Re: [PATCH] gimple-fold: fold fwrite of a single byte to fputc

On Wed, Aug 19, 2026 at 2:34 PM Kyrylo Tkachov <[email protected]> wrote:
>
>
>
> > On 19 Aug 2026, at 14:06, Richard Biener <[email protected]> wrote:
> >
> > On Wed, Aug 19, 2026 at 1:03 PM <[email protected]> wrote:
> >>
> >> From: Kyrylo Tkachov <[email protected]>
> >>
> >> fwrite (p, 1, 1, f) transfers one byte and, when nothing looks at its result,
> >> writes exactly what fputc (*p, f) writes.  fputc reaches the stream directly
> >> while fwrite goes through the generic buffered-write path, so the fold is
> >> worth making.  GCC already treats the two as interchangeable in the other
> >> direction, since gimple_fold_builtin_fputs turns a one-character fputs into
> >> fputc and a longer one into fwrite.  There was simply no folder for fwrite.
> >>
> >> The unlocked entry points fold the same way, into fputc_unlocked, on the
> >> assumption gimple_fold_builtin_fputs already makes: a program that calls one
> >> unlocked stdio function has the others available too.
> >>
> >> jq emits its output one character at a time through a helper that ends in
> >> fwrite (&c, 1, 1, f), which is the shape this fold is aimed at:
> >>
> >>    before                              after
> >>        stp     x29, x30, [sp, -32]!    and     w0, w0, 255
> >>        mov     w2, 1                   b       fputc
> >>        mov     x29, sp
> >>        mov     x3, x1
> >>        uxtw    x1, w2
> >>        strb    w0, [sp, 31]
> >>        add     x0, sp, 31
> >>        bl      fwrite
> >>        ldp     x29, x30, [sp], 32
> >>        ret
> >>
> >> Measured with jq 1.8.2 over a 57 MB NDJSON corpus on Grace with
> >> -mcpu=grace -O3 gets these speedups.
> >>
> >>    jq -c .                             -30.30% instructions, -44.80% cycles
> >>    jq -c -f <a map/select filter>       -2.92% instructions,  -5.67% cycles
> >>
> >> Bootstrapped and tested on aarch64-none-linux-gnu.
> >> Ok for trunk?
> >> Thanks,
> >> Kyrill
> >>
> >> gcc/ChangeLog:
> >>
> >>        * gimple-fold.cc (gimple_fold_builtin_fwrite): New function.
> >>        (gimple_fold_builtin): Call it for BUILT_IN_FWRITE and
> >>        BUILT_IN_FWRITE_UNLOCKED.
> >>
> >> gcc/testsuite/ChangeLog:
> >>
> >>        * gcc.dg/fwrite-1.c: New test.
> >>        * gcc.dg/fwrite-2.c: New test.
> >>        * gcc.dg/fwrite-3.c: New test.
> >>
> >> Signed-off-by: Kyrylo Tkachov <[email protected]>
> >> ---
> >> gcc/gimple-fold.cc              | 47 +++++++++++++++++++++++++++++++++
> >> gcc/testsuite/gcc.dg/fwrite-1.c | 15 +++++++++++
> >> gcc/testsuite/gcc.dg/fwrite-2.c | 32 ++++++++++++++++++++++
> >> gcc/testsuite/gcc.dg/fwrite-3.c | 25 ++++++++++++++++++
> >> 4 files changed, 119 insertions(+)
> >> create mode 100644 gcc/testsuite/gcc.dg/fwrite-1.c
> >> create mode 100644 gcc/testsuite/gcc.dg/fwrite-2.c
> >> create mode 100644 gcc/testsuite/gcc.dg/fwrite-3.c
> >>
> >> diff --git a/gcc/gimple-fold.cc b/gcc/gimple-fold.cc
> >> index 24041466108..583afe119e7 100644
> >> --- a/gcc/gimple-fold.cc
> >> +++ b/gcc/gimple-fold.cc
> >> @@ -3099,6 +3099,49 @@ gimple_fold_builtin_fputs (gimple_stmt_iterator *gsi,
> >>     }
> >> }
> >>
> >> +/* Fold a call to fwrite (PTR, SIZE, N, STREAM) at *GSI.  UNLOCKED says whether
> >> +   the callee is fwrite_unlocked rather than fwrite.  A call that transfers a
> >> +   single byte and whose result is nobody's business writes the same byte as
> >> +   fputc (*PTR, STREAM), which reaches the stream without going through the
