Skip to content

gergo-/missed-optimizations

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

27 Commits
 
 
 
 
 
 
 
 

Repository files navigation

Missed optimizations in C compilers

This is a list of some missed optimizations in C compilers. To my knowledge, these have not been reported by others (but it is difficult to search for such things, and people may not have bothered to write these up before). I have reported some of these in bug trackers and/or developed more or less mature patches for some; see links below.

These are not correctness bugs, only cases where a compiler misses an opportunity to generate simpler or otherwise (seemingly) more efficient code. None of these are likely to affect the actual, measurable performance of real applications. Also, I have not measured most of these: The seemingly inefficient code may somehow turn out to be faster on real hardware.

I have tested GCC, Clang, and CompCert, mostly targeting ARM at -O3 and with -fomit-frame-pointer. The exact versions tested are specified below. For each example, at least one of these three compilers (typically GCC or Clang) produces the "right" code.

This list was put together by Gergö Barany [email protected]. I'm interested in feedback. I have described how I found these missed optimizations in a paper (PDF) that I presented at CC'18, where it won the best paper award. The software described in the paper is not released yet. I also made a quick-and-dirty poster (PDF) and some more informative presentation slides (PDF).

Since initial publication, there have been some insightful comments on this list on Reddit and on Hacker News.

Added 2018-03-25

Unless noted otherwise, the examples below can be reproduced with the following compiler versions and configurations:

Compiler Version Revision Date Configuration
GCC 8.0.1 20180325 r258845 2018-03-25 --target=armv7a-eabihf --with-arch=armv7-a --with-fpu=vfpv3-d16 --with-float-abi=hard --with-float=hard
Clang 7.0.0 (trunk 328450) LLVM r328450, Clang r328447 2018-03-25 --target=armv7a-eabihf -march=armv7-a -mfpu=vfpv3-d16 -mfloat-abi=hard

GCC

Loop vectorization regression

This issue concerns GCC for x86-64, configured with --target=x86_64-pc-linux-gnu. It used to unroll and vectorize the loop in the following function:

int N;
long fn1(void) {
  short i;
  long a;
  i = a = 0;
  while (i < N)
    a -= i++;
  return a;
}

After revision r258036 (2018-02-27), this was no longer the case, and the non-vectorized version showed a slowdown of about 1.8x on my machine.

Reported at https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84986, fixed.

Clang

Spill despite available registers

double fn1(double p1, double p2, double p3, double p4, double p5, double p6,
           double p7, double p8) {
  double a, b, c, d, v, e;
  a = p4 - 5 + 9 * p1 * 7;
  c = 2 - p6 + 1;
  b = 3 * p6 * 4 - (p2 + 10) * a;
  d = (8 + p3) * p5 * 5 * p7 - p4;
  v = p1 - p8 - 2 - (p8 + p2 - b) * ((p3 + 6) * c);
  e = v * p5 + d;
  return e;
}

I will omit the full code generated by Clang, but you can see it at https://godbolt.org/g/8ZuYbz (generated by a slightly older version, but the code is the same).

The interesting issue is the treatment of register d6, which is spilled in the middle of the function and used for temporaries in some calculations, before its value is reloaded into d1 towards the end of the function:

    ... some computations omitted ...
    vmov.f64    d13, #4.000000e+00
    vmov.f64    d14, #1.000000e+01
    vmul.f64    d10, d10, d13           @ last use of d13
    vadd.f64    d11, d1, d14            @ last use of d14
    vmov.f64    d12, #2.000000e+00
    vmov.f64    d15, #8.000000e+00
    vsub.f64    d5, d12, d5
    vadd.f64    d12, d2, d15            @ last use of d15
    vmls.f64    d10, d11, d9
    vstr    d6, [sp]                    @ spill original value of d6
    vmov.f64    d6, #6.000000e+00       @ use d6 for constant 6.0
    vmov.f64    d8, #1.000000e+00
    vadd.f64    d1, d1, d7
    vadd.f64    d2, d2, d6              @ last use of constant 6.0 in d6
    vadd.f64    d5, d5, d8
    vsub.f64    d0, d0, d7
    vmul.f64    d7, d12, d4
    vmov.f64    d8, #5.000000e+00
    vmov.f64    d6, #-2.000000e+00      @ use d6 for constant -2.0
    vmul.f64    d2, d2, d5
    vsub.f64    d1, d1, d10
    vmul.f64    d5, d7, d8
    vadd.f64    d0, d0, d6              @ last use of constant -2.0 in d6
    vmls.f64    d0, d2, d1
    vldr    d1, [sp]                    @ reload original value of d6
    vnmls.f64    d3, d5, d1
    ...

