmm6", "type" => 1 },
{ "name" => "xmm7", "type" => 1 },
- { "name" => "xmm_NOREG", "type" => 6 }, # we need a dummy register for NoReg nodes
- { "name" => "xmm_UKNWN", "type" => 6 }, # we need a dummy register for Unknown nodes
+ { "name" => "xmm_NOREG", "type" => 2 | 4 | 16 }, # we need a dummy register for NoReg nodes
+ { "name" => "xmm_UKNWN", "type" => 2 | 4 | 8 | 16}, # we need a dummy register for Unknown nodes
{ "mode" => "mode_D" }
],
"vfp" => [
- { "name" => "vf0", "type" => 1 },
- { "name" => "vf1", "type" => 1 },
- { "name" => "vf2", "type" => 1 },
- { "name" => "vf3", "type" => 1 },
- { "name" => "vf4", "type" => 1 },
- { "name" => "vf5", "type" => 1 },
- { "name" => "vf6", "type" => 1 },
- { "name" => "vf7", "type" => 1 },
- { "name" => "vfp_NOREG", "type" => 6 }, # we need a dummy register for NoReg nodes
- { "name" => "vfp_UKNWN", "type" => 6 }, # we need a dummy register for Unknown nodes
+ { "name" => "vf0", "type" => 1 | 16 },
+ { "name" => "vf1", "type" => 1 | 16 },
+ { "name" => "vf2", "type" => 1 | 16 },
+ { "name" => "vf3", "type" => 1 | 16 },
+ { "name" => "vf4", "type" => 1 | 16 },
+ { "name" => "vf5", "type" => 1 | 16 },
+ { "name" => "vf6", "type" => 1 | 16 },
+ { "name" => "vf7", "type" => 1 | 16 },
+ { "name" => "vfp_NOREG", "type" => 2 | 4 | 16 }, # we need a dummy register for NoReg nodes
+ { "name" => "vfp_UKNWN", "type" => 2 | 4 | 8 | 16}, # we need a dummy register for Unknown nodes
{ "mode" => "mode_E" }
],
"st" => [
},
"MulS" => {
+ # we should not rematrialize this node. It produces 2 results and has
+ # very strict constrains
"comment" => "construct MulS: MulS(a, b) = MulS(b, a) = a * b",
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
"reg_req" => { "in" => [ "gp", "gp", "eax", "gp", "none" ], "out" => [ "eax", "edx" ] },
"emit" => '. mul %ia32_emit_unop /* Mul(%A1, %A2) -> %D1 */',
"outs" => [ "EAX", "EDX", "M" ],
+ "latency" => 10,
},
"l_MulS" => {
+ # we should not rematrialize this node. It produces 2 results and has
+ # very strict constrains
"op_flags" => "C",
"cmp_attr" => " return 1;\n",
"comment" => "construct lowered MulS: MulS(a, b) = MulS(b, a) = a * b",
"reg_req" => { "in" => [ "gp", "gp", "gp", "gp", "none" ], "out" => [ "in_r3" ] },
"emit" => '. imul %ia32_emit_binop /* Mul(%A1, %A2) -> %D1 */',
"outs" => [ "res", "M" ],
+ "latency" => 5,
},
"l_Mul" => {
# Mulh is an exception from the 4 INs with AM because the target is always EAX:EDX
"Mulh" => {
+ # we should not rematrialize this node. It produces 2 results and has
+ # very strict constrains
"comment" => "construct Mul: Mul(a, b) = Mul(b, a) = a * b",
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
"reg_req" => { "in" => [ "gp", "gp", "eax", "gp", "none" ], "out" => [ "eax", "edx" ] },
"emit" => '. imul %ia32_emit_unop /* Mulh(%A1, %A2) -> %D1 */',
"outs" => [ "EAX", "EDX", "M" ],
+ "latency" => 5,
},
"And" => {
else {
4. cmovb %D1, %S2 /* %S1 is below %S2 */
}
-'
+',
+ "latency" => 2,
},
"Min" => {
else {
2. cmova %D1, %S2, %D1 /* %S1 is above %S2 */
}
-'
+',
+ "latency" => 2,
},
# not commutative operations
"SubC" => {
"comment" => "construct Sub with Carry: SubC(a, b) = a - b - carry",
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
- "reg_req" => { "in" => [ "gp", "gp", "gp", "gp", "none" ], "out" => [ "in_r3" ] },
+ "reg_req" => { "in" => [ "gp", "gp", "gp", "gp", "none" ], "out" => [ "in_r3 !in_r4" ] },
"emit" => '. sbb %ia32_emit_binop /* SubC(%A3, %A4) -> %D1 */',
"outs" => [ "res", "M" ],
},
}
',
"outs" => [ "div_res", "mod_res", "M" ],
+ "latency" => 25,
},
"Shl" => {
"irn_flags" => "R",
"comment" => "construct Shl: Shl(a, b) = a << b",
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
- "reg_req" => { "in" => [ "gp", "gp", "gp", "ecx", "none" ], "out" => [ "in_r3 !in_r4" ] },
+ "reg_req" => { "in" => [ "gp", "gp", "gp", "ecx gp_NOREG", "none" ], "out" => [ "in_r3 !in_r4" ] },
"emit" => '. shl %ia32_emit_binop /* Shl(%A1, %A2) -> %D1 */',
"outs" => [ "res", "M" ],
},
"irn_flags" => "R",
"comment" => "construct ShlD: ShlD(a, b, c) = a, b << count (shift left count bits from b into a)",
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
- "reg_req" => { "in" => [ "gp", "gp", "gp", "gp", "ecx", "none" ], "out" => [ "in_r3 !in_r5" ] },
+ # Out requirements is: different from all in
+ # This is because, out must be different from LowPart and ShiftCount.
