[RISCV][compiler-rt] Add __riscv_restore_tailcall_N entry points

[edward-jones]

Summary

This introduces entry points to compiler-rt which allow the -msave-restore optimization to be applied to tail calls as well as ordinary calls.

[edward-jones]

Updated the implementation to closely match the __riscv_restore_N entry points.

[edward-jones]

It was pointed out in the RISC-V LLVM sync-up that these entry points could be in separate sections to allow -gc-sections to remove any unnecessary entry points. This would mitigate the code size penalty for small programs as only the required entry points would be pulled in, however it would require the linker script to behave and not reorder the sections, and would require some changes to ensure that the fallthrough function does not accidentally get collected.

[luismarques]

Jim Wilson's [draft GNU implementation][1] has CFI directives. IIRC we also didn't add them to the non-tail Compiler-RT implementation but it would be good to take this opportunity to add them here.

[1]: https://github.com/riscv-non-isa/riscv-toolchain-conventions/files/7800168/sibling-call-save-restore.txt

compiler-rt/lib/builtins/riscv/restore_tailcall.S

104

Remove the "/9/8"?

[edward-jones]

Added CFI directives to the __riscv_restore_tailcall_<N> functions

[luismarques]

LGTM. (Don't forget to address my old nit comment, "Remove the /9/8", before merging)
Let's discuss on D91719 when this work should be merged.

compiler-rt/lib/builtins/riscv/restore_tailcall.S

1

Maybe update the description to reflect tail call, if it fits.

[edward-jones]

Fix comment

compiler-rt/lib/builtins/CMakeLists.txt

set(powerpc64_SOURCES
     ${powerpc64_SOURCES}
   )
 endif()
 set(powerpc64le_SOURCES ${powerpc64_SOURCES})
 
 set(riscv_SOURCES
   riscv/save.S
   riscv/restore.S
+  riscv/restore_tailcall.S
   ${GENERIC_SOURCES}
   ${GENERIC_TF_SOURCES}
 )
 set(riscv32_SOURCES
   riscv/mulsi3.S
   ${riscv_SOURCES}
 )
 set(riscv64_SOURCES