It is reasonable to spill d6 and to use it for temporaries during the computation if no other registers are free. But that is not the case here: The registers d13, d14, and d15 have short previous uses but are then unused for the rest of the function. The live ranges allocated to d6 in the above code could have been allocated to these registers instead, saving the spill of d6 and its reload into d1.

In this particular case this spill causes an overhead of the store, the load, and one instruction each to allocate and free the stack frame. GCC generates fewer instructions overall, with lower register use and without this inline spill.

Reported at https://bugs.llvm.org/show_bug.cgi?id=37073. It turns out that this is due to bad spilling decisions before register allocation that are not visible in the example above because they are undone by smart scheduling after register allocation. Can be fixed by selecting a different scheduler to run before allocation instead of the (bad) default choice.

Excessive copying

double fn1(double p1, double p2, double p3, double p4, double p5, double p6,
           double p7, double p8) {
  double a, b, c, d, e, f;
  a = d = b = p1 + p3;
  e = p3 + p2 * 8.7 + ((p4 - p4) * (a * p7) - 1.7);
  c = b - d - (p2 - 7.4) * 8.1;
  f = e - p2 * p3 - d * (4.6 + c) * p4 + p5 * p6 - p7 - p8 + p1;
  return f;
}

GCC's code starts with an add, a multiply, and a subtract (interspersed with constant loads I omit here):

    vadd.f64    d8, d0, d2
    vmul.f64    d9, d8, d6
    vsub.f64    d10, d3, d3

The subtraction is independent of the other instructions, it corresponds to the expression (p4 - p4) in the source code.

Clang starts with the same operations, but for unclear reasons it inserts four (!) register copy instructions, at least two of which have the effect of "setting up" the register d4 for use by the subtraction, although it could just use the values from d3.

    vadd.f64    d8, d0, d2
    vmov.f64    d10, d7
    vmov.f64    d7, d5
    vmov.f64    d5, d4
    vmov.f64    d4, d3
    vmul.f64    d11, d8, d6
    vsub.f64    d12, d4, d4

This seems to be related to scheduling as well; selecting a different scheduler with -mllvm -pre-RA-sched=list-burr makes it go away.

Added 2017-11-04

Unless noted otherwise, the examples below can be reproduced with the following compiler versions and configurations:

Compiler Version Revision Date Configuration
GCC 8.0.0 20171101 r254306 2017-11-01 --target=armv7a-eabihf --with-arch=armv7-a --with-fpu=vfpv3-d16 --with-float-abi=hard --with-float=hard
Clang 6.0.0 (trunk 317088) LLVM r317088, Clang r317076 2017-11-01 --target=armv7a-eabihf -march=armv7-a -mfpu=vfpv3-d16 -mfloat-abi=hard
CompCert 3.1 0c00ace 2017-10-26 armv7a-linux

GCC

Two dead stores for type conversion

Similar to a case below, but maybe more striking:

unsigned fn1(float p1, double p2) {
  unsigned a = p2;
  if (p1)
    a = 8;
  return a;
}

GCC predicates the entire function, nicely moving the conversion of p2 to unsigned into a branch. But for some reason, in both branches it spills the value for a to the stack, without ever reloading it:

    vcmp.f32    s0, #0
    sub sp, sp, #8
    vmrs    APSR_nzcv, FPSCR
    vcvteq.u32.f64  s15, d1
    movne   r3, #8
    strne   r3, [sp, #4]
    movne   r0, r3
    vmoveq  r0, s15 @ int
    vstreq.32   s15, [sp, #4]   @ int
    add sp, sp, #8
    @ sp needed
    bx  lr

