user/glm: test failure with GCC 13

awilcox on gwyn ~/Code/awilfox/gcc-next/user/glm/src/glm/build % ctest --rerun-failed --output-on-failure
Test project /home/awilcox/Code/awilfox/gcc-next/user/glm/src/glm/build
    Start  24: test-core_type_mat4x4
1/3 Test  #24: test-core_type_mat4x4 ............***Failed    0.00 sec

    Start  32: test-core_func_integer
2/3 Test  #32: test-core_func_integer ...........***Failed    0.00 sec
bitCount - TimeIf 18
bitCount - TimeVec 29
bitCount - TimeDefault 5
bitCount - TimeVec4 132
bitCount - bitfield 5
glm::bitfieldReverse: 3 clocks
bitfieldReverseLoop: 18 clocks
bitfieldReverseUint32: 4 clocks
bitfieldReverseOps: 4 clocks
glm::bitfieldReverse - 64: 4 clocks
bitfieldReverseLoop - 64: 57 clocks
bitfieldReverseUint - 64: 3 clocks
bitfieldReverseOps - 64: 4 clocks
glm::findMSB: 1 clocks
findMSB - nlz1: 1 clocks
findMSB - nlz2: 0 clocks
findMSB - 0.9.5: 1 clocks
findMSB - pop: 1 clocks
glm::findLSB: 1 clocks
findLSB - 0.9.5: 1 clocks
findLSB - ntz2: 1 clocks
findLSB - branchfree: 1 clocks

    Start 100: test-gtx_euler_angle
3/3 Test #100: test-gtx_euler_angle .............***Failed    0.00 sec
mat3x3((0.050484, -0.613396, -0.788160), (0.658332, 0.613885, -0.435597), (0.751033, -0.496880, 0.434809))
mat3x3((0.000000, 0.000000, 0.000000), (0.000000, 0.000000, 0.000000), (0.000000, 0.000000, 0.000000))
mat3x3((0.658372, -0.613396, 0.436224), (0.049959, 0.613885, 0.787813), (-0.751033, -0.496880, 0.434809))

RESIDUAL
mat3x3((0.050484, -0.613396, -0.788160), (0.658332, 0.613885, -0.435597), (0.751033, -0.496880, 0.434809))
mat3x3((-0.607888, 0.000000, -1.224384), (0.608372, 0.000000, -1.223410), (1.502067, 0.000000, 0.000000))


0% tests passed, 3 tests failed out of 3

Total Test time (real) =   0.02 sec

The following tests FAILED:
	 24 - test-core_type_mat4x4 (Failed)
	 32 - test-core_func_integer (Failed)
	100 - test-gtx_euler_angle (Failed)
Errors while running CTest

A few notes about this package:

• We are on 0.9.9.5 because shortly after the release of 0.9.9.5, before 0.9.9.6, they removed the .pc file generation which breaks builds of some of the dependents of glm. Reported upstream to deaf ears.
• The tests do not explain what is going on. I have no idea what is failing in core_type_mat4x4 for instance.
• One of the test failures, core_func_integer, was reported upstream by Debian and Gentoo. One fix is -fno-ipa-modref; the proper fix includes a bunch of other changes that make it impossible to backport easily.

We will need to figure out what to do. For now I am disabling tests during the Big GCC 13 Trybuild, so that we can keep things moving.

added GCC 13 TEST labels

mentioned in commit 776db488

mentioned in commit 36478002

mentioned in commit 757fd21e

mentioned in commit f975117c

mentioned in commit 23d432a3

I found this on pmmx today (all other tests seem to pass):

31/45 Testing: test-core_func_integer
31/45 Test: test-core_func_integer
Command: "/usr/src/packages/user/glm/src/glm/build/test/core/test-core_func_integer"
Directory: /usr/src/packages/user/glm/src/glm/build/test/core
"test-core_func_integer" start time: Jun 10 21:47 UTC
Output:
----------------------------------------------------------
bitCount - TimeIf 34
bitCount - TimeVec 25
bitCount - TimeDefault 3
bitCount - TimeVec4 19
bitCount - bitfield 4
glm::bitfieldReverse: 4 clocks
bitfieldReverseLoop: 45 clocks
bitfieldReverseUint32: 3 clocks
bitfieldReverseOps: 4 clocks
glm::bitfieldReverse - 64: 15 clocks
bitfieldReverseLoop - 64: 306 clocks
bitfieldReverseUint - 64: 15 clocks
bitfieldReverseOps - 64: 15 clocks
glm::findMSB: 0 clocks
findMSB - nlz1: 0 clocks
findMSB - nlz2: 1 clocks
findMSB - 0.9.5: 0 clocks
findMSB - pop: 1 clocks
glm::findLSB: 1 clocks
findLSB - 0.9.5: 0 clocks
findLSB - ntz2: 1 clocks
findLSB - branchfree: 0 clocks
<end of output>
Test time =   0.00 sec
----------------------------------------------------------
Test Failed.
"test-core_func_integer" end time: Jun 10 21:47 UTC
"test-core_func_integer" time elapsed: 00:00:00
----------------------------------------------------------

Yes, that is what will be fixed with -fno-ipa-modref. The other two seem to be Power-specific.

mentioned in commit 9d5552d1

mentioned in commit 892f9e73

mentioned in commit 6cafcf6a

The Power specific bugs bisect to:

[gentoo64] awilcox@gwyn ~/gcc $ git bisect good
985a8962712a56d1c9418cf05e1761f8bd20cf8f is the first bad commit

gcc@985a8962712a is in the tree-sra optimisation system. However, -fno-tree-sra does not fix any test failure (and actually regresses core_func_matrix as well).

-fno-stack-protector is an odd workaround that causes all tests to pass including on our present GCC 13 build. -mcpu=power9 also works around the issue and causes all tests to pass, but that is because it uses newer ISA instructions that are not available on all processors we target.

Diff in generated assembler between the bisected commit and the immediately preceding commit is attached in glm-gcc12-bisection.diff

added UPSTREAM label

Filed as GCC PR tree-optimization/115977.

mentioned in commit 8ca157a8