[PowerPC] dyn_cast should be dyn_cast_or_null in MASSV pass

[masoud.ataei]

Summary

It is possible that we have different constants in different slots of a vector double (float).
So, in this case Exp->getSplatValue() will return nullptr.
Here, I am proposing to handle it properly.

[steven.zhang]

llvm/lib/Target/PowerPC/PPCLowerMASSVEntries.cpp

108

It seems that some of the test change is not relative with this.(i.e. rename the func name) And the only test change I can imagine is that, some case that we hit ice before now it is passed. That is, if Exp->getSplatValue() return null, dyn_cast will hit ice before. But I didn't see such kind of test.

[masoud.ataei]

llvm/lib/Target/PowerPC/PPCLowerMASSVEntries.cpp

108

I agree the rename the func name in the test is irrelevant to this change but the one with

@__powd2_massv(<2 x double> %wide.load, <2 x double> <double 7.700000e-01, double 7.600000e-01>)

and similar ones that tests the two slots of the second vector double has non-equal constants are relevant. For example in this test Exp->getSplatValue() returns nullptr which is handled here.

llvm/test/CodeGen/PowerPC/pow_massv_075_025exp.ll

74

Two slots of the second vector double in __powd2_massv arguments has non-equal constants.

100

Two slots of the second vector double in __powd2_massv arguments has non-equal constants.

126

Two slots of the second vector double in __powd2_massv arguments has non-equal constants.

llvm/test/CodeGen/PowerPC/powf_massv_075_025exp.ll

74

Four slots of the second vector float in __powf4_massv arguments has non-equal constants.

100

Four slots of the second vector float in __powf4_massv arguments has non-equal constants.

126

Four slots of the second vector float in __powf4_massv arguments has non-equal constants.

[steven.zhang]

LGTM as I indeed see the crash fixed.

llvm/test/CodeGen/PowerPC/powf_massv_075_025exp.ll

74

Usually, we will add some comments in the test to indicate that no crash any more.

llvm/lib/Target/PowerPC/PPCLowerMASSVEntries.cpp

 /// If there are proper fast-math flags, this function creates llvm.pow
 /// intrinsics when the exponent is 0.25 or 0.75.
 bool PPCLowerMASSVEntries::handlePowSpecialCases(CallInst *CI, Function &Func,
                                                  Module &M) {
   if (Func.getName() != "__powf4_massv" && Func.getName() != "__powd2_massv")
     return false;
 
   if (Constant *Exp = dyn_cast<Constant>(CI->getArgOperand(1)))
-    if (ConstantFP *CFP = dyn_cast<ConstantFP>(Exp->getSplatValue())) {
+    if (ConstantFP *CFP = dyn_cast_or_null<ConstantFP>(Exp->getSplatValue())) {
       // If the argument is 0.75 or 0.25 it is cheaper to turn it into pow
       // intrinsic so that it could be optimzed as sequence of sqrt's.
       if (!CI->hasNoInfs() || !CI->hasApproxFunc())
         return false;
 
       if (!CFP->isExactlyValue(0.75) && !CFP->isExactlyValue(0.25))
         return false;
 

llvm/test/CodeGen/PowerPC/pow_massv_075_025exp.ll

 ; RUN: llc -vector-library=MASSV < %s -mtriple=powerpc64le-unknown-unknown -mcpu=pwr9 | FileCheck -check-prefixes=CHECK-PWR9 %s
 ; RUN: llc -vector-library=MASSV < %s -mtriple=powerpc64le-unknown-unknown -mcpu=pwr8 | FileCheck -check-prefixes=CHECK-PWR8 %s
 
 ; Exponent is a variable
-define void @my_vpow_var(double* nocapture %z, double* nocapture readonly %y, double* nocapture readonly %x) {
-; CHECK-LABEL:       @vspow_var
+define void @vpow_var(double* nocapture %z, double* nocapture readonly %y, double* nocapture readonly %x) {
+; CHECK-LABEL:       @vpow_var
 ; CHECK-PWR9:        bl __powd2_P9
 ; CHECK-PWR8:        bl __powd2_P8
 ; CHECK:             blr
 entry:
   br label %vector.body
 
 vector.body:
   %index = phi i64 [ %index.next, %vector.body ], [ 0, %entry ]
vector.body:
   %4 = icmp eq i64 %index.next, 1024
   br i1 %4, label %for.end, label %vector.body
 
 for.end:
   ret void
 }
 
 ; Exponent is a constant != 0.75 and !=0.25
-define void @my_vpow_const(double* nocapture %y, double* nocapture readonly %x) {
-; CHECK-LABEL:       @vspow_const
+define void @vpow_const(double* nocapture %y, double* nocapture readonly %x) {
+; CHECK-LABEL:       @vpow_const
 ; CHECK-PWR9:        bl __powd2_P9
 ; CHECK-PWR8:        bl __powd2_P8
 ; CHECK:             blr
 entry:
   br label %vector.body
 
