Academia.eduAcademia.edu

Figure 1 – uploaded by Camil Demetrescu

See full PDFdownloadDownload figure
We now address each step in detail using the example given in Listing 1.5:

Figure 1 We now address each step in detail using the example given in Listing 1.5:

Related Figures (30)

The following instructions correspond to the lines 3-6 of the program in Listing 1.1:
The following instructions correspond to the lines 3-6 of the program in Listing 1.1:
The code generated by gcc -S first2.c for lines 3-6 of Listing 1.4 is:
The code generated by gcc -S first2.c for lines 3-6 of Listing 1.4 is:
Table 1.4: System V AMD64 calling conventions: parameter passing.
Table 1.4: System V AMD64 calling conventions: parameter passing.
Notice that the non-negative constant actual parameters 1, 2, 3, etc. are written in the lowest 32 bits of 64-bit destination registers, while the upper 32 bits are automatically filled with zero (see word-level register write access in Section 1.1.2). 1. Parameter passing. Parameters are passed by the caller to the callee in accordance with the con- ventions listed in Table 1.4 (left). Up to 6 arguments (each of size up to 8 bytes) can be passed using registers rdi, rsi, rdx, rcx, r8, and r9. Further parameters are pushed on stack in reverse order so that the 7th argument is pointed to by register rsp prior to the procedure call. For instance, the eight parameters of the call f(1, 2, 3, 4, 5, 6, 7, 8) atline 6 of Listing 1.5
Notice that the non-negative constant actual parameters 1, 2, 3, etc. are written in the lowest 32 bits of 64-bit destination registers, while the upper 32 bits are automatically filled with zero (see word-level register write access in Section 1.1.2). 1. Parameter passing. Parameters are passed by the caller to the callee in accordance with the con- ventions listed in Table 1.4 (left). Up to 6 arguments (each of size up to 8 bytes) can be passed using registers rdi, rsi, rdx, rcx, r8, and r9. Further parameters are pushed on stack in reverse order so that the 7th argument is pointed to by register rsp prior to the procedure call. For instance, the eight parameters of the call f(1, 2, 3, 4, 5, 6, 7, 8) atline 6 of Listing 1.5
The setup is the same as in the case of the main function, except that no room below rbp is allocated (see Figure 1.2 on the right).
The setup is the same as in the case of the main function, except that no room below rbp is allocated (see Figure 1.2 on the right).
4, Callee execution. The callee executes the instructions in its body. In our example, the body returns in rax the sum of the arguments:
4, Callee execution. The callee executes the instructions in its body. In our example, the body returns in rax the sum of the arguments:
The setup consists of saving the current value of rbp on stack, letting rbp point to the new stack frame, and making room for parameters to be passed to f (see Figure 1.2 on the left). The callee setup of function f of Listing 1.5 is: In our example, the callee setup of function main of Listing 1.5 is done as follows:
The setup consists of saving the current value of rbp on stack, letting rbp point to the new stack frame, and making room for parameters to be passed to f (see Figure 1.2 on the left). The callee setup of function f of Listing 1.5 is: In our example, the callee setup of function main of Listing 1.5 is done as follows:
Listing 1.6: x86-64 code for functions f and main of Listing 1.5 Listing 1.6 summarizes the full x86-64 code generated by gcc -S -02! for functions f and main of Listing 1.5. 1_Q2 means optimization level 2 and will be discussed later in this chapter.
Listing 1.6: x86-64 code for functions f and main of Listing 1.5 Listing 1.6 summarizes the full x86-64 code generated by gcc -S -02! for functions f and main of Listing 1.5. 1_Q2 means optimization level 2 and will be discussed later in this chapter.
Listing 1.8: swap -00. S without register allocation (gcc -00)
Listing 1.8: swap -00. S without register allocation (gcc -00)
LO Experimental Analysis To assess the performance boost given by register allocation on the swap function, we measured the time required to execute 1 billion swaps of two int variables:
LO Experimental Analysis To assess the performance boost given by register allocation on the swap function, we measured the time required to execute 1 billion swaps of two int variables:
Listing 1.11; Test program for Swap defined as a macrc Preprocessing the program with gcc -E yields the following code in which the occurrence of the macro name is replaced by the token sequence of the macro definition, which is inlined in the code of main: We illustrate these ideas by considering again the example of Section 1.2.1 and replacing the swap function with a C macro that performs the same task: Calling a function incurs some overhead due to stack, registers, and control flow operations. If a function is frequently called and its body is reasonably short, it may be convenient to expand each call with the instructions in the body itself, saving the time required for function activations at the price of increasing the code size. This technique is called function inlining and can be done by programmers at source code level using C macros. Code inlining has the additional benefits of allowing further local optimizations that would not be applied by the compiler across function calls, such as register allocation. As we will see, under some circumstances, compilers can do inlining of regular functions automatically. OF ee eco ee: Dives 2A scutes “kes coe dimetleecss monnifos des sccm « acl? Deondkccws 4&4 G4 weed) coms) wcteess Dh cee wm ce
