nvptx: PTX 'alloca' for '-mptx=7.3'+, '-march=sm_52'+ [PR65181]

..., and use it for '-mno-soft-stack': PTX "native" stacks.

	PR target/65181
	gcc/
	* config/nvptx/nvptx.cc (nvptx_get_drap_rtx): Handle
	'!TARGET_SOFT_STACK'.
	* config/nvptx/nvptx.md (define_c_enum "unspec"): Add
	'UNSPEC_STACKSAVE', 'UNSPEC_STACKRESTORE'.
	(define_expand "allocate_stack", define_expand "save_stack_block")
	(define_expand "save_stack_block"): Handle '!TARGET_SOFT_STACK',
	PTX 'alloca'.
	(define_insn "@nvptx_alloca_<mode>")
	(define_insn "@nvptx_stacksave_<mode>")
	(define_insn "@nvptx_stackrestore_<mode>"): New.
	* doc/invoke.texi (Nvidia PTX Options): Update '-msoft-stack',
	'-mno-soft-stack'.
	* doc/sourcebuild.texi (nvptx-specific attributes): Document
	'nvptx_runtime_alloca_ptx'.
	(Add Options): Document 'nvptx_alloca_ptx'.
	gcc/testsuite/
	* gcc.target/nvptx/alloca-1.c: Evolve into...
	* gcc.target/nvptx/alloca-1-O0.c: ... this, ...
	* gcc.target/nvptx/alloca-1-O1.c: ... this, and...
	* gcc.target/nvptx/alloca-1-sm_30.c: ... this.
	* gcc.target/nvptx/vla-1.c: Evolve into...
	* gcc.target/nvptx/vla-1-O0.c: ... this, ...
	* gcc.target/nvptx/vla-1-O1.c: ... this, and...
	* gcc.target/nvptx/vla-1-sm_30.c: ... this.
	* gcc.c-torture/execute/pr36321.c: Adjust.
	* gcc.target/nvptx/__builtin_alloca_0-1-O0.c: Likewise.
	* gcc.target/nvptx/__builtin_alloca_0-1-O1.c: Likewise.
	* gcc.target/nvptx/__builtin_stack_save___builtin_stack_restore-1.c:
	Likewise.
	* gcc.target/nvptx/softstack.c: Likewise.
	* gcc.target/nvptx/__builtin_stack_save___builtin_stack_restore-1-sm_30.c:
	New.
	* gcc.target/nvptx/alloca-2-O0.c: Likewise.
	* gcc.target/nvptx/alloca-3-O1.c: Likewise.
	* gcc.target/nvptx/alloca-4-O3.c: Likewise.
	* gcc.target/nvptx/alloca-5.c: Likewise.
	* lib/target-supports.exp (check_effective_target_alloca): Adjust.
	(check_nvptx_default_ptx_isa_target_architecture_at_least)
	(check_nvptx_runtime_ptx_isa_target_architecture_at_least)
	(check_effective_target_nvptx_runtime_alloca_ptx)
	(add_options_for_nvptx_alloca_ptx): New.
	libgomp/
	* fortran.c (omp_get_device_from_uid_): Adjust.
	* testsuite/libgomp.oacc-fortran/privatized-ref-2.f90: Likewise.

2 files changed, 78 insertions, 18 deletions

diff --git a/gcc/config/nvptx/nvptx.cc b/gcc/config/nvptx/nvptx.cc
index 5860b3d..060f453 100644
--- a/gcc/config/nvptx/nvptx.cc
+++ b/gcc/config/nvptx/nvptx.cc
@@ -245,7 +245,7 @@ default_ptx_version_option (void)
      warp convergence.  */
   res = MAX (res, PTX_VERSION_6_0);
 
-  /* For sm_52+, pick at least 7.3.  */
+  /* For sm_52+, pick at least 7.3, to enable PTX 'alloca'.  */
   if (ptx_isa_option >= PTX_ISA_SM52)
     res = MAX (res, PTX_VERSION_7_3);
 
@@ -1797,7 +1797,7 @@ nvptx_function_ok_for_sibcall (tree, tree)
 static rtx
 nvptx_get_drap_rtx (void)
 {
-  if (TARGET_SOFT_STACK && stack_realign_drap)
+  if (stack_realign_drap)
     return arg_pointer_rtx;
   return NULL_RTX;
 }
diff --git a/gcc/config/nvptx/nvptx.md b/gcc/config/nvptx/nvptx.md
index b300f2e..a22a088 100644
--- a/gcc/config/nvptx/nvptx.md
+++ b/gcc/config/nvptx/nvptx.md
@@ -35,8 +35,9 @@
    UNSPEC_FPINT_NEARBYINT
 
