feat: partial SSE scalar op support

[pnwamk]

No description provided.

[pnwamk]

C code used to sanity check SSE-related assembly call in LLVM:

#include <stdio.h>

// clang -S -emit-llvm sse_ops.c

int main()
{
  float x1 = 3.0;
  float y1 = 0.14159;
  float res1;

  asm ( "addss %2,%1" : "=x,x" (res1) : "x,x" (x1), "x,m" (y1) );
  printf("addss: x = %f, y = %f, res = %f\n", x1, y1, res1);
  asm ( "subss %2,%1" : "=x,x" (res1) : "x,x" (x1), "x,m" (y1) );
  printf("subss: x = %f, y = %f, res = %f\n", x1, y1, res1);
  asm ( "mulss %2,%1" : "=x,x" (res1) : "x,x" (x1), "x,m" (y1) );
  printf("mulss: x = %f, y = %f, res = %f\n", x1, y1, res1);
  asm ( "divss %2,%1" : "=x,x" (res1) : "x,x" (x1), "x,m" (y1) );
  printf("divss: x = %f, y = %f, res = %f\n", x1, y1, res1);
  asm ( "minss %2,%1" : "=x,x" (res1) : "x,x" (x1), "x,m" (y1) );
  printf("minss: x = %f, y = %f, res = %f\n", x1, y1, res1);
  asm ( "maxss %2,%1" : "=x,x" (res1) : "x,x" (x1), "x,m" (y1) );
  printf("maxss: x = %f, y = %f, res = %f\n", x1, y1, res1);

  double x2 = 3.0;
  double y2 = 0.14159;
  double res2;

  asm ( "addsd %2,%1" : "=x,x" (res2) : "x,x" (x2), "x,m" (y2) );
  printf("addsd: x = %f, y = %f, res = %f\n", x2, y2, res2);
  asm ( "subsd %2,%1" : "=x,x" (res2) : "x,x" (x2), "x,m" (y2) );
  printf("subsd: x = %f, y = %f, res = %f\n", x2, y2, res2);
  asm ( "mulsd %2,%1" : "=x,x" (res2) : "x,x" (x2), "x,m" (y2) );
  printf("mulsd: x = %f, y = %f, res = %f\n", x2, y2, res2);
  asm ( "divsd %2,%1" : "=x,x" (res2) : "x,x" (x2), "x,m" (y2) );
  printf("divsd: x = %f, y = %f, res = %f\n", x2, y2, res2);
  asm ( "minsd %2,%1" : "=x,x" (res2) : "x,x" (x2), "x,m" (y2) );
  printf("minsd: x = %f, y = %f, res = %f\n", x2, y2, res2);
  asm ( "maxsd %2,%1" : "=x,x" (res2) : "x,x" (x2), "x,m" (y2) );
  printf("maxsd: x = %f, y = %f, res = %f\n", x2, y2, res2);

  return 0;
}

[pnwamk]

These changes make a small increase in the number of functions successfully recovered:

diff -U1 old-reopt-explore-stats.txt new-reopt-explore-stats.txt 
--- old-stats.txt	2021-07-30 13:13:17.000000000 -0700
+++ new-stats.txt	2021-07-30 13:13:33.000000000 -0700
@@ -28,6 +28,6 @@
 Recovery:
-  Succeeded: 21,819 (72%)
-  Failed: 8,402 (28%)
-    Unsupported function value: 2,336 (8%)
-    Unimplemented LLVM backend feature: 4,990 (17%)
+  Succeeded: 21,842 (72%)
+  Failed: 8,379 (28%)
+    Unsupported function value: 2,396 (8%)
+    Unimplemented LLVM backend feature: 4,907 (16%)
     Stack offset escape: 83 (0%)

Click to see full stats before and after.

Full reopt-explore stats w/o PR changes:

reopt analyzed 394 binaries:
Generated LLVM bitcode for 394 out of 394 binaries.
Initialization:
  Code segment: 42,933,178 bytes
  Initial entry points: 79776
  Warnings: 0
Discovery:
  Bytes discovered: 23,025,164 (54%)
  Succeeded: 64,494 (81%)
  Failed: 15,500 (19%)
    Unhandled instruction: 425 (1%)
    Unidentified control flow: 15,075 (19%)
Argument Analysis:
  Succeeded: 40,429 (63%)
  Failed: 24,065 (37%)
  Header Warnings: 0
  DWARF Warnings: 0
  Code Warnings: 38,681
Invariant Inference:
  Succeeded: 30,221 (75%)
  Failed: 10,208 (25%)
    Symbolic call stack height: 1 (0%)
    Unresolved stack read: 13 (0%)
    Indirect call target: 526 (1%)
    Call target not function entry point: 41 (0%)
    Unresolved call target arguments: 9,614 (24%)
    Could not resolve varargs args: 13 (0%)
Recovery:
  Succeeded: 21,819 (72%)
  Failed: 8,402 (28%)
    Unsupported function value: 2,336 (8%)
    Unimplemented LLVM backend feature: 4,990 (17%)
    Stack offset escape: 83 (0%)
    Stack read overlapping offset: 1 (0%)
    Unresolved return value: 8 (0%)
    Missing variable value: 984 (3%)