Listing 1.11; Test program for Swap defined as a macrc Preprocessing the program with gcc -E yields the following code in which the occurrence of the macro name is replaced by the token sequence of the macro definition, which is inlined in the code of main: We illustrate these ideas by considering again the example of Section 1.2.1 and replacing the swap function with a C macro that performs the same task: Calling a function incurs some overhead due to stack, registers, and control flow operations. If a function is frequently called and its body is reasonably short, it may be convenient to expand each call with the instructions in the body itself, saving the time required for function activations at the price of increasing the code size. This technique is called function inlining and can be done by programmers at source code level using C macros. Code inlining has the additional benefits of allowing further local optimizations that would not be applied by the compiler across function calls, such as register allocation. As we will see, under some circumstances, compilers can do inlining of regular functions automatically. OF ee eco ee: Dives 2A scutes “kes coe dimetleecss monnifos des sccm « acl? Deondkccws 4&4 G4 weed) coms) wcteess Dh cee wm ce
Listing 1.13: Assembly code for the program of Listing 1.11 generated by gcc -S -O1
Listing 1.13: Assembly code for the program of Listing 1.11 generated by gcc -S -O1
Listing 1.15: Assembly code for the program of Listing 1.14 generated by gcc -S -O1
Listing 1.15: Assembly code for the program of Listing 1.14 generated by gcc -S -O1
Compiling the program with gcc -S -01yields the following assembly code for the main function:
Compiling the program with gcc -S -01yields the following assembly code for the main function:
Notice that in the code above we could also apply constant folding, replacing 8 - 2 with 6. Constant propagation is not done by default by gcc:
Notice that in the code above we could also apply constant folding, replacing 8 - 2 with 6. Constant propagation is not done by default by gcc:
Listing 1.23: Test program for the f function of Listing 1.19
Listing 1.23: Test program for the f function of Listing 1.19
Notice that the code computes the expression (x + y) twice at lines 6-8 and 9-10. Compiling al optimization level 1 (-01) enables register allocation and common subexpression elimination in gcc: 1.2.6 Expression Simplification
Notice that the code computes the expression (x + y) twice at lines 6-8 and 9-10. Compiling al optimization level 1 (-01) enables register allocation and common subexpression elimination in gcc: 1.2.6 Expression Simplification
Listing 1.26: Assembly code of Listing 1.24 generated by gcc -S -00
Listing 1.26: Assembly code of Listing 1.24 generated by gcc -S -00
1.2.7 Dead Code Elimination Dead code elimination consists of removing portions of code that is either unnecessary, or unreachable Consider for instance the following code fragment:
1.2.7 Dead Code Elimination Dead code elimination consists of removing portions of code that is either unnecessary, or unreachable Consider for instance the following code fragment:
Figure 2.2: Blocked matrix multiplication algorithm with blocking parameter s. Symbols H and &X denote standard addition and multiplication between matrices: any internal memory algorithm can be used to implement H and &.
Figure 2.2: Blocked matrix multiplication algorithm with blocking parameter s. Symbols H and &X denote standard addition and multiplication between matrices: any internal memory algorithm can be used to implement H and &.
Table 3.1: Analisi sperimentale del tempo totale e del tempo medio per operazione su una sequenza di k = 2” incrementi di un contatore binario a n bit.
Table 3.1: Analisi sperimentale del tempo totale e del tempo medio per operazione su una sequenza di k = 2” incrementi di un contatore binario a n bit.
Figure 3.5: Dati mantenuti dall’implementazione 2 del tipo di dato DistGrafoDinamico: lista [n(v) degli entranti in v, puntatore curr(v) alla posizione corrente in n(v), e insieme Out(v) degli uscenti da v. I predecessori del nodo corrente nella lista Jn(v) hanno una distanza dalla sorgente strettamente maggiore di quella del nodo corrente stesso, e pertanto non sono utili se si cerca un cammino alternativo che preservi la distanza di v a fronte di un aumento di distanza del nodo puntato da curr(v).
Figure 3.5: Dati mantenuti dall’implementazione 2 del tipo di dato DistGrafoDinamico: lista [n(v) degli entranti in v, puntatore curr(v) alla posizione corrente in n(v), e insieme Out(v) degli uscenti da v. I predecessori del nodo corrente nella lista Jn(v) hanno una distanza dalla sorgente strettamente maggiore di quella del nodo corrente stesso, e pertanto non sono utili se si cerca un cammino alternativo che preservi la distanza di v a fronte di un aumento di distanza del nodo puntato da curr(v).
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved