Magnus pushed to branch wip/mangoiv/ghc-9.12-bp at Glasgow Haskell Compiler / GHC Commits: 9f20f0e9 by ARATA Mizuki at 2026-05-14T17:08:40+02:00 rts: Align stack to 64-byte boundary in StgRun on x86 When LLVM spills AVX/AVX-512 vector registers to the stack, it requires 32-byte (__m256) or 64-byte (__m512) alignment. If the stack is not sufficiently aligned, LLVM inserts a realignment prologue that reserves %rbp as a frame pointer, conflicting with GHC's use of %rbp as an STG callee-saved register and breaking the tail-call-based calling convention. Previously, GHC worked around this by lying to LLVM about the stack alignment and rewriting aligned vector loads/stores (VMOVDQA, VMOVAPS) to unaligned ones (VMOVDQU, VMOVUPS) in the LLVM Mangler. This had two problems: - It did not extend to AVX-512, which requires 64-byte alignment. (#26595) - When Haskell calls a C function that takes __m256/__m512 arguments on the stack, the callee requires genuine alignment, which could cause a segfault. (#26822) This patch genuinely aligns the stack to 64 bytes in StgRun by saving the original stack pointer before alignment and restoring it in StgReturn. We now unconditionally advertise 64-byte stack alignment to LLVM for all x86 targets, making rewriteAVX in the LLVM Mangler unnecessary. STG_RUN_STACK_FRAME_SIZE is increased from 48 to 56 bytes on non-Windows x86-64 to store the saved stack pointer. Closes #26595 and #26822 Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> (cherry picked from commit a5ec467ee3d4e77c026437a545981269acde3434) - - - - - 13 changed files: - compiler/GHC/CmmToLlvm.hs - compiler/GHC/CmmToLlvm/Config.hs - compiler/GHC/CmmToLlvm/Mangler.hs - compiler/GHC/Driver/Config/CmmToLlvm.hs - rts/StgCRun.c - rts/include/rts/Constants.h - + testsuite/tests/simd/should_run/StackAlignment32.hs - + testsuite/tests/simd/should_run/StackAlignment32.stdout - + testsuite/tests/simd/should_run/StackAlignment32_main.c - + testsuite/tests/simd/should_run/StackAlignment64.hs - + testsuite/tests/simd/should_run/StackAlignment64.stdout - + testsuite/tests/simd/should_run/StackAlignment64_main.c - testsuite/tests/simd/should_run/all.T Changes: ===================================== compiler/GHC/CmmToLlvm.hs ===================================== @@ -217,11 +217,15 @@ cmmMetaLlvmPrelude = do Nothing -> [ MetaStr name ] platform <- getPlatform - cfg <- getConfig let stack_alignment_metas = case platformArch platform of - ArchX86_64 | llvmCgAvxEnabled cfg -> [mkStackAlignmentMeta 32] - _ -> [] + -- LLVM inserts stack realignment prologue/epilogue when it wants to place __m256 or __m512 on the stack. + -- However, it reserves %rbp as the frame pointer, conflicting with our use of it. + -- Therefore, we tell LLVM that the stack is already aligned to avoid stack realignment. + -- See also Note [Stack Alignment on X86] and https://gitlab.haskell.org/ghc/ghc/-/issues/26595. + ArchX86 -> [mkStackAlignmentMeta 64] + ArchX86_64 -> [mkStackAlignmentMeta 64] + _ -> [] let codel_model_metas = case platformArch platform of -- FIXME: We should not rely on LLVM ===================================== compiler/GHC/CmmToLlvm/Config.hs ===================================== @@ -22,7 +22,6 @@ data LlvmCgConfig = LlvmCgConfig , llvmCgContext :: !SDocContext -- ^ Context for LLVM code generation , llvmCgFillUndefWithGarbage :: !Bool -- ^ Fill undefined literals with garbage values , llvmCgSplitSection :: !Bool -- ^ Split sections - , llvmCgAvxEnabled :: !Bool , llvmCgBmiVersion :: Maybe BmiVersion -- ^ (x86) BMI instructions , llvmCgLlvmVersion :: Maybe LlvmVersion -- ^ version of Llvm we're using , llvmCgDoWarn :: !Bool -- ^ True ==> warn unsupported Llvm version ===================================== compiler/GHC/CmmToLlvm/Mangler.hs ===================================== @@ -38,7 +38,7 @@ llvmFixupAsm platform f1 f2 = {-# SCC "llvm_mangler" #-} -- | These are the rewrites that the mangler will perform rewrites :: [Rewrite] -rewrites = [rewriteSymType, rewriteAVX, rewriteCall, rewriteJump] +rewrites = [rewriteSymType, rewriteCall, rewriteJump] type Rewrite = Platform -> B.ByteString -> Maybe B.ByteString @@ -85,23 +85,6 @@ rewriteSymType _ l funcType = prefix `B.cons` B.pack "function" objType = prefix `B.cons` B.pack "object" --- | This rewrites aligned AVX instructions to their unaligned counterparts on --- x86-64. This is necessary because the stack is not adequately aligned for --- aligned AVX spills, so LLVM would emit code that adjusts the stack pointer --- and disable tail call optimization. Both would be catastrophic here so GHC --- tells LLVM that the stack is 32-byte aligned (even though it isn't) and then --- rewrites the instructions in the mangler. -rewriteAVX :: Rewrite -rewriteAVX platform s - | not isX86_64 = Nothing - | isVmovdqa s = Just $ replaceOnce (B.pack "vmovdqa") (B.pack "vmovdqu") s - | isVmovap s = Just $ replaceOnce (B.pack "vmovap") (B.pack "vmovup") s - | otherwise = Nothing - where - isX86_64 = platformArch platform == ArchX86_64 - isVmovdqa = B.isPrefixOf (B.pack "vmovdqa") - isVmovap = B.isPrefixOf (B.pack "vmovap") - -- | This rewrites (tail) calls to avoid creating PLT entries for -- functions on riscv64. The replacement will load the address from the -- GOT, which is resolved to point to the real address of the function. ===================================== compiler/GHC/Driver/Config/CmmToLlvm.hs ===================================== @@ -23,7 +23,6 @@ initLlvmCgConfig logger config_cache dflags = do , llvmCgContext = initSDocContext dflags PprCode , llvmCgFillUndefWithGarbage = gopt Opt_LlvmFillUndefWithGarbage dflags , llvmCgSplitSection = gopt Opt_SplitSections dflags - , llvmCgAvxEnabled = isAvxEnabled dflags , llvmCgBmiVersion = case platformArch (targetPlatform dflags) of ArchX86_64 -> bmiVersion dflags ArchX86 -> bmiVersion dflags ===================================== rts/StgCRun.c ===================================== @@ -108,19 +108,24 @@ StgFunPtr StgReturn(void) /* * Note [Stack Alignment on X86] * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ - * On X86 (both 32bit and 64bit) we keep the stack aligned on function calls at - * a 16-byte boundary. This is done because on a number of architectures the - * ABI requires this (e.g. the System V AMD64 ABI, Mac OS X 32-bit/64-bit ABIs, - * and the Win64 ABI) as well as interfacing with * other libraries through the - * FFI. + * On X86, we keep the stack aligned on function calls. We use 64-byte alignment + * because it is required for passing AVX-512 vector types (__m512 and friends) + * as arguments. See #26822 for a problem caused by insufficient stack alignment. * - * As part of this arrangement we must maintain the stack at a 16-byte boundary - * - word_size-bytes (so 16n - 4 for i386 and 16n - 8 for x64) on entry to a - * procedure since both GCC and LLVM expect this. This is because the stack - * should have been 16-byte boundary aligned and then a call made which pushes - * a return address onto the stack (so word_size more space used). In STG code - * we only jump to other STG procedures, so we maintain the 16n - word_size - * alignment for these jumps. + * Strictly speaking, we only need 16-byte alignment when we do not use AVX or + * AVX-512, but dispatching based on the available ISA extensions would + * complicate things. Always using 64-byte alignment is simpler. + * + * However, the C ABI only requires 16-byte alignment, which is just sufficient + * for SSE2 vectors. Therefore, we dynamically align the stack with + * `and{l,q} $-64`, and save the original stack pointer. + * + * On entry to a procedure, we must maintain the stack at a 64-byte boundary + * - word_size-bytes (so 64n - 4 for i386 and 64n - 8 for x64) since both GCC + * and LLVM expect this. This is because the stack should have been aligned to + * a 64-byte boundary, and then a call made which pushes a return address onto + * the stack (so word_size more space used). In STG code we only jump to other + * STG procedures, so we maintain the 64n - word_size alignment for these jumps. * * This gives us binary compatibility with LLVM and GCC as well as dealing * with the FFI. Previously we just maintained a 16n byte alignment for @@ -157,51 +162,84 @@ StgFunPtr StgReturn(void) * Concretely this means we must always keep the stack valid. * */ +/* Stack layout on x86-32: + +-----------------------------+ <------ esp + | 4-byte padding | + |-----------------------------| <------ esp + 4 (64-byte aligned) + | | + | RESERVED_C_STACK_BYTES ~16k | + | | + |-----------------------------| <------ 64-byte aligned + | ebx | + |-----------------------------| + | esi | + |-----------------------------| + | edi | + |-----------------------------| + | ebp | + |-----------------------------| + | saved esp |----+ + |-----------------------------| | + | (padding for alignment) | | + |-----------------------------| <--+ + | eip saved by call StgRun | + | in schedule() | + +-----------------------------+ + | the function pointer | + +-----------------------------+ + | BaseReg | + +-----------------------------+ + ... + schedule() stack frame + + Lower addresses on the top + */ + +#if RESERVED_C_STACK_BYTES % 64 != 0 +#error "RESERVED_C_STACK_BYTES must be a multiple of 64. If you are changing it, please make sure %esp+4 is a multiple of 64" +#endif static void STG_USED StgRunIsImplementedInAssembler(void) { __asm__ volatile ( STG_GLOBAL STG_RUN "\n" -#if !defined(mingw32_HOST_OS) STG_HIDDEN STG_RUN "\n" -#endif STG_RUN ":\n\t" /* * move %esp down to reserve an area for temporary storage * during the execution of STG code. * - * The stack pointer has to be aligned to a multiple of 16 - * bytes from here - this is a requirement of the C ABI, so - * that C code can assign SSE2 registers directly to/from - * stack locations. - * - * See Note [Windows Stack allocations] + * We want the stack pointer to be aligned to a multiple of 64 + * bytes from here - this is a requirement of the C ABI for + * AVX-512, so that C code can assign ZMM registers directly + * to/from stack locations. */ -#if defined(mingw32_HOST_OS) - "movl %0, %%eax\n\t" - "call ___chkstk_ms\n\t" -#endif - "subl %0, %%esp\n\t" + /* We no longer support i386 Windows, so