 vector.body:
   %index = phi i64 [ %index.next, %vector.body ], [ 0, %entry ]
   %next.gep = getelementptr double, double* %y, i64 %index
   %next.gep19 = getelementptr double, double* %x, i64 %index
   %0 = bitcast double* %next.gep19 to <2 x double>*
   %wide.load = load <2 x double>, <2 x double>* %0, align 8
   %1 = call ninf afn nsz <2 x double> @__powd2_massv(<2 x double> %wide.load, <2 x double> <double 7.600000e-01, double 7.600000e-01>)
   %2 = bitcast double* %next.gep to <2 x double>*
   store <2 x double> %1, <2 x double>* %2, align 8
   %index.next = add i64 %index, 2
   %3 = icmp eq i64 %index.next, 1024
   br i1 %3, label %for.end, label %vector.body
 
 for.end:
   ret void
 }
 
+; Exponent is a constant != 0.75 and !=0.25 and they are different 
+define void @vpow_noeq_const(double* nocapture %y, double* nocapture readonly %x) {
+; CHECK-LABEL:       @vpow_noeq_const
+; CHECK-PWR9:        bl __powd2_P9
+; CHECK-PWR8:        bl __powd2_P8
+; CHECK:             blr
+entry:
+  br label %vector.body
+
+vector.body:
+  %index = phi i64 [ %index.next, %vector.body ], [ 0, %entry ]
+  %next.gep = getelementptr double, double* %y, i64 %index
+  %next.gep19 = getelementptr double, double* %x, i64 %index
+  %0 = bitcast double* %next.gep19 to <2 x double>*
+  %wide.load = load <2 x double>, <2 x double>* %0, align 8
+  %1 = call ninf afn nsz <2 x double> @__powd2_massv(<2 x double> %wide.load, <2 x double> <double 7.700000e-01, double 7.600000e-01>)
+  %2 = bitcast double* %next.gep to <2 x double>*
+  store <2 x double> %1, <2 x double>* %2, align 8
+  %index.next = add i64 %index, 2
+  %3 = icmp eq i64 %index.next, 1024
+  br i1 %3, label %for.end, label %vector.body
+
+for.end:
+  ret void
+}
+
+; Exponent is a constant != 0.75 and !=0.25 and they are different 
+define void @vpow_noeq075_const(double* nocapture %y, double* nocapture readonly %x) {
+; CHECK-LABEL:       @vpow_noeq075_const
+; CHECK-PWR9:        bl __powd2_P9
+; CHECK-PWR8:        bl __powd2_P8
+; CHECK:             blr
+entry:
+  br label %vector.body
+
+vector.body:
+  %index = phi i64 [ %index.next, %vector.body ], [ 0, %entry ]
+  %next.gep = getelementptr double, double* %y, i64 %index
+  %next.gep19 = getelementptr double, double* %x, i64 %index
+  %0 = bitcast double* %next.gep19 to <2 x double>*
+  %wide.load = load <2 x double>, <2 x double>* %0, align 8
+  %1 = call ninf afn nsz <2 x double> @__powd2_massv(<2 x double> %wide.load, <2 x double> <double 7.700000e-01, double 7.500000e-01>)
+  %2 = bitcast double* %next.gep to <2 x double>*
+  store <2 x double> %1, <2 x double>* %2, align 8
+  %index.next = add i64 %index, 2
+  %3 = icmp eq i64 %index.next, 1024
+  br i1 %3, label %for.end, label %vector.body
+
+for.end:
+  ret void
+}
+
+; Exponent is a constant != 0.75 and !=0.25 and they are different 
+define void @vpow_noeq025_const(double* nocapture %y, double* nocapture readonly %x) {
+; CHECK-LABEL:       @vpow_noeq025_const
+; CHECK-PWR9:        bl __powd2_P9
+; CHECK-PWR8:        bl __powd2_P8
+; CHECK:             blr
+entry:
+  br label %vector.body
+
+vector.body:
+  %index = phi i64 [ %index.next, %vector.body ], [ 0, %entry ]
+  %next.gep = getelementptr double, double* %y, i64 %index
+  %next.gep19 = getelementptr double, double* %x, i64 %index
+  %0 = bitcast double* %next.gep19 to <2 x double>*
+  %wide.load = load <2 x double>, <2 x double>* %0, align 8
+  %1 = call ninf afn nsz <2 x double> @__powd2_massv(<2 x double> %wide.load, <2 x double> <double 7.700000e-01, double 2.500000e-01>)
+  %2 = bitcast double* %next.gep to <2 x double>*
+  store <2 x double> %1, <2 x double>* %2, align 8
+  %index.next = add i64 %index, 2
+  %3 = icmp eq i64 %index.next, 1024
+  br i1 %3, label %for.end, label %vector.body
+
+for.end:
+  ret void
+}
+
 ; Exponent is 0.75
-define void @my_vpow_075(double* nocapture %y, double* nocapture readonly %x) {
-; CHECK-LABEL:       @vspow_075
+define void @vpow_075(double* nocapture %y, double* nocapture readonly %x) {
+; CHECK-LABEL:       @vpow_075
 ; CHECK-NOT:         bl __powd2_P{{[8,9]}}
 ; CHECK:             xvrsqrtesp
 ; CHECK:             blr
 entry:
   br label %vector.body
 