> >> +   generic buffered-write path.  Return true if the call was folded.  */
> >> +
> >> +static bool
> >> +gimple_fold_builtin_fwrite (gimple_stmt_iterator *gsi, bool unlocked)
> >> +{
> >> +  gimple *stmt = gsi_stmt (*gsi);
> >> +
> >> +  /* fwrite reports the number of items transferred and fputc the character
> >> +     written, so only fold when nothing looks at the result.  */
> >> +  if (gimple_call_lhs (stmt))
> >> +    return false;
> >> +
> >> +  /* fwrite transfers SIZE * N bytes, so writing a single byte needs both
> >> +     counts to be one: no other pair of non-negative values multiplies to
> >> +     one.  */
> >> +  if (!integer_onep (gimple_call_arg (stmt, 1))
> >> +      || !integer_onep (gimple_call_arg (stmt, 2)))
> >> +    return false;
> >> +
> >> +  /* If we're using an unlocked function, assume the other unlocked
> >> +     functions exist explicitly.  */
> >> +  tree const fn_fputc = (unlocked
> >> +                        ? builtin_decl_explicit (BUILT_IN_FPUTC_UNLOCKED)
> >> +                        : builtin_decl_implicit (BUILT_IN_FPUTC));
> >> +  if (!fn_fputc || (!gimple_vdef (stmt) && gimple_in_ssa_p (cfun)))
> >> +    return false;
> >> +
> >> +  location_t loc = gimple_location (stmt);
> >> +  gimple_seq stmts = NULL;
> >> +  tree byte = gimple_load_first_char (loc, gimple_call_arg (stmt, 0), &stmts);
> >> +  tree c = make_ssa_name (integer_type_node);
> >> +  gimple_seq_add_stmt_without_update (&stmts,
> >> +                                     gimple_build_assign (c, NOP_EXPR, byte));
> >
> >   tree c = gimple_convert (&stmts, integer_type_node, byte);
>
> Will do, thanks.
>
> >
> >> +  gimple_seq_add_stmt_without_update
> >> +    (&stmts, gimple_build_call (fn_fputc, 2, c, gimple_call_arg (stmt, 3)));
> >
> >  gimple_build (&stmts, fn_putc, void_type_node, c, gimple_call_arg (stmt, 3));
> >
> > might work here (unsure about calls with no LHS).
>
> There's no gimple_build overload taking a function decl so that wouldn’t compile, unless I’m not looking in the right place.

Ah, true, it would be on BUILT_IN_FPUTC directly, no need to lookup
a decl here.

Richard.

> Thanks,
> Kyrill
>
> >
> >> +  gsi_replace_with_seq_vops (gsi, stmts);
> >> +  return true;
> >> +}
> >> +
> >> /* Fold a call to the __mem{cpy,pcpy,move,set}_chk builtin.
> >>    DEST, SRC, LEN, and SIZE are the arguments to the call.
> >>    IGNORE is true, if return value can be ignored.  FCODE is the BUILT_IN_*
> >> @@ -5583,6 +5626,10 @@ gimple_fold_builtin (gimple_stmt_iterator *gsi)
> >>     case BUILT_IN_FPUTS_UNLOCKED:
> >>       return gimple_fold_builtin_fputs (gsi, gimple_call_arg (stmt, 0),
> >>                                        gimple_call_arg (stmt, 1), true);
> >> +    case BUILT_IN_FWRITE:
> >> +      return gimple_fold_builtin_fwrite (gsi, false);
> >> +    case BUILT_IN_FWRITE_UNLOCKED:
> >> +      return gimple_fold_builtin_fwrite (gsi, true);
> >>     case BUILT_IN_MEMCPY_CHK:
> >>     case BUILT_IN_MEMPCPY_CHK:
> >>     case BUILT_IN_MEMMOVE_CHK:
> >> diff --git a/gcc/testsuite/gcc.dg/fwrite-1.c b/gcc/testsuite/gcc.dg/fwrite-1.c
> >> new file mode 100644
> >> index 00000000000..5e3a5043e9a
> >> --- /dev/null
> >> +++ b/gcc/testsuite/gcc.dg/fwrite-1.c
> >> @@ -0,0 +1,15 @@
> >> +/* fwrite of a single byte whose result is unused writes the same byte as
> >> +   fputc, so it should be folded.  */
> >> +/* { dg-do compile } */
> >> +/* { dg-options "-O2 -Wno-unused-result -fdump-tree-optimized" } */
> >> +