+ # We could say "!ecx !in_r4" but it can occur, that all values live through
+ # this Shift and the only value dying is the ShiftCount. Then there would be a
+ # register missing, as result must not be ecx and all other registers are
+ # occupied. What we should write is "!in_r4 !in_r5", but this is not supported
+ # (and probably never will). So we create artificial interferences of the result
+ # with all inputs, so the spiller can always assure a free register.
+ "reg_req" => { "in" => [ "gp", "gp", "gp", "gp", "ecx", "none" ], "out" => [ "!in" ] },
"emit" =>
'
if (get_ia32_immop_type(n) == ia32_ImmNone) {
}
',
"outs" => [ "res", "M" ],
+ "latency" => 6,
},
"l_ShlD" => {
"irn_flags" => "R",
"comment" => "construct Shr: Shr(a, b) = a >> b",
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
- "reg_req" => { "in" => [ "gp", "gp", "gp", "ecx", "none" ], "out" => [ "in_r3 !in_r4" ] },
+ "reg_req" => { "in" => [ "gp", "gp", "gp", "ecx gp_NOREG", "none" ], "out" => [ "in_r3 !in_r4" ] },
"emit" => '. shr %ia32_emit_binop /* Shr(%A1, %A2) -> %D1 */',
"outs" => [ "res", "M" ],
},
"irn_flags" => "R",
"comment" => "construct ShrD: ShrD(a, b, c) = a, b >> count (shift rigth count bits from a into b)",
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
- "reg_req" => { "in" => [ "gp", "gp", "gp", "gp", "ecx", "none" ], "out" => [ "in_r3 !in_r5" ] },
+ # Out requirements is: different from all in
+ # This is because, out must be different from LowPart and ShiftCount.
+ # We could say "!ecx !in_r4" but it can occur, that all values live through
+ # this Shift and the only value dying is the ShiftCount. Then there would be a
+ # register missing, as result must not be ecx and all other registers are
+ # occupied. What we should write is "!in_r4 !in_r5", but this is not supported
+ # (and probably never will). So we create artificial interferences of the result
+ # with all inputs, so the spiller can always assure a free register.
+ "reg_req" => { "in" => [ "gp", "gp", "gp", "gp", "ecx", "none" ], "out" => [ "!in" ] },
"emit" =>
'
if (get_ia32_immop_type(n) == ia32_ImmNone) {
}
',
"outs" => [ "res", "M" ],
+ "latency" => 6,
},
"l_ShrD" => {
"irn_flags" => "R",
"comment" => "construct Shrs: Shrs(a, b) = a >> b",
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
- "reg_req" => { "in" => [ "gp", "gp", "gp", "ecx", "none" ], "out" => [ "in_r3 !in_r4" ] },
+ "reg_req" => { "in" => [ "gp", "gp", "gp", "ecx gp_NOREG", "none" ], "out" => [ "in_r3 !in_r4" ] },
"emit" => '. sar %ia32_emit_binop /* Shrs(%A1, %A2) -> %D1 */',
"outs" => [ "res", "M" ],
},
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
"reg_req" => { "in" => [ "gp", "gp", "gp", "gp", "none" ] },
"outs" => [ "false", "true" ],
+ "latency" => 3,
},
"TestJmp" => {
"reg_req" => { "in" => [ "gp", "gp" ] },
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
"outs" => [ "false", "true" ],
+ "latency" => 3,
},
"CJmpAM" => {
"comment" => "construct switch",
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
"reg_req" => { "in" => [ "gp" ], "out" => [ "none" ] },
+ "latency" => 3,
},
"Const" => {
"irn_flags" => "R",
"comment" => "represents an integer constant",
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
- "reg_req" => { "in" => [ "none" ], "out" => [ "gp" ] },
+ "reg_req" => { "out" => [ "gp" ] },
},
"Cdq" => {
- "irn_flags" => "R",
+ # we should not rematrialize this node. It produces 2 results and has
+ # very strict constrains
"comment" => "construct CDQ: sign extend EAX -> EDX:EAX",
"reg_req" => { "in" => [ "gp" ], "out" => [ "eax in_r1", "edx" ] },
"emit" => '. cdq /* sign extend EAX -> EDX:EAX, (%A1) */',
"comment" => "construct Load: Load(ptr, mem) = LD ptr -> reg",
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
"reg_req" => { "in" => [ "gp", "gp", "none" ], "out" => [ "gp" ] },
+ "latency" => 3,
"emit" =>
