1Pull in r199028 from upstream llvm trunk (by Jakob Stoklund Olesen): 2 3 The SPARCv9 ABI returns a float in %f0. 4 5 This is different from the argument passing convention which puts the 6 first float argument in %f1. 7 8 With this patch, all returned floats are treated as if the 'inreg' flag 9 were set. This means multiple float return values get packed in %f0, 10 %f1, %f2, ... 11 12 Note that when returning a struct in registers, clang will set the 13 'inreg' flag on the return value, so that behavior is unchanged. This 14 also happens when returning a float _Complex. 15 16Introduced here: http://svnweb.freebsd.org/changeset/base/262261 17 18Index: test/CodeGen/SPARC/64abi.ll 19=================================================================== 20--- test/CodeGen/SPARC/64abi.ll 21+++ test/CodeGen/SPARC/64abi.ll 22@@ -180,7 +180,7 @@ define void @call_inreg_fi(i32* %p, i32 %i1, float 23 } 24 25 ; CHECK: inreg_ff 26-; CHECK: fsubs %f0, %f1, %f1 27+; CHECK: fsubs %f0, %f1, %f0 28 define float @inreg_ff(float inreg %a0, ; %f0 29 float inreg %a1) { ; %f1 30 %rv = fsub float %a0, %a1 31@@ -262,10 +262,10 @@ define void @call_ret_i64_pair(i64* %i0) { 32 ret void 33 } 34 35-; This is not a C struct, each member uses 8 bytes. 36+; This is not a C struct, the i32 member uses 8 bytes, but the float only 4. 37 ; CHECK: ret_i32_float_pair 38 ; CHECK: ld [%i2], %i0 39-; CHECK: ld [%i3], %f3 40+; CHECK: ld [%i3], %f2 41 define { i32, float } @ret_i32_float_pair(i32 %a0, i32 %a1, 42 i32* %p, float* %q) { 43 %r1 = load i32* %p 44@@ -279,7 +279,7 @@ define { i32, float } @ret_i32_float_pair(i32 %a0, 45 ; CHECK: call_ret_i32_float_pair 46 ; CHECK: call ret_i32_float_pair 47 ; CHECK: st %o0, [%i0] 48-; CHECK: st %f3, [%i1] 49+; CHECK: st %f2, [%i1] 50 define void @call_ret_i32_float_pair(i32* %i0, float* %i1) { 51 %rv = call { i32, float } @ret_i32_float_pair(i32 undef, i32 undef, 52 i32* undef, float* undef) 53Index: test/CodeGen/SPARC/constpool.ll 54=================================================================== 55--- test/CodeGen/SPARC/constpool.ll 56+++ test/CodeGen/SPARC/constpool.ll 57@@ -21,7 +21,7 @@ entry: 58 ; abs44: add %[[R1]], %m44(.LCPI0_0), %[[R2:[gilo][0-7]]] 59 ; abs44: sllx %[[R2]], 12, %[[R3:[gilo][0-7]]] 60 ; abs44: retl 61-; abs44: ld [%[[R3]]+%l44(.LCPI0_0)], %f1 62+; abs44: ld [%[[R3]]+%l44(.LCPI0_0)], %f0 63 64 65 ; abs64: floatCP 66@@ -31,7 +31,7 @@ entry: 67 ; abs64: add %[[R3]], %hm(.LCPI0_0), %[[R4:[gilo][0-7]]] 68 ; abs64: sllx %[[R4]], 32, %[[R5:[gilo][0-7]]] 69 ; abs64: retl 70-; abs64: ld [%[[R5]]+%[[R2]]], %f1 71+; abs64: ld [%[[R5]]+%[[R2]]], %f0 72 73 74 ; v8pic32: floatCP 75@@ -50,7 +50,7 @@ entry: 76 ; v9pic32: sethi %hi(.LCPI0_0), %[[R1:[gilo][0-7]]] 77 ; v9pic32: add %[[R1]], %lo(.LCPI0_0), %[[Goffs:[gilo][0-7]]] 78 ; v9pic32: ldx [%[[GOT:[gilo][0-7]]]+%[[Goffs]]], %[[Gaddr:[gilo][0-7]]] 