    UNSPEC_ALLOCA
-
    UNSPEC_SET_SOFTSTACK
+   UNSPEC_STACKSAVE
+   UNSPEC_STACKRESTORE
 
    UNSPEC_DIM_SIZE
 
@@ -1663,22 +1664,47 @@
    (match_operand 1 "nvptx_register_operand")]
   ""
 {
-  if (TARGET_SOFT_STACK)
+  if (!TARGET_SOFT_STACK
+      && TARGET_PTX_7_3
+      && TARGET_SM52)
+    emit_insn (gen_nvptx_alloca (Pmode, operands[0], operands[1]));
+  else if (!TARGET_SOFT_STACK)
+    {
+      sorry ("target cannot support alloca");
+      emit_insn (gen_nop ());
+    }
+  else if (TARGET_SOFT_STACK)
     {
       emit_move_insn (stack_pointer_rtx,
 		      gen_rtx_MINUS (Pmode, stack_pointer_rtx, operands[1]));
       emit_insn (gen_set_softstack (Pmode, stack_pointer_rtx));
       emit_move_insn (operands[0], virtual_stack_dynamic_rtx);
-      DONE;
     }
-  /* The ptx documentation specifies an alloca intrinsic (for 32 bit
-     only)  but notes it is not implemented.  The assembler emits a
-     confused error message.  Issue a blunt one now instead.  */
-  sorry ("target cannot support alloca");
-  emit_insn (gen_nop ());
+  else
+    gcc_unreachable ();
   DONE;
 })
 
+(define_insn "@nvptx_alloca_<mode>"
+  [(set (match_operand:P 0 "nvptx_register_operand" "=R")
+        (unspec:P [(match_operand:P 1 "nvptx_nonmemory_operand" "Ri")]
+		  UNSPEC_ALLOCA))]
+  "TARGET_PTX_7_3
+   && TARGET_SM52"
+  {
+    /* Convert the address from '.local' state space to generic.  That way,
+       we don't have to use 'st.local', 'ld.local', and can easily pass the
+       address to other "generic functions".
+       TODO 'gcc.target/nvptx/alloca-5.c' */
+    output_asm_insn ("{", NULL);
+    output_asm_insn ("\\t.reg%t0\\t%0_local;", operands);
+    output_asm_insn ("\\talloca%u0\\t%0_local, %1;", operands);
+    output_asm_insn ("\\tcvta.local%u0\\t%0, %0_local;", operands);
+    output_asm_insn ("}", NULL);
+    return "";
+  }
+  [(set_attr "predicable" "no")])
+
 (define_insn "@set_softstack_<mode>"
   [(unspec [(match_operand:P 0 "nvptx_register_operand" "R")]
 	   UNSPEC_SET_SOFTSTACK)]
@@ -1692,30 +1718,64 @@
    (match_operand 1 "register_operand" "")]
   "!TARGET_SOFT_STACK"
 {
-  /* The concept of a '%stack' pointer doesn't apply like this for
-     PTX "native" stacks.  GCC however occasionally synthesizes
-     '__builtin_stack_save ()', '__builtin_stack_restore ()', and isn't able to
-     optimize them all away.  Just submit a dummy -- user code shouldn't be
-     able to observe this.  */
-  emit_move_insn (operands[0], GEN_INT (0xdeadbeef));
+  if (TARGET_PTX_7_3
+      && TARGET_SM52)
+    {
+      gcc_checking_assert (REG_P (operands[0]));
+      emit_insn (gen_nvptx_stacksave (Pmode, operands[0], operands[1]));
+    }
+  else
+    {
+      /* The concept of a '%stack' pointer doesn't apply like this.
+         GCC however occasionally synthesizes '__builtin_stack_save ()',
+	 '__builtin_stack_restore ()', and isn't able to optimize them all
+	 away.  Just submit a dummy -- user code shouldn't be able to observe
+	 this.  */
+      emit_move_insn (operands[0], GEN_INT (0xdeadbeef));
+    }
   DONE;
 })
 
+(define_insn "@nvptx_stacksave_<mode>"
+  [(set (match_operand:P 0 "nvptx_register_operand" "=R")
+        (unspec:P [(match_operand:P 1 "register_operand" "R")]
+	 UNSPEC_STACKSAVE))]
+  "TARGET_PTX_7_3
+   && TARGET_SM52"
+  "%.\\tstacksave%u0\\t%0;")
+
 (define_expand "restore_stack_block"
   [(match_operand 0 "register_operand" "")
    (match_operand 1 "register_operand" "")]
   ""
 {
-  if (!TARGET_SOFT_STACK)
+  if (!TARGET_SOFT_STACK
+      && TARGET_PTX_7_3
+      && TARGET_SM52)
+    {
+      operands[1] = force_reg (Pmode, operands[1]);
+      emit_insn (gen_nvptx_stackrestore (Pmode, operands[0], operands[1]));
+    }
+  else if (!TARGET_SOFT_STACK)
     ; /* See 'save_stack_block'.  */
-  else
+  else if (TARGET_SOFT_STACK)
     {
       emit_move_insn (operands[0], operands[1]);
       emit_insn (gen_set_softstack (Pmode, operands[0]));
     }
+  else
+    gcc_unreachable ();
   DONE;
 })
 
+(define_insn "@nvptx_stackrestore_<mode>"
+  [(set (match_operand:P 0 "nvptx_register_operand" "=R")
+        (unspec:P [(match_operand:P 1 "nvptx_register_operand" "R")]
+         UNSPEC_STACKRESTORE))]
+  "TARGET_PTX_7_3
+   && TARGET_SM52"
+  "%.\\tstackrestore%u1\\t%1;")
+
 (define_expand "save_stack_function"
   [(match_operand 0 "register_operand" "")
    (match_operand 1 "register_operand" "")]