Missed simplification of modulo-and

short fn1(int p1) {
  signed a = (p1 % 4) & 1;
  return a;
}

GCC computes modulo 4 as bitwise-and 3, carefully handling the case where p1 might be negative, then does the bitwise-and 1:

    rsbs    r3, r0, #0
    and r0, r0, #3
    and r3, r3, #3
    rsbpl   r0, r3, #0
    and r0, r0, #1

But the result of (p1 % 4) & 1 is 1 iff p1 % 4 is odd iff p1 is odd iff p1 & 1 is 1, so Clang just computes this directly:

    and r0, r0, #1

Missed simplification of division in condition

In the function:

char fn1(char p1) {
  signed a = 4;
  if (p1 / (p1 + 3))
    a = 1;
  return a;
}

GCC misses the fact that the condition is always false because p1 + 3 is always greater than p1 (as char is unsigned on ARM), and generates code to evaluate it:

    add r1, r0, #3
    bl  __aeabi_idiv
    cmp r0, #0
    moveq   r0, #4
    movne   r0, #1

Clang returns 4 unconditionally.

Missed simplification/floating-point constant propagation?

float fn1(short p1) {
  char a;
  float b, c;
  a = p1;
  b = c = 765;
  if (a == b)
    c = 0;
  return c;
}

At the comparison a == b, a is a char, and it cannot hold a value that compares equal to b's value of 765.

GCC computes the comparison anyway, Clang returns 765.0f unconditionally.

Missed bitwise tricks

char fn1(short p1, int p2) {
  long a;
  int v;
  a = p2 & (p1 | 415615);
  v = a << 1;
  return v;
}

The least significant byte of 415615 is 0x7f, so the lowest 7 bits of a are the same as the lowest 7 bits of p2, hence the lowest 8 bits of v are the same as the lowest 8 bits of p2 << 1, and that is all that is needed for the result. GCC generates all the operations present at the source level, while Clang simplifies nicely:

    lsl r0, r1, #1
    uxtb    r0, r0

Unnecessary spilling due to badly scheduled move-immediates

This is just a much smaller example of an issue with the same title below.

long fn1(char p1, short p2, long p3, long p4, short p5, long p6) {
  long a, b = 1000893;
  a = 600 - p2 * 3;
  if (p4 >= p5 * !(p6 * a))
    b = p3;
  return b;
}

GCC loads the constant 1000893 into a register early on, causing it to spill a register, but only uses the value much later and conditionally:

    str lr, [sp, #-4]!      @ spill lr
    movw    lr, #17853      @ move lower part of 1000893 into lr
    ldr r1, [sp, #8]
    movt    lr, 15          @ move upper part of 1000893 into lr
    ...
    cmp r1, r3
    movle   r0, r2
    movgt   r0, lr          @ use 1000893

Clang just loads the constant conditionally, at the point where it is actually needed, and avoids spilling:

    ...
    movwgt  r2, #17853
    movtgt  r2, #15
    mov r0, r2

Clang

Missed simplification of integer-valued floating-point arithmetic

In some cases (see one further below), Clang can perform floating-point arithmetic in integers where the result is an exact integer; GCC doesn't usually do this. But here is a counterexample:

char fn1(char p1) {
  float a;
  short b;
  a = !p1;
  b = -a;
  return b;
}

Clang converts to float, negates, and converts back:

    vmov.f32    s2, #1.000000e+00
    vldr    s0, .LCPI0_0        @ load 0.0
    cmp r0, #0
    vmoveq.f32  s0, s2
    vneg.f32    s0, s0
    vcvt.s32.f32    s0, s0
    vmov    r0, s0

GCC doesn't:

    cmp r0, #0
    moveq   r0, #255
    movne   r0, #0

Missed conditional constant propagation on double

short fn1(double p1) {
  unsigned a = 4;
  if (p1 == 604884)
    if (0 * p1)
      a = 7;
  return a;
}

If p1 is 604884, then 0 * p1 is false. Clang evaluates both conditions, GCC returns 4 unconditionally.

Missed simplification of bitwise operations

float fn1(short p1) {
  float a = 516.403544893f;
  if (p1 & 6 >> (p1 & 31) ^ ~0)
    a = 1;
  return a;
}

Clang evaluates the branch condition because it misses that ((p1 & (6 >> (p1 & 31))) ^ ~0) is nonzero because almost all the bits of the inner expression are zero.

