diff options
author | Gabe Black <gblack@eecs.umich.edu> | 2007-09-13 16:34:46 -0700 |
---|---|---|
committer | Gabe Black <gblack@eecs.umich.edu> | 2007-09-13 16:34:46 -0700 |
commit | f7b6230d99e102f3a6195687fed0617005a70424 (patch) | |
tree | 360dd76c0b7d7f159e816af90ab8fb3e62f3c36c /src/arch/x86/isa/insts | |
parent | 0f57b407a3df68f93e73e0635569d7bf5dd151b0 (diff) | |
download | gem5-f7b6230d99e102f3a6195687fed0617005a70424.tar.xz |
X86: Total overhaul of the division instructions and microops.
--HG--
extra : convert_revision : 303ea45f69f7805361ad877fe6bb43fbc3dfd7a6
Diffstat (limited to 'src/arch/x86/isa/insts')
-rw-r--r-- | src/arch/x86/isa/insts/arithmetic/multiply_and_divide.py | 579 |
1 files changed, 552 insertions, 27 deletions
diff --git a/src/arch/x86/isa/insts/arithmetic/multiply_and_divide.py b/src/arch/x86/isa/insts/arithmetic/multiply_and_divide.py index 197de5fb3..a9b53acac 100644 --- a/src/arch/x86/isa/insts/arithmetic/multiply_and_divide.py +++ b/src/arch/x86/isa/insts/arithmetic/multiply_and_divide.py @@ -234,20 +234,70 @@ def macroop IMUL_R_P_I def macroop DIV_B_R { - div1 rax, rax, reg + # Do the initial part of the division + div1 rsi, reg, dataSize=1 + + #These are split out so we can initialize the number of bits in the + #second register + div2i t1, rax, 8, dataSize=1 + div2 t1, rax, t1, dataSize=1 + + #Loop until we're out of bits to shift in +divLoopTop: + div2 t1, rax, t1, dataSize=1 + div2 t1, rax, t1, flags=(EZF,), dataSize=1 + bri t0, label("divLoopTop"), flags=(nCEZF,) + + #Unload the answer + divq rax, dataSize=1 + divr rsi, dataSize=1 }; def macroop DIV_B_M { - ld t1, seg, sib, disp - div1 rax, rax, t1 + ld t2, seg, sib, disp + + # Do the initial part of the division + div1 rsi, t2, dataSize=1 + + #These are split out so we can initialize the number of bits in the + #second register + div2i t1, rax, 8, dataSize=1 + div2 t1, rax, t1, dataSize=1 + + #Loop until we're out of bits to shift in +divLoopTop: + div2 t1, rax, t1, dataSize=1 + div2 t1, rax, t1, flags=(EZF,), dataSize=1 + bri t0, label("divLoopTop"), flags=(nCEZF,) + + #Unload the answer + divq rax, dataSize=1 + divr rsi, dataSize=1 }; def macroop DIV_B_P { rdip t7 - ld t1, seg, riprel, disp - div1 rax, rax, t1 + ld t2, seg, riprel, disp + + # Do the initial part of the division + div1 rsi, t2, dataSize=1 + + #These are split out so we can initialize the number of bits in the + #second register + div2i t1, rax, 8, dataSize=1 + div2 t1, rax, t1, dataSize=1 + + #Loop until we're out of bits to shift in +divLoopTop: + div2 t1, rax, t1, dataSize=1 + div2 t1, rax, t1, flags=(EZF,), dataSize=1 + bri t0, label("divLoopTop"), flags=(nCEZF,) + + #Unload the answer + divq rax, dataSize=1 + divr rsi, dataSize=1 }; # @@ -256,24 +306,301 @@ def macroop DIV_B_P def macroop DIV_R { - divr t1, rax, reg - divq rax, rax, reg - mov rdx, rdx, t1 + # Do the initial part of the division + div1 rdx, reg + + #These are split out so we can initialize the number of bits in the + #second register + div2i t1, rax, "env.dataSize * 8" + div2 t1, rax, t1 + + #Loop until we're out of bits to shift in + #The amount of unrolling here could stand some tuning +divLoopTop: + div2 t1, rax, t1 + div2 t1, rax, t1 + div2 t1, rax, t1 + div2 t1, rax, t1, flags=(EZF,) + bri t0, label("divLoopTop"), flags=(nCEZF,) + + #Unload the answer + divq rax + divr rdx }; def macroop DIV_M { - ld t1, seg, sib, disp - divr rdx, rax, t1 - divq rax, rax, t1 + ld t2, seg, sib, disp + + # Do the initial part of the division + div1 rdx, t2 + + #These are split