12/22/2016 Control Flow Computers execute instructions in sequence. - PDF document

12/22/2016 Control Flow Computers execute instructions in sequence. Except when we change the flow of control  Jump and Call instructions Controlling Program Flow  Unconditional jump  Direct jump: jmp Label » Jump target is specified by a label (e.g., jmp .L1 )  Indirect jump: jmp *Operand » Jump target is specified by a register or memory location (e.g., jmp *%rax ) – 2 – Conditional statements Condition codes Processor flag register eflags Some jumps are conditional (extended flags)  A computer needs to jump if certain a condition is true Registers  In C, if, for , and while statements Flags are set or cleared by %rax %r8 if (x) {…} else {…} depending on the result of %rbx %r9 an instruction %rcx %r10 while (x) {…} %rdx %r11 Each bit is a flag, or condition %rsi %r12 code do {…} while (x) %rdi %r13 CF Carry Flag SF Sign Flag %rsp %r14 for (i=0; i<max; i++) {…} ZF Zero Flag OF Overflow Flag %rbp %r15 %rip switch (x) { case 1: … case 2: … Condition codes } CF ZF SF OF – 3 – – 4 – 1

12/22/2016 Implicit setting Explicit setting via compare Automatically Set By Arithmetic and Logical Operations Setting condition codes via compare instruction Example: addq Src , Dest cmpq b , a C analog: t = a + b  Computes a-b without setting destination  CF (for unsigned integers)  CF set if carry out from most significant bit  set if carry out from most significant bit (unsigned overflow)  Used for unsigned comparisons (unsigned long t) < (unsigned long a)  ZF set if a == b  ZF (zero flag)  SF set if (a-b) < 0  set if t == 0  SF (for signed integers)  OF set if two’s complement (signed) overflow  set if t < 0 (a>0 && b<0 && (a-b)<0) || (a<0 && b>0 && (a-b)>0)  OF (for signed integers)  Byte, word, and double word versions cmpb, cmpw, cmpl  set if signed (two’s complement) overflow (a>0 && b>0 && t<0) || (a<0 && b<0 && t>=0) Not set by lea , push, pop, mov instructions – 5 – – 6 – Explicit setting via test Conditional jump instrcutions Setting condition codes via test instruction Jump to different part of code based on condition codes testq b , a jX Condition Description  Computes a&b without setting destination jmp 1 Unconditional  Sets condition codes based on result je, jz ZF Equal / Zero  Useful to have one of the operands be a mask jne,jnz ~ZF Not Equal / Not Zero  Often used to test zero, positive js SF Negative testq %rax, %rax jns ~SF Nonnegative  ZF set when a&b == 0  SF set when a&b < 0 jg ~(SF^OF)&~ZF Greater (Signed)  Byte, word and double word versions testb, testw, jge ~(SF^OF) Greater or Equal (Signed) testl jl (SF^OF) Less (Signed) jle (SF^OF)|ZF Less or Equal (Signed) ja ~CF&~ZF Above (unsigned) jb CF Below (unsigned) – 7 – – 8 – Overflow flips result 2

12/22/2016 Jump instructions Conditional jump example Non-optimized What’s the difference between jg and ja ? gcc –Og -S –fno-if-conversion control.c absdiff: cmpq %rsi, %rdi # x:y Which one would you use to compare two pointers? long absdiff(long x, long y) jle .L4 { movq %rdi, %rax long result; subq %rsi, %rax if (x > y) ret result = x-y; .L4: # x <= y else movq %rsi, %rax result = y-x; subq %rdi, %rax return result; ret } Register Use(s) %rdi Argument x %rsi Argument y %rax Return value – 9 – – 10 – General Conditional Expression Practice problem 3.18 Translation (Using Branches) /* x in %rdi, y in %rsi, z in %rdx */ long test(long x, long y, long z) test: { C Code leaq (%rdi,%rsi), %rax addq %rdx, %rax x+y+z long val = ____________ ; val = Test ? Then_Expr : Else_Expr ; cmpq $-3, %rdi if ( ____________ ) { x < -3 jge .L2 cmp %rdx,%rsi y < z if ( ____________ ) jge .L3 val = x>y ? x-y : y-x; x*y movq %rdi, %rax val = ____________ ; imulq %rsi, %rax else ret y*z .L3: val = ____________ ; movq %rsi, %rax Goto Version } else if ( ____________ ) x > 2 imulq %rdx,%rax ret ntest = ! Test ; val = ____________ ; x*z Create separate code regions for .L2 if (ntest) goto Else; then & else expressions cmpq $2, %rdi return val; val = Then_Expr ; jle .L4 } movq %rdi, %rax goto Done; imulq %rdx, %rax Else: Execute appropriate one .L4 ret val = Else_Expr ; Done: . . . – 11 – – 12 – 3

