[gcc(refs/vendors/ARM/heads/morello)] PCS use correct offset when loading into registers

[Matthew Malcomson <matmal01@gcc.gnu.org>]

https://gcc.gnu.org/g:3d5dfab2db1550d1cce437cc76b006d3c4dd882f

commit 3d5dfab2db1550d1cce437cc76b006d3c4dd882f
Author: Matthew Malcomson <matthew.malcomson@arm.com>
Date:   Thu Sep 16 11:33:58 2021 +0100

    PCS use correct offset when loading into registers
    
    When loading an argument into registers in order to pass to a function
    we were previously miscalculating the offset to load the second
    capability register from.  We used 8 since this was UNITS_PER_WORD, this
    commit adjusts that to use GET_MODE_SIZE (CADImode) so the offset when
    loading capability registers is 16.
    
    This can be seen in the offset used to load the second register holding
    the structure argument in the below testcase.
    
    typedef __SIZE_TYPE__ size_t;
    struct alt_v { void *a; unsigned short int b; };
    extern struct alt_v check_alt_v (struct alt_v);
    struct alt_v altv;
    size_t test_alt_v ()
    {
      struct alt_v ret = check_alt_v (altv);
      return sizeof (struct alt_v) + ret.b;
    }
    
    Testcases added in another patch.

Diff:
---
 gcc/config/aarch64/aarch64.c | 18 +++++++++++-------
 1 file changed, 11 insertions(+), 7 deletions(-)

diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index ad4e2739576..8e25d7da8cf 100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -6090,7 +6090,7 @@ aarch64_layout_arg (cumulative_args_t pcum_v, const function_arg_info &arg)
   machine_mode mode = arg.mode;
   int ncrn, nvrn, nregs;
   bool allocate_ncrn, allocate_nvrn;
-  HOST_WIDE_INT size;
+  HOST_WIDE_INT size, units_per_reg = UNITS_PER_WORD;
   bool abi_break;
   bool in_cap_regs = false;
 
@@ -6210,8 +6210,6 @@ aarch64_layout_arg (cumulative_args_t pcum_v, const function_arg_info &arg)
 	}
     }
 
-  ncrn = pcum->aapcs_ncrn;
-  nregs = size / UNITS_PER_WORD;
   if (type)
     switch (aarch64_classify_capability_contents (type))
       {
@@ -6220,7 +6218,7 @@ aarch64_layout_arg (cumulative_args_t pcum_v, const function_arg_info &arg)
 	break;
       case CAPCOM_SOME:
 	in_cap_regs = true;
-	nregs = ROUND_UP (nregs, 2) / 2;
+	units_per_reg = GET_MODE_SIZE (CADImode);
 	break;
       case CAPCOM_OVERLAP:
 	goto on_stack;
@@ -6230,9 +6228,15 @@ aarch64_layout_arg (cumulative_args_t pcum_v, const function_arg_info &arg)
       }
   if (TARGET_MORELLO && mode == CADImode)
     {
+      /* If there was a type then it should have been identified as something
+	 that should be passed in capability registers.  May as well add an
+	 assertion here to ensure this is correct.  */
+      gcc_assert (!type || in_cap_regs == true);
       in_cap_regs = true;
-      nregs = ROUND_UP (nregs, 2) / 2;
+      units_per_reg = GET_MODE_SIZE (CADImode);
     }
+  ncrn = pcum->aapcs_ncrn;
+  nregs = size / units_per_reg;
 
   /* C6 - C9.  though the sign and zero extension semantics are
      handled elsewhere.  This is the case where the argument fits
@@ -6265,7 +6269,7 @@ aarch64_layout_arg (cumulative_args_t pcum_v, const function_arg_info &arg)
 	 Using the normal (parallel [...]) would suppress the shifting.  */
       if (sve_p
 	  && BYTES_BIG_ENDIAN
-	  && maybe_ne (GET_MODE_SIZE (mode), nregs * UNITS_PER_WORD)
+	  && maybe_ne (GET_MODE_SIZE (mode), nregs * units_per_reg)
 	  && aarch64_pad_reg_upward (mode, type, false))
 	{
 	  mode = int_mode_for_mode (mode).require ();
@@ -6294,7 +6298,7 @@ aarch64_layout_arg (cumulative_args_t pcum_v, const function_arg_info &arg)
 		reg_mode = int_mode_for_mode (mode).require ();
 	      rtx tmp = gen_rtx_REG (reg_mode, R0_REGNUM + ncrn + i);
 	      tmp = gen_rtx_EXPR_LIST (VOIDmode, tmp,
-				       GEN_INT (i * UNITS_PER_WORD));
+				       GEN_INT (i * units_per_reg));
 	      XVECEXP (par, 0, i) = tmp;
 	    }
 	  pcum->aapcs_reg = par;