test_sse1 fails with -mnontrapping-fptoint

[sbc100]

$ EMCC_CFLAGS=-mnontrapping-fptoint ./test/runner core0.test_sse1 2> output
Test suites:
['test_core']
Running test_core: (1 tests)
(checking sanity from test runner)
...
--- expected
+++ actual
@@ -15701,17 +15701,17 @@
 _mm_cvtt_ss2si([0.800000012,1,1.5,2.5]) = 0x00000002
 _mm_cvtt_ss2si([1,1.5,2.5,3.5]) = 0x00000003
 _mm_cvtt_ss2si([1.5,2.5,3.5,3.5999999]) = 0x00000003
-_mm_cvtt_ss2si([2.5,3.5,3.5999999,3.40282347e+38]) = 0x80000000
-_mm_cvtt_ss2si([3.5,3.5999999,3.40282347e+38,inf]) = 0x80000000
-_mm_cvtt_ss2si([3.5999999,3.40282347e+38,inf,NaN(0x7FC00000)]) = 0x80000000
+_mm_cvtt_ss2si([2.5,3.5,3.5999999,3.40282347e+38]) = 0x7FFFFFFF
+_mm_cvtt_ss2si([3.5,3.5999999,3.40282347e+38,inf]) = 0x7FFFFFFF
+_mm_cvtt_ss2si([3.5999999,3.40282347e+38,inf,NaN(0x7FC00000)]) = 0x00000000
 _mm_cvtt_ss2si([3.40282347e+38,inf,NaN(0x7FC00000),2.38793926e-38]) = 0x00000000
 _mm_cvtt_ss2si([inf,NaN(0x7FC00000),2.38793926e-38,-0]) = 0x00000000
-_mm_cvtt_ss2si([NaN(0x7FC00000),2.38793926e-38,-0,NaN(0x7FFFFFFF)]) = 0x80000000
-_mm_cvtt_ss2si([2.38793926e-38,-0,NaN(0x7FFFFFFF),NaN(0xFFFFFFFF)]) = 0x80000000
+_mm_cvtt_ss2si([NaN(0x7FC00000),2.38793926e-38,-0,NaN(0x7FFFFFFF)]) = 0x00000000
+_mm_cvtt_ss2si([2.38793926e-38,-0,NaN(0x7FFFFFFF),NaN(0xFFFFFFFF)]) = 0x00000000
 _mm_cvtt_ss2si([-0,NaN(0x7FFFFFFF),NaN(0xFFFFFFFF),-2.70497005e+38]) = 0x80000000
 _mm_cvtt_ss2si([NaN(0x7FFFFFFF),NaN(0xFFFFFFFF),-2.70497005e+38,-3.29950007e-21]) = 0x00000000
-_mm_cvtt_ss2si([NaN(0xFFFFFFFF),-2.70497005e+38,-3.29950007e-21,3.40282002e+38]) = 0x80000000
-_mm_cvtt_ss2si([-2.70497005e+38,-3.29950007e-21,3.40282002e+38,3.38211008e+19]) = 0x80000000
+_mm_cvtt_ss2si([NaN(0xFFFFFFFF),-2.70497005e+38,-3.29950007e-21,3.40282002e+38]) = 0x7FFFFFFF
+_mm_cvtt_ss2si([-2.70497005e+38,-3.29950007e-21,3.40282002e+38,3.38211008e+19]) = 0x7FFFFFFF
 _mm_cvtt_ss2si([-3.29950007e-21,3.40282002e+38,3.38211008e+19,-inf]) = 0x80000000
 _mm_cvtt_ss2si([3.40282002e+38,3.38211008e+19,-inf,-3.40282347e+38]) = 0x80000000
 _mm_cvtt_ss2si([3.38211008e+19,-inf,-3.40282347e+38,-2.5]) = 0xFFFFFFFE
@@ -15733,17 +15733,17 @@
 _mm_cvttss_si32([0.800000012,1,1.5,2.5]) = 0x00000002
 _mm_cvttss_si32([1,1.5,2.5,3.5]) = 0x00000003
 _mm_cvttss_si32([1.5,2.5,3.5,3.5999999]) = 0x00000003
-_mm_cvttss_si32([2.5,3.5,3.5999999,3.40282347e+38]) = 0x80000000
-_mm_cvttss_si32([3.5,3.5999999,3.40282347e+38,inf]) = 0x80000000
-_mm_cvttss_si32([3.5999999,3.40282347e+38,inf,NaN(0x7FC00000)]) = 0x80000000
+_mm_cvttss_si32([2.5,3.5,3.5999999,3.40282347e+38]) = 0x7FFFFFFF
+_mm_cvttss_si32([3.5,3.5999999,3.40282347e+38,inf]) = 0x7FFFFFFF
+_mm_cvttss_si32([3.5999999,3.40282347e+38,inf,NaN(0x7FC00000)]) = 0x00000000
 _mm_cvttss_si32([3.40282347e+38,inf,NaN(0x7FC00000),2.38793926e-38]) = 0x00000000