12/22/2016 Avoiding conditional branches Conditional moves Conditional instruction exectuion Modern CPUs with deep pipelines cmovXX Src, Dest  Instructions fetched far in advance of execution  Move value from src to dest if condition XX holds  Mask the latency going to memory  No branching  Problem: What if you hit a conditional branch?  Handled as operation within Execution Unit  Must predict which branch to take!  Added with P6 microarchitecture (PentiumPro onward, 1995)  Branch prediction in CPUs well-studied, fairly effective Example  But, best to avoid conditional branching altogether # %rdi = x, %rsi = y # return value in %rax returns max(x,y) movq %rdi,%rdx # Get x movq %rsi,%rax # rval=y cmpq %rdx, %rax # rval:x cmovl %rdx,%rax # If <, rval=x Performance  14 cycles on all data  More efficient than conditional branching (single control flow path)  But overhead: both branches are evaluated – 13 – – 14 – General Conditional Expression Conditional Move example Translation (Using conditional move) Branch version long absdiff(long x, long y) absdiff: Conditional Move template { C Code cmpq %rsi, %rdi # x:y long result;  Instruction supports jle .L4 val = Test ? Then_Expr : Else_Expr ; if (x > y)  if (Test) Dest  Src movq %rdi, %rax result = x-y; subq %rsi, %rax  GCC attempts to restructure else result = Then_Expr ; ret execution to avoid disruptive result = y-x; .L4: # x <= y eval = Else_Expr ; conditional branch return result; movq %rsi, %rax nt = ! Test ; }  Both values computed subq %rdi, %rax if (nt) result = eval; ret  Overwrite “then”-value with return result; “else”-value if condition doesn’t hold Branch version absdiff: Register Use(s) movq %rdi, %rax # x  Conditional moves do not ntest = ! Test ; %rdi Argument x subq %rsi, %rax # result = x-y transfer control if (ntest) goto Else; movq %rsi, %rdx %rsi Argument y val = Then_Expr ; subq %rdi, %rdx # eval = y-x goto Done; %rax Return value cmpq %rsi, %rdi # x:y Else: cmovle %rdx, %rax # if <=, result = eval val = Else_Expr ; ret Done: – 15 – – 16 – 4

12/22/2016 Practice problem 3.21 When not to use Conditional Move /* x in %rdi, y in %rsi */ long test(long x, long y) Expensive computations test: { leaq 0(,%rdi,8), %rax val = Test(x) ? Hard1(x) : Hard2(x); testq %rsi, %rsi 8*x long val = ____________ ; jle .L2 y > 0 if ( ____________ ) { movq %rsi, %rax subq %rdi, %rax  Both Hard1(x) and Hard2(x) computed x < y if ( ____________ ) movq %rdi, %rdx y-x andq %rsi, %rdx val = ____________ ;  Use branching when “then” and “else” expressions are more cmpq %rsi, %rdi expensive than branch misprediction else cmovge %rdx, %rax ret x&y val = ____________ ; Computations with side effects .L2: y <= -2 } else if ( ____________ ) addq %rsi, %rdi val = x > 0 ? x*=7 : x+=3; cmpq $-2, %rsi x+y val = ____________ ; cmovle %rdi, %rax return val; ret  Executing both values causes incorrect behavior } Condition must hold to prevent fault  Null pointer check – 17 – – 18 – Loops C example long factorial_do(long x) Implemented in assembly via tests and jumps { long result = 1;  Compilers implement most loops as do-while do {  Add additional check at beginning to get “ while-do ” result *= x; x = x-1; } while (x > 1); do { return result; body-statements } } while (test-expr); factorial_do: movq $1, %rax ; result = 1 .L2: imulq %rdi, %rax ; result *= x subq $1, %rdi ; x = x - 1 cmpq $1, %rdi ; if x > 1 jg .L2 ; goto loop ret ; return result http://thefengs.com/wuchang/courses/cs201/class/07 – 19 – – 20 – 5

12/22/2016 Are these equivalent? C code of while-do Assembly of while-do C code of do-while Assembly of do-while long factorial_do(long x) long factorial_while(long x) factorial_do: factorial_while: movq $1, %rax movq $1, %rax { { .L2: jmp .L2 long result = 1; long result = 1; imulq %rdi, %rax .L3: do { while (x > 1) { subq $1, %rdi imulq %rdi, %rax cmpq $1, %rdi subq $1, %rdi result *= x; result *= x; jg .L2 .L2: x = x-1; x = x-1; ret cmpq $1, %rdi } while (x > 1); } jg .L3 return result; return result; rep ret } } http://thefengs.com/wuchang/courses/cs201/class/07 diff factorial_do.s factorial_while.s – 21 – – 22 – “For” Loop Example “For” Loop Example General Form long factorial_for(long x) factorial_while: factorial_for: movq $1, %rax movq $1, %rax { for ( Init ; Test ; Update ) jmp .L2 jmp .L2 long result; .L3: .L3: Body for (result=1; x > 1; x=x-1) { imulq %rdi, %rax imulq %rdi, %rax result *= x; subq $1, %rdi subq $1, %rdi .L2: .L2: } cmpq $1, %rdi cmpq $1, %rdi return result; jg .L3 jg .L3 } ret ret Init Test Update result = 1 x > 1 x = x - 1 { http://thefengs.com/wuchang/courses/cs201/class/07 result *= x; Body diff factorial_for.s factorial_while.s } Is this code equivalent to the do-while version or the while-do version? – 23 – – 24 – 6