[ppc64le] Added float32 support and implemented all R4 opcodes (#86092)

* [ppc64le] Fixed thunk address 8 byte alignment issue

* Fixed FSharp crash issue

* [ppc64le] Implementation of mono_arch_get_delegate_virtual_invoke_impl method for ppc64le architecture

* Fixed clang15 build issues and returning address of sc_sp instead of value

* Added float32 support and implemented related opcodes

* Correction in OP_RCONV_TO_R cases

* Corrected code for few opcodes

diff --git a/src/mono/mono/arch/ppc/ppc-codegen.h b/src/mono/mono/arch/ppc/ppc-codegen.h
index 291337f..52e2b17 100644 (file)
--- a/src/mono/mono/arch/ppc/ppc-codegen.h
+++ b/src/mono/mono/arch/ppc/ppc-codegen.h
@@ -793,6 +793,10 @@ my and Ximian's copyright to this code. ;)
 #define ppc_fcfid(c,D,B)  ppc_fcfidx(c,D,B,0)
 #define ppc_fcfidd(c,D,B) ppc_fcfidx(c,D,B,1)
 
+#define ppc_fcfidsx(c,D,B,Rc) ppc_emit32(c, (59 << 26) | ((D) << 21) | (0 << 16) | ((B) << 11) | (846 << 1) | (Rc))
+#define ppc_fcfids(c,D,B)  ppc_fcfidsx(c,D,B,0)
+#define ppc_fcfidsd(c,D,B) ppc_fcfidsx(c,D,B,1)
+
 #define ppc_fctidx(c,D,B,Rc) ppc_emit32(c, (63 << 26) | ((D) << 21) | (0 << 16) | ((B) << 11) | (814 << 1) | (Rc))
 #define ppc_fctid(c,D,B)  ppc_fctidx(c,D,B,0)
 #define ppc_fctidd(c,D,B) ppc_fctidx(c,D,B,1)

diff --git a/src/mono/mono/mini/cpu-ppc64.mdesc b/src/mono/mono/mini/cpu-ppc64.mdesc
index 77d9965..8729c30 100644 (file)
--- a/src/mono/mono/mini/cpu-ppc64.mdesc
+++ b/src/mono/mono/mini/cpu-ppc64.mdesc
@@ -229,6 +229,37 @@ ppc_subfze: dest:i src1:i len:4
 bigmul: len:12 dest:i src1:i src2:i
 bigmul_un: len:12 dest:i src1:i src2:i
 
+# R4 opcodes
+r4_conv_to_i1: dest:i src1:f len:40
+r4_conv_to_u1: dest:i src1:f len:40
+r4_conv_to_i2: dest:i src1:f len:40
+r4_conv_to_u2: dest:i src1:f len:40
+r4_conv_to_i4: dest:i src1:f len:40
+r4_conv_to_u4: dest:i src1:f len:40
+r4_conv_to_i8: dest:i src1:f len:40
+r4_conv_to_u8: dest:i src1:f len:40
+r4_conv_to_r8: dest:f src1:f len:4
+r4_conv_to_r4: dest:f src1:f len:4
+r4_add: dest:f src1:f src2:f len:4
+r4_sub: dest:f src1:f src2:f len:4
+r4_mul: dest:f src1:f src2:f len:4
+r4_div: dest:f src1:f src2:f len:4
+r4_rem: dest:f src1:f src2:f len:16
+r4_neg: dest:f src1:f len:4
+r4_ceq: dest:i src1:f src2:f len:16
+r4_cgt: dest:i src1:f src2:f len:16
+r4_cgt_un: dest:i src1:f src2:f len:20
+r4_clt: dest:i src1:f src2:f len:16
+r4_clt_un: dest:i src1:f src2:f len:20
+r4_cneq: dest:i src1:f src2:f len:16
+r4_cge: dest:i src1:f src2:f len:16
+r4_cle: dest:i src1:f src2:f len:16
+rmove: dest:f src1:f len:4
+rcompare: src1:f src2:f len:12
+rcall_membase: dest:g src1:b len:16 clob:c
+rcall_reg: dest:g src1:i len:16 clob:c
+rcall: dest:g len:40 clob:c
+
 # Linear IR opcodes
 dummy_use: src1:i len:0
 dummy_iconst: dest:i len:0
@@ -258,7 +289,7 @@ int_conv_to_i1: dest:i src1:i len:8
 int_conv_to_i2: dest:i src1:i len:8
 int_conv_to_i4: dest:i src1:i len:4
 sext_i4: dest:i src1:i len:4
-int_conv_to_r4: dest:f src1:i len:20
+int_conv_to_r4: dest:f src1:i len:16
 int_conv_to_r8: dest:f src1:i len:16
 int_conv_to_u4: dest:i src1:i len:4
 int_conv_to_u2: dest:i src1:i len:8
@@ -354,7 +385,7 @@ long_not: dest:i src1:i len:4
 long_conv_to_i1: dest:i src1:i len:4
 long_conv_to_i2: dest:i src1:i len:4
 long_conv_to_i4: dest:i src1:i len:4
-long_conv_to_r4: dest:f src1:i len:16
+long_conv_to_r4: dest:f src1:i len:12
 long_conv_to_r8: dest:f src1:i len:12
 long_conv_to_u4: dest:i src1:i
 long_conv_to_u2: dest:i src1:i len:4