' if (get_mode_size_bits(get_ia32_ls_mode(n)) < 32) {
4. mov%Mx %D1, %ia32_emit_am /* Load((%A1)) -> %D1 */
"reg_req" => { "in" => [ "gp", "gp", "gp", "none" ] },
"emit" => '. mov %ia32_emit_binop /* Store(%A3) -> (%A1) */',
"outs" => [ "M" ],
+ "latency" => 3,
},
"Store8Bit" => {
"state" => "exc_pinned",
"comment" => "construct 8Bit Store: Store(ptr, val, mem) = ST ptr,val",
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
- "reg_req" => { "in" => [ "gp", "gp", "eax ebx ecx edx", "none" ] },
+ "reg_req" => { "in" => [ "gp", "gp", "eax ebx ecx edx gp_NOREG", "none" ] },
"emit" => '. mov %ia32_emit_binop /* Store(%A3) -> (%A1) */',
"outs" => [ "M" ],
+ "latency" => 3,
},
"Lea" => {
"comment" => "construct Lea: Lea(a,b) = lea [a+b*const+offs] | res = a + b * const + offs with const = 0,1,2,4,8",
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
"reg_req" => { "in" => [ "gp", "gp" ], "out" => [ "in_r1" ] },
- "emit" => '. lea %D1, %ia32_emit_am /* LEA(%A1, %A2) */'
+ "emit" => '. lea %D1, %ia32_emit_am /* LEA(%A1, %A2) */',
+ "latency" => 2,
},
"Push" => {
- "comment" => "push a gp register on the stack",
- "reg_req" => { "in" => [ "esp", "gp", "none" ], "out" => [ "esp" ] },
- "emit" => '
-if (get_ia32_id_cnst(n)) {
- if (get_ia32_immop_type(n) == ia32_ImmConst) {
-4. push %C /* Push const on stack */
-} else {
-4. push OFFSET FLAT:%C /* Push symconst on stack */
- }
-}
-else if (get_ia32_op_type(n) == ia32_Normal) {
-2. push %S2 /* Push(%A2) */
-}
-else {
-2. push %ia32_emit_am /* Push memory to stack */
-};
-',
- "outs" => [ "stack", "M" ],
+ "comment" => "push on the stack",
+ "reg_req" => { "in" => [ "gp", "gp", "gp", "esp", "none" ], "out" => [ "esp" ] },
+ "emit" => '. push %ia32_emit_unop /* PUSH(%A1) */',
+ "outs" => [ "stack:I|S", "M" ],
+ "latency" => 3,
},
"Pop" => {
+ # We don't set class modify stack here (but we will do this on proj 1)
"comment" => "pop a gp register from the stack",
- "reg_req" => { "in" => [ "esp", "none" ], "out" => [ "gp", "esp" ] },
- "emit" => '
-if (get_ia32_op_type(n) == ia32_Normal) {
-2. pop %D1 /* Pop from stack into %D1 */
-}
-else {
-2. pop %ia32_emit_am /* Pop from stack into memory */
-}
-',
- "outs" => [ "res", "stack", "M" ],
+ "reg_req" => { "in" => [ "gp", "gp", "esp", "none" ], "out" => [ "gp", "esp" ] },
+ "emit" => '. pop %ia32_emit_unop /* POP(%A1) */',
+ "outs" => [ "res", "stack:I|S", "M" ],
+ "latency" => 4,
},
"Enter" => {
"comment" => "create stack frame",
"reg_req" => { "in" => [ "esp" ], "out" => [ "ebp", "esp" ] },
"emit" => '. enter /* Enter */',
- "outs" => [ "frame", "stack", "M" ],
+ "outs" => [ "frame:I", "stack:I|S", "M" ],
+ "latency" => 15,
},
"Leave" => {
"comment" => "destroy stack frame",
"reg_req" => { "in" => [ "esp", "ebp" ], "out" => [ "ebp", "esp" ] },
"emit" => '. leave /* Leave */',
- "outs" => [ "frame", "stack", "M" ],
+ "outs" => [ "frame:I", "stack:I|S", "M" ],
+ "latency" => 3,
},
"AddSP" => {
- "irn_flags" => "S|I",
+ "irn_flags" => "I",
"comment" => "allocate space on stack",
"reg_req" => { "in" => [ "esp", "gp" ], "out" => [ "esp", "none" ] },
- "outs" => [ "stack", "M" ],
+ "outs" => [ "stack:S", "M" ],
+},
+
+"SubSP" => {
+ "irn_flags" => "I",
+ "comment" => "free space on stack",
+ "reg_req" => { "in" => [ "esp", "gp" ], "out" => [ "esp", "none" ] },
+ "outs" => [ "stack:S", "M" ],
+},
+
+"LdTls" => {
+ "irn_flags" => "R",
+ "comment" => "get the TLS base address",
+ "reg_req" => { "out" => [ "gp" ] },
},
+
+
#-----------------------------------------------------------------------------#
# _____ _____ ______ __ _ _ _ #
# / ____/ ____| ____| / _| | | | | | #
"reg_req" => { "in" => [ "gp", "gp", "xmm", "xmm", "none" ], "out" => [ "in_r3" ] },
"emit" => '. adds%M %ia32_emit_binop /* SSE Add(%A3, %A4) -> %D1 */',
"outs" => [ "res", "M" ],
+ "latency" => 4,
},
"xMul" => {