no need to call __chkstk_ms */ + + /* Save the original esp in eax */ + "movl %%esp, %%eax\n\t" + "subl $20, %%esp\n\t" /* area to save 5 registers */ + "andl $-64, %%esp\n\t" /* * save callee-saves registers on behalf of the STG code. */ - "movl %%esp, %%eax\n\t" - "addl %0-16, %%eax\n\t" - "movl %%ebx,0(%%eax)\n\t" - "movl %%esi,4(%%eax)\n\t" - "movl %%edi,8(%%eax)\n\t" - "movl %%ebp,12(%%eax)\n\t" + "movl %%ebx, 0(%%esp)\n\t" + "movl %%esi, 4(%%esp)\n\t" + "movl %%edi, 8(%%esp)\n\t" + "movl %%ebp, 12(%%esp)\n\t" + "movl %%eax, 16(%%esp)\n\t" + "subl %0, %%esp\n\t" /* * Set BaseReg */ - "movl 24(%%eax),%%ebx\n\t" + "movl 8(%%eax),%%ebx\n\t" /* * grab the function argument from the stack */ - "movl 20(%%eax),%%eax\n\t" + "movl 4(%%eax),%%eax\n\t" /* * jump to it */ @@ -216,17 +254,16 @@ StgRunIsImplementedInAssembler(void) * restore callee-saves registers. (Don't stomp on %%eax!) */ "movl %%esp, %%edx\n\t" - "addl %0-16, %%edx\n\t" + "addl %0, %%edx\n\t" "movl 0(%%edx),%%ebx\n\t" /* restore the registers saved above */ "movl 4(%%edx),%%esi\n\t" "movl 8(%%edx),%%edi\n\t" "movl 12(%%edx),%%ebp\n\t" - - "addl %0, %%esp\n\t" + "movl 16(%%edx),%%esp\n\t" "ret" - : : "i" (RESERVED_C_STACK_BYTES + 16) - // + 16 to make room for the 4 registers we have to save + : : "i" (RESERVED_C_STACK_BYTES + 4) + // + 4 to mimic `calll` // See Note [Stack Alignment on X86] ); } @@ -302,10 +339,12 @@ is as follows: C STACK "ADDRESS SPACE" | v +-----------------------------+ <------ rsp + | 8-byte padding | + |-----------------------------| <------ rsp + 8 (64-byte aligned) | | | RESERVED_C_STACK_BYTES ~16k | | | - |-----------------------------| + |-----------------------------| <------ 64-byte aligned | rbx || |-----------------------------| \ | rbp | | @@ -315,9 +354,13 @@ is as follows: | r13 | | STG_RUN_STACK_FRAME_SIZE |-----------------------------| / | r14 | | - |-----------------------------| / - | r15 | | - |-----------------------------|/ + |-----------------------------| | + | r15 | / + |-----------------------------| | + | saved rsp |-+--- + |-----------------------------|/ | + | (padding for alignment) | | + |-----------------------------| <--- | rip saved by call StgRun | | in schedule() | +-----------------------------+ @@ -367,6 +410,10 @@ stack unwinding. */ +#if RESERVED_C_STACK_BYTES % 64 != 0 +#error "RESERVED_C_STACK_BYTES must be a multiple of 64. If you are changing it, please make sure %rsp+8 is a multiple of 64" +#endif + static void STG_USED StgRunIsImplementedInAssembler(void) { @@ -385,9 +432,22 @@ StgRunIsImplementedInAssembler(void) #if defined(mingw32_HOST_OS) "movq %1, %%rax\n\t" "addq %0, %%rax\n\t" + "addq $63, %%rax\n\t" /* extra space for 64-byte alignment */ "callq ___chkstk_ms\n\t" #endif + /* + * Save the original rsp in r11 (caller-saved, so we don't need to + * preserve it across calls). We need this to restore rsp after + * the 64-byte alignment. + */ + "movq %%rsp, %%r11\n\t" + /* + * Allocate space for saved registers, then align to 64-byte boundary. + * The alignment is required for AVX-512 instructions. + * See Note [Stack Alignment on X86]. + */ "subq %1, %%rsp\n\t" + "andq $-64, %%rsp\n\t" "movq %%rsp, %%rax\n\t" "subq %0, %%rsp\n\t" "movq %%rbx,0(%%rax)\n\t" @@ -396,6 +456,7 @@ StgRunIsImplementedInAssembler(void) "movq %%r13,24(%%rax)\n\t" "movq %%r14,32(%%rax)\n\t" "movq %%r15,40(%%rax)\n\t" + "movq %%r11,48(%%rax)\n\t" /* save