out so we can initialize the number of bits in the + #second register + div2i t1, rax, "env.dataSize * 8" + div2 t1, rax, t1 + + #Loop until we're out of bits to shift in + #The amount of unrolling here could stand some tuning +divLoopTop: + div2 t1, rax, t1 + div2 t1, rax, t1 + div2 t1, rax, t1 + div2 t1, rax, t1, flags=(EZF,) + bri t0, label("divLoopTop"), flags=(nCEZF,) + + #Unload the answer + divq rax + divr rdx }; def macroop DIV_P { rdip t7 - ld t1, seg, riprel, disp - divr rdx, rax, t1 - divq rax, rax, t1 + ld t2, seg, riprel, disp + + # Do the initial part of the division + div1 rdx, t2 + + #These are split out so we can initialize the number of bits in the + #second register + div2i t1, rax, "env.dataSize * 8" + div2 t1, rax, t1 + + #Loop until we're out of bits to shift in + #The amount of unrolling here could stand some tuning +divLoopTop: + div2 t1, rax, t1 + div2 t1, rax, t1 + div2 t1, rax, t1 + div2 t1, rax, t1, flags=(EZF,) + bri t0, label("divLoopTop"), flags=(nCEZF,) + + #Unload the answer + divq rax + divr rdx +}; + +# +# One byte version of signed division +# + +def macroop IDIV_B_R +{ + # Negate dividend + sub t1, t0, rax, flags=(ECF,), dataSize=1 + ruflag t4, 3 + sub t2, t0, rsi, dataSize=1 + sub t2, t2, t4 + + #Find the sign of the divisor + #FIXME!!! This depends on shifts setting the carry flag correctly. + slli t0, reg, 1, flags=(ECF,), dataSize=1 + + # Negate divisor + sub t3, t0, reg, dataSize=1 + # Put the divisor's absolute value into t3 + mov t3, t3, reg, flags=(nCECF,), dataSize=1 + + #Find the sign of the dividend + #FIXME!!! This depends on shifts setting the carry flag correctly. + slli t0, rsi, 1, flags=(ECF,), dataSize=1 + + # Put the dividend's absolute value into t1 and t2 + mov t1, t1, rax, flags=(nCECF,), dataSize=1 + mov t2, t2, rsi, flags=(nCECF,), dataSize=1 + + # Do the initial part of the division + div1 t2, t3, dataSize=1 + + #These are split out so we can initialize the number of bits in the + #second register + div2i t4, t1, 8, dataSize=1 + div2 t4, t1, t4, dataSize=1 + + #Loop until we're out of bits to shift in +divLoopTop: + div2 t4, t1, t4, dataSize=1 + div2 t4, t1, t4, flags=(EZF,), dataSize=1 + bri t0, label("divLoopTop"), flags=(nCEZF,) + + #Unload the answer + divq t5, dataSize=1 + divr t6, dataSize=1 + + # Fix up signs. The sign of the dividend is still lying around in ECF. + # The sign of the remainder, ah, is the same as the dividend. The sign + # of the quotient is negated if the signs of the divisor and dividend + # were different. + + # Negate the remainder + sub t4, t0, t6, dataSize=1 + # If the dividend was negitive, put the negated remainder in rsi. + mov rsi, rsi, t4, (CECF,), dataSize=1 + # Otherwise put the regular remainder in rsi. + mov rsi, rsi, t6, (nCECF,), dataSize=1 + + # Negate the quotient. + sub t4, t0, t5, dataSize=1 + # If the dividend was negative, start using the negated quotient + mov t5, t5, t4, (CECF,), dataSize=1 + + # Check the sign of the divisor + slli t0, t3, 1, flags=(ECF,), dataSize=1 + + # Negate the (possibly already negated) quotient + sub t4, t0, t5, dataSize=1 + # If the divisor was negative, put the negated quotient in rax. + mov rax, rax, t4, (CECF,), dataSize=1 + # Otherwise put the one that wasn't negated (at least here) in rax. + mov rax, rax, t5, (nCECF,), dataSize=1 +}; + +def macroop IDIV_B_M +{ + # Negate dividend + sub t1, t0, rax, flags=(ECF,), dataSize=1 + ruflag t4, 3 + sub t2, t0, rsi, dataSize=1 + sub t2, t2, t4 + + ld t3, seg, sib, disp + + #Find the sign of the divisor + #FIXME!!! This depends on shifts setting the carry flag correctly. + slli t0, t3, 1, flags=(ECF,), dataSize=1 + + # Negate divisor + sub t4, t0, t3, dataSize=1 + # Put the divisor's absolute value into t3 + mov t3, t3, t4, flags=(CECF,), dataSize=1 + + #Find the sign of the dividend + #FIXME!!! This depends on shifts setting the carry flag correctly. + slli t0, rsi, 1, flags=(ECF,), dataSize=1 + + # Put the dividend's absolute value into t1 and t2 + mov t1, t1, rax, flags=(nCECF,), dataSize=1 + mov t2, t2, rsi, flags=(nCECF,), dataSize=1 + + # Do the initial part of the division + div1 t2, t3, dataSize=1 + + #These are split out so we can initialize the number of bits in the + #second register + div2i t4, t1, 8, dataSize=1 + div2 t4, t1, t4, dataSize=1 + + #Loop until we're out of bits to shift in +divLoopTop: + div2 t4, t1, t4, dataSize=1 + div2 t4, t1, t4, flags=(EZF,), dataSize=1 + bri t0, label("divLoopTop"), flags=(nCEZF,) + + #Unload the answer + divq t5, dataSize=1 + divr t6, dataSize=1 + + # Fix up signs. The sign of the dividend is still lying around in ECF. + # The sign of the remainder, ah, is the same as the dividend. The sign + # of the quotient is negated if the signs of the divisor and dividend + # were different. + + # Negate the remainder + sub t4, t0, t6, dataSize=1 + # If the dividend was negitive, put the negated remainder in rsi. + mov rsi, rsi, t4, (CECF,), dataSize=1 + # Otherwise put the regular remainder in rsi. + mov rsi, rsi, t6, (nCECF,), dataSize=1 + + # Negate the quotient. + sub t4, t0, t5, dataSize=1 + # If the dividend was negative, start using the negated quotient + mov t5, t5, t4, (CECF,), dataSize=1 + + # Check the sign of the divisor + slli t0, t3, 1, flags=(ECF,), dataSize=1 + + # Negate the (possibly already negated) quotient + sub t4, t0, t5, dataSize=1 + # If the divisor was negative, put the negated quotient in rax. + mov rax, rax, t4, (CECF,), dataSize=1 + # Otherwise put the one that wasn't negated (at least here) in rax. + mov rax, rax, t5, (nCECF,), dataSize=1 +}; + +def macroop IDIV_B_P +{ + # Negate dividend + sub t1, t0, rax, flags=(ECF,), dataSize=1 + ruflag t4, 3 + sub t2, t0, rsi, dataSize=1 + sub t2, t2, t4 + + rdip t7 + ld t3, seg, riprel, disp + + #Find the sign of the divisor + #FIXME!!! This depends on shifts setting the carry flag correctly. + slli t0, t3, 1, flags=(ECF,), dataSize=1 + + # Negate divisor + sub t4, t0, t3, dataSize=1 + # Put the divisor's absolute value into t3 + mov t3, t3, t4, flags=(CECF,), dataSize=1 + + #Find the sign of the dividend + #FIXME!!! This depends on shifts setting the carry flag correctly. + slli t0, rsi, 1, flags=(ECF,), dataSize=1 + + # Put the dividend's absolute value into t1 and t2 + mov t1, t1, rax, flags=(nCECF,), dataSize=1 + mov t2, t2, rsi, flags=(nCECF,), dataSize=1 + + # Do the initial part of the division + div1 t2, t3, dataSize=1 + + #These are split out so we can initialize the number of bits in the + #second register + div2i t4, t1, 8, dataSize=1 + div2 t4, t1, t4, dataSize=1 + + #Loop until we're out of bits to shift in +divLoopTop: + div2 t4, t1, t4, dataSize=1 + div2 t4, t1, t4, flags=(EZF,), dataSize=1 + bri t0, label("divLoopTop"), flags=(nCEZF,) + + #Unload the answer + divq t5, dataSize=1 + divr t6, dataSize=1 + + # Fix up signs. The sign of the dividend is still lying around in ECF. + # The sign of the remainder, ah, is the same as the dividend. The sign + # of the quotient is negated if the signs of the divisor and dividend + # were different. + + # Negate the remainder + sub