"reg_req" => { "in" => [ "gp", "gp", "xmm", "xmm", "none" ], "out" => [ "in_r3" ] },
"emit" => '. muls%M %ia32_emit_binop /* SSE Mul(%A3, %A4) -> %D1 */',
"outs" => [ "res", "M" ],
+ "latency" => 4,
},
"xMax" => {
"reg_req" => { "in" => [ "gp", "gp", "xmm", "xmm", "none" ], "out" => [ "in_r3" ] },
"emit" => '. maxs%M %ia32_emit_binop /* SSE Max(%A3, %A4) -> %D1 */',
"outs" => [ "res", "M" ],
+ "latency" => 2,
},
"xMin" => {
"reg_req" => { "in" => [ "gp", "gp", "xmm", "xmm", "none" ], "out" => [ "in_r3" ] },
"emit" => '. mins%M %ia32_emit_binop /* SSE Min(%A3, %A4) -> %D1 */',
"outs" => [ "res", "M" ],
+ "latency" => 2,
},
"xAnd" => {
"reg_req" => { "in" => [ "gp", "gp", "xmm", "xmm", "none" ], "out" => [ "in_r3" ] },
"emit" => '. andp%M %ia32_emit_binop /* SSE And(%A3, %A4) -> %D1 */',
"outs" => [ "res", "M" ],
+ "latency" => 3,
},
"xOr" => {
"reg_req" => { "in" => [ "gp", "gp", "xmm", "xmm", "none" ], "out" => [ "in_r3" ] },
"emit" => '. xorp%M %ia32_emit_binop /* SSE Xor(%A3, %A4) -> %D1 */',
"outs" => [ "res", "M" ],
+ "latency" => 3,
},
# not commutative operations
"reg_req" => { "in" => [ "gp", "gp", "xmm", "xmm", "none" ], "out" => [ "in_r3 !in_r4" ] },
"emit" => '. andnp%M %ia32_emit_binop /* SSE AndNot(%A3, %A4) -> %D1 */',
"outs" => [ "res", "M" ],
+ "latency" => 3,
},
"xSub" => {
"reg_req" => { "in" => [ "gp", "gp", "xmm", "xmm", "none" ], "out" => [ "in_r3" ] },
"emit" => '. subs%M %ia32_emit_binop /* SSE Sub(%A1, %A2) -> %D1 */',
"outs" => [ "res", "M" ],
+ "latency" => 4,
},
"xDiv" => {
"reg_req" => { "in" => [ "gp", "gp", "xmm", "xmm", "none" ], "out" => [ "in_r3 !in_r4" ] },
"emit" => '. divs%M %ia32_emit_binop /* SSE Div(%A1, %A2) -> %D1 */',
"outs" => [ "res", "M" ],
+ "latency" => 16,
},
# other operations
"comment" => "construct SSE Compare: Cmp(a, b) == a = a cmp b",
"reg_req" => { "in" => [ "gp", "gp", "xmm", "xmm", "none" ], "out" => [ "in_r3 !in_r4" ] },
"outs" => [ "res", "M" ],
+ "latency" => 3,
},
"xCondJmp" => {
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
"reg_req" => { "in" => [ "gp", "gp", "xmm", "xmm", "none" ], "out" => [ "none", "none" ] },
"outs" => [ "false", "true" ],
+ "latency" => 5,
},
"xConst" => {
"irn_flags" => "R",
"comment" => "represents a SSE constant",
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
- "reg_req" => { "in" => [ "none" ], "out" => [ "xmm" ] },
+ "reg_req" => { "out" => [ "xmm" ] },
"emit" => '. movs%M %D1, %C /* Load fConst into register */',
+ "latency" => 2,
},
# Load / Store
"reg_req" => { "in" => [ "gp", "gp", "none" ], "out" => [ "xmm" ] },
"emit" => '. movs%M %D1, %ia32_emit_am /* Load((%A1)) -> %D1 */',
"outs" => [ "res", "M" ],
+ "latency" => 2,
},
"xStore" => {
"reg_req" => { "in" => [ "gp", "gp", "xmm", "none" ] },
"emit" => '. movs%M %ia32_emit_binop /* Store(%S3) -> (%A1) */',
"outs" => [ "M" ],
+ "latency" => 2,
},
"xStoreSimple" => {
"reg_req" => { "in" => [ "gp", "xmm", "none" ] },
"emit" => '. movs%M %ia32_emit_am, %S2 /* store XMM0 onto stack */',
"outs" => [ "M" ],
+ "latency" => 2,
},
"l_X87toSSE" => {
"reg_req" => { "in" => [ "gp", "none" ] },
"emit" => '. fstp %ia32_emit_am /* store ST0 onto stack */',
"outs" => [ "M" ],
+ "latency" => 4,
},
"SetST0" => {
"state" => "exc_pinned",
"comment" => "load ST0 from stack",
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
- "reg_req" => { "in" => [ "gp", "none" ], "out" => [ "st0", "none" ] },
+ "reg_req" => { "in" => [ "gp", "none" ], "out" => [ "vf0", "none" ] },
"emit" => '. fld %ia32_emit_am /* load ST0 from stack */',
"outs" => [ "res", "M" ],
+ "latency" => 2,
},
# CopyB
"op_flags" => "F|H",
"state" => "pinned",
"comment" => "implements a memcopy: CopyB(dst, src, mem) == memcpy(dst, src, attr(size))",
- "cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
+ "cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
"reg_req" => { "in" => [ "edi", "esi", "none" ], "out" => [ "edi", "esi", "none" ] },
