On 15/10/2025 19:58, Tobias Burnus wrote:
Hi Andrew, hi all,

Andrew Stubbs wrote:
The GCN bits look OK, although it would be nice if unsupported
tests didn't FAIL.

I actually wanted to have a look at this for a longer while,
but now I finally do so.

This patch now adds all missing context-selector -march tests
for AMD GCN – and it also handles the absent multilib issue,
that I was also running into before.

(Note that libgomp.c/declare-variant-4.c is with the default
options, i.e. for that test, I expect that linking is always
successful.)

Downside is only that we might miss some fails due to
UNRESOLVED if something fails early, but there is probably
nothing we can do ...

Does this look OK for mainline?

It feels like a bodge, but if it removes the FAIL then that's a step forward.

OK, thanks.

(I suppose the perfect solution would have an "effective target" test check the output from -print-multi-lib on the offload compiler, but that's non-trivial in the general case, and probably overkill.)

Andrew


Tobias

PS: On my system, I get:
XPASS: libgomp.c/declare-variant-4-gfx1036.c (test for excess errors)
XPASS: libgomp.c/declare-variant-4-gfx11-generic.c (test for excess errors)
XPASS: libgomp.c/declare-variant-4-gfx1100.c (test for excess errors)
XPASS: libgomp.c/declare-variant-4-gfx900.c (test for excess errors)
XPASS: libgomp.c/declare-variant-4-gfx906.c (test for excess errors)
XPASS: libgomp.c/declare-variant-4-gfx908.c (test for excess errors)
XPASS: libgomp.c/declare-variant-4-gfx90a.c (test for excess errors)
                 === libgomp Summary ===
# of expected passes            27
# of unexpected successes       7
# of expected failures          22
# of unresolved testcases       7
# of untested testcases         1

And unresolved are the following ones - because mkoffload
fails early with:
    gcn mkoffload: fatal error: GCC was built without library support
    for '-march=gfx1101'; consider compiling for the associated
    generic architecture '-march=gfx11-generic' instead
UNRESOLVED: libgomp.c/declare-variant-4-gfx1101.c scan-amdgcn-amdhsa-offload-tree-dump 
optimized "= gfx1101 \\(\\);"
UNRESOLVED: libgomp.c/declare-variant-4-gfx1102.c scan-amdgcn-amdhsa-offload-tree-dump 
optimized "= gfx1102 \\(\\);"
UNRESOLVED: libgomp.c/declare-variant-4-gfx1103.c scan-amdgcn-amdhsa-offload-tree-dump 
optimized "= gfx1103 \\(\\);"
UNRESOLVED: libgomp.c/declare-variant-4-gfx1150.c scan-amdgcn-amdhsa-offload-tree-dump 
optimized "= gfx1150 \\(\\);"
UNRESOLVED: libgomp.c/declare-variant-4-gfx1151.c scan-amdgcn-amdhsa-offload-tree-dump 
optimized "= gfx1151 \\(\\);"
UNRESOLVED: libgomp.c/declare-variant-4-gfx1152.c scan-amdgcn-amdhsa-offload-tree-dump 
optimized "= gfx1152 \\(\\);"
UNRESOLVED: libgomp.c/declare-variant-4-gfx1153.c scan-amdgcn-amdhsa-offload-tree-dump 
optimized "= gfx1153 \\(\\);"


Reply via email to