 _mm_cvttss_si32([inf,NaN(0x7FC00000),2.38793926e-38,-0]) = 0x00000000
-_mm_cvttss_si32([NaN(0x7FC00000),2.38793926e-38,-0,NaN(0x7FFFFFFF)]) = 0x80000000
-_mm_cvttss_si32([2.38793926e-38,-0,NaN(0x7FFFFFFF),NaN(0xFFFFFFFF)]) = 0x80000000
+_mm_cvttss_si32([NaN(0x7FC00000),2.38793926e-38,-0,NaN(0x7FFFFFFF)]) = 0x00000000
+_mm_cvttss_si32([2.38793926e-38,-0,NaN(0x7FFFFFFF),NaN(0xFFFFFFFF)]) = 0x00000000
 _mm_cvttss_si32([-0,NaN(0x7FFFFFFF),NaN(0xFFFFFFFF),-2.70497005e+38]) = 0x80000000
 _mm_cvttss_si32([NaN(0x7FFFFFFF),NaN(0xFFFFFFFF),-2.70497005e+38,-3.29950007e-21]) = 0x00000000
-_mm_cvttss_si32([NaN(0xFFFFFFFF),-2.70497005e+38,-3.29950007e-21,3.40282002e+38]) = 0x80000000
-_mm_cvttss_si32([-2.70497005e+38,-3.29950007e-21,3.40282002e+38,3.38211008e+19]) = 0x80000000
+_mm_cvttss_si32([NaN(0xFFFFFFFF),-2.70497005e+38,-3.29950007e-21,3.40282002e+38]) = 0x7FFFFFFF
+_mm_cvttss_si32([-2.70497005e+38,-3.29950007e-21,3.40282002e+38,3.38211008e+19]) = 0x7FFFFFFF
 _mm_cvttss_si32([-3.29950007e-21,3.40282002e+38,3.38211008e+19,-inf]) = 0x80000000
 _mm_cvttss_si32([3.40282002e+38,3.38211008e+19,-inf,-3.40282347e+38]) = 0x80000000
 _mm_cvttss_si32([3.38211008e+19,-inf,-3.40282347e+38,-2.5]) = 0xFFFFFFFE
@@ -15765,7 +15765,7 @@
 _mm_cvttss_si64([0.800000012,1,1.5,2.5]) = 0x0000000000000002
 _mm_cvttss_si64([1,1.5,2.5,3.5]) = 0x0000000000000003
 _mm_cvttss_si64([1.5,2.5,3.5,3.5999999]) = 0x0000000000000003
-_mm_cvttss_si64([2.5,3.5,3.5999999,3.40282347e+38]) = 0x8000000000000000
+_mm_cvttss_si64([2.5,3.5,3.5999999,3.40282347e+38]) = 0x7FFFFFFFFFFFFFFF
 _mm_cvttss_si64([3.5,3.5999999,3.40282347e+38,inf]) = 0x8000000000000000
 _mm_cvttss_si64([3.5999999,3.40282347e+38,inf,NaN(0x7FC00000)]) = 0x8000000000000000
 _mm_cvttss_si64([3.40282347e+38,inf,NaN(0x7FC00000),2.38793926e-38]) = 0x0000000000000000
@@ -15774,8 +15774,8 @@
 _mm_cvttss_si64([2.38793926e-38,-0,NaN(0x7FFFFFFF),NaN(0xFFFFFFFF)]) = 0x8000000000000000
 _mm_cvttss_si64([-0,NaN(0x7FFFFFFF),NaN(0xFFFFFFFF),-2.70497005e+38]) = 0x8000000000000000
 _mm_cvttss_si64([NaN(0x7FFFFFFF),NaN(0xFFFFFFFF),-2.70497005e+38,-3.29950007e-21]) = 0x0000000000000000
-_mm_cvttss_si64([NaN(0xFFFFFFFF),-2.70497005e+38,-3.29950007e-21,3.40282002e+38]) = 0x8000000000000000
-_mm_cvttss_si64([-2.70497005e+38,-3.29950007e-21,3.40282002e+38,3.38211008e+19]) = 0x8000000000000000
+_mm_cvttss_si64([NaN(0xFFFFFFFF),-2.70497005e+38,-3.29950007e-21,3.40282002e+38]) = 0x7FFFFFFFFFFFFFFF
+_mm_cvttss_si64([-2.70497005e+38,-3.29950007e-21,3.40282002e+38,3.38211008e+19]) = 0x7FFFFFFFFFFFFFFF
 _mm_cvttss_si64([-3.29950007e-21,3.40282002e+38,3.38211008e+19,-inf]) = 0x8000000000000000
 _mm_cvttss_si64([3.40282002e+38,3.38211008e+19,-inf,-3.40282347e+38]) = 0x8000000000000000
 _mm_cvttss_si64([3.38211008e+19,-inf,-3.40282347e+38,-2.5]) = 0xFFFFFFFFFFFFFFFE