"outs" => [ "DST", "SRC", "M" ],
},
"Conv_I2I" => {
"reg_req" => { "in" => [ "gp", "gp", "gp", "none" ], "out" => [ "in_r3", "none" ] },
- "cmp_attr" => " return ia32_compare_conv_attr(attr_a, attr_b);\n",
+ "cmp_attr" => " return ia32_compare_conv_attr(attr_a, attr_b);\n",
"comment" => "construct Conv Int -> Int",
- "outs" => [ "res", "M" ],
+ "outs" => [ "res", "M" ],
},
"Conv_I2I8Bit" => {
"reg_req" => { "in" => [ "gp", "gp", "eax ebx ecx edx", "none" ], "out" => [ "in_r3", "none" ] },
- "cmp_attr" => " return ia32_compare_conv_attr(attr_a, attr_b);\n",
+ "cmp_attr" => " return ia32_compare_conv_attr(attr_a, attr_b);\n",
"comment" => "construct Conv Int -> Int",
- "outs" => [ "res", "M" ],
+ "outs" => [ "res", "M" ],
},
"Conv_I2FP" => {
"reg_req" => { "in" => [ "gp", "gp", "gp", "none" ], "out" => [ "xmm", "none" ] },
- "cmp_attr" => " return ia32_compare_conv_attr(attr_a, attr_b);\n",
+ "cmp_attr" => " return ia32_compare_conv_attr(attr_a, attr_b);\n",
"comment" => "construct Conv Int -> Floating Point",
- "outs" => [ "res", "M" ],
+ "outs" => [ "res", "M" ],
+ "latency" => 10,
},
"Conv_FP2I" => {
"reg_req" => { "in" => [ "gp", "gp", "xmm", "none" ], "out" => [ "gp", "none" ] },
- "cmp_attr" => " return ia32_compare_conv_attr(attr_a, attr_b);\n",
+ "cmp_attr" => " return ia32_compare_conv_attr(attr_a, attr_b);\n",
"comment" => "construct Conv Floating Point -> Int",
- "outs" => [ "res", "M" ],
+ "outs" => [ "res", "M" ],
+ "latency" => 10,
},
"Conv_FP2FP" => {
"reg_req" => { "in" => [ "gp", "gp", "xmm", "none" ], "out" => [ "xmm", "none" ] },
- "cmp_attr" => " return ia32_compare_conv_attr(attr_a, attr_b);\n",
+ "cmp_attr" => " return ia32_compare_conv_attr(attr_a, attr_b);\n",
"comment" => "construct Conv Floating Point -> Floating Point",
- "outs" => [ "res", "M" ],
+ "outs" => [ "res", "M" ],
+ "latency" => 8,
},
"CmpCMov" => {
"irn_flags" => "R",
"comment" => "construct Conditional Move: CMov(sel, a, b) == sel ? a : b",
- "reg_req" => { "in" => [ "gp", "gp", "gp", "gp" ], "out" => [ "in_r4" ] }
+ "reg_req" => { "in" => [ "gp", "gp", "gp", "gp" ], "out" => [ "in_r4" ] },
+ "latency" => 2,
},
"PsiCondCMov" => {
"irn_flags" => "R",
"comment" => "check if Psi condition tree evaluates to true and move result accordingly",
- "reg_req" => { "in" => [ "gp", "gp", "gp" ], "out" => [ "in_r3" ] }
+ "reg_req" => { "in" => [ "gp", "gp", "gp" ], "out" => [ "in_r3" ] },
+ "latency" => 2,
},
"xCmpCMov" => {
"irn_flags" => "R",
"comment" => "construct Conditional Move: SSE Compare + int CMov ",
- "reg_req" => { "in" => [ "xmm", "xmm", "gp", "gp" ], "out" => [ "in_r4" ] }
+ "reg_req" => { "in" => [ "xmm", "xmm", "gp", "gp" ], "out" => [ "in_r4" ] },
+ "latency" => 5,
},
"vfCmpCMov" => {
"irn_flags" => "R",
"comment" => "construct Conditional Move: x87 Compare + int CMov",
- "reg_req" => { "in" => [ "vfp", "vfp", "gp", "gp" ], "out" => [ "in_r4" ] }
+ "reg_req" => { "in" => [ "vfp", "vfp", "gp", "gp" ], "out" => [ "in_r4" ] },
+ "latency" => 10,
},
"CmpSet" => {
"comment" => "construct Set: Set(sel) == sel ? 1 : 0",
"reg_req" => { "in" => [ "gp", "gp", "gp", "gp", "none" ], "out" => [ "eax ebx ecx edx", "none" ] },
"outs" => [ "res", "M" ],
+ "latency" => 2,
},
"PsiCondSet" => {
"irn_flags" => "R",
"comment" => "check if Psi condition tree evaluates to true and set result accordingly",
"reg_req" => { "in" => [ "gp" ], "out" => [ "eax ebx ecx edx" ] },
+ "latency" => 2,
},
"xCmpSet" => {
"comment" => "construct Set: SSE Compare + int Set",
"reg_req" => { "in" => [ "gp", "gp", "xmm", "xmm", "none" ], "out" => [ "eax ebx ecx edx", "none" ] },
"outs" => [ "res", "M" ],
+ "latency" => 5,
},
"vfCmpSet" => {
"comment" => "construct Set: x87 Compare + int Set",
"reg_req" => { "in" => [ "gp", "gp", "vfp", "vfp", "none" ], "out" => [ "eax ebx ecx edx", "none" ] },
"outs" => [ "res", "M" ],