original rsp */ #if defined(mingw32_HOST_OS) /* * Additional callee saved registers on Win64. This must match @@ -403,25 +464,35 @@ StgRunIsImplementedInAssembler(void) * both represent the Win64 calling convention. * * Note that we must save the entire 128-bit width of the XMM - * registers, as noted in #21465. Moreover, note that, due to the - * presence of the return address on the stack, %rsp+8 is - * 16-byte aligned. Since MOVAPS requires memory operands to be aligned - * to 16-bytes, we must add a word of padding here. + * registers, as noted in #21465. Since we align the stack to + * 64 bytes, MOVAPS alignment requirements are satisfied. + * + * The Win64 calling convention says the upper YMM/ZMM parts are + * volatile, so we do not need to save them. See Microsoft Learn + * for the description of the calling convention: + * https://learn.microsoft.com/en-us/cpp/build/x64-calling-convention?view=msvc... + * + * Layout (offsets from rax): + * 0-47: rbx, rbp, r12, r13, r14, r15 + * 48: original rsp + * 56: rdi + * 64: rsi + * 72: padding for 16-byte alignment of XMM registers + * 80-239: xmm6-xmm15 (10 * 16 bytes) */ - "movq %%rdi, 48(%%rax)\n\t" - "movq %%rsi, 56(%%rax)\n\t" - /* 8 bytes of padding for alignment */ - "movaps %%xmm6, 72(%%rax)\n\t" - "movaps %%xmm7, 88(%%rax)\n\t" - "movaps %%xmm8, 104(%%rax)\n\t" - "movaps %%xmm9, 120(%%rax)\n\t" - "movaps %%xmm10,136(%%rax)\n\t" - "movaps %%xmm11,152(%%rax)\n\t" - "movaps %%xmm12,168(%%rax)\n\t" - "movaps %%xmm13,184(%%rax)\n\t" - "movaps %%xmm14,200(%%rax)\n\t" - "movaps %%xmm15,216(%%rax)\n\t" + "movq %%rdi, 56(%%rax)\n\t" + "movq %%rsi, 64(%%rax)\n\t" /* 8 bytes of padding for alignment */ + "movaps %%xmm6, 80(%%rax)\n\t" + "movaps %%xmm7, 96(%%rax)\n\t" + "movaps %%xmm8, 112(%%rax)\n\t" + "movaps %%xmm9, 128(%%rax)\n\t" + "movaps %%xmm10,144(%%rax)\n\t" + "movaps %%xmm11,160(%%rax)\n\t" + "movaps %%xmm12,176(%%rax)\n\t" + "movaps %%xmm13,192(%%rax)\n\t" + "movaps %%xmm14,208(%%rax)\n\t" + "movaps %%xmm15,224(%%rax)\n\t" #endif #if defined(ENABLE_UNWINDING) @@ -431,6 +502,11 @@ StgRunIsImplementedInAssembler(void) * * N.B. We don't support unwinding on Darwin due to * various toolchain insanity. + * + * Note: Due to dynamic 64-byte alignment, the exact stack offset + * varies at runtime. The original rsp is saved at offset 48 in the + * register save area. We use a DWARF expression to compute the + * CFA from the saved rsp value. */ ".cfi_def_cfa rsp, 0\n\t" ".cfi_offset rbx, %c2\n\t" @@ -439,30 +515,69 @@ StgRunIsImplementedInAssembler(void) ".cfi_offset r13, %c5\n\t" ".cfi_offset r14, %c6\n\t" ".cfi_offset r15, %c7\n\t" - ".cfi_offset rip, %c8\n\t" - ".cfi_escape " // DW_CFA_val_expression is not expressible otherwise - "0x16, " // DW_CFA_val_expression - "0x07, " // register num 7 - rsp - "0x04, " // block length - "0x77, " // DW_OP_breg7 - signed LEB128 offset from rsp -#define RSP_DELTA (RESERVED_C_STACK_BYTES + STG_RUN_STACK_FRAME_SIZE + 8) - "%c9" // signed LEB128 encoded delta - byte 1 -#if (RSP_DELTA >> 7) > 0 - ", %c10" // signed LEB128 encoded delta - byte 2 + /* + * The original rsp is saved at RESERVED_C_STACK_BYTES + 56 from + * current rsp. We use DW_CFA_expression to indicate that rsp's + * value can be found by loading from that stack location. + */ +#define RSP_DELTA (RESERVED_C_STACK_BYTES + 56) + ".cfi_escape " + "0x10, " // DW_CFA_expression + "0x07, " // register