----------------------------------------------------------------------
Ran 1 test in 2.866s

FAILED (failures=1)

[sbc100]

@dschuff

[sbc100]

The differences all seem to come from the convert test. I narrowed it down to just:

#include <xmmintrin.h>                                                           
#include "test_sse.h"                                                            
                                                                                 
bool testNaNBits = false;                                                        
                                                                                 
float* interesting_floats;                                                       
int numInterestingFloats;                                                        
                                                                                 
uint32_t* interesting_ints;                                                      
int numInterestingInts;                                                          
                                                                                 
void convert(void) {                                                             
  // SSE1 Convert instructions:                                                  
  Ret_M128(int, _mm_cvtt_ss2si);                                                 
  Ret_M128(int, _mm_cvttss_si32);                                                
  Ret_M128(int64_t, _mm_cvttss_si64);                                            
  return;                                                                        
}                                                                                
                                                                                 
int main() {                                                                     
  interesting_floats = get_interesting_floats();                                 
  numInterestingFloats =                                                         
    sizeof(interesting_floats_) / sizeof(interesting_floats_[0]);                
  assert(numInterestingFloats % 4 == 0);                                         
                                                                                 
  interesting_ints = get_interesting_ints();                                     
  numInterestingInts = sizeof(interesting_ints_) / sizeof(interesting_ints_[0]); 
  assert(numInterestingInts % 4 == 0);                                           
                                                                                 
  convert();                                                                     
}          

[sbc100]

So it looks like the function which is returning different results is:

static __inline__ int __attribute__((__always_inline__, __nodebug__, DIAGNOSE_SLOW)) _mm_cvttss_si32(__m128 __a)
{                                                                                
  int x = lrint(((__f32x4)__a)[0]);                                              
  if (x != 0 || fabsf(((__f32x4)__a)[0]) < 2.f)                                  
    return (int)((__f32x4)__a)[0];                                               
  else                                                                           
    return (int)0x80000000;                                                      
}   

Here you see a bunch of different results where we expect 0x80000000:

--- expected
+++ actual
@@ -9,9 +9,9 @@
 [4] _mm_cvtt_ss2si([1,1.5,2.5,3.5]) =
   0x00000003
 [5] _mm_cvtt_ss2si([3.5999999,3.40282347e+38,inf,NaN]) =
-  0x80000000
+  0x00000000
 [6] _mm_cvtt_ss2si([2.38793926e-38,-0,NaN,NaN]) =
-  0x80000000
+  0x00000000
 [7] _mm_cvtt_ss2si([-2.70497005e+38,-3.29950007e-21,3.40282002e+38,3.38211008e+19]) =
-  0x80000000
+  0x7FFFFFFF

[sbc100]

It looks like maybe sign related.

I've reduced it to the following program:

#include <stdio.h>

float f = 33821100815692070912.0f;

int main() {
  printf("%f %d\n", f, (int)f);
  return 0;
}

$ clang ./test.c && ./a.out
33821100815692070912.000000 -2147483648
$ ./emcc ./test.c && node ./a.out.js 
33821100815692070912.000000 -2147483648
$ ./emcc -mnontrapping-fptoint ./test.c && node ./a.out.js 
33821100815692070912.000000 2147483647