[Bug target/116085] RISC-V: Miscompile at -O2 with zbb

[law at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116085

--- Comment #5 from Jeffrey A. Law <law at gcc dot gnu.org> ---
Bisection landed on this change from 2022:

commit 3142265dedd84c2f3dbf824f2d1b0c182e3c8b3c
Author: Philipp Tomsich <philipp.toms...@vrull.eu>
Date:   Sun Oct 16 10:51:47 2022 +0200

    RISC-V: No extensions for SImode min/max against safe constant

    Optimize the common case of a SImode min/max against a constant
    that is safe both for sign- and zero-extension.
    E.g., consider the case
      int f(unsigned int* a)
      {
        const int C = 1000;
        return *a * 3 > C ? C : *a * 3;
      }
    where the constant C will yield the same result in DImode whether
    sign- or zero-extended.

    This should eventually go away once the lowering to RTL smartens up
    and considers the precision/signedness and the value-ranges of the
    operands to MIN_EXPR and MAX_EXPR.

    gcc/ChangeLog:

            * config/riscv/bitmanip.md (*minmax): Additional pattern for
              min/max against constants that are extension-invariant.
            * config/riscv/iterators.md (minmax_optab): Add an iterator
              that has only min and max rtl.

    gcc/testsuite/ChangeLog:

            * gcc.target/riscv/zbb-min-max-02.c: New test.

diff --git a/gcc/config/riscv/bitmanip.md b/gcc/config/riscv/bitmanip.md
index 58bac7231e3..d17133d58c1 100644
--- a/gcc/config/riscv/bitmanip.md
+++ b/gcc/config/riscv/bitmanip.md
@@ -368,6 +368,24 @@ (define_insn "<bitmanip_optab><mode>3"
   "<bitmanip_insn>\t%0,%1,%2"
   [(set_attr "type" "bitmanip")])

+;; Optimize the common case of a SImode min/max against a constant
+;; that is safe both for sign- and zero-extension.
+(define_insn_and_split "*minmax"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+       (sign_extend:DI
+         (subreg:SI
+           (bitmanip_minmax:DI (zero_extend:DI (match_operand:SI 1
"register_operand" "r"))
+                                               (match_operand:DI 2
"immediate_operand" "i"))
+          0)))
+   (clobber (match_scratch:DI 3 "=&r"))
+   (clobber (match_scratch:DI 4 "=&r"))]
+  "TARGET_64BIT && TARGET_ZBB && sext_hwi (INTVAL (operands[2]), 32) >= 0"
+  "#"
+  "&& reload_completed"
+  [(set (match_dup 3) (sign_extend:DI (match_dup 1)))
+   (set (match_dup 4) (match_dup 2))
+   (set (match_dup 0) (<minmax_optab>:DI (match_dup 3) (match_dup 4)))])



Note carefully the sign_extend in the split code.  That looks wrong to me at
first glance.