411 lines · plain
1// z_AIX_asm.S: - microtasking routines specifically2// written for Power platforms running AIX OS3 4//5////===----------------------------------------------------------------------===//6////7//// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.8//// See https://llvm.org/LICENSE.txt for license information.9//// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception10////11////===----------------------------------------------------------------------===//12//13 14// -----------------------------------------------------------------------15// macros16// -----------------------------------------------------------------------17 18#include "kmp_config.h"19 20#if KMP_OS_AIX21//------------------------------------------------------------------------22// int23// __kmp_invoke_microtask( void (*pkfn) (int *gtid, int *tid, ...),24// int gtid, int tid,25// int argc, void *p_argv[]26// #if OMPT_SUPPORT27// ,28// void **exit_frame_ptr29// #endif30// ) {31// #if OMPT_SUPPORT32// *exit_frame_ptr = OMPT_GET_FRAME_ADDRESS(0);33// #endif34//35// (*pkfn)( & gtid, & tid, p_argv[0], ... );36//37// // FIXME: This is done at call-site and can be removed here.38// #if OMPT_SUPPORT39// *exit_frame_ptr = 0;40// #endif41//42// return 1;43// }44//45// parameters:46// r3: pkfn47// r4: gtid48// r5: tid49// r6: argc50// r7: p_argv51// r8: &exit_frame52//53// return: r3 (always 1/TRUE)54//55 56#if KMP_ARCH_PPC64_XCOFF57 58 .globl __kmp_invoke_microtask[DS]59 .globl .__kmp_invoke_microtask60 .align 461 .csect __kmp_invoke_microtask[DS],362 .vbyte 8, .__kmp_invoke_microtask63 .vbyte 8, TOC[TC0]64 .vbyte 8, 065 .csect .text[PR],266 .machine "pwr7"67.__kmp_invoke_microtask:68 69 70// -- Begin __kmp_invoke_microtask71// mark_begin;72 73// We need to allocate a stack frame large enough to hold all of the parameters74// on the stack for the microtask plus what this function needs. That's 4875// bytes under the XCOFF64 ABI, plus max(64, 8*(2 + argc)) for76// the parameters to the microtask (gtid, tid, argc elements of p_argv),77// plus 8 bytes to store the values of r4 and r5, and 8 bytes to store r31.78// With OMP-T support, we need an additional 8 bytes to save r30 to hold79// a copy of r8.80// Stack offsets relative to stack pointer:81// r31: -8, r30: -16, gtid: -20, tid: -2482 83 mflr 084 std 31, -8(1) # Save r31 to the stack85 std 0, 16(1) # Save LR to the linkage area86 87// This is unusual because normally we'd set r31 equal to r1 after the stack88// frame is established. In this case, however, we need to dynamically compute89// the stack frame size, and so we keep a direct copy of r1 to access our90// register save areas and restore the r1 value before returning.91 mr 31, 192 93// Compute the size of the "argc" portion of the parameter save area.94// The parameter save area is always at least 64 bytes long (i.e. 8 regs)95// The microtask has (2 + argc) parameters, so if argc <= 6, we need to96// to allocate 8*6 bytes, not 8*argc.97 li 0, 698 cmpwi 0, 6, 699 iselgt 0, 6, 0 # r0 = (argc > 6)? argc : 6100 sldi 0, 0, 3 # r0 = 8 * max(argc, 6)101 102// Compute the size necessary for the local stack frame.103// 88 = 48 + 4 (for r4) + 4 (for r5) + 8 (for r31) + 8 (for OMP-T r30) +104// 8 (parameter gtid) + 8 (parameter tid)105 li 12, 88106 add 12, 0, 12107 neg 12, 12108 109// We need to make sure that the stack frame stays aligned (to 16 bytes).110 li 0, -16111 and 12, 0, 12112 113// Establish the local stack frame.114 stdux 1, 1, 12115 116#if OMPT_SUPPORT117 std 30, -16(31) # Save r30 to the stack118 std 1, 0(8)119 mr 30, 8120#endif121 122// Store gtid and tid to the stack because they're passed by reference to the microtask.123 stw 4, -20(31) # Save gtid to the stack124 stw 5, -24(31) # Save tid to the stack125 126 mr 12, 6 # r12 = argc127 mr 4, 7 # r4 = p_argv128 129 cmpwi 0, 12, 1130 blt 0, .Lcall # if (argc < 1) goto .Lcall131 132 ld 5, 0(4) # r5 = p_argv[0]133 134 cmpwi 0, 12, 2135 blt 0, .Lcall # if (argc < 2) goto .Lcall136 137 ld 6, 8(4) # r6 = p_argv[1]138 139 cmpwi 0, 12, 3140 blt 0, .Lcall # if (argc < 3) goto .Lcall141 142 ld 7, 16(4) # r7 = p_argv[2]143 144 cmpwi 0, 12, 4145 blt 0, .Lcall # if (argc < 4) goto .Lcall146 147 ld 8, 24(4) # r8 = p_argv[3]148 149 cmpwi 0, 12, 5150 blt 0, .Lcall # if (argc < 5) goto .Lcall151 152 ld 9, 32(4) # r9 = p_argv[4]153 154 cmpwi 0, 12, 6155 blt 0, .Lcall # if (argc < 6) goto .Lcall156 157 ld 10, 40(4) # r10 = p_argv[5]158 159 cmpwi 0, 12, 7160 blt 0, .Lcall # if (argc < 7) goto .Lcall161 162// There are more than 6 microtask parameters, so we need to store the163// remainder to the stack.164 addi 12, 12, -6 # argc -= 6165 mtctr 12166 167// These are set to 8 bytes before the first desired store address (we're using168// pre-increment loads and stores in the loop below). The parameter save area169// for the microtask begins 48 + 8*8 == 112 bytes above r1 for XCOFF64.170 addi 4, 4, 40 # p_argv = p_argv + 5171 # (i.e. skip the 5 elements we already processed)172 addi 12, 1, 104 # r12 = stack offset (112 - 8)173 174.Lnext:175 ldu 0, 8(4)176 stdu 0, 8(12)177 bdnz .Lnext178 179.Lcall:180 std 2, 40(1) # Save the TOC pointer to the linkage area181// Load the actual function address from the function descriptor.182 ld 12, 0(3) # Function address183 ld 2, 8(3) # TOC pointer184 ld 11, 16(3) # Environment pointer185 186 addi 3, 31, -20 # r3 = >id187 addi 4, 31, -24 # r4 = &tid188 189 mtctr 12 # CTR = function address190 bctrl # Branch to CTR191 ld 2, 40(1) # Restore TOC pointer from linkage area192 193#if OMPT_SUPPORT194 li 3, 0195 std 3, 0(30)196#endif197 198 li 3, 1199 200#if OMPT_SUPPORT201 ld 30, -16(31) # Restore r30 from the saved value on the stack202#endif203 204 mr 1, 31205 ld 31, -8(1) # Restore r31 from the saved value on the stack206 ld 0, 16(1)207 mtlr 0 # Restore LR from the linkage area208 blr # Branch to LR209 210#else // KMP_ARCH_PPC_XCOFF211 212 .globl __kmp_invoke_microtask[DS]213 .globl .__kmp_invoke_microtask214 .align 4215 .csect __kmp_invoke_microtask[DS],2216 .vbyte 4, .__kmp_invoke_microtask217 .vbyte 4, TOC[TC0]218 .vbyte 4, 0219 .csect .text[PR],2220 .machine "pwr7"221.__kmp_invoke_microtask:222 223 224// -- Begin __kmp_invoke_microtask225// mark_begin;226 227// We need to allocate a stack frame large enough to hold all of the parameters228// on the stack for the microtask plus what this function needs. That's 24229// bytes under the XCOFF ABI, plus max(32, 8*(2 + argc)) for230// the parameters to the microtask (gtid, tid, argc elements of p_argv),231// plus 8 bytes to store the values of r4 and r5, and 4 bytes to store r31.232// With OMP-T support, we need an additional 4 bytes to save r30 to hold233// a copy of r8.234// Stack offsets relative to stack pointer:235// r31: -4, r30: -8, gtid: -12, tid: -16236 237 mflr 0238 stw 31, -4(1) # Save r31 to the stack239 stw 0, 8(1) # Save LR to the linkage area240 241// This is unusual because normally we'd set r31 equal to r1 after the stack242// frame is established. In this case, however, we need to dynamically compute243// the stack frame size, and so we keep a direct copy of r1 to access our244// register save areas and restore the r1 value before returning.245 mr 31, 1246 247// Compute the size of the "argc" portion of the parameter save area.248// The parameter save area is always at least 32 bytes long (i.e. 8 regs)249// The microtask has (2 + argc) parameters, so if argc <= 6, we need to250// to allocate 4*6 bytes, not 4*argc.251 li 0, 6252 cmpwi 0, 6, 6253 iselgt 0, 6, 0 # r0 = (argc > 6)? argc : 6254 slwi 0, 0, 2 # r0 = 4 * max(argc, 6)255 256// Compute the size necessary for the local stack frame.257// 56 = 32 + 4 (for r4) + 4 (for r5) + 4 (for r31) + 4 (for OMP-T r30) +258// 4 (parameter gtid) + 4 (parameter tid)259 li 12, 56260 add 12, 0, 12261 neg 12, 12262 263// We need to make sure that the stack frame stays aligned (to 16 bytes).264 li 0, -16265 and 12, 0, 12266 267// Establish the local stack