aboutsummaryrefslogtreecommitdiffstats
path: root/lib/hipe/regalloc/hipe_amd64_specific_sse2.erl
diff options
context:
space:
mode:
authorBjörn Gustavsson <[email protected]>2015-12-17 14:50:43 +0100
committerBjörn Gustavsson <[email protected]>2016-01-07 13:48:59 +0100
commit7c72e25926e153811ff099057bea649afa0be376 (patch)
tree710f2b5c7ce9f1e390474e0b3c7a03c2008b7b6f /lib/hipe/regalloc/hipe_amd64_specific_sse2.erl
parent82a835d94be7ee5e98d101a29999fedaf6cd75fe (diff)
downloadotp-7c72e25926e153811ff099057bea649afa0be376.tar.gz
otp-7c72e25926e153811ff099057bea649afa0be376.tar.bz2
otp-7c72e25926e153811ff099057bea649afa0be376.zip
beam_bool: Fix unsafe optimization
beam_bool would make the following code unsafe (which would be reported by beam_validator): scotland(Echo) -> found(case Echo of Echo when true; Echo, Echo, Echo -> Echo; echo -> [] end, Echo = placed). found(_, _) -> million. Basically, beam_bool would see that the 'case' would always return the value of Echo. Thus: scotland(Echo) -> found(Echo, Echo = placed). The only problem is that beam_bool would also remove a 'move' instruction that would save Echo to the stack. Here is the assembly code for part of the function: {allocate_zero,1,1}. {move,{x,0},{y,0}}. %% Save Echo on stack. {bif,'=:=',{f,7},[{x,0},{atom,true}],{x,1}}. {bif,'=:=',{f,7},[{x,0},{atom,true}],{x,2}}. {bif,'=:=',{f,7},[{x,0},{atom,true}],{x,3}}. {bif,'and',{f,7},[{x,2},{x,3}],{x,2}}. {bif,'and',{f,7},[{x,1},{x,2}],{x,1}}. {jump,{f,8}}. {label,7}. {move,{atom,false},{x,1}}. {label,8}. {bif,'or',{f,6},[{atom,true},{x,1}],{x,1}}. {test,is_eq_exact,{f,6},[{x,1},{atom,true}]}. %% Jump never taken. {jump,{f,5}}. {label,6}. {test,is_eq_exact,{f,9},[{x,0},{atom,echo}]}. {move,nil,{x,0}}. {jump,{f,5}}. {label,9}. {test_heap,3,0}. {put_tuple,2,{x,0}}. {put,{atom,case_clause}}. {put,{y,0}}. {line,[{location,"t.erl",5}]}. {call_ext,1,{extfunc,erlang,error,1}}. {jump,{f,5}}. {label,5}. {test,is_eq_exact,{f,12},[{atom,placed},{y,0}]}. beam_bool would see that the is_eq_exact test at label 8 would always succeed. It could therefore remove most of the code before the jump to label 5. Unfortunately it also removed the essential move of Echo to the stack: {allocate_zero,1,1}. %% Instruction incorrectly removed: {move,{x,0},{y,0}}. {jump,{f,5}}. {label,5}. {test,is_eq_exact,{f,12},[{atom,placed},{y,0}]}. The root cause of the problem is that the 'move' instruction is included in the block of 'bif' instructions before label 8. Normally the 'move' instruction would not have been discarded, but because the left operand to the 'or' BIF is 'true', the entire block with 'bif' instructions are dropped. As far as I can see, there is no gain by including 'move' instructions in the first place. There is no way that better code will be produced. In fact, the entire optimization can be given up if 'move' instructions are found in the block. Thus we can fix this bug by never including any 'move' instructions in the block of 'bif' instructions. We can also remove all the code that deals with 'move' instructions within blocks. Reported-by: Thomas Arts
Diffstat (limited to 'lib/hipe/regalloc/hipe_amd64_specific_sse2.erl')
0 files changed, 0 insertions, 0 deletions