[PATCH v2] cgen: Compute correct mask and values when offset in define-ifield is not 0.

[Guillermo Martinez]

On Thursday, November 25, 2021 5:02:11 PM CST Alan Modra wrote:
> On Wed, Sep 15, 2021 at 06:04:44PM +0000, Guillermo Martinez wrote:
> > On Wednesday, September 15, 2021 11:49:12 AM CDT Jose E. Marchesi wrote:
> > > I just installed this version of the patch on your behalf.
> > > Thanks!
> > > 
> > > > If an instruction field is defined in a long form, assigning
> > > > an offset different to 0 the mask and constant values are not
> > > > computed appropriately:
> [snip]
> 
> It's been a while since I refreshed the binutils opcodes files that
> are cgen generated.  On doing so it appears this patch is responsible
> for build errors when using mainline gcc.  First example:
> 
> /home/alan/src/binutils-gdb/opcodes/bpf-opc.c:57:11: error: conversion from ‘long unsigned int’ to ‘unsigned int’ changes value from ‘18446744073709486335’ to ‘4294902015’ [-Werror=overflow]
>    57 |   64, 64, 0xffffffffffff00ff, { { F (F_IMM32) }, { F (F_OFFSET16) }, { F (F_SRCLE) }, { F (F_OP_CODE) }, { F (F_DSTLE) }, { F (F_OP_SRC) }, { F (F_OP_CLASS) }, { 0 } }
>       |           ^~~~~~~~~~~~~~~~~~
> 
> I tested an obvious workaround,
> 
> diff --git a/include/opcode/cgen.h b/include/opcode/cgen.h
> index 8b7d2a4b547..0e9571d19b0 100644
> --- a/include/opcode/cgen.h
> +++ b/include/opcode/cgen.h
> @@ -914,7 +914,7 @@ typedef struct
>       Each insn's value is stored with the insn.
>       The first step in recognizing an insn for disassembly is
>       (opcode & mask) == value.  */
> -  CGEN_INSN_INT mask;
> +  uint64_t mask;
>  #define CGEN_IFMT_MASK(ifmt) ((ifmt)->mask)
>  
>    /* Instruction fields.
> diff --git a/opcodes/cgen-dis.c b/opcodes/cgen-dis.c
> index 1a5d1ae8459..37ee5a23564 100644
> --- a/opcodes/cgen-dis.c
> +++ b/opcodes/cgen-dis.c
> @@ -39,7 +39,7 @@ static void		 add_insn_to_hash_chain (CGEN_INSN_LIST *,
>  static int
>  count_decodable_bits (const CGEN_INSN *insn)
>  {
> -  unsigned mask = CGEN_INSN_BASE_MASK (insn);
> +  uint64_t mask = CGEN_INSN_BASE_MASK (insn);
>  #if GCC_VERSION >= 3004
>    return __builtin_popcount (mask);
>  #else
> 
> Quite possibly the above isn't a complete fix, I just threw it
> together to see what happens.  Regressions:
correct, it's not a complete fix:
  https://sourceware.org/pipermail/binutils/2021-October/118284.html

However applying patch v2 on upstream pitifully it won't work because there is
an error in BPF simulator (sim/bpf/decode-le.c (extract_sfmt_lddwle))
(not introduced by this patch)  but now raised by adding -Werror=shift-count-overflow
in binutils-gdb builder (currently im debugging this :-))

> +FAIL: eBPF CALL instruction
> +FAIL: eBPF CALL instruction, big endian
> +FAIL: CALL with disp32 reloc
> +FAIL: CALL with disp32 reloc and addend
> +FAIL: CALL check unsigned underflow
> 
> gas/testsuite/gas.log for the first one shows:
> regexp_diff match failure
> regexp "^  20:  85 10 00 00 00 00 00 00         call 0$"
> line   "  20:   85 10 00 00 00 00 00 00         *unknown*"
> regexp_diff match failure
> regexp "^  28:  85 10 00 00 ff ff ff ff         call -1$"
> line   "  28:   85 10 00 00 ff ff ff ff         *unknown*"
> regexp_diff match failure
> regexp "^  30:  85 10 00 00 fe ff ff ff         call -2$"
> line   "  30:   85 10 00 00 fe ff ff ff         *unknown*"
> regexp_diff match failure
> regexp "^  38:  85 10 00 00 fd ff ff ff         call -3$"
> line   "  38:   85 10 00 00 fd ff ff ff         *unknown*"
> FAIL: eBPF CALL instruction
> 
> At this point I don't intend to look any further into the problem.
> The opcodes/ refresh can wait.
I'll work on this ASAP.

Thanks Alan!
Guillermo