79-; v9pic32: ld [%[[Gaddr]]], %f1 80+; v9pic32: ld [%[[Gaddr]]], %f0 81 ; v9pic32: ret 82 ; v9pic32: restore 83 84Index: test/CodeGen/SPARC/64cond.ll 85=================================================================== 86--- test/CodeGen/SPARC/64cond.ll 87+++ test/CodeGen/SPARC/64cond.ll 88@@ -80,7 +80,7 @@ entry: 89 ; CHECK: selectf32_xcc 90 ; CHECK: cmp %i0, %i1 91 ; CHECK: fmovsg %xcc, %f5, %f7 92-; CHECK: fmovs %f7, %f1 93+; CHECK: fmovs %f7, %f0 94 define float @selectf32_xcc(i64 %x, i64 %y, float %a, float %b) { 95 entry: 96 %tobool = icmp sgt i64 %x, %y 97Index: lib/Target/Sparc/SparcISelLowering.cpp 98=================================================================== 99--- lib/Target/Sparc/SparcISelLowering.cpp 100+++ lib/Target/Sparc/SparcISelLowering.cpp 101@@ -254,7 +254,7 @@ SparcTargetLowering::LowerReturn_64(SDValue Chain, 102 DAG.getTarget(), RVLocs, *DAG.getContext()); 103 104 // Analyze return values. 105- CCInfo.AnalyzeReturn(Outs, CC_Sparc64); 106+ CCInfo.AnalyzeReturn(Outs, RetCC_Sparc64); 107 108 SDValue Flag; 109 SmallVector<SDValue, 4> RetOps(1, Chain); 110@@ -1258,7 +1258,7 @@ SparcTargetLowering::LowerCall_64(TargetLowering:: 111 if (CLI.Ins.size() == 1 && CLI.Ins[0].VT == MVT::f32 && CLI.CS == 0) 112 CLI.Ins[0].Flags.setInReg(); 113 114- RVInfo.AnalyzeCallResult(CLI.Ins, CC_Sparc64); 115+ RVInfo.AnalyzeCallResult(CLI.Ins, RetCC_Sparc64); 116 117 // Copy all of the result registers out of their specified physreg. 118 for (unsigned i = 0; i != RVLocs.size(); ++i) { 119Index: lib/Target/Sparc/SparcCallingConv.td 120=================================================================== 121--- lib/Target/Sparc/SparcCallingConv.td 122+++ lib/Target/Sparc/SparcCallingConv.td 123@@ -103,7 +103,7 @@ def RetCC_Sparc32 : CallingConv<[ 124 // Function return values are passed exactly like function arguments, except a 125 // struct up to 32 bytes in size can be returned in registers. 126 127-// Function arguments AND return values. 128+// Function arguments AND most return values. 129 def CC_Sparc64 : CallingConv<[ 130 // The frontend uses the inreg flag to indicate i32 and float arguments from 131 // structs. These arguments are not promoted to 64 bits, but they can still 132@@ -118,6 +118,15 @@ def CC_Sparc64 : CallingConv<[ 133 CCCustom<"CC_Sparc64_Full"> 134 ]>; 135 136+def RetCC_Sparc64 : CallingConv<[ 137+ // A single f32 return value always goes in %f0. The ABI doesn't specify what 138+ // happens to multiple f32 return values outside a struct. 139+ CCIfType<[f32], CCCustom<"CC_Sparc64_Half">>, 140+ 141+ // Otherwise, return values are passed exactly like arguments. 142+ CCDelegateTo<CC_Sparc64> 143+]>; 144+ 145 // Callee-saved registers are handled by the register window mechanism. 146 def CSR : CalleeSavedRegs<(add)> { 147 let OtherPreserved = (add (sequence "I%u", 0, 7), 148