+ "latency" => 10,
},
"vfCMov" => {
"irn_flags" => "R",
"comment" => "construct x87 Conditional Move: vfCMov(sel, a, b) = sel ? a : b",
- "reg_req" => { "in" => [ "vfp", "vfp", "vfp", "vfp" ], "out" => [ "vfp" ] }
+ "reg_req" => { "in" => [ "vfp", "vfp", "vfp", "vfp" ], "out" => [ "vfp" ] },
+ "latency" => 10,
},
#----------------------------------------------------------#
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
"reg_req" => { "in" => [ "gp", "gp", "vfp", "vfp", "none" ], "out" => [ "vfp" ] },
"outs" => [ "res", "M" ],
+ "latency" => 4,
},
"vfmul" => {
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
"reg_req" => { "in" => [ "gp", "gp", "vfp", "vfp", "none" ], "out" => [ "vfp" ] },
"outs" => [ "res", "M" ],
+ "latency" => 4,
},
"l_vfmul" => {
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
"reg_req" => { "in" => [ "gp", "gp", "vfp", "vfp", "none" ], "out" => [ "vfp" ] },
"outs" => [ "res", "M" ],
+ "latency" => 4,
},
"l_vfsub" => {
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
"reg_req" => { "in" => [ "gp", "gp", "vfp", "vfp", "none" ], "out" => [ "vfp" ] },
"outs" => [ "res", "M" ],
+ "latency" => 20,
},
"l_vfdiv" => {
"irn_flags" => "R",
"comment" => "virtual fp Abs: Abs(a) = |a|",
"reg_req" => { "in" => [ "vfp"], "out" => [ "vfp" ] },
+ "latency" => 2,
},
"vfchs" => {
"irn_flags" => "R",
"comment" => "virtual fp Chs: Chs(a) = -a",
"reg_req" => { "in" => [ "vfp"], "out" => [ "vfp" ] },
+ "latency" => 2,
},
"vfsin" => {
"irn_flags" => "R",
"comment" => "virtual fp Sin: Sin(a) = sin(a)",
"reg_req" => { "in" => [ "vfp"], "out" => [ "vfp" ] },
+ "latency" => 150,
},
"vfcos" => {
"irn_flags" => "R",
"comment" => "virtual fp Cos: Cos(a) = cos(a)",
"reg_req" => { "in" => [ "vfp"], "out" => [ "vfp" ] },
+ "latency" => 150,
},
"vfsqrt" => {
"irn_flags" => "R",
"comment" => "virtual fp Sqrt: Sqrt(a) = a ^ 0.5",
"reg_req" => { "in" => [ "vfp"], "out" => [ "vfp" ] },
+ "latency" => 30,
},
# virtual Load and Store
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
"reg_req" => { "in" => [ "gp", "gp", "none" ], "out" => [ "vfp", "none" ] },
"outs" => [ "res", "M" ],
+ "latency" => 2,
},
"vfst" => {
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
"reg_req" => { "in" => [ "gp", "gp", "vfp", "none" ] },
"outs" => [ "M" ],
+ "latency" => 2,
},
# Conversions
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
"reg_req" => { "in" => [ "gp", "gp", "none" ], "out" => [ "vfp", "none" ] },
"outs" => [ "res", "M" ],
+ "latency" => 4,
},
"l_vfild" => {
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
"reg_req" => { "in" => [ "gp", "gp", "vfp", "none" ] },
"outs" => [ "M" ],
+ "latency" => 4,
},
"l_vfist" => {
"irn_flags" => "R",
"comment" => "virtual fp Load 0.0: Ld 0.0 -> reg",
"reg_req" => { "out" => [ "vfp" ] },
+ "latency" => 4,
},
"vfld1" => {
"irn_flags" => "R",
"comment" => "virtual fp Load 1.0: Ld 1.0 -> reg",
"reg_req" => { "out" => [ "vfp" ] },
+ "latency" => 4,
},
"vfldpi" => {
"irn_flags" => "R",
"comment" => "virtual fp Load pi: Ld pi -> reg",
"reg_req" => { "out" => [ "vfp" ] },
+ "latency" => 4,
},
"vfldln2" => {
"irn_flags" => "R",
"comment" => "virtual fp Load ln 2: Ld ln 2 -> reg",
"reg_req" => { "out" => [ "vfp" ] },
+ "latency" => 4,
},
"vfldlg2" => {
"irn_flags" => "R",
"comment" => "virtual fp Load lg 2: Ld lg 2 -> reg",
"reg_req" => { "out" => [ "vfp" ] },
+ "latency" => 4,
},
"vfldl2t" => {
"irn_flags" => "R",
"comment" => "virtual fp Load ld 10: Ld ld 10 -> reg",
"reg_req" => { "out" => [ "vfp" ] },
+ "latency" => 4,
},
"vfldl2e" => {
"irn_flags" => "R",
"comment" => "virtual fp Load ld e: Ld ld e -> reg",
"reg_req" => { "out" => [ "vfp" ] },
+ "latency" => 4,
},
"vfConst" => {
"init_attr" => " set_ia32_ls_mode(res, mode);",
"comment" => "represents a virtual floating point constant",
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