CompCert

Reluctance to select mvn instruction

In this function:

int fn1(int p1) {
  return -7 - p1;
}

CompCert computes the subtraction by adding a sequence of large constants:

    rsb r0, r0, #249
    add r0, r0, #65280
    add r0, r0, #16711680
    add r0, r0, #-16777216

It's simpler to synthesize the constant -7 as the negation of 6 and subtracting, as GCC does:

    mvn r3, #6
    sub r0, r3, r0

Added 2017-09-06

Unless noted otherwise, the examples below can be reproduced with the following compiler versions and configurations:

Compiler Version Revision Date Configuration
GCC 8.0.0 20170906 r251801 2017-09-06 --target=armv7a-eabihf --with-arch=armv7-a --with-fpu=vfpv3-d16 --with-float-abi=hard --with-float=hard
Clang 6.0.0 (trunk 312637) LLVM r312635, Clang r312634 2017-09-06 --target=armv7a-eabihf -march=armv7-a -mfpu=vfpv3-d16 -mfloat-abi=hard
CompCert 3.0.1 f3c60be 2017-08-28 armv7a-linux

GCC

Useless initialization of struct passed by value

struct S0 {
  int f0;
  int f1;
  int f2;
  int f3;
};

int f1(struct S0 p) {
    return p.f0;
}

The struct is passed in registers, and the function's result is already in r0, which is also the return register. The function could return immediately, but GCC first stores all the struct fields to the stack and reloads the first field:

    sub sp, sp, #16
    add ip, sp, #16
    stmdb   ip, {r0, r1, r2, r3}
    ldr r0, [sp]
    add sp, sp, #16
    bx  lr

Reported at https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80905.

float to char type conversion goes through memory

char fn1(float p1) {
  return (char) p1;
}

results in:

    vcvt.u32.f32    s15, s0
    sub sp, sp, #8
    vstr.32 s15, [sp, #4]   @ int
    ldrb    r0, [sp, #4]    @ zero_extendqisi2

i.e., the result of the conversion in s15 is stored to the stack and then reloaded into r0 instead of just copying it between the registers (and possibly truncating to char's range). Same for double instead of float, but not for short instead of char.

Reported at https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80861.

If you are worried about the conversion overflowing, you might prefer this version:

#include <math.h>
#include <limits.h>

char fn1(float p1) {
  if (isfinite(p1) && CHAR_MIN <= p1 && p1 <= CHAR_MAX) {
    return (char) p1;
  }
  return 0;
}

GCC still goes through memory to perform the zero extension.

Spill instead of register copy

int N;
int fn2(int p1, int p2) {
  int a = p2;
  if (N)
    a *= 10.5;
  return a;
}

GCC spills a although enough integer registers would be available:

    str r1, [sp, #4]    // initialize a to p2
    cmp r3, #0
    beq .L13
    ... compute multiplication, put into d7 ...
    vcvt.s32.f64    s15, d7
    vstr.32 s15, [sp, #4]   @ int   // update a in branch
.L13:
    ldr r0, [sp, #4]    // reload a

Reported at https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81012 (together with another issue further below).

Missed simplification of multiplication by integer-valued floating-point constant

Variant of the above code with the constant changed slightly:

int N;
int fn5(int p1, int p2) {
  int a = p2;
  if (N)
    a *= 10.0;
  return a;
}

GCC converts a to double and back as above, but the result must be the same as simply multiplying by the integer 10. Clang realizes this and generates an integer multiply, removing all floating-point operations.

If you are worried about the multiplication overflowing, you might prefer this version:

#include <limits.h>

int N;
int fn5(int p1, int p2) {
  int a = p2;
  if (N && INT_MIN / 10.0 <= a && a <= INT_MAX / 10.0)
    a *= 10.0;
  return a;
}

Clang still generates code for multiplying integers, while GCC multiplies as double.

Dead store for int to double conversion

double fn3(int p1, short p2) {
  double a = p1;
  if (1881 * p2 + 5)
    a = 2;
  return a;
}

a is initialized to the value of p1, and GCC spills this value. The spill store is dead, the value is never reloaded:

    ...
    str r0, [sp, #4]
    cmn r1, #5
    vmovne.f64  d0, #2.0e+0
    vmoveq  s15, r0 @ int
    vcvteq.f64.s32  d0, s15
    add sp, sp, #8
    @ sp needed
    bx  lr

Reported at https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81012 (together with another issue above).