num 7 - rsp +#if (RSP_DELTA >> 20) > 0 + "0x05, " // block length = 5 +#elif (RSP_DELTA >> 13) > 0 + "0x04, " // block length = 4 +#elif (RSP_DELTA >> 6) > 0 + "0x03, " // block length = 3 +#else + "0x02, " // block length = 2 #endif - -#if (RSP_DELTA >> 14) > 0 - ", %c11" // signed LEB128 encoded delta - byte 3 + "0x77, " // DW_OP_breg7 (rsp + offset) + "%c8" // signed LEB128 offset to saved rsp (RESERVED_C_STACK_BYTES + 56) +#if (RSP_DELTA >> 6) > 0 + ", %c9" // signed LEB128 encoded delta - byte 2 #endif - -#if (RSP_DELTA >> 21) > 0 - ", %c12" // signed LEB128 encoded delta - byte 4 +#if (RSP_DELTA >> 13) > 0 + ", %c10" // signed LEB128 encoded delta - byte 3 #endif - -#if (RSP_DELTA >> 28) > 0 +#if (RSP_DELTA >> 20) > 0 + ", %c11" // signed LEB128 encoded delta - byte 4 +#endif +#if (RSP_DELTA >> 27) > 0 #error "RSP_DELTA too big" #endif "\n\t" + /* + * The return address (rip) is at the original rsp location. + * Since original rsp is saved at offset 48 in register save area, + * rip = *saved_rsp = **(rsp + 8 + RESERVED_C_STACK_BYTES + 48). + */ + ".cfi_escape " + "0x10, " // DW_CFA_expression + "0x10, " // register num 16 - rip +#if (RSP_DELTA >> 20) > 0 + "0x06, " // block length = 6 +#elif (RSP_DELTA >> 13) > 0 + "0x05, " // block length = 5 +#elif (RSP_DELTA >> 6) > 0 + "0x04, " // block length = 4 +#else + "0x03, " // block length = 3 +#endif + "0x77, " // DW_OP_breg7 (rsp + offset) + "%c8" // signed LEB128 offset to saved rsp +#if (RSP_DELTA >> 6) > 0 + ", %c9" // signed LEB128 encoded delta - byte 2 +#endif +#if (RSP_DELTA >> 13) > 0 + ", %c10" // signed LEB128 encoded delta - byte 3 +#endif +#if (RSP_DELTA >> 20) > 0 + ", %c11" // signed LEB128 encoded delta - byte 4 +#endif + ", 0x06" // DW_OP_deref + "\n\t" #endif /* defined(ENABLE_UNWINDING) */ /* @@ -510,55 +625,50 @@ StgRunIsImplementedInAssembler(void) "movq 32(%%rsp),%%r14\n\t" "movq 40(%%rsp),%%r15\n\t" #if defined(mingw32_HOST_OS) - "movq 48(%%rsp),%%rdi\n\t" - "movq 56(%%rsp),%%rsi\n\t" + "movq 56(%%rsp),%%rdi\n\t" + "movq 64(%%rsp),%%rsi\n\t" /* 8 bytes of padding for alignment */ - "movaps 72(%%rsp),%%xmm6\n\t" - "movaps 88(%%rsp),%%xmm7\n\t" - "movaps 104(%%rsp),%%xmm8\n\t" - "movaps 120(%%rsp),%%xmm9\n\t" - "movaps 136(%%rsp),%%xmm10\n\t" - "movaps 152(%%rsp),%%xmm11\n\t" - "movaps 168(%%rsp),%%xmm12\n\t" - "movaps 184(%%rsp),%%xmm13\n\t" - "movaps 200(%%rsp),%%xmm14\n\t" - "movaps 216(%%rsp),%%xmm15\n\t" - /* 8 bytes of padding for alignment */ -#endif - "addq %1, %%rsp\n\t" + "movaps 80(%%rsp),%%xmm6\n\t" + "movaps 96(%%rsp),%%xmm7\n\t" + "movaps 112(%%rsp),%%xmm8\n\t" + "movaps 128(%%rsp),%%xmm9\n\t" + "movaps 144(%%rsp),%%xmm10\n\t" + "movaps 160(%%rsp),%%xmm11\n\t" + "movaps 176(%%rsp),%%xmm12\n\t" + "movaps 192(%%rsp),%%xmm13\n\t" + "movaps 208(%%rsp),%%xmm14\n\t" + "movaps 224(%%rsp),%%xmm15\n\t" +#endif + "movq 48(%%rsp),%%rsp\n\t" /* restore original rsp */ "retq" : - : "i"(RESERVED_C_STACK_BYTES), + : "i"(RESERVED_C_STACK_BYTES + 8), "i"(STG_RUN_STACK_FRAME_SIZE /* stack frame size */), - "i"(RESERVED_C_STACK_BYTES /* rbx relative to cfa (rsp) */), - "i"(RESERVED_C_STACK_BYTES + 8 /* rbp relative to cfa (rsp) */), - "i"(RESERVED_C_STACK_BYTES + 16 /* r12 relative to cfa (rsp) */), - "i"(RESERVED_C_STACK_BYTES + 24 /* r13 relative to cfa (rsp) */), - "i"(RESERVED_C_STACK_BYTES + 32 /* r14 relative to cfa (rsp) */), - "i"(RESERVED_C_STACK_BYTES + 40 /* r15 