RFA: gimplify vector constructors that aren't initialiser constants

[Richard Sandiford <rsandifo at nildram dot co dot uk>]

This is a second attempt at fixing PR 33614.  We currently don't gimplify
the contents of TREE_CONSTANT vector constructors, even if they need
run-time evaluation.

My first attempt at fixing this bug was based on the wrong assumption
that we had two types of vector constructors: those that could be turned
into VECTOR_CSTs and those that need run-time evaluation.  Andrew pointed
out that there's a third category: vector constructors that can't be
reduced to a VECTOR_CST but that are legitimate initialiser constants
(e.g. those involving ADDR_EXPRs).  We want to continue to treat those
constructors as TREE_CONSTANT.

The patch below therefore adds an initializer_constant_valid_p check
to the code that is trying to keep TREE_CONSTANT constructors.  If this
check fails, we turn the constructor into a nonconstant value and
gimplify the contents.

Bootstrapped & regression-tested on x86_64-linux-gnu.  Also regression
tested on mipsisa32-elf and mipsisa64-elf.  It fixes an ICE in
compile/20050113-1.c for mipsisa32-elf and an ICE the attached
testcase for x86_64-linux-gnu.  OK to install?

Richard


gcc/
	PR tree-optimization/33614
	* gimplify.c (gimplify_init_constructor): Gimplify vector constructors
	if they can't be reduced to VECTOR_CSTs and aren't legitimate
	initializer constants.

gcc/testsuite/
	PR tree-optimization/33614
	* gcc.c-torture/compile/pr33614.c: New test.
Index: gcc/gimplify.c
===================================================================
--- gcc/gimplify.c	2007-10-29 21:31:03.000000000 +0000
+++ gcc/gimplify.c	2007-10-29 21:31:07.000000000 +0000
@@ -3290,8 +3290,9 @@ gimplify_init_constructor (tree *expr_p,
 	    tree value;
 
 	    /* Even when ctor is constant, it might contain non-*_CST
-	      elements (e.g. { 1.0/0.0 - 1.0/0.0, 0.0 }) and those don't
-	      belong into VECTOR_CST nodes.  */
+	       elements, such as addresses or trapping values like
+	       1.0/0.0 - 1.0/0.0.  Such expressions don't belong
+	       in VECTOR_CST nodes.  */
 	    FOR_EACH_CONSTRUCTOR_VALUE (elts, ix, value)
 	      if (!CONSTANT_CLASS_P (value))
 		{
@@ -3305,10 +3306,14 @@ gimplify_init_constructor (tree *expr_p,
 		break;
 	      }
 
-	    /* Don't reduce a TREE_CONSTANT vector ctor even if we can't
+	    /* Don't reduce an initializer constant even if we can't
 	       make a VECTOR_CST.  It won't do anything for us, and it'll
 	       prevent us from representing it as a single constant.  */
-	    break;
+	    if (initializer_constant_valid_p (ctor, type))
+	      break;
+
+	    TREE_CONSTANT (ctor) = 0;
+	    TREE_INVARIANT (ctor) = 0;
 	  }
 
 	/* Vector types use CONSTRUCTOR all the way through gimple
Index: gcc/testsuite/gcc.c-torture/compile/pr33614.c
===================================================================
--- /dev/null	2007-10-29 07:32:19.552097000 +0000
+++ gcc/testsuite/gcc.c-torture/compile/pr33614.c	2007-10-29 21:31:09.000000000 +0000
@@ -0,0 +1,9 @@
+typedef float V2SF __attribute__ ((vector_size (8)));
+
+V2SF
+foo (int x, V2SF a)
+{
+  while (x--)
+    a += (V2SF) {1.0f/0.0f - 1.0f/0.0f, 1.0f/0.0f - 1.0f/0.0f};
+  return a;
+}