t4, t0, t6, dataSize=1 + # If the dividend was negitive, put the negated remainder in rsi. + mov rsi, rsi, t4, (CECF,), dataSize=1 + # Otherwise put the regular remainder in rsi. + mov rsi, rsi, t6, (nCECF,), dataSize=1 + + # Negate the quotient. + sub t4, t0, t5, dataSize=1 + # If the dividend was negative, start using the negated quotient + mov t5, t5, t4, (CECF,), dataSize=1 + + # Check the sign of the divisor + slli t0, t3, 1, flags=(ECF,), dataSize=1 + + # Negate the (possibly already negated) quotient + sub t4, t0, t5, dataSize=1 + # If the divisor was negative, put the negated quotient in rax. + mov rax, rax, t4, (CECF,), dataSize=1 + # Otherwise put the one that wasn't negated (at least here) in rax. + mov rax, rax, t5, (nCECF,), dataSize=1 }; # @@ -282,27 +609,225 @@ def macroop DIV_P def macroop IDIV_R { - divr t1, rax, reg - divq rax, rax, reg - mov rdx, rdx, t1 + # Negate dividend + sub t1, t0, rax, flags=(ECF,) + ruflag t4, 3 + sub t2, t0, rdx + sub t2, t2, t4 + + #Find the sign of the divisor + #FIXME!!! This depends on shifts setting the carry flag correctly. + slli t0, reg, 1, flags=(ECF,) + + # Negate divisor + sub t3, t0, reg + # Put the divisor's absolute value into t3 + mov t3, t3, reg, flags=(nCECF,) + + #Find the sign of the dividend + #FIXME!!! This depends on shifts setting the carry flag correctly. + slli t0, rdx, 1, flags=(ECF,) + + # Put the dividend's absolute value into t1 and t2 + mov t1, t1, rax, flags=(nCECF,) + mov t2, t2, rdx, flags=(nCECF,) + + # Do the initial part of the division + div1 t2, t3 + + #These are split out so we can initialize the number of bits in the + #second register + div2i t4, t1, "env.dataSize * 8" + div2 t4, t1, t4 + + #Loop until we're out of bits to shift in +divLoopTop: + div2 t4, t1, t4 + div2 t4, t1, t4 + div2 t4, t1, t4 + div2 t4, t1, t4, flags=(EZF,) + bri t0, label("divLoopTop"), flags=(nCEZF,) + + #Unload the answer + divq t5 + divr t6 + + # Fix up signs. The sign of the dividend is still lying around in ECF. + # The sign of the remainder, ah, is the same as the dividend. The sign + # of the quotient is negated if the signs of the divisor and dividend + # were different. + + # Negate the remainder + sub t4, t0, t6 + # If the dividend was negitive, put the negated remainder in rdx. + mov rdx, rdx, t4, (CECF,) + # Otherwise put the regular remainder in rdx. + mov rdx, rdx, t6, (nCECF,) + + # Negate the quotient. + sub t4, t0, t5 + # If the dividend was negative, start using the negated quotient + mov t5, t5, t4, (CECF,) + + # Check the sign of the divisor + slli t0, t3, 1, flags=(ECF,) + + # Negate the (possibly already negated) quotient + sub t4, t0, t5 + # If the divisor was negative, put the negated quotient in rax. + mov rax, rax, t4, (CECF,) + # Otherwise put the one that wasn't negated (at least here) in rax. + mov rax, rax, t5, (nCECF,) }; def macroop IDIV_M { - ld t1, seg, sib, disp - divr rdx, rax, t1 - divq rax, rax, t1 + # Negate dividend + sub t1, t0, rax, flags=(ECF,) + ruflag t4, 3 + sub t2, t0, rdx + sub t2, t2, t4 + + ld t3, seg, sib, disp + + #Find the sign of the divisor + #FIXME!!! This depends on shifts setting the carry flag correctly. + slli t0, t3, 1, flags=(ECF,) + + # Negate divisor + sub t4, t0, t3 + # Put the divisor's absolute value into t3 + mov t3, t3, t4, flags=(CECF,) + + #Find the sign of the dividend + #FIXME!!! This depends on shifts setting the carry flag correctly. + slli t0, rdx, 1, flags=(ECF,) + + # Put the dividend's absolute value into t1 and t2 + mov t1, t1, rax, flags=(nCECF,) + mov