compiler-rt/lib/builtins/riscv/restore_tailcall.S

+//===-- restore_tailcall.S - restore up to 12 callee-save registers -------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// Multiple entry points depending on number of registers to restore. Restores
+// registers and then performs a tail call.
+//
+//===----------------------------------------------------------------------===//
+
+// These are alternative entry points to the normal __riscv_restore_<N>
+// functions which allow the save-restore mechnism to be used even when the
+// caller is making a tail call. These entry points take a single argument
+// in `t1` which is the address of the function to jump to after restoring
+// the registers and stack pointer.
+
+// All of the entry points are in the same section since we rely on many of
+// them falling through into each other and don't want the linker to
+// accidentally split them up, garbage collect, or reorder them.
+//
+// The entry points are grouped up into 2s for rv64 and 4s for rv32 since this
+// is the minimum grouping which will maintain the required 16-byte stack
+// alignment.
+
+  .text
+
+#if __riscv_xlen == 32
+
+  .globl  __riscv_restore_tailcall_12
+  .type   __riscv_restore_tailcall_12,@function
+__riscv_restore_tailcall_12:
+  .cfi_startproc
+  .cfi_def_cfa_offset 64
+  .cfi_offset s11, -64+12
+  .cfi_offset s10, -64+16
+  .cfi_offset s9, -64+20
+  .cfi_offset s8, -64+24
+  .cfi_offset s7, -64+28
+  .cfi_offset s6, -64+32
+  .cfi_offset s5, -64+36
+  .cfi_offset s4, -64+40
+  .cfi_offset s3, -64+44
+  .cfi_offset s2, -64+48
+  .cfi_offset s1, -64+52
+  .cfi_offset s0, -64+56
+  .cfi_offset ra, -64+60
+  lw      s11, 12(sp)
+  .cfi_restore s11
+  addi    sp, sp, 16
+  // fallthrough into __riscv_restore_tailcall_11/10/9/8
+
+  .globl  __riscv_restore_tailcall_11
+  .type   __riscv_restore_tailcall_11,@function
+  .globl  __riscv_restore_tailcall_10
+  .type   __riscv_restore_tailcall_10,@function
+  .globl  __riscv_restore_tailcall_9
+  .type   __riscv_restore_tailcall_9,@function
+  .globl  __riscv_restore_tailcall_8
+  .type   __riscv_restore_tailcall_8,@function
+__riscv_restore_tailcall_11:
+__riscv_restore_tailcall_10:
+__riscv_restore_tailcall_9:
+__riscv_restore_tailcall_8:
+  .cfi_restore s11
+  .cfi_def_cfa_offset 48
+  lw      s10, 0(sp)
+  .cfi_restore s10
+  lw      s9,  4(sp)
+  .cfi_restore s9
+  lw      s8,  8(sp)
+  .cfi_restore s8
+  lw      s7,  12(sp)
+  .cfi_restore s7
+  addi    sp, sp, 16
+  // fallthrough into __riscv_restore_tailcall_7/6/5/4
+
+  .globl  __riscv_restore_tailcall_7
+  .type   __riscv_restore_tailcall_7,@function
+  .globl  __riscv_restore_tailcall_6
+  .type   __riscv_restore_tailcall_6,@function
+  .globl  __riscv_restore_tailcall_5
+  .type   __riscv_restore_tailcall_5,@function
+  .globl  __riscv_restore_tailcall_4
+  .type   __riscv_restore_tailcall_4,@function
+__riscv_restore_tailcall_7:
+__riscv_restore_tailcall_6:
+__riscv_restore_tailcall_5:
+__riscv_restore_tailcall_4:
+  .cfi_restore s7
+  .cfi_restore s8
+  .cfi_restore s9
+  .cfi_restore s10
+  .cfi_restore s11
+  .cfi_def_cfa_offset 32
+  lw      s6,  0(sp)
+  .cfi_restore s6
+  lw      s5,  4(sp)
+  .cfi_restore s5
+  lw      s4,  8(sp)
+  .cfi_restore s4
+  lw      s3,  12(sp)
+  .cfi_restore s3
+  addi    sp, sp, 16
+  // fallthrough into __riscv_restore_tailcall_3/2/1/0
+
+  .globl  __riscv_restore_tailcall_3
+  .type   __riscv_restore_tailcall_3,@function
+  .globl  __riscv_restore_tailcall_2
+  .type   __riscv_restore_tailcall_2,@function
+  .globl  __riscv_restore_tailcall_1
+  .type   __riscv_restore_tailcall_1,@function
+  .globl  __riscv_restore_tailcall_0
+  .type   __riscv_restore_tailcall_0,@function
+__riscv_restore_tailcall_3:
+__riscv_restore_tailcall_2:
+__riscv_restore_tailcall_1:
+__riscv_restore_tailcall_0:
+  .cfi_restore s3
+  .cfi_restore s4
+  .cfi_restore s5
+  .cfi_restore s6
+  .cfi_restore s7
+  .cfi_restore s8
+  .cfi_restore s9
+  .cfi_restore s10
+  .cfi_restore s11
+  .cfi_def_cfa_offset 16
+  lw      s2,  0(sp)
+  .cfi_restore s2
+  lw      s1,  4(sp)
+  .cfi_restore s1
