363c14ae0cd2baa624d85e8c9db12cd1677190ea is the first bad commit
commit 363c14ae0cd2baa624d85e8c9db12cd1677190ea
Author: Olivier Galibert <galibert@pobox.com>
Date:   Wed May 2 23:11:42 2012 +0200

    glsl: Change built-in constant expression evaluation to run the IR.
    
    This removes code duplication with
    ir_expression::constant_expression_value and builtins/ir/*.
    
    Signed-off-by: Olivier Galibert <galibert@pobox.com>
    Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

:040000 040000 b5544fbb108c323c6f87b67cf2ed88b52c80acd3 554bf308d67557ed0ac39762d1b4e9f59b28e032 M	src
bisect run success

Yeah, this is due to a slight difference in precision.  We concluded that either:

1. Our built-ins need to become more precise, or
2. The test needs to be made less stringent.

I'd prefer the former, and it's probably not actually very hard to do...

Following cases fail on Ironlake, Sandybridge and Ivybridge with mesa master branch, and have same bisect commit.
Piglit 
shaders_glsl-const-folding-01
Oglc
glsl-constexpr(positive.asin)
glsl-constexpr(positive.acos)
glsl-constexpr(positive.atan)
glsl-constexpr(positive.atan_y_over_x)
glsl-autointconv(function.builtIn.trigonometry)

Oglc cases 'glsl-constexpr(positive.tanh)' and 'glsl-bif-trig(advanced.tanh.vec2)' fail on Sandybridge and Ivybridge with mesa master branch, and have same bisect commit.

Piglit patch for atan(1):
  http://lists.freedesktop.org/archives/piglit/2012-May/002342.html

Mesa patch for acos(1):
  http://lists.freedesktop.org/archives/mesa-dev/2012-May/021586.html

The test passes once those two are in.

These were sent before the patch series second version.

  OG.

The glsl-const-builtin-normalize piglit failure with r300g driver is also caused by the aforementioned commit, it looks like some precision problem too. Should I open a new bug for that?

./bin/shader_runner tests/shaders/glsl-const-builtin-normalize.shader_test -auto -fbo
r300: DRM version: 2.15.0, Name: ATI RV530, ID: 0x71c5, GB: 1, Z: 2
r300: GART size: 509 MB, VRAM size: 256 MB
r300: AA compression RAM: YES, Z compression RAM: YES, HiZ RAM: YES
Probe at (15,15)
  Expected: 0.707107 0.000000 0.000000
  Observed: 0.705882 1.000000 1.000000
PIGLIT: {'result': 'fail' }

(In reply to comment #5)
> The glsl-const-builtin-normalize piglit failure with r300g driver is also
> caused by the aforementioned commit, it looks like some precision problem too.
> Should I open a new bug for that?
> 
> ./bin/shader_runner tests/shaders/glsl-const-builtin-normalize.shader_test
> -auto -fbo
> r300: DRM version: 2.15.0, Name: ATI RV530, ID: 0x71c5, GB: 1, Z: 2
> r300: GART size: 509 MB, VRAM size: 256 MB
> r300: AA compression RAM: YES, Z compression RAM: YES, HiZ RAM: YES
> Probe at (15,15)
>   Expected: 0.707107 0.000000 0.000000
>   Observed: 0.705882 1.000000 1.000000
> PIGLIT: {'result': 'fail' }

Well, the constant folding gives:
      (assign  (xyzw) (var_ref gl_FragColor)  (constant vec4 (0.707107 -nan -nan 1.000000)) ) 

The initial 0.707107 is sqrt(2), i.e. the result of normalizing (1, 1).  Then it gets multiplied by 255, truncated, then re-divided by 255, giving 0.705882.  So that slot is good.

The other two are the result of normalizing (0, 0), which is undefined (the test itself says so), because it does a pair of 0/0.  I'm not convinced pessimizing normalize just to return (0, 0) in that case is worth it.

It also happens on mesa 9.0 branch(commit:6886da783ac2fc549b4ffc1f4)

mesa: 1a316af0343b1c1b345d6209a687ce858b47c438 (master)

glsl-const-folding-01 regression is still present.

Created attachment 78390 [details] [review]
glsl: Improve the accuracy of the asin() builtin function.

I extended the Series by two degrees to 

asin(x) = sign(x) * asin_pos(abs(x))

asin_pos(x) = pi/2 - sqrt(1 - x) * T5(x)

T5(x) = (a0 + x * (a1 + x * (a2 + x * (a3 + x * (a4 + x * a5)))))

a0 = pi/2
a1 = (pi/4 - 1)
a2, a3, a4 and a5 found by numerical fitting.

This approximation has a worst case absolute error <4e-6. Unfortunately, due to round off errors of single precision floats the relative error is of the order of 100% near 0 (for |x|<1e-3). Otherwise the relative error is <2e-5.

The current implementation has similar problems with round of errors near 0. A possible way to fix this would be a piecewise linear approximation near 0.

mesa: fc25956badb8e1932cc19d8c97b4be16e92dfc65 (master 10.2.0-devel)

glsl-const-folding-01 regression is still present.

Mesa 10.3.0-devel (git-5b8f1a0), r600g

Failed to link:
error: unresolved reference to function `bad_constant_folding'

mesa: d13d2fd16132f351ec7c8184f165faeac3b31bb4 (master 10.4.0-devel)

Regression is still present.

Should be fixed by 326e303175ab700d59602f0e344e8d84e15f1aa8