t2, t2, rdx, flags=(nCECF,) + + # Do the initial part of the division + div1 t2, t3 + + #These are split out so we can initialize the number of bits in the + #second register + div2i t4, t1, "env.dataSize * 8" + div2 t4, t1, t4 + + #Loop until we're out of bits to shift in +divLoopTop: + div2 t4, t1, t4 + div2 t4, t1, t4 + div2 t4, t1, t4 + div2 t4, t1, t4, flags=(EZF,) + bri t0, label("divLoopTop"), flags=(nCEZF,) + + #Unload the answer + divq t5 + divr t6 + + # Fix up signs. The sign of the dividend is still lying around in ECF. + # The sign of the remainder, ah, is the same as the dividend. The sign + # of the quotient is negated if the signs of the divisor and dividend + # were different. + + # Negate the remainder + sub t4, t0, t6 + # If the dividend was negitive, put the negated remainder in rdx. + mov rdx, rdx, t4, (CECF,) + # Otherwise put the regular remainder in rdx. + mov rdx, rdx, t6, (nCECF,) + + # Negate the quotient. + sub t4, t0, t5 + # If the dividend was negative, start using the negated quotient + mov t5, t5, t4, (CECF,) + + # Check the sign of the divisor + slli t0, t3, 1, flags=(ECF,) + + # Negate the (possibly already negated) quotient + sub t4, t0, t5 + # If the divisor was negative, put the negated quotient in rax. + mov rax, rax, t4, (CECF,) + # Otherwise put the one that wasn't negated (at least here) in rax. + mov rax, rax, t5, (nCECF,) }; def macroop IDIV_P { + # Negate dividend + sub t1, t0, rax, flags=(ECF,) + ruflag t4, 3 + sub t2, t0, rdx + sub t2, t2, t4 + rdip t7 - ld t1, seg, riprel, disp - divr rdx, rax, t1 - divq rax, rax, t1 + ld t3, seg, riprel, disp + + #Find the sign of the divisor + #FIXME!!! This depends on shifts setting the carry flag correctly. + slli t0, t3, 1, flags=(ECF,) + + # Negate divisor + sub t4, t0, t3 + # Put the divisor's absolute value into t3 + mov t3, t3, t4, flags=(CECF,) + + #Find the sign of the dividend + #FIXME!!! This depends on shifts setting the carry flag correctly. + slli t0, rdx, 1, flags=(ECF,) + + # Put the dividend's absolute value into t1 and t2 + mov t1, t1, rax, flags=(nCECF,) + mov t2, t2, rdx, flags=(nCECF,) + + # Do the initial part of the division + div1 t2, t3 + + #These are split out so we can initialize the number of bits in the + #second register + div2i t4, t1, "env.dataSize * 8" + div2 t4, t1, t4 + + #Loop until we're out of bits to shift in +divLoopTop: + div2 t4, t1, t4 + div2 t4, t1, t4 + div2 t4, t1, t4 + div2 t4, t1, t4, flags=(EZF,) + bri t0, label("divLoopTop"), flags=(nCEZF,) + + #Unload the answer + divq t5 + divr t6 + + # Fix up signs. The sign of the dividend is still lying around in ECF. + # The sign of the remainder, ah, is the same as the dividend. The sign + # of the quotient is negated if the signs of the divisor and dividend + # were different. + + # Negate the remainder + sub t4, t0, t6 + # If the dividend was negitive, put the negated remainder in rdx. + mov rdx, rdx, t4, (CECF,) + # Otherwise put the regular remainder in rdx. + mov rdx, rdx, t6, (nCECF,) + + # Negate the quotient. + sub t4, t0, t5 + # If the dividend was negative, start using the negated quotient + mov t5, t5, t4, (CECF,) + + # Check the sign of the divisor + slli t0, t3, 1, flags=(ECF,) + + # Negate the (possibly already negated) quotient + sub t4, t0, t5 + # If the divisor was negative, put the negated quotient in rax. + mov rax, rax, t4, (CECF,) + # Otherwise put the one that wasn't negated (at least here) in rax. + mov rax, rax, t5, (nCECF,) }; ''' -#let {{ -# class IDIV(Inst): -# "GenFault ${new UnimpInstFault}" -#}}; |