+  lw      s0,  8(sp)
+  .cfi_restore s0
+  lw      ra,  12(sp)
+  .cfi_restore ra
+  addi    sp, sp, 16
+  .cfi_def_cfa_offset 0
+  jr      t1
+  .cfi_endproc
+
+#elif __riscv_xlen == 64
+
+  .globl  __riscv_restore_tailcall_12
+  .type   __riscv_restore_tailcall_12,@function
+__riscv_restore_tailcall_12:
+  .cfi_startproc
+  .cfi_def_cfa_offset 112
+  .cfi_offset s11, -112+8
+  .cfi_offset s10, -112+16
+  .cfi_offset s9, -112+24
+  .cfi_offset s8, -112+32
+  .cfi_offset s7, -112+40
+  .cfi_offset s6, -112+48
+  .cfi_offset s5, -112+56
+  .cfi_offset s4, -112+64
+  .cfi_offset s3, -112+72
+  .cfi_offset s2, -112+80
+  .cfi_offset s1, -112+88
+  .cfi_offset s0, -112+96
+  .cfi_offset ra, -112+104
+  ld      s11, 8(sp)
+  .cfi_restore s11
+  addi    sp, sp, 16
+  // fallthrough into __riscv_restore_tailcall_11/10/9/8
+
+  .globl  __riscv_restore_tailcall_11
+  .type   __riscv_restore_tailcall_11,@function
+  .globl  __riscv_restore_tailcall_10
+  .type   __riscv_restore_tailcall_10,@function
+__riscv_restore_tailcall_11:
+__riscv_restore_tailcall_10:
+  .cfi_restore s11
+  .cfi_def_cfa_offset 96
+  ld      s10, 0(sp)
+  .cfi_restore s10
+  ld      s9,  8(sp)
+  .cfi_restore s9
+  addi    sp, sp, 16
+  // fallthrough into __riscv_restore_tailcall_9/8
+
+  .globl  __riscv_restore_tailcall_9
+  .type   __riscv_restore_tailcall_9,@function
+  .globl  __riscv_restore_tailcall_8
+  .type   __riscv_restore_tailcall_8,@function
+__riscv_restore_tailcall_9:
+__riscv_restore_tailcall_8:
+  .cfi_restore s9
+  .cfi_restore s10
+  .cfi_restore s11
+  .cfi_def_cfa_offset 80
+  ld      s8,  0(sp)
+  .cfi_restore s8
+  ld      s7,  8(sp)
+  .cfi_restore s7
+  addi    sp, sp, 16
+  // fallthrough into __riscv_restore_tailcall_7/6
+
+  .globl  __riscv_restore_tailcall_7
+  .type   __riscv_restore_tailcall_7,@function
+  .globl  __riscv_restore_tailcall_6
+  .type   __riscv_restore_tailcall_6,@function
+__riscv_restore_tailcall_7:
+__riscv_restore_tailcall_6:
+  .cfi_restore s7
+  .cfi_restore s8
+  .cfi_restore s9
+  .cfi_restore s10
+  .cfi_restore s11
+  .cfi_def_cfa_offset 64
+  ld      s6,  0(sp)
+  .cfi_restore s6
+  ld      s5,  8(sp)
+  .cfi_restore s5
+  addi    sp, sp, 16
+  // fallthrough into __riscv_restore_tailcall_5/4
+
+  .globl  __riscv_restore_tailcall_5
+  .type   __riscv_restore_tailcall_5,@function
+  .globl  __riscv_restore_tailcall_4
+  .type   __riscv_restore_tailcall_4,@function
+__riscv_restore_tailcall_5:
+__riscv_restore_tailcall_4:
+  .cfi_restore s5
+  .cfi_restore s6
+  .cfi_restore s7
+  .cfi_restore s8
+  .cfi_restore s9
+  .cfi_restore s10
+  .cfi_restore s11
+  .cfi_def_cfa_offset 48
+  ld      s4,  0(sp)
+  .cfi_restore s4
+  ld      s3,  8(sp)
+  .cfi_restore s3
+  addi    sp, sp, 16
+  // fallthrough into __riscv_restore_tailcall_3/2
+
+  .globl  __riscv_restore_tailcall_3
+  .type   __riscv_restore_tailcall_3,@function
+  .globl  __riscv_restore_tailcall_2
+  .type   __riscv_restore_tailcall_2,@function
+  .globl  __riscv_restore_tailcall_1
+  .type   __riscv_restore_tailcall_1,@function
+  .globl  __riscv_restore_tailcall_0
+  .type   __riscv_restore_tailcall_0,@function
+__riscv_restore_tailcall_3:
+__riscv_restore_tailcall_2:
+  .cfi_restore s3
+  .cfi_restore s4
+  .cfi_restore s5
+  .cfi_restore s6
+  .cfi_restore s7
+  .cfi_restore s8
+  .cfi_restore s9
+  .cfi_restore s10
+  .cfi_restore s11
+  .cfi_def_cfa_offset 32
+  ld      s2,  0(sp)
+  .cfi_restore s2
+  ld      s1,  8(sp)
+  .cfi_restore s1
+  addi    sp, sp, 16
+  // fallthrough into __riscv_restore_tailcall_1/0
+
+__riscv_restore_tailcall_1:
+__riscv_restore_tailcall_0:
+  .cfi_restore s1
+  .cfi_restore s2
+  .cfi_restore s3
+  .cfi_restore s4
+  .cfi_restore s5
+  .cfi_restore s6
+  .cfi_restore s7
+  .cfi_restore s8
+  .cfi_restore s9
+  .cfi_restore s10
+  .cfi_restore s11
+  .cfi_def_cfa_offset 16
+  ld      s0,  0(sp)
+  .cfi_restore s0
+  ld      ra,  8(sp)
+  .cfi_restore ra
+  addi    sp, sp, 16
+  .cfi_def_cfa_offset 0
+  jr      t1
+  .cfi_endproc
+
+#else
+# error "xlen must be 32 or 64 for save-restore implementation
+#endif