Missed strength reduction of division to comparison

unsigned int fn1(unsigned int a, unsigned int p2) {
  int b;
  b = 3;
  if (p2 / a)
    b = 7;
  return b;
}

The condition p2 / a is true (nonzero) iff p2 >= a. Clang compares, GCC generates a division (on ARM, x86, and PowerPC).

Missed simplification of floating-point addition

long fn1(int p1) {
  double a;
  long b = a = p1;
  if (0 + a + 0)
    b = 9;
  return b;
}

It is not correct in general to replace a + 0 by a in floating-point arithmetic due to NaNs and negative zeros and whatnot, but here a is initialized from an int's value, so none of these cases can happen. GCC generates all the floating-point operations, while Clang just compares p1 to zero.

Missed reassociation of multiplications

double fn1(double p1) {
  double v;
  v = p1 * p1 * (-p1 * p1);
  return v;
}

int fn2(int p1) {
  int v;
  v = p1 * p1 * (-p1 * p1);
  return v;
}

For each function, GCC generates three multiplications:

    vnmul.f64   d7, d0, d0
    vmul.f64    d0, d0, d0
    vmul.f64    d0, d7, d0

and

    rsb r2, r0, #0
    mul r3, r0, r0
    mul r0, r0, r2
    mul r0, r3, r0

Clang can do the same with two multiplications each:

    vmul.f64    d0, d0, d0
    vnmul.f64   d0, d0, d0

and

    mul r0, r0, r0
    rsb r1, r0, #0
    mul r0, r0, r1

Heroic loop optimization instead of removing the loop

int N;
double fn1(double *p1) {
  int i = 0;
  double v = 0;
  while (i < N) {
    v = p1[i];
    i++;
  }
  return v;
}

This function returns 0 if N is 0 or negative; otherwise, it returns p1[N-1]. On x86-64, GCC generates a complex unrolled loop (and a simpler loop on ARM). Clang removes the loop and replaces it by a simple branch. GCC's code is too long and boring to show here, but it's similar enough to https://godbolt.org/g/RYwgq4.

Note: An earlier version of this document claimed that GCC also removed the loop on ARM. It was pointed out to me that this claim was false, GCC does generate a loop on ARM too. My (stupid) mistake, I had misread the assembly code.

Unnecessary spilling due to badly scheduled move-immediates

double fn1(double p1, float p2, double p3, float p4, double *p5, double p6)
{
  double a, c, d, e;
  float b;
  a = p3 / p1 / 38 / 2.56380695236e38f * 5;
  c = 946293016.021f / a + 9;
  b = p1 - 6023397303.74f / p3 * 909 * *p5;
  d = c / 683596028.301 + p3 - b + 701.296936339 - p4 * p6;
  e = p2 + d;
  return e;
}

Sorry about the mess. GCC generates a bunch of code for this, including:

    vstr.64 d10, [sp]
    vmov.f64    d11, #5.0e+0
    vmov.f64    d13, #9.0e+0
    ... computations involving d10 but not d11 or d13 ...
    vldr.64 d10, [sp]
    vcvt.f64.f32    d6, s0
    vmul.f64    d11, d7, d11
    vdiv.f64    d5, d12, d11
    vadd.f64    d13, d5, d13
    vdiv.f64    d0, d13, d14

The register d10 must be spilled to make room for the constants 5 and 9 being loaded into d11 and d13, respectively. But these loads are much too early: Their values are only needed after d10 is restored. These move instructions (at least one of them) should be closer to their uses, freeing up registers and avoiding the spill.

Missed constant propagation (or so it seemed)

int fn1(int p1) {
  int a = 6;
  int b = (p1 / 12 == a);
  return b;
}
int fn2(int p1) {
  int b = (p1 / 12 == 6);
  return b;
}

These functions should be compiled identically. The condition is equivalent to 6 * 12 <= p1 < 7 * 12, and the two signed comparisons can be replaced by a subtraction and a single unsigned comparison:

    sub r0, r0, #72
    cmp r0, #11
    movhi   r0, #0
    movls   r0, #1
    bx  lr

GCC used to do this for the first function, but not the second one. It seemed like constant propagation failed to replace a by 6.