diff --git a/src/mono/mono/mini/mini-ppc.c b/src/mono/mono/mini/mini-ppc.c
index ced67cc..3b9bdf2 100644 (file)
--- a/src/mono/mono/mini/mini-ppc.c
+++ b/src/mono/mono/mini/mini-ppc.c
@@ -3033,8 +3033,11 @@ emit_move_return_value (MonoCompile *cfg, MonoInst *ins, guint8 *code)
 {
        switch (ins->opcode) {
        case OP_FCALL:
+       case OP_RCALL:
        case OP_FCALL_REG:
+       case OP_RCALL_REG:
        case OP_FCALL_MEMBASE:
+       case OP_RCALL_MEMBASE:
                if (ins->dreg != ppc_f1)
                        ppc_fmr (code, ins->dreg, ppc_f1);
                break;
@@ -3763,6 +3766,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                                ppc_mr (code, ppc_r4, saved);
                        break;
                }
+               case OP_RMOVE:
                case OP_FMOVE:
                        if (ins->dreg != ins->sreg1)
                                ppc_fmr (code, ins->dreg, ins->sreg1);
@@ -3894,6 +3898,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        ppc_stptr (code, ppc_r0, 0, ins->sreg1);
                        break;
                }
+               case OP_RCALL:
                case OP_FCALL:
                case OP_LCALL:
                case OP_VCALL:
@@ -3916,6 +3921,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        /* FIXME: this should be handled somewhere else in the new jit */
                        code = emit_move_return_value (cfg, ins, code);
                        break;
+               case OP_RCALL_REG:
                case OP_FCALL_REG:
                case OP_LCALL_REG:
                case OP_VCALL_REG:
@@ -3945,6 +3951,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        /* FIXME: this should be handled somewhere else in the new jit */
                        code = emit_move_return_value (cfg, ins, code);
                        break;
+               case OP_RCALL_MEMBASE:
                case OP_FCALL_MEMBASE:
                case OP_LCALL_MEMBASE:
                case OP_VCALL_MEMBASE:
@@ -4266,21 +4273,27 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case CEE_CONV_R4: /* FIXME: change precision */
                case CEE_CONV_R8:
                        g_assert_not_reached ();
+               case OP_RCONV_TO_I1:
                case OP_FCONV_TO_I1:
                        code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 1, TRUE);
                        break;
+               case OP_RCONV_TO_U1:
                case OP_FCONV_TO_U1:
                        code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 1, FALSE);
                        break;
+               case OP_RCONV_TO_I2:
                case OP_FCONV_TO_I2:
                        code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 2, TRUE);
                        break;
+               case OP_RCONV_TO_U2:
                case OP_FCONV_TO_U2:
                        code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 2, FALSE);
                        break;
+               case OP_RCONV_TO_I4:
                case OP_FCONV_TO_I4:
                        code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 4, TRUE);
                        break;
+               case OP_RCONV_TO_U4:
                case OP_FCONV_TO_U4:
                        code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 4, FALSE);
                        break;
@@ -4339,21 +4352,35 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_SQRT:
                        ppc_fsqrtd (code, ins->dreg, ins->sreg1);
                        break;
+               case OP_RADD:
+                       ppc_fadds (code, ins->dreg, ins->sreg1, ins->sreg2);
+                       break;
                case OP_FADD:
                        ppc_fadd (code, ins->dreg, ins->sreg1, ins->sreg2);
                        break;
+               case OP_RSUB:
+                       ppc_fsubs (code, ins->dreg, ins->sreg1, ins->sreg2);
+                       break;
                case OP_FSUB:
                        ppc_fsub (code, ins->dreg, ins->sreg1, ins->sreg2);
                        break;
+               case OP_RMUL:
+                       ppc_fmuls (code, ins->dreg, ins->sreg1, ins->sreg2);
+                       break;
                case OP_FMUL:
                        ppc_fmul (code, ins->dreg, ins->sreg1, ins->sreg2);
                        break;
+               case OP_RDIV:
+                       ppc_fdivs (code, ins->dreg, ins->sreg1, ins->sreg2);
+                       break;
                case OP_FDIV:
                        ppc_fdiv (code, ins->dreg, ins->sreg1, ins->sreg2);
                        break;
+               case OP_RNEG:
                case OP_FNEG:
                        ppc_fneg (code, ins->dreg, ins->sreg1);
                        break;
+               case OP_RREM:
                case OP_FREM:
                        /* emulated */
                        g_assert_not_reached ();
