[nvptx] Fix ASM_SPEC workaround for sm_30
authorTom de Vries <tdevries@suse.de>
Wed, 30 Mar 2022 11:26:52 +0000 (13:26 +0200)
committerTom de Vries <tdevries@suse.de>
Thu, 31 Mar 2022 07:34:42 +0000 (09:34 +0200)
Newer versions of CUDA no longer support sm_30, and nvptx-tools as
currently doesn't handle that gracefully when verifying
( https://github.com/MentorEmbedded/nvptx-tools/issues/30 ).

There's a --no-verify work-around in place in ASM_SPEC, but that one doesn't
work when using -Wa,--verify on the command line.

Use a more robust workaround: verify using sm_35 when misa=sm_30 is specified
(either implicitly or explicitly).

Tested on nvptx.

gcc/ChangeLog:

2022-03-30  Tom de Vries  <tdevries@suse.de>

* config/nvptx/nvptx.h (ASM_SPEC): Use "-m sm_35" for -misa=sm_30.

gcc/config/nvptx/nvptx.h

index 75ac7a6..3b06f33 100644 (file)
 
 #define STARTFILE_SPEC "%{mmainkernel:crt0.o}"
 
-/* Default needs to be in sync with default for misa in nvptx.opt.
-   We add a default here to work around a hard-coded sm_30 default in
-   nvptx-as.  */
-#define ASM_SPEC "%{misa=*:-m %*; :-m sm_35}%{misa=sm_30:--no-verify}"
+/* Newer versions of CUDA no longer support sm_30, and nvptx-tools as
+   currently doesn't handle that gracefully when verifying
+   ( https://github.com/MentorEmbedded/nvptx-tools/issues/30 ).  Work around
+   this by verifying with sm_35 when having misa=sm_30 (either implicitly
+   or explicitly).  */
+#define ASM_SPEC                               \
+  "%{"                                         \
+  /* Explict misa=sm_30.  */                   \
+  "misa=sm_30:-m sm_35"                                \
+  /* Separator.         */                             \
+  "; "                                         \
+  /* Catch-all.         */                             \
+  "misa=*:-m %*"                               \
+  /* Separator.         */                             \
+  "; "                                         \
+  /* Implicit misa=sm_30.  */                  \
+  ":-m sm_35"                                  \
+  "}"
 
 #define TARGET_CPU_CPP_BUILTINS() nvptx_cpu_cpp_builtins ()