Reported at https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81346 and fixed in revision r250342; it turns out that this pattern used to be implemented (only) in a very early folding pass that ran before constant propagation.

Missed tail-call optimization for compiler-generated call to intrinsic

int fn1(int a, int b) {
  return a / b;
}

On ARM, integer division is turned into a function call:

    push    {r4, lr}
    bl  __aeabi_idiv
    pop {r4, pc}

In the code above, GCC allocates a stack frame, not realizing that this call in tail position could be simply

    b   __aeabi_idiv

(When the call, even to __aeabi_idiv, appears in the source code, GCC is perfectly capable of compiling it as a tail call.)

Clang

Incomplete range analysis for certain types

short fn1(short p) {
  short c = p / 32769;
  return c;
}

int fn2(signed char p) {
  int c = p / 129;
  return c;
}

In both cases, the divisor is outside the possible range of values for p, so the function's result is always 0. Clang doesn't realize this and generates code to compute the result. (It does optimize corresponding cases for unsigned short and char.)

A report with this and all the other range analysis examples below was filed by @gnzlbg at https://bugs.llvm.org/show_bug.cgi?id=34517, and most cases have been fixed.

More incomplete range analysis

int fn1(short p4, int p5) {
  int v = 0;
  if ((p4 - 32779) - !p5)
    v = 7;
  return v;
}

The condition is always true, and the function always returns 7. Clang generates code to evaluate the condition. Similar to the case above, but seems to be more complex: If the - !p5 is removed, Clang also realizes that the condition is always true.

Even more incomplete range analysis

int fn1(int p, int q) {
  int a = q + (p % 6) / 9;
  return a;
}

int fn2(int p) {
  return (1 / p) / 100;
}
int fn3(int p5, int p6) {
  int a, c, v;
  v = -!p5;
  a = v + 7;
  c = (4 % a);
  return c;
}
int fn4(int p4, int p5) {
  int a = 5;
  if ((5 % p5) / 9)
    a = p4;
  return a;
}

The first function always returns q, the others always return 0, 4, and 5, respectively. Clang evaluates at least one division or modulo operation in each case.

(I have a lot more examples like this, but this list is boring enough as it is.)

Redundant code generated for double to int conversion

int fn1(double c, int *p, int *q, int *r, int *s) {
  int i = (int) c;
  *p = i;
  *q = i;
  *r = i;
  *s = i;
  return i;
}

There is a single type conversion on the source level. Clang duplicates the conversion operation (vcvt) for each use of i:

    vcvt.s32.f64    s2, d0
    vstr    s2, [r0]
    vcvt.s32.f64    s2, d0
    vstr    s2, [r1]
    vcvt.s32.f64    s2, d0
    vstr    s2, [r2]
    vcvt.s32.f64    s2, d0
    vcvt.s32.f64    s0, d0
    vmov    r0, s0
    vstr    s2, [r3]

Several times, the result of the conversion is written into the register that already holds that same value (s2).

Reported at https://bugs.llvm.org/show_bug.cgi?id=33199, fixed.

Missing instruction selection pattern for vnmla

double fn1(double p1, double p2, double p3) {
  double a = -p1 - p2 * p3;
  return a;
}

Clang generates

    vneg.f64    d0, d0
    vmls.f64    d0, d1, d2

but this could just be

    vnmla.f64   d0, d1, d2

The instruction selector has a pattern for selecting vnmla for the expression -(a * b) - dst, but not for the symmetric -dst - (a * b).

Patch submitted: https://reviews.llvm.org/D35911

CompCert

No dead code elimination for useless branches

void fn1(int r0) {
  if (r0 + 42)
    0;
}

The conditional branch due to the if statement is removed, but the computation for its condition remains in the generated code:

    add r0, r0, #42

This appears to be because dead code elimination runs before useless branches are identified and eliminated.

Fixed in https://github.com/AbsInt/CompCert/commit/93d2fc9e3b23d69c0c97d229f9fa4ab36079f507 due to my report.