> >> +#include <stdio.h>
> >> +
> >> +void one (FILE *f, const char *p) { fwrite (p, 1, 1, f); }
> >> +size_t used (FILE *f, const char *p) { return fwrite (p, 1, 1, f); }
> >> +void two_items (FILE *f, const char *p) { fwrite (p, 1, 2, f); }
> >> +void two_bytes (FILE *f, const char *p) { fwrite (p, 2, 1, f); }
> >> +void unknown (FILE *f, const char *p, size_t n) { fwrite (p, 1, n, f); }
> >> +
> >> +/* { dg-final { scan-tree-dump-times "fputc" 1 "optimized" } } */
> >> +/* { dg-final { scan-tree-dump-times "fwrite" 4 "optimized" } } */
> >> diff --git a/gcc/testsuite/gcc.dg/fwrite-2.c b/gcc/testsuite/gcc.dg/fwrite-2.c
> >> new file mode 100644
> >> index 00000000000..7565f52def8
> >> --- /dev/null
> >> +++ b/gcc/testsuite/gcc.dg/fwrite-2.c
> >> @@ -0,0 +1,32 @@
> >> +/* Check that folding fwrite of a single byte to fputc keeps the output the
> >> +   same, including the bytes that are not folded, and that a side effect in the
> >> +   stream argument is still evaluated exactly once.  */
> >> +/* { dg-do run } */
> >> +/* { dg-options "-O2 -Wno-unused-result" } */
> >> +
> >> +#include <stdio.h>
> >> +
> >> +extern void abort (void);
> >> +
> >> +const char s[] = "abcd";
> >> +
> >> +int
> >> +main (void)
> >> +{
> >> +  FILE *streams[] = { stdout, NULL }, **p = streams;
> >> +
> >> +  fwrite (s, 1, 1, stdout);
> >> +  fwrite (s + 1, 1, 1, stdout);
> >> +  fwrite (s + 2, 1, 2, stdout);
> >> +  fwrite (s, 2, 1, stdout);
> >> +
> >> +  /* The folded call must still advance P exactly once.  */
> >> +  fwrite (s, 1, 1, *p++);
> >> +  if (p != streams + 1 || *p != NULL)
> >> +    abort ();
> >> +
> >> +  fflush (stdout);
> >> +  return 0;
> >> +}
> >> +
> >> +/* { dg-output "abcdaba" } */
> >> diff --git a/gcc/testsuite/gcc.dg/fwrite-3.c b/gcc/testsuite/gcc.dg/fwrite-3.c
> >> new file mode 100644
> >> index 00000000000..6ee0f58cb59
> >> --- /dev/null
> >> +++ b/gcc/testsuite/gcc.dg/fwrite-3.c
> >> @@ -0,0 +1,25 @@
> >> +/* fwrite_unlocked of a single byte whose result is unused writes the same byte
> >> +   as fputc_unlocked, so it is folded the same way as the locked form, and into
> >> +   the unlocked entry point rather than the locked one.  */
> >> +/* { dg-do compile } */
> >> +/* { dg-options "-O2 -Wno-unused-result -fdump-tree-optimized" } */
> >> +
> >> +#include <stdio.h>
> >> +
> >> +/* Declared here rather than through _GNU_SOURCE so that the test does not
> >> +   depend on the host header exposing the unlocked entry points.  */
> >> +extern int (fputc_unlocked) (int, FILE *);
> >> +extern size_t (fwrite_unlocked) (const void *, size_t, size_t, FILE *);
> >> +
> >> +void one (FILE *f, const char *p) { fwrite_unlocked (p, 1, 1, f); }
> >> +size_t used (FILE *f, const char *p) { return fwrite_unlocked (p, 1, 1, f); }
> >> +void two_items (FILE *f, const char *p) { fwrite_unlocked (p, 1, 2, f); }
> >> +void two_bytes (FILE *f, const char *p) { fwrite_unlocked (p, 2, 1, f); }
> >> +void unknown (FILE *f, const char *p, size_t n) { fwrite_unlocked (p, 1, n, f); }
> >> +
> >> +/* { dg-final { scan-tree-dump-times "fputc_unlocked" 1 "optimized" } } */
> >> +/* { dg-final { scan-tree-dump-times "fwrite_unlocked" 4 "optimized" } } */
> >> +/* The unlocked form must not be folded into the locked fputc.  A call with no
> >> +   result is dumped as "  fputc (...", which the leading space matches without
> >> +   also matching fputc_unlocked.  */
> >> +/* { dg-final { scan-tree-dump-not " fputc \\(" "optimized" } } */
> >> --
> >> 2.50.1 (Apple Git-155)
>
>