- "reg_req" => { "in" => [ "none" ], "out" => [ "vfp" ] },
+ "reg_req" => { "out" => [ "vfp" ] },
+ "latency" => 3,
},
# other
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
"reg_req" => { "in" => [ "gp", "gp", "vfp", "vfp", "none" ], "out" => [ "none", "none", "eax" ] },
"outs" => [ "false", "true", "temp_reg_eax" ],
+ "latency" => 10,
},
#------------------------------------------------------------------------#
"rd_constructor" => "NONE",
"comment" => "x87 Add: Add(a, b) = Add(b, a) = a + b",
"reg_req" => { },
- "emit" => '. fadd %ia32_emit_x87_binop /* x87 fadd(%A1, %A2) -> %D1 */',
+ "emit" => '. fadd %ia32_emit_x87_binop /* x87 fadd(%A3, %A4) -> %D1 */',
},
"faddp" => {
"rd_constructor" => "NONE",
"comment" => "x87 Add: Add(a, b) = Add(b, a) = a + b",
"reg_req" => { },
- "emit" => '. faddp %ia32_emit_x87_binop /* x87 fadd(%A1, %A2) -> %D1 */',
+ "emit" => '. faddp %ia32_emit_x87_binop /* x87 fadd(%A3, %A4) -> %D1 */',
},
"fmul" => {
"rd_constructor" => "NONE",
"comment" => "x87 fp Mul: Mul(a, b) = Mul(b, a) = a + b",
"reg_req" => { },
- "emit" => '. fmul %ia32_emit_x87_binop /* x87 fmul(%A1, %A2) -> %D1 */',
+ "emit" => '. fmul %ia32_emit_x87_binop /* x87 fmul(%A3, %A4) -> %D1 */',
},
"fmulp" => {
"rd_constructor" => "NONE",
"comment" => "x87 fp Mul: Mul(a, b) = Mul(b, a) = a + b",
"reg_req" => { },
- "emit" => '. fmulp %ia32_emit_x87_binop /* x87 fmul(%A1, %A2) -> %D1 */',,
+ "emit" => '. fmulp %ia32_emit_x87_binop /* x87 fmul(%A3, %A4) -> %D1 */',,
},
"fsub" => {
"rd_constructor" => "NONE",
"comment" => "x87 fp Sub: Sub(a, b) = a - b",
"reg_req" => { },
- "emit" => '. fsub %ia32_emit_x87_binop /* x87 fsub(%A1, %A2) -> %D1 */',
+ "emit" => '. fsub %ia32_emit_x87_binop /* x87 fsub(%A3, %A4) -> %D1 */',
},
"fsubp" => {
"rd_constructor" => "NONE",
"comment" => "x87 fp Sub: Sub(a, b) = a - b",
"reg_req" => { },
- "emit" => '. fsubp %ia32_emit_x87_binop /* x87 fsub(%A1, %A2) -> %D1 */',
+ "emit" => '. fsubp %ia32_emit_x87_binop /* x87 fsub(%A3, %A4) -> %D1 */',
},
"fsubr" => {
"irn_flags" => "R",
"comment" => "x87 fp SubR: SubR(a, b) = b - a",
"reg_req" => { },
- "emit" => '. fsubr %ia32_emit_x87_binop /* x87 fsubr(%A1, %A2) -> %D1 */',
+ "emit" => '. fsubr %ia32_emit_x87_binop /* x87 fsubr(%A3, %A4) -> %D1 */',
},
"fsubrp" => {
"irn_flags" => "R",
"comment" => "x87 fp SubR: SubR(a, b) = b - a",
"reg_req" => { },
- "emit" => '. fsubrp %ia32_emit_x87_binop /* x87 fsubr(%A1, %A2) -> %D1 */',
+ "emit" => '. fsubrp %ia32_emit_x87_binop /* x87 fsubr(%A3, %A4) -> %D1 */',
},
"fdiv" => {
"rd_constructor" => "NONE",
"comment" => "x87 fp Div: Div(a, b) = a / b",
"reg_req" => { },
- "emit" => '. fdiv %ia32_emit_x87_binop /* x87 fdiv(%A1, %A2) -> %D1 */',
+ "emit" => '. fdiv %ia32_emit_x87_binop /* x87 fdiv(%A3, %A4) -> %D1 */',
},
"fdivp" => {
"rd_constructor" => "NONE",
"comment" => "x87 fp Div: Div(a, b) = a / b",
"reg_req" => { },
- "emit" => '. fdivp %ia32_emit_x87_binop /* x87 fdiv(%A1, %A2) -> %D1 */',
+ "emit" => '. fdivp %ia32_emit_x87_binop /* x87 fdiv(%A3, %A4) -> %D1 */',
},
"fdivr" => {
"rd_constructor" => "NONE",
"comment" => "x87 fp DivR: DivR(a, b) = b / a",
"reg_req" => { },
- "emit" => '. fdivr %ia32_emit_x87_binop /* x87 fdivr(%A1, %A2) -> %D1 */',
+ "emit" => '. fdivr %ia32_emit_x87_binop /* x87 fdivr(%A3, %A4) -> %D1 */',
},
"fdivrp" => {
"rd_constructor" => "NONE",
"comment" => "x87 fp DivR: DivR(a, b) = b / a",
"reg_req" => { },
- "emit" => '. fdivrp %ia32_emit_x87_binop /* x87 fdivr(%A1, %A2) -> %D1 */',
+ "emit" => '. fdivrp %ia32_emit_x87_binop /* x87 fdivr(%A3, %A4) -> %D1 */',
},
"fabs" => {
"rd_constructor" => "NONE",
"comment" => "x87 fp Abs: Abs(a) = |a|",
"reg_req" => { },
- "emit" => '. fabs /* x87 fabs(%S1) -> %D1 */',
+ "emit" => '. fabs /* x87 fabs(%A1) -> %D1 */',
},
"fchs" => {
"rd_constructor" => "NONE",
"comment" => "x87 fp Chs: Chs(a) = -a",
"reg_req" => { },
