[SLP] Polyhedron/fatigue2 16% slowdown after 88e7b8b

[vzakhari]

Reverting 88e7b8b commit brings performance back.

It is reproducible on AMD EPYC 9684X: time changes from 36.8 seconds to 42.7 seconds.

The following lines of code slowed down by more than twice:

 950       deviatoric_stress_tensor(:,:) = stress_tensor(:,:)
 951       deviatoric_stress_tensor(1,1) = stress_tensor(1,1) - hydrostatic_stress
 952       deviatoric_stress_tensor(2,2) = stress_tensor(2,2) - hydrostatic_stress
 953       deviatoric_stress_tensor(3,3) = stress_tensor(3,3) - hydrostatic_stress
 954 !
 955       damaged_dev_stress_tensor(:,:) = deviatoric_stress_tensor(:,:)/(1.0_LONGreal -        &
 956                                                                     crack_parameter * damage)
 957 !
 958 !        compute the yield function
 959 !
 960       equivalent_stress = sqrt(1.5_LONGreal* sum((damaged_dev_stress_tensor(:,:)            &
 961                                                               - back_stress_tensor(:,:))**2))

default.log.gz - before/after SLP with the current Flang compiler

revert.log.gz - before/after SLP with the reverted commit

Reproducer command: opt -mcpu=znver4 -print-after=slp-vectorizer --passes=slp-vectorizer

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Slava Zakharin (vzakhari)

Reverting https://github.com/llvm/llvm-project/commit/88e7b8b81c061113399637f936937ffaf5a9bc08 commit brings performance back.

It is reproducible on AMD EPYC 9684X: time changes from 36.8 seconds to 42.7 seconds.

The following lines of code slowed down by more than twice:

 950       deviatoric_stress_tensor(:,:) = stress_tensor(:,:)
 951       deviatoric_stress_tensor(1,1) = stress_tensor(1,1) - hydrostatic_stress
 952       deviatoric_stress_tensor(2,2) = stress_tensor(2,2) - hydrostatic_stress
 953       deviatoric_stress_tensor(3,3) = stress_tensor(3,3) - hydrostatic_stress
 954 !
 955       damaged_dev_stress_tensor(:,:) = deviatoric_stress_tensor(:,:)/(1.0_LONGreal -        &
 956                                                                     crack_parameter * damage)
 957 !
 958 !        compute the yield function
 959 !
 960       equivalent_stress = sqrt(1.5_LONGreal* sum((damaged_dev_stress_tensor(:,:)            &
 961                                                               - back_stress_tensor(:,:))**2))

default.log.gz - before/after SLP with the current Flang compiler

revert.log.gz - before/after SLP with the reverted commit

Reproducer command: opt -mcpu=znver4 -print-after=slp-vectorizer --passes=slp-vectorizer

[alexey-bataev]

@RKSimon something related to x86 cost model?

[RKSimon]

Possibly, I'll take a look next week

[RKSimon]

@vzakhari Sorry for not getting back to this - are you still seeing the regression? The getScalarizationOverhead code in SLP has changed since this was reported.

[vzakhari]

Hi @RKSimon! The benchmark has been showing the "slow" time until today. My guess is that #139682 could have improved it.

It is hard to say now whether it is still an issue. I will check what SLP does now for this loop. If you could create a revert patch of 88e7b8b based on the current head, I would be able to check it as well.

[vzakhari]

Hi @RKSimon! The benchmark has been showing the "slow" time until today. My guess is that #139682 could have improved it.

It is hard to say now whether it is still an issue. I will check what SLP does now for this loop. If you could create a revert patch of 88e7b8b based on the current head, I would be able to check it as well.

Nope, it was my machine issue. After a clean run, I can still reproduce the same slow time, and the same SLP behavior for the loop as after 88e7b8b.

Please let me know if you need anything from me to reproduce.