This patch fixes the Altivec addend construction for the fused multiply-add

instruction (vmaddfp) to conform with IEEE to ensure the sign of a zero
result when resulting product is -0.0.

The -0.0 vector addend to vmaddfp is generated by a creating a vector
with full bits sets and then shifting each elements by 31-bits to the
left, resulting in a vector of 0x80000000 (or -0.0 as float).

The 'buildvec_canonicalize.ll' was adjusted to reflect this change and
the 'vec_mul.ll' was complemented with the float vector multiplication
test.

llvm-svn: 168998

llvm/lib/Target/PowerPC/PPCInstrAltivec.td

@@ -158,10 +158,6 @@ def vecspltisw : PatLeaf<(build_vector), [{
   return PPC::get_VSPLTI_elt(N, 4, *CurDAG).getNode() != 0;
 }], VSPLTISW_get_imm>;
 
-def V_immneg0 : PatLeaf<(build_vector), [{
-  return PPC::isAllNegativeZeroVector(N);
-}]>;
-
 //===----------------------------------------------------------------------===//
 // Helpers for defining instructions that directly correspond to intrinsics.
 
@@ -585,7 +581,12 @@ def VCMPGTUWo : VCMPo<646, "vcmpgtuw. $vD, $vA, $vB", v4i32>;
 def V_SET0 : VXForm_setzero<1220, (outs VRRC:$vD), (ins),
                       "vxor $vD, $vD, $vD", VecFP,
                       [(set VRRC:$vD, (v4i32 immAllZerosV))]>;
+let IMM=-1 in {
+def V_SETALLONES : VXForm_3<908, (outs VRRC:$vD), (ins),
+                      "vspltisw $vD, -1", VecFP,
+                      [(set VRRC:$vD, (v4i32 immAllOnesV))]>;
 }
+} // VALU Operations.
 
 //===----------------------------------------------------------------------===//
 // Additional Altivec Patterns
@@ -672,7 +673,8 @@ def : Pat<(v4i32 (and VRRC:$A, (vnot_ppc VRRC:$B))),
           (VANDC VRRC:$A, VRRC:$B)>;
 
 def : Pat<(fmul VRRC:$vA, VRRC:$vB),
-          (VMADDFP VRRC:$vA, VRRC:$vB, (v4i32 (V_SET0)))>; 
+          (VMADDFP VRRC:$vA, VRRC:$vB,
+             (v4i32 (VSLW (V_SETALLONES), (V_SETALLONES))))>; 
 
 // Fused multiply add and multiply sub for packed float.  These are represented
 // separately from the real instructions above, for operations that must have

llvm/test/CodeGen/PowerPC/buildvec_canonicalize.ll

@@ -1,10 +1,4 @@
-; There should be exactly one vxor here.
-; RUN: llc < %s -march=ppc32 -mcpu=g5 --enable-unsafe-fp-math | \
-; RUN:   grep vxor | count 1
-
-; There should be exactly one vsplti here.
-; RUN: llc < %s -march=ppc32 -mcpu=g5 --enable-unsafe-fp-math | \
-; RUN:   grep vsplti | count 1
+; RUN: llc < %s -mattr=+altivec --enable-unsafe-fp-math | FileCheck %s
 
 define void @VXOR(<4 x float>* %P1, <4 x i32>* %P2, <4 x float>* %P3) {
         %tmp = load <4 x float>* %P3            ; <<4 x float>> [#uses=1]
@@ -15,10 +9,16 @@ define void @VXOR(<4 x float>* %P1, <4 x i32>* %P2, <4 x float>* %P3) {
         store <4 x i32> zeroinitializer, <4 x i32>* %P2
         ret void
 }
+; The fmul will spill a vspltisw to create a -0.0 vector used as the addend
+; to vmaddfp (so it would IEEE compliant with zero sign propagation).
+; CHECK: @VXOR
+; CHECK: vsplti
+; CHECK: vxor
 
 define void @VSPLTI(<4 x i32>* %P2, <8 x i16>* %P3) {
         store <4 x i32> bitcast (<16 x i8> < i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1 > to <4 x i32>), <4 x i32>* %P2
         store <8 x i16> < i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1 >, <8 x i16>* %P3
         ret void
 }
-
+; CHECK: @VSPLTI
+; CHECK: vsplti

llvm/test/CodeGen/PowerPC/vec_mul.ll

@@ -1,5 +1,4 @@
-; RUN: llc < %s -march=ppc32 -mcpu=g5 | not grep mullw
-; RUN: llc < %s -march=ppc32 -mcpu=g5 | grep vmsumuhm
+; RUN: llc < %s -mattr=+altivec | FileCheck %s
 
 define <4 x i32> @test_v4i32(<4 x i32>* %X, <4 x i32>* %Y) {
 	%tmp = load <4 x i32>* %X		; <<4 x i32>> [#uses=1]
@@ -7,6 +6,9 @@ define <4 x i32> @test_v4i32(<4 x i32>* %X, <4 x i32>* %Y) {
 	%tmp3 = mul <4 x i32> %tmp, %tmp2		; <<4 x i32>> [#uses=1]
 	ret <4 x i32> %tmp3
 }
+; CHECK: test_v4i32:
+; CHECK: vmsumuhm
+; CHECK-NOT: mullw
 
 define <8 x i16> @test_v8i16(<8 x i16>* %X, <8 x i16>* %Y) {
 	%tmp = load <8 x i16>* %X		; <<8 x i16>> [#uses=1]
@@ -14,6 +16,9 @@ define <8 x i16> @test_v8i16(<8 x i16>* %X, <8 x i16>* %Y) {
 	%tmp3 = mul <8 x i16> %tmp, %tmp2		; <<8 x i16>> [#uses=1]
 	ret <8 x i16> %tmp3
 }
+; CHECK: test_v8i16:
+; CHECK: vmladduhm
+; CHECK-NOT: mullw
 
 define <16 x i8> @test_v16i8(<16 x i8>* %X, <16 x i8>* %Y) {
 	%tmp = load <16 x i8>* %X		; <<16 x i8>> [#uses=1]
@@ -21,3 +26,21 @@ define <16 x i8> @test_v16i8(<16 x i8>* %X, <16 x i8>* %Y) {
 	%tmp3 = mul <16 x i8> %tmp, %tmp2		; <<16 x i8>> [#uses=1]
 	ret <16 x i8> %tmp3
 }
+; CHECK: test_v16i8:
+; CHECK: vmuloub
+; CHECK: vmuleub
+; CHECK-NOT: mullw
+
+define <4 x float> @test_float(<4 x float>* %X, <4 x float>* %Y) {
+	%tmp = load <4 x float>* %X
+	%tmp2 = load <4 x float>* %Y
+	%tmp3 = fmul <4 x float> %tmp, %tmp2
+	ret <4 x float> %tmp3
+}
+; Check the creation of a negative zero float vector by creating a vector of
+; all bits set and shifting it 31 bits to left, resulting a an vector of 
+; 4 x 0x80000000 (-0.0 as float).
+; CHECK: test_float:
+; CHECK: vspltisw [[ZNEG:[0-9]+]], -1
+; CHECK: vslw     {{[0-9]+}}, [[ZNEG]], [[ZNEG]]
+; CHECK: vmaddfp