 vector.body:
   %index = phi i64 [ %index.next, %vector.body ], [ 0, %entry ]
   %next.gep = getelementptr double, double* %y, i64 %index
   %next.gep19 = getelementptr double, double* %x, i64 %index
   %0 = bitcast double* %next.gep19 to <2 x double>*
   %wide.load = load <2 x double>, <2 x double>* %0, align 8
   %1 = call ninf afn <2 x double> @__powd2_massv(<2 x double> %wide.load, <2 x double> <double 7.500000e-01, double 7.500000e-01>)
   %2 = bitcast double* %next.gep to <2 x double>*
   store <2 x double> %1, <2 x double>* %2, align 8
   %index.next = add i64 %index, 2
   %3 = icmp eq i64 %index.next, 1024
   br i1 %3, label %for.end, label %vector.body
 
 for.end:
   ret void
 }
 
 ; Exponent is 0.25
-define void @my_vpow_025(double* nocapture %y, double* nocapture readonly %x) {
-; CHECK-LABEL:       @vspow_025
+define void @vpow_025(double* nocapture %y, double* nocapture readonly %x) {
+; CHECK-LABEL:       @vpow_025
 ; CHECK-NOT:         bl __powd2_P{{[8,9]}}
 ; CHECK:             xvrsqrtesp
 ; CHECK:             blr
 entry:
   br label %vector.body
 
 vector.body:
   %index = phi i64 [ %index.next, %vector.body ], [ 0, %entry ]
   %next.gep = getelementptr double, double* %y, i64 %index
   %next.gep19 = getelementptr double, double* %x, i64 %index
   %0 = bitcast double* %next.gep19 to <2 x double>*
   %wide.load = load <2 x double>, <2 x double>* %0, align 8
   %1 = call ninf afn nsz <2 x double> @__powd2_massv(<2 x double> %wide.load, <2 x double> <double 2.500000e-01, double 2.500000e-01>)
   %2 = bitcast double* %next.gep to <2 x double>*
   store <2 x double> %1, <2 x double>* %2, align 8
   %index.next = add i64 %index, 2
   %3 = icmp eq i64 %index.next, 1024
   br i1 %3, label %for.end, label %vector.body
 
 for.end:
   ret void
 }
 
 ; Exponent is 0.75 but no proper fast-math flags
-define void @my_vpow_075_nofast(double* nocapture %y, double* nocapture readonly %x) {
-; CHECK-LABEL:       @vspow_075_nofast
+define void @vpow_075_nofast(double* nocapture %y, double* nocapture readonly %x) {
+; CHECK-LABEL:       @vpow_075_nofast
 ; CHECK-PWR9:        bl __powd2_P9
 ; CHECK-PWR8:        bl __powd2_P8
 ; CHECK-NOT:         xvrsqrtesp
 ; CHECK:             blr
 entry:
   br label %vector.body
 
 vector.body:
vector.body:
   %3 = icmp eq i64 %index.next, 1024
   br i1 %3, label %for.end, label %vector.body
 
 for.end:
   ret void
 }
 
 ; Exponent is 0.25 but no proper fast-math flags
-define void @my_vpow_025_nofast(double* nocapture %y, double* nocapture readonly %x) {
-; CHECK-LABEL:       @vspow_025_nofast
+define void @vpow_025_nofast(double* nocapture %y, double* nocapture readonly %x) {
+; CHECK-LABEL:       @vpow_025_nofast
 ; CHECK-PWR9:        bl __powd2_P9
 ; CHECK-PWR8:        bl __powd2_P8
 ; CHECK-NOT:         xvrsqrtesp
 ; CHECK:             blr
 entry:
   br label %vector.body
 
 vector.body:

