summaryrefslogtreecommitdiffstats
path: root/contrib/llvm/lib/Target/SystemZ/SystemZInstrFP.td
diff options
context:
space:
mode:
Diffstat (limited to 'contrib/llvm/lib/Target/SystemZ/SystemZInstrFP.td')
-rw-r--r--contrib/llvm/lib/Target/SystemZ/SystemZInstrFP.td55
1 files changed, 44 insertions, 11 deletions
diff --git a/contrib/llvm/lib/Target/SystemZ/SystemZInstrFP.td b/contrib/llvm/lib/Target/SystemZ/SystemZInstrFP.td
index 27fbd7d..0cb2672 100644
--- a/contrib/llvm/lib/Target/SystemZ/SystemZInstrFP.td
+++ b/contrib/llvm/lib/Target/SystemZ/SystemZInstrFP.td
@@ -46,15 +46,28 @@ let Defs = [CC], CCValues = 0xF, CompareZeroCCMask = 0xF in {
defm LTDBR : LoadAndTestRRE<"ltdb", 0xB312, FP64>;
defm LTXBR : LoadAndTestRRE<"ltxb", 0xB342, FP128>;
}
-// Note that the comparison against zero operation is not available if we
-// have vector support, since load-and-test instructions will partially
-// clobber the target (vector) register.
+// Note that LTxBRCompare is not available if we have vector support,
+// since load-and-test instructions will partially clobber the target
+// (vector) register.
let Predicates = [FeatureNoVector] in {
defm : CompareZeroFP<LTEBRCompare, FP32>;
defm : CompareZeroFP<LTDBRCompare, FP64>;
defm : CompareZeroFP<LTXBRCompare, FP128>;
}
+// Use a normal load-and-test for compare against zero in case of
+// vector support (via a pseudo to simplify instruction selection).
+let Defs = [CC], usesCustomInserter = 1 in {
+ def LTEBRCompare_VecPseudo : Pseudo<(outs), (ins FP32:$R1, FP32:$R2), []>;
+ def LTDBRCompare_VecPseudo : Pseudo<(outs), (ins FP64:$R1, FP64:$R2), []>;
+ def LTXBRCompare_VecPseudo : Pseudo<(outs), (ins FP128:$R1, FP128:$R2), []>;
+}
+let Predicates = [FeatureVector] in {
+ defm : CompareZeroFP<LTEBRCompare_VecPseudo, FP32>;
+ defm : CompareZeroFP<LTDBRCompare_VecPseudo, FP64>;
+ defm : CompareZeroFP<LTXBRCompare_VecPseudo, FP128>;
+}
+
// Moves between 64-bit integer and floating-point registers.
def LGDR : UnaryRRE<"lgd", 0xB3CD, bitconvert, GR64, FP64>;
def LDGR : UnaryRRE<"ldg", 0xB3C1, bitconvert, FP64, GR64>;
@@ -238,26 +251,46 @@ let Predicates = [FeatureFPExtension] in {
// Unary arithmetic
//===----------------------------------------------------------------------===//
+// We prefer generic instructions during isel, because they do not
+// clobber CC and therefore give the scheduler more freedom. In cases
+// the CC is actually useful, the SystemZElimCompare pass will try to
+// convert generic instructions into opcodes that also set CC. Note
+// that lcdf / lpdf / lndf only affect the sign bit, and can therefore
+// be used with fp32 as well. This could be done for fp128, in which
+// case the operands would have to be tied.
+
// Negation (Load Complement).
let Defs = [CC], CCValues = 0xF, CompareZeroCCMask = 0xF in {
- def LCEBR : UnaryRRE<"lceb", 0xB303, fneg, FP32, FP32>;
- def LCDBR : UnaryRRE<"lcdb", 0xB313, fneg, FP64, FP64>;
+ def LCEBR : UnaryRRE<"lceb", 0xB303, null_frag, FP32, FP32>;
+ def LCDBR : UnaryRRE<"lcdb", 0xB313, null_frag, FP64, FP64>;
def LCXBR : UnaryRRE<"lcxb", 0xB343, fneg, FP128, FP128>;
}
+// Generic form, which does not set CC.
+def LCDFR : UnaryRRE<"lcdf", 0xB373, fneg, FP64, FP64>;
+let isCodeGenOnly = 1 in
+ def LCDFR_32 : UnaryRRE<"lcdf", 0xB373, fneg, FP32, FP32>;
// Absolute value (Load Positive).
let Defs = [CC], CCValues = 0xF, CompareZeroCCMask = 0xF in {
- def LPEBR : UnaryRRE<"lpeb", 0xB300, fabs, FP32, FP32>;
- def LPDBR : UnaryRRE<"lpdb", 0xB310, fabs, FP64, FP64>;
+ def LPEBR : UnaryRRE<"lpeb", 0xB300, null_frag, FP32, FP32>;
+ def LPDBR : UnaryRRE<"lpdb", 0xB310, null_frag, FP64, FP64>;
def LPXBR : UnaryRRE<"lpxb", 0xB340, fabs, FP128, FP128>;
}
+// Generic form, which does not set CC.
+def LPDFR : UnaryRRE<"lpdf", 0xB370, fabs, FP64, FP64>;
+let isCodeGenOnly = 1 in
+ def LPDFR_32 : UnaryRRE<"lpdf", 0xB370, fabs, FP32, FP32>;
// Negative absolute value (Load Negative).
let Defs = [CC], CCValues = 0xF, CompareZeroCCMask = 0xF in {
- def LNEBR : UnaryRRE<"lneb", 0xB301, fnabs, FP32, FP32>;
- def LNDBR : UnaryRRE<"lndb", 0xB311, fnabs, FP64, FP64>;
+ def LNEBR : UnaryRRE<"lneb", 0xB301, null_frag, FP32, FP32>;
+ def LNDBR : UnaryRRE<"lndb", 0xB311, null_frag, FP64, FP64>;
def LNXBR : UnaryRRE<"lnxb", 0xB341, fnabs, FP128, FP128>;
}
+// Generic form, which does not set CC.
+def LNDFR : UnaryRRE<"lndf", 0xB371, fnabs, FP64, FP64>;
+let isCodeGenOnly = 1 in
+ def LNDFR_32 : UnaryRRE<"lndf", 0xB371, fnabs, FP32, FP32>;
// Square root.
def SQEBR : UnaryRRE<"sqeb", 0xB314, fsqrt, FP32, FP32>;
@@ -414,6 +447,6 @@ let Defs = [CC], CCValues = 0xF in {
// Peepholes
//===----------------------------------------------------------------------===//
-def : Pat<(f32 fpimmneg0), (LCEBR (LZER))>;
-def : Pat<(f64 fpimmneg0), (LCDBR (LZDR))>;
+def : Pat<(f32 fpimmneg0), (LCDFR_32 (LZER))>;
+def : Pat<(f64 fpimmneg0), (LCDFR (LZDR))>;
def : Pat<(f128 fpimmneg0), (LCXBR (LZXR))>;
OpenPOWER on IntegriCloud