@@ -4391,23 +4418,29 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        ppc_cmpl (code, 0, 1, ins->sreg1, ins->sreg2);
                        ppc_iselgt (code, ins->dreg, ins->sreg1, ins->sreg2);
                        break;
+               case OP_RCOMPARE:
                case OP_FCOMPARE:
                        ppc_fcmpu (code, 0, ins->sreg1, ins->sreg2);
                        break;
+               case OP_RCEQ:
+               case OP_RCNEQ:
                case OP_FCEQ:
                case OP_FCNEQ:
                        ppc_fcmpo (code, 0, ins->sreg1, ins->sreg2);
                        ppc_li (code, ins->dreg, 1);
-                       ppc_bc (code, ins->opcode == OP_FCEQ ? PPC_BR_TRUE : PPC_BR_FALSE, PPC_BR_EQ, 2);
+                       ppc_bc (code, ins->opcode == OP_FCEQ || ins->opcode == OP_RCEQ ? PPC_BR_TRUE : PPC_BR_FALSE, PPC_BR_EQ, 2);
                        ppc_li (code, ins->dreg, 0);
                        break;
+               case OP_RCLT:
+               case OP_RCGE:
                case OP_FCLT:
                case OP_FCGE:
                        ppc_fcmpo (code, 0, ins->sreg1, ins->sreg2);
                        ppc_li (code, ins->dreg, 1);
-                       ppc_bc (code, ins->opcode == OP_FCLT ? PPC_BR_TRUE : PPC_BR_FALSE, PPC_BR_LT, 2);
+                       ppc_bc (code, ins->opcode == OP_FCLT || ins->opcode == OP_RCLT ? PPC_BR_TRUE : PPC_BR_FALSE, PPC_BR_LT, 2);
                        ppc_li (code, ins->dreg, 0);
                        break;
+               case OP_RCLT_UN:
                case OP_FCLT_UN:
                        ppc_fcmpu (code, 0, ins->sreg1, ins->sreg2);
                        ppc_li (code, ins->dreg, 1);
@@ -4415,13 +4448,16 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        ppc_bc (code, PPC_BR_TRUE, PPC_BR_LT, 2);
                        ppc_li (code, ins->dreg, 0);
                        break;
+               case OP_RCGT:
+               case OP_RCLE:
                case OP_FCGT:
                case OP_FCLE:
                        ppc_fcmpo (code, 0, ins->sreg1, ins->sreg2);
                        ppc_li (code, ins->dreg, 1);
-                       ppc_bc (code, ins->opcode == OP_FCGT ? PPC_BR_TRUE : PPC_BR_FALSE, PPC_BR_GT, 2);
+                       ppc_bc (code, ins->opcode == OP_FCGT || ins->opcode == OP_RCGT ? PPC_BR_TRUE : PPC_BR_FALSE, PPC_BR_GT, 2);
                        ppc_li (code, ins->dreg, 0);
                        break;
+               case OP_RCGT_UN:
                case OP_FCGT_UN:
                        ppc_fcmpu (code, 0, ins->sreg1, ins->sreg2);
                        ppc_li (code, ins->dreg, 1);
@@ -4492,6 +4528,10 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_ZEXT_I4:
                        ppc_clrldi (code, ins->dreg, ins->sreg1, 32);
                        break;
+               case OP_RCONV_TO_R4:
+               case OP_RCONV_TO_R8:
+                       ppc_fmr (code, ins->dreg, ins->sreg1);
+                       break;
                case OP_ICONV_TO_R4:
                case OP_ICONV_TO_R8:
                case OP_LCONV_TO_R4:
@@ -4509,9 +4549,11 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                                ppc_str (code, tmp, -8, ppc_r1);
                                ppc_lfd (code, ins->dreg, -8, ppc_r1);
                        }
-                       ppc_fcfid (code, ins->dreg, ins->dreg);
-                       if (ins->opcode == OP_ICONV_TO_R4 || ins->opcode == OP_LCONV_TO_R4)
-                               ppc_frsp (code, ins->dreg, ins->dreg);
+                       if (ins->opcode == OP_ICONV_TO_R4 || ins->opcode == OP_LCONV_TO_R4) {
+                               ppc_fcfids (code, ins->dreg, ins->dreg);
+                       } else {
+                               ppc_fcfid (code, ins->dreg, ins->dreg);
+                       }
                        break;
                }
                case OP_LSHR:
@@ -4544,9 +4586,11 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_LBLE_UN:
                        EMIT_COND_BRANCH (ins, ins->opcode - OP_LBEQ);
                        break;
+               case OP_RCONV_TO_I8:
                case OP_FCONV_TO_I8:
                        code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 8, TRUE);
                        break;
+               case OP_RCONV_TO_U8:
                case OP_FCONV_TO_U8:
                        code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 8, FALSE);
                        break;

diff --git a/src/mono/mono/mini/mini-ppc.h b/src/mono/mono/mini/mini-ppc.h
index e872c4b..5fd50f4 100644 (file)
--- a/src/mono/mono/mini/mini-ppc.h
+++ b/src/mono/mono/mini/mini-ppc.h
@@ -261,6 +261,7 @@ typedef struct MonoCompileArch {
 #define MONO_ARCH_HAVE_FULL_AOT_TRAMPOLINES 1
 
 #define MONO_ARCH_GSHARED_SUPPORTED 1
+#define MONO_ARCH_FLOAT32_SUPPORTED 1
 
 #define MONO_ARCH_NEED_DIV_CHECK 1
 #define MONO_ARCH_AOT_SUPPORTED 1