Hi Richard, Thanks for the review and explanation.
The previous fix adding if condition of TARGET_FLOAT does crash glibc-2.29. I checked the past log of writing the function aarch64_init_cumulative_args, and did not find the reason why Alan Lawrence added TREE_PUBLIC (fndecl) as one condition for entering the function type check. Maybe Alan could clarify? I tried to delete TREE_PUBLIC (fndecl), which turns out could solve both the glibc problem and the previous ICE problem. A new fix is made as following, passed bootstrap and deja test. I believe this fix is reasonable, since the function type should be checked no matter if it has external linkage or not. The function aarch64_init_cumulative_args checks the function types and should catch the error that "-mgeneral-regs-only" is incompatible with the use of SIMD/FP registers. In the test case on PR96479, the function myfunc2 returns one vector of 4 integers, while it is defined static type. TREE_PUBLIC (fndecl) is set as false and it prevents from entering if statement and checking function types. I delete "TREE_PUBLIC (fndecl)" so that gcc can catch the error through the function aarch64_init_cumulative_args now. The ICE on PR96479 can report the diagnostic error with this fix. The patch for the fix is attached as following: diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c index b7f5bc76f1b..9ce83dce131 100644 --- a/gcc/config/aarch64/aarch64.c +++ b/gcc/config/aarch64/aarch64.c @@ -6017,7 +6017,7 @@ aarch64_init_cumulative_args (CUMULATIVE_ARGS *pcum, if (!silent_p && !TARGET_FLOAT - && fndecl && TREE_PUBLIC (fndecl) + && fndecl && fntype && fntype != error_mark_node) { const_tree type = TREE_TYPE (fntype); Christophe, thanks for your tests on glibc-2.29. With the above fix, I built glibc-2.29, and the previous error does not show up now. Could you please check if this fix works? Do you have any suggestions on this fix? All the best, Peixin -----Original Message----- From: Richard Sandiford [mailto:richard.sandif...@arm.com] Sent: Thursday, August 13, 2020 8:19 PM To: Christophe Lyon <christophe.l...@linaro.org> Cc: qiaopeixin <qiaopei...@huawei.com>; gcc-patches@gcc.gnu.org Subject: Re: [PATCH] AArch64: Add if condition in aarch64_function_value [PR96479] Christophe Lyon <christophe.l...@linaro.org> writes: > On Thu, 13 Aug 2020 at 03:54, qiaopeixin <qiaopei...@huawei.com> wrote: >> >> Thanks for the review and commit. >> >> All the best, >> Peixin >> >> -----Original Message----- >> From: Richard Sandiford [mailto:richard.sandif...@arm.com] >> Sent: 2020年8月13日 0:25 >> To: qiaopeixin <qiaopei...@huawei.com> >> Cc: gcc-patches@gcc.gnu.org >> Subject: Re: [PATCH] AArch64: Add if condition in >> aarch64_function_value [PR96479] >> >> qiaopeixin <qiaopei...@huawei.com> writes: >> > Hi, >> > >> > The test case vector-subscript-2.c in the gcc testsuit will report an ICE >> > in the expand pass since '-mgeneral-regs-only' is incompatible with the >> > use of V4SI mode. I propose to report the diagnostic information instead >> > of ICE, and the problem has been discussed on PR 96479. >> > >> > I attached the patch to solve the problem. Bootstrapped and tested on >> > aarch64-linux-gnu. Any suggestions? >> >> Thanks, pushed. I was initially sceptical because raising an error here and >> in aarch64_layout_arg is a hack. Both functions are just query functions >> and shouldn't have any side effects. >> >> The approach we took for FP modes seemed better: we define the FP move >> patterns unconditionally, and raise an error if we try to emit an FP move >> with !TARGET_FLOAT. This defers any error reporting until we actually try >> to generate code that depends on TARGET_FLOAT. >> >> But I guess SIMD stuff is different. There's no reason in principle why you >> can't use: >> >> unsigned short __attribute__((vector_size(8))) >> >> *within* a function with -mgeneral-regs-only. It would just need to be >> emulated, in the same way as for: >> >> unsigned short __attribute__((vector_size(4))) >> >> So it would be wrong to define the SIMD move patterns unconditionally and >> raise an error there. >> >> So all in all, I agree this is the best we can do given the current >> infrastructure. >> > > Since this patch was committed my buildbot is broken for > aarch64-linux-gnu because it now fails to build glibc-2.29: > ../stdlib/bits/stdlib-float.h: In function 'atof': > ../stdlib/bits/stdlib-float.h:26:1: error: '-mgeneral-regs-only' is > incompatible with the use of floating-point types Thanks for the heads-up. I've reverted the patch for now. Looking more closely, it seems like aarch64_init_cumulative_args already tries to catch the problem that the patch was fixing: if (!silent_p && !TARGET_FLOAT && fndecl && TREE_PUBLIC (fndecl) && fntype && fntype != error_mark_node) { const_tree type = TREE_TYPE (fntype); machine_mode mode ATTRIBUTE_UNUSED; /* To pass pointer as argument. */ int nregs ATTRIBUTE_UNUSED; /* Likewise. */ if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (type), type, &mode, &nregs, NULL, false)) aarch64_err_no_fpadvsimd (TYPE_MODE (type)); } The only reason it doesn't work for the testcase is that TREE_PUBLIC condition. TBH I'm not sure why it or the fndecl test is there: this is just as problematic when calling via a function pointer or when calling a static function. Richard