[SystemZ] Take better care when computing needed vector registers in TTI.

A new function getNumVectorRegs() is better to use for the number of needed
vector registers instead of getNumberOfParts(). This is to make sure that the
number of vector registers (and typically operations) required for a vector
type is accurate.

getNumberOfParts() which was previously used works by splitting the vector
type until it is legal gives incorrect results for types with a non
power of two number of elements (rare).

A new static function getScalarSizeInBits() that also checks for a pointer
type and returns 64U for it since otherwise it gets a value of 0). Used in a
few places where Ty may be pointer.

Review: Ulrich Weigand

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344115 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 21 insertions, 0 deletions

diff --git a/test/Analysis/CostModel/SystemZ/load-ptr-cmp-fold.ll b/test/Analysis/CostModel/SystemZ/load-ptr-cmp-fold.ll
new file mode 100644
index 00000000000..68227973d06
--- /dev/null
+++ b/test/Analysis/CostModel/SystemZ/load-ptr-cmp-fold.ll
@@ -0,0 +1,11 @@
+; RUN: opt < %s -cost-model -analyze -mtriple=systemz-unknown -mcpu=z13 | FileCheck %s
+
+; Test that the cost heuristic for a folded load works also for a pointer operand.
+define void @fun0(i64* %lhs, i64** %rhs_ptr) {
+  %rhs = load i64*, i64** %rhs_ptr
+  %c = icmp eq i64* %lhs, %rhs
+  ret void
+; CHECK: Printing analysis 'Cost Model Analysis' for function 'fun0':
+; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   %rhs = load
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %c = icmp
+}
diff --git a/test/Analysis/CostModel/SystemZ/numvectorregs.ll b/test/Analysis/CostModel/SystemZ/numvectorregs.ll
new file mode 100644
index 00000000000..35b58e33704
--- /dev/null
+++ b/test/Analysis/CostModel/SystemZ/numvectorregs.ll
@@ -0,0 +1,10 @@
+; RUN: opt < %s -cost-model -analyze -mtriple=systemz-unknown -mcpu=z13 | FileCheck %s
+
+; Test that the cost for the number of vector registers is returned for a
+; non-power-of-two vector type.
+define <6 x double> @fun0(<6 x double> %lhs, <6 x double> %rhs) {
+  %a = fadd <6 x double> %lhs, %rhs
+  ret <6 x double> %a
+; CHECK: Printing analysis 'Cost Model Analysis' for function 'fun0':
+; CHECK: Cost Model: Found an estimated cost of 3 for instruction:   %a = fadd <6 x double>
+}