llvm/test/CodeGen/PowerPC/powf_massv_075_025exp.ll

vector.body:
   %index.next = add i64 %index, 4
   %3 = icmp eq i64 %index.next, 1024
   br i1 %3, label %for.end, label %vector.body
 
 for.end:
   ret void
 }
 
+; Exponent is a constant != 0.75 and !=0.25 and they are different 
+define void @vspow_neq_const(float* nocapture %y, float* nocapture readonly %x)  {
+; CHECK-LABEL:       @vspow_neq_const
+; CHECK-PWR9:        bl __powf4_P9
+; CHECK-PWR8:        bl __powf4_P8
+; CHECK:             blr
+entry:
+  br label %vector.body
+
+vector.body:
+  %index = phi i64 [ %index.next, %vector.body ], [ 0, %entry ]
+  %next.gep = getelementptr float, float* %y, i64 %index
+  %next.gep19 = getelementptr float, float* %x, i64 %index
+  %0 = bitcast float* %next.gep19 to <4 x float>*
+  %wide.load = load <4 x float>, <4 x float>* %0, align 4
+  %1 = call ninf afn nsz <4 x float> @__powf4_massv(<4 x float> %wide.load, <4 x float> <float 0x3FE861EB80000000, float 0x3FE871EB80000000, float 0x3FE851EB80000000, float 0x3FE851EB80000000>)
+  %2 = bitcast float* %next.gep to <4 x float>*
+  store <4 x float> %1, <4 x float>* %2, align 4
+  %index.next = add i64 %index, 4
+  %3 = icmp eq i64 %index.next, 1024
+  br i1 %3, label %for.end, label %vector.body
+
+for.end:
+  ret void
+}
+
+; Exponent is a constant != 0.75 and !=0.25
+define void @vspow_neq075_const(float* nocapture %y, float* nocapture readonly %x)  {
+; CHECK-LABEL:       @vspow_neq075_const
+; CHECK-PWR9:        bl __powf4_P9
+; CHECK-PWR8:        bl __powf4_P8
+; CHECK:             blr
+entry:
+  br label %vector.body
+
+vector.body:
+  %index = phi i64 [ %index.next, %vector.body ], [ 0, %entry ]
+  %next.gep = getelementptr float, float* %y, i64 %index
+  %next.gep19 = getelementptr float, float* %x, i64 %index
+  %0 = bitcast float* %next.gep19 to <4 x float>*
+  %wide.load = load <4 x float>, <4 x float>* %0, align 4
+  %1 = call ninf afn nsz <4 x float> @__powf4_massv(<4 x float> %wide.load, <4 x float> <float 7.500000e-01, float 7.500000e-01, float 7.500000e-01, float 0x3FE851EB80000000>)
+  %2 = bitcast float* %next.gep to <4 x float>*
+  store <4 x float> %1, <4 x float>* %2, align 4
+  %index.next = add i64 %index, 4
+  %3 = icmp eq i64 %index.next, 1024
+  br i1 %3, label %for.end, label %vector.body
+
+for.end:
+  ret void
+}
+
+; Exponent is a constant != 0.75 and !=0.25
+define void @vspow_neq025_const(float* nocapture %y, float* nocapture readonly %x)  {
+; CHECK-LABEL:       @vspow_neq025_const
+; CHECK-PWR9:        bl __powf4_P9
+; CHECK-PWR8:        bl __powf4_P8
+; CHECK:             blr
+entry:
+  br label %vector.body
+
+vector.body:
+  %index = phi i64 [ %index.next, %vector.body ], [ 0, %entry ]
+  %next.gep = getelementptr float, float* %y, i64 %index
+  %next.gep19 = getelementptr float, float* %x, i64 %index
+  %0 = bitcast float* %next.gep19 to <4 x float>*
+  %wide.load = load <4 x float>, <4 x float>* %0, align 4
+  %1 = call ninf afn nsz <4 x float> @__powf4_massv(<4 x float> %wide.load, <4 x float> <float 0x3FE851EB80000000, float 2.500000e-01, float 0x3FE851EB80000000, float 2.500000e-01>)
+  %2 = bitcast float* %next.gep to <4 x float>*
+  store <4 x float> %1, <4 x float>* %2, align 4
+  %index.next = add i64 %index, 4
+  %3 = icmp eq i64 %index.next, 1024
+  br i1 %3, label %for.end, label %vector.body
+
+for.end:
+  ret void
+}
+
 ; Exponent is 0.75
 define void @vspow_075(float* nocapture %y, float* nocapture readonly %x)  {
 ; CHECK-LABEL:       @vspow_075
 ; CHECK-NOT:         bl __powf4_P{{[8,9]}}
 ; CHECK:             xvrsqrtesp
 ; CHECK:             blr
 entry:
   br label %vector.body