- "emit" => '. fchs /* x87 fchs(%S1) -> %D1 */',
+ "emit" => '. fchs /* x87 fchs(%A1) -> %D1 */',
},
"fsin" => {
"rd_constructor" => "NONE",
"comment" => "x87 fp Sin: Sin(a) = sin(a)",
"reg_req" => { },
- "emit" => '. fsin /* x87 sin(%S1) -> %D1 */',
+ "emit" => '. fsin /* x87 sin(%A1) -> %D1 */',
},
"fcos" => {
"rd_constructor" => "NONE",
"comment" => "x87 fp Cos: Cos(a) = cos(a)",
"reg_req" => { },
- "emit" => '. fcos /* x87 cos(%S1) -> %D1 */',
+ "emit" => '. fcos /* x87 cos(%A1) -> %D1 */',
},
"fsqrt" => {
"rd_constructor" => "NONE",
"comment" => "x87 fp Sqrt: Sqrt(a) = a ^ 0.5",
"reg_req" => { },
- "emit" => '. fsqrt $ /* x87 sqrt(%S1) -> %D1 */',
+ "emit" => '. fsqrt $ /* x87 sqrt(%A1) -> %D1 */',
},
# x87 Load and Store
# constants
"fldz" => {
- "op_flags" => "R",
+ "op_flags" => "R|c",
"irn_flags" => "R",
- "rd_constructor" => "NONE",
"comment" => "x87 fp Load 0.0: Ld 0.0 -> reg",
- "reg_req" => { },
+ "reg_req" => { "out" => [ "vfp" ] },
"emit" => '. fldz /* x87 0.0 -> %D1 */',
},
"fld1" => {
- "op_flags" => "R",
+ "op_flags" => "R|c",
"irn_flags" => "R",
- "rd_constructor" => "NONE",
"comment" => "x87 fp Load 1.0: Ld 1.0 -> reg",
- "reg_req" => { },
+ "reg_req" => { "out" => [ "vfp" ] },
"emit" => '. fld1 /* x87 1.0 -> %D1 */',
},
"fldpi" => {
- "op_flags" => "R",
+ "op_flags" => "R|c",
"irn_flags" => "R",
- "rd_constructor" => "NONE",
"comment" => "x87 fp Load pi: Ld pi -> reg",
- "reg_req" => { },
+ "reg_req" => { "out" => [ "vfp" ] },
"emit" => '. fldpi /* x87 pi -> %D1 */',
},
"fldln2" => {
- "op_flags" => "R",
+ "op_flags" => "R|c",
"irn_flags" => "R",
- "rd_constructor" => "NONE",
"comment" => "x87 fp Load ln 2: Ld ln 2 -> reg",
- "reg_req" => { },
+ "reg_req" => { "out" => [ "vfp" ] },
"emit" => '. fldln2 /* x87 ln(2) -> %D1 */',
},
"fldlg2" => {
- "op_flags" => "R",
+ "op_flags" => "R|c",
"irn_flags" => "R",
- "rd_constructor" => "NONE",
"comment" => "x87 fp Load lg 2: Ld lg 2 -> reg",
- "reg_req" => { },
+ "reg_req" => { "out" => [ "vfp" ] },
"emit" => '. fldlg2 /* x87 log(2) -> %D1 */',
},
"fldl2t" => {
- "op_flags" => "R",
+ "op_flags" => "R|c",
"irn_flags" => "R",
- "rd_constructor" => "NONE",
"comment" => "x87 fp Load ld 10: Ld ld 10 -> reg",
- "reg_req" => { },
+ "reg_req" => { "out" => [ "vfp" ] },
"emit" => '. fldll2t /* x87 ld(10) -> %D1 */',
},
"fldl2e" => {
- "op_flags" => "R",
+ "op_flags" => "R|c",
"irn_flags" => "R",
- "rd_constructor" => "NONE",
"comment" => "x87 fp Load ld e: Ld ld e -> reg",
- "reg_req" => { },
+ "reg_req" => { "out" => [ "vfp" ] },
"emit" => '. fldl2e /* x87 ld(e) -> %D1 */',
},
"rd_constructor" => "NONE",
"comment" => "represents a x87 constant",
"cmp_attr" => " return ia32_compare_immop_attr(attr_a, attr_b);\n",
- "reg_req" => { "out" => [ "st" ] },
+ "reg_req" => { "out" => [ "vfp" ] },
"emit" => '. fld %ia32_emit_adr /* Load fConst into register -> %D1 */',
},
# fxch, fpush, fpop
# Note that it is NEVER allowed to do CSE on these nodes
+# Moreover, note the virtual register requierements!
"fxch" => {
"op_flags" => "R|K",
"comment" => "x87 stack exchange",
- "reg_req" => { "in" => [ "st"], "out" => [ "st" ] },
+ "reg_req" => { },
"cmp_attr" => " return 1;\n",
"emit" => '. fxch %X1 /* x87 swap %X1, %X3 */',
},
"fpush" => {
+ "op_flags" => "R|K",
+ "comment" => "x87 stack push",
+ "reg_req" => {},
+ "cmp_attr" => " return 1;\n",
+ "emit" => '. fld %X1 /* x87 push %X1 */',
+},
+
+"fpushCopy" => {
"op_flags" => "R",
"comment" => "x87 stack push",
- "reg_req" => { "in" => [ "st"], "out" => [ "st" ] },
+ "reg_req" => { "in" => [ "vfp"], "out" => [ "vfp" ] },
"cmp_attr" => " return 1;\n",
"emit" => '. fld %X1 /* x87 push %X1 */',
},
"fpop" => {
"op_flags" => "R|K",
"comment" => "x87 stack pop",
- "reg_req" => { "in" => [ "st"], "out" => [ "st" ] },
+ "reg_req" => { },
"cmp_attr" => " return 1;\n",
"emit" => '. fstp %X1 /* x87 pop %X1 */',
},