On 6/27/24 06:58, Akihiko Odaki wrote:
sme-smopa-2.c requires sme-i16i64 but the compiler option used not to
specify it. Instead, the extension was specified with the inline
assembly, resulting in mixing assembly code targeting sme-i1664 and C
code that does not target sme-i1664.
clang version 18.1.6 does not support such mixing so properly specify
the extension with the compiler option instead.
Signed-off-by: Akihiko Odaki
---
tests/tcg/aarch64/sme-smopa-2.c | 2 +-
tests/tcg/aarch64/Makefile.target | 11 +--
2 files changed, 10 insertions(+), 3 deletions(-)
diff --git a/tests/tcg/aarch64/sme-smopa-2.c b/tests/tcg/aarch64/sme-smopa-2.c
index c9f48c3bfca2..2c9707065992 100644
--- a/tests/tcg/aarch64/sme-smopa-2.c
+++ b/tests/tcg/aarch64/sme-smopa-2.c
@@ -14,7 +14,7 @@ int main()
long svl;
/* Validate that we have a wide enough vector for 4 elements. */
-asm(".arch armv8-r+sme-i64\n\trdsvl %0, #1" : "=r"(svl));
+asm("rdsvl %0, #1" : "=r"(svl));
if (svl < 32) {
return 0;
}
diff --git a/tests/tcg/aarch64/Makefile.target
b/tests/tcg/aarch64/Makefile.target
index 70d728ae9af7..ad99e0e3b198 100644
--- a/tests/tcg/aarch64/Makefile.target
+++ b/tests/tcg/aarch64/Makefile.target
@@ -27,7 +27,8 @@ config-cc.mak: Makefile
$(call cc-option,-march=armv8.5-a,
CROSS_CC_HAS_ARMV8_5); \
$(call cc-option,-mbranch-protection=standard,
CROSS_CC_HAS_ARMV8_BTI); \
$(call cc-option,-march=armv8.5-a+memtag,
CROSS_CC_HAS_ARMV8_MTE); \
- $(call cc-option,-Wa$(COMMA)-march=armv9-a+sme,
CROSS_AS_HAS_ARMV9_SME)) 3> config-cc.mak
+ $(call cc-option,-Wa$(COMMA)-march=armv9-a+sme,
CROSS_AS_HAS_ARMV9_SME); \
+ $(call cc-option,-march=armv9-a+sme-i16i64,
CROSS_AS_HAS_ARMV9_SME_I1664)) 3> config-cc.mak
-include config-cc.mak
ifneq ($(CROSS_CC_HAS_ARMV8_2),)
@@ -68,7 +69,13 @@ endif
# SME Tests
ifneq ($(CROSS_AS_HAS_ARMV9_SME),)
-AARCH64_TESTS += sme-outprod1 sme-smopa-1 sme-smopa-2
+AARCH64_TESTS += sme-outprod1 sme-smopa-1
+endif
+
+# SME I16I64 Tests
+ifneq ($(CROSS_AS_HAS_ARMV9_SME_I1664),)
+AARCH64_TESTS += sme-smopa-2
+sme-smopa-2: CFLAGS += -march=armv9-a+sme-i16i64
endif
How interesting. We were not actually passing -march=armv9-a+sme to the assembler
previously. Lack of this is what is causing sme-outprod1 to fail to build, as reported by
Alex.
That said, if we use compiler directives we must have gcc-14 or newer to test this,
whereas binutils supported sme (and extensions) much earlier. Given that this is all
inline assembly, we do not really need compiler support.
I think we should continue to pass assembler options (-Wa,...) and detect and use clang's
-no-integrated-as option as well, at least for the SME tests.
r~