frame.268 stwux 1, 1, 12269 270#if OMPT_SUPPORT271 stw 30, -8(31) # Save r30 to the stack272 stw 1, 0(8)273 mr 30, 8274#endif275 276// Store gtid and tid to the stack because they're passed by reference to the microtask.277 stw 4, -12(31) # Save gtid to the stack278 stw 5, -16(31) # Save tid to the stack279 280 mr 12, 6 # r12 = argc281 mr 4, 7 # r4 = p_argv282 283 cmpwi 0, 12, 1284 blt 0, .Lcall # if (argc < 1) goto .Lcall285 286 lwz 5, 0(4) # r5 = p_argv[0]287 288 cmpwi 0, 12, 2289 blt 0, .Lcall # if (argc < 2) goto .Lcall290 291 lwz 6, 4(4) # r6 = p_argv[1]292 293 cmpwi 0, 12, 3294 blt 0, .Lcall # if (argc < 3) goto .Lcall295 296 lwz 7, 8(4) # r7 = p_argv[2]297 298 cmpwi 0, 12, 4299 blt 0, .Lcall # if (argc < 4) goto .Lcall300 301 lwz 8, 12(4) # r8 = p_argv[3]302 303 cmpwi 0, 12, 5304 blt 0, .Lcall # if (argc < 5) goto .Lcall305 306 lwz 9, 16(4) # r9 = p_argv[4]307 308 cmpwi 0, 12, 6309 blt 0, .Lcall # if (argc < 6) goto .Lcall310 311 lwz 10, 20(4) # r10 = p_argv[5]312 313 cmpwi 0, 12, 7314 blt 0, .Lcall # if (argc < 7) goto .Lcall315 316// There are more than 6 microtask parameters, so we need to store the317// remainder to the stack.318 addi 12, 12, -6 # argc -= 6319 mtctr 12320 321// These are set to 4 bytes before the first desired store address (we're using322// pre-increment loads and stores in the loop below). The parameter save area323// for the microtask begins 24 + 4*8 == 56 bytes above r1 for XCOFF.324 addi 4, 4, 20 # p_argv = p_argv + 5325 # (i.e. skip the 5 elements we already processed)326 addi 12, 1, 52 # r12 = stack offset (56 - 4)327 328.Lnext:329 lwzu 0, 4(4)330 stwu 0, 4(12)331 bdnz .Lnext332 333.Lcall:334 stw 2, 20(1) # Save the TOC pointer to the linkage area335// Load the actual function address from the function descriptor.336 lwz 12, 0(3) # Function address337 lwz 2, 4(3) # TOC pointer338 lwz 11, 8(3) # Environment pointer339 340 addi 3, 31, -12 # r3 = >id341 addi 4, 31, -16 # r4 = &tid342 343 mtctr 12 # CTR = function address344 bctrl # Branch to CTR345 lwz 2, 20(1) # Restore TOC pointer from linkage area346 347#if OMPT_SUPPORT348 li 3, 0349 stw 3, 0(30)350#endif351 352 li 3, 1353 354#if OMPT_SUPPORT355 lwz 30, -8(31) # Restore r30 from the saved value on the stack356#endif357 358 mr 1, 31359 lwz 31, -4(1) # Restore r31 from the saved value on the stack360 lwz 0, 8(1)361 mtlr 0 # Restore LR from the linkage area362 blr # Branch to LR363 364#endif // KMP_ARCH_PPC64_XCOFF365 366.Lfunc_end0:367 .vbyte 4, 0x00000000 # Traceback table begin368 .byte 0x00 # Version = 0369 .byte 0x09 # Language = CPlusPlus370 .byte 0x20 # -IsGlobalLinkage, -IsOutOfLineEpilogOrPrologue371 # +HasTraceBackTableOffset, -IsInternalProcedure372 # -HasControlledStorage, -IsTOCless373 # -IsFloatingPointPresent374 # -IsFloatingPointOperationLogOrAbortEnabled375 .byte 0x61 # -IsInterruptHandler, +IsFunctionNamePresent, +IsAllocaUsed376 # OnConditionDirective = 0, -IsCRSaved, +IsLRSaved377 .byte 0x80 # +IsBackChainStored, -IsFixup, NumOfFPRsSaved = 0378#if OMPT_SUPPORT379 .byte 0x02 # -HasExtensionTable, -HasVectorInfo, NumOfGPRsSaved = 2380 .byte 0x06 # NumberOfFixedParms = 6381#else382 .byte 0x01 # -HasExtensionTable, -HasVectorInfo, NumOfGPRsSaved = 1383 .byte 0x05 # NumberOfFixedParms = 5384#endif385 .byte 0x01 # NumberOfFPParms = 0, +HasParmsOnStack386 .vbyte 4, 0x00000000 # Parameter type = i, i, i, i, i387 .vbyte 4, .Lfunc_end0-.__kmp_invoke_microtask # Function size388 .vbyte 2, 0x0016 # Function name len = 22389 .byte "__kmp_invoke_microtask" # Function Name390 .byte 0x1f # AllocaRegister = 31391 # -- End function392 393// -- End __kmp_invoke_microtask394 395// Support for unnamed common blocks.396 397 .comm .gomp_critical_user_, 32, 3398#if KMP_ARCH_PPC64_XCOFF399 .csect __kmp_unnamed_critical_addr[RW],3400#else401 .csect __kmp_unnamed_critical_addr[RW],2402#endif403 .globl __kmp_unnamed_critical_addr[RW]404 .ptr .gomp_critical_user_405 406// -- End unnamed common block407 408 .toc409 410#endif // KMP_OS_AIX411