Missing constant folding patterns for modulo

int fn4(int p1) {
  int a = 0 % p1;
  return a;
}

int fn5(int a) {
  int b = a % a;
  return b; 
} 

In both cases, CompCert generates a modulo operation (on ARM, it generates a library call, a multiply, and a subtraction). In both cases, this operation could be constant folded to zero.

My patch at https://github.com/gergo-/CompCert/commit/78555eb7afacac184d94db43c9d4438d20502be8 fixes the first one, but not the second. It also fails to deal with the equivalent case of

  int zero = 0;
  int a = zero % p1;

because, by the time the constant propagator runs, the modulo operation has been mangled into a complex call/multiply/subtract sequence that is inconvenient to recognize.

Status: CompCert developers are (understandably) uninterested in this trivial issue.

Failure to generate movw instruction

int fn1(int p1) {
  int a = 506 * p1;
  return a;
}

results in:

    mov r1, #250
    orr r1, r1, #256
    mul r0, r0, r1

Two instructions are spent on loading the constant. However, ARM's movw instruction can take an immediate between 0 and 65535, so this can be simplified to:

    movw r1, #506

Fixed in a series of patches from me, merged at https://github.com/AbsInt/CompCert/commit/c4dcf7c08016f175ba6c06d20c530ebaaad67749.

Failure to constant fold mla

Call this code mla.c:

int fn1(int p1) {
  int a, b, c, d, e, v, f;
  a = 0;
  b = c = 0;
  d = e = p1;
  v = 4;
  f = e * d | a * p1 + b;
  return f;
}

The expression a * p1 + b evaluates to 0 * p1 + 0, i.e., 0. CompCert is able to fold both 0 * x and x + 0 in isolation, but on ARM it generates an mla (multiply-add) instruction, for which it has no constant folding rules:

    mov r4, #0
    mov r12, #0
    ...
    mla r2, r4, r0, r12

Fixed in a patch from me, merged at https://github.com/AbsInt/CompCert/commit/4f46e57884a909d2f62fa7cea58b3d933a6a5e58.

Unnecessary spilling due to dead code

On mla.c from above without the mla constant folding patch, CompCert spills the r4 register which it then uses to store a temporary:

    str r4, [sp, #8]
    mov r4, #0
    mov r12, #0
    mov r1, r0
    mov r2, r1
    mul r3, r2, r1
    mla r2, r4, r0, r12
    orr r0, r3, r2
    ldr r4, [sp, #8]

However, the spill goes away if the line v = 4; is removed from the program:

    mov r3, #0
    mov r12, #0
    mov r1, r0
    mov r2, r1
    mul r1, r1, r2
    mla r12, r3, r0, r12
    orr r0, r1, r12

The value of v is never used. This assignment is dead code, and it is compiled away. It should not affect register allocation.

Missed copy propagation

In the previous example, the copies for the arguments to the mul operation (r0 to r1, then on to r2) are redundant. They could be removed, and the multiplication written as just mul r1, r0, r0.

Note: This entry previously spoke of copy coalescing, but Sebastian Hack pointed out that it's actually copy propagation that is missed here.

Failure to propagate folded constants

Continuing the mla.c example further, but this time after applying the mla constant folding patch, CompCert generates:

    mov r1, r0
    mul r2, r1, r0
    mov r1, #0
    orr r0, r2, r1

The x | 0 operation is redundant. CompCert is able to fold this operation if it appears in the source code, but in this case the 0 comes from the previously folded mla operation. Such constants are not propagated by the "constant propagation" (in reality, only local folding) pass. Values are only propagated by the value analysis that runs before, but for technical reasons the value analysis cannot currently take advantage of the algebraic properties of operators' neutral and zero elements.

Missed reassociation and folding of constants

int fn1(int p1) {
  int a, v, b;
  v = 1;
  a = 3 * p1 * 2 * 3;
  b = v + a * 83;
  return b;
}

All the multiplications could be folded into a single multiplication of p1 by 1494, but CompCert generates a series of adds and shifts before a multiplication by 83:

    add r3, r0, r0, lsl #1
    mov r0, r3, lsl #1
    add r1, r0, r0, lsl #1
    mov r2, #83
    mla r0, r2, r1, r12

About

Missed optimizations in C compilers

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published