Full reopt-explore stats w/ PR changes:

reopt analyzed 394 binaries:
Generated LLVM bitcode for 394 out of 394 binaries.
Initialization:
  Code segment: 42,933,178 bytes
  Initial entry points: 79776
  Warnings: 0
Discovery:
  Bytes discovered: 23,025,164 (54%)
  Succeeded: 64,494 (81%)
  Failed: 15,500 (19%)
    Unhandled instruction: 425 (1%)
    Unidentified control flow: 15,075 (19%)
Argument Analysis:
  Succeeded: 40,429 (63%)
  Failed: 24,065 (37%)
  Header Warnings: 0
  DWARF Warnings: 0
  Code Warnings: 38,681
Invariant Inference:
  Succeeded: 30,221 (75%)
  Failed: 10,208 (25%)
    Symbolic call stack height: 1 (0%)
    Unresolved stack read: 13 (0%)
    Indirect call target: 526 (1%)
    Call target not function entry point: 41 (0%)
    Unresolved call target arguments: 9,614 (24%)
    Could not resolve varargs args: 13 (0%)
Recovery:
  Succeeded: 21,842 (72%)
  Failed: 8,379 (28%)
    Unsupported function value: 2,396 (8%)
    Unimplemented LLVM backend feature: 4,907 (16%)
    Stack offset escape: 83 (0%)
    Stack read overlapping offset: 1 (0%)
    Unresolved return value: 8 (0%)
    Missing variable value: 984 (3%)

[pnwamk]

C code for extracting LLVM for ucomisx instructions:

#include <stdio.h>

// clang -S -emit-llvm ucomsx.c

int main()
{
  float x1 = 2.0;
  float y1 = 1.0;
  short res1;
  short cf = 0; // bit 0
  short pf = 0; // bit 2
  short zf = 0; // bit 6

  asm ( "ucomiss %1,%2\n\t"
        "pushfw\n\t"
        "popw %0\n\t"
        : "=r,r" (res1)
        : "x,x" (x1), "x,M" (y1) );
  // cf = (res1 & 0b1) ? 1 : 0;
  // pf = (res1 & 0b100) ? 1 : 0;
  // zf = (res1 & 0b1000000) ? 1 : 0;

  // printf("ucomiss: x = %f, y = %f, zf,pf,cf =%i,%i,%i\n",
  //         x1, y1, zf, pf, cf);

  double x2 = 3.0;
  double y2 = 1.0;
  short res2;

  asm ( "ucomisd %1,%2\n\t"
        "pushfw\n\t"
        "popw %0\n\t"
        : "=r,r" (res2)
        : "x,x" (x2), "x,N" (y2) );
  // cf = (res2 & 0b1) ? 1 : 0;
  // pf = (res2 & 0b100) ? 1 : 0;
  // zf = (res2 & 0b1000000) ? 1 : 0;

  // printf("ucomiss: x = %f, y = %f, zf,pf,cf =%i,%i,%i\n",
  //         x1, y1, zf, pf, cf);

  return 0;
}

[pnwamk]

Metric changes with latest improvements:

$ diff -U1 pre_additions.txt post_additions.txt 
--- pre_additions.txt	2021-08-09 12:18:49.000000000 -0700
+++ post_additions.txt	2021-08-09 12:18:18.000000000 -0700
@@ -28,6 +28,6 @@
 Recovery:
-  Succeeded: 21,819 (72%)
-  Failed: 8,402 (28%)
-    Unsupported function value: 2,336 (8%)
-    Unimplemented LLVM backend feature: 4,990 (17%)
+  Succeeded: 21,868 (72%)
+  Failed: 8,353 (28%)
+    Unsupported function value: 2,396 (8%)
+    Unimplemented LLVM backend feature: 4,881 (16%)
     Stack offset escape: 83 (0%)