relative to cfa (rsp) */), - "i"(RESERVED_C_STACK_BYTES + STG_RUN_STACK_FRAME_SIZE - /* rip relative to cfa */) + "i"(RESERVED_C_STACK_BYTES + 8 /* rbx relative to cfa (rsp) */), + "i"(RESERVED_C_STACK_BYTES + 16 /* rbp relative to cfa (rsp) */), + "i"(RESERVED_C_STACK_BYTES + 24 /* r12 relative to cfa (rsp) */), + "i"(RESERVED_C_STACK_BYTES + 32 /* r13 relative to cfa (rsp) */), + "i"(RESERVED_C_STACK_BYTES + 40 /* r14 relative to cfa (rsp) */), + "i"(RESERVED_C_STACK_BYTES + 48 /* r15 relative to cfa (rsp) */) #if defined(ENABLE_UNWINDING) - , "i"((RSP_DELTA & 127) | (128 * ((RSP_DELTA >> 7) > 0))) + /* LEB128-encoded offset to saved rsp (RESERVED_C_STACK_BYTES + 56) */ + , "i"((RSP_DELTA & 127) | (128 * ((RSP_DELTA >> 6) > 0))) /* signed LEB128-encoded delta from rsp - byte 1 */ -#if (RSP_DELTA >> 7) > 0 - , "i"(((RSP_DELTA >> 7) & 127) | (128 * ((RSP_DELTA >> 14) > 0))) +#if (RSP_DELTA >> 6) > 0 + , "i"(((RSP_DELTA >> 7) & 127) | (128 * ((RSP_DELTA >> 13) > 0))) /* signed LEB128-encoded delta from rsp - byte 2 */ #endif - -#if (RSP_DELTA >> 14) > 0 - , "i"(((RSP_DELTA >> 14) & 127) | (128 * ((RSP_DELTA >> 21) > 0))) +#if (RSP_DELTA >> 13) > 0 + , "i"(((RSP_DELTA >> 14) & 127) | (128 * ((RSP_DELTA >> 20) > 0))) /* signed LEB128-encoded delta from rsp - byte 3 */ #endif - -#if (RSP_DELTA >> 21) > 0 - , "i"(((RSP_DELTA >> 21) & 127) | (128 * ((RSP_DELTA >> 28) > 0))) +#if (RSP_DELTA >> 20) > 0 + , "i"(((RSP_DELTA >> 21) & 127) | (128 * ((RSP_DELTA >> 27) > 0))) /* signed LEB128-encoded delta from rsp - byte 4 */ #endif #undef RSP_DELTA - #endif /* defined(ENABLE_UNWINDING) */ ); ===================================== rts/include/rts/Constants.h ===================================== @@ -117,16 +117,14 @@ How large is the stack frame saved by StgRun? world. Used in StgCRun.c. - The size has to be enough to save the registers (see StgCRun) - plus padding if the result is not 16 byte aligned. - See the Note [Stack Alignment on X86] in StgCRun.c for details. + The size has to be enough to save the registers (see StgCRun). -------------------------------------------------------------------------- */ #if defined(x86_64_HOST_ARCH) # if defined(mingw32_HOST_OS) # define STG_RUN_STACK_FRAME_SIZE 240 # else -# define STG_RUN_STACK_FRAME_SIZE 48 +# define STG_RUN_STACK_FRAME_SIZE 56 # endif #endif ===================================== testsuite/tests/simd/should_run/StackAlignment32.hs ===================================== @@ -0,0 +1,15 @@ +{-# LANGUAGE MagicHash, UnboxedTuples, UnliftedFFITypes #-} +module StackAlignment32 where +import GHC.Exts + +foreign import ccall unsafe add10 :: DoubleX4# -> DoubleX4# -> DoubleX4# -> DoubleX4# -> DoubleX4# -> DoubleX4# -> DoubleX4# -> DoubleX4# -> DoubleX4# -> DoubleX4# -> DoubleX4# + +foo :: Double -> IO () +foo (D# x) = do + let a = broadcastDoubleX4# x + b = packDoubleX4# (# 1.0##, 2.0##, 3.0##, 4.0## #) + c = add10 a a a a a a a a a b + (# c0, c1, c2, c3 #) = unpackDoubleX4# c + print (D# c0, D# c1, D# c2, D# c3) + +foreign export ccall foo :: Double -> IO () ===================================== testsuite/tests/simd/should_run/StackAlignment32.stdout ===================================== @@ -0,0 +1,3 @@ +(1.1,2.1,3.1,4.1) +(2.1,3.1,4.1,5.1) +(3.2,4.2,5.2,6.2) ===================================== testsuite/tests/simd/should_run/StackAlignment32_main.c ===================================== @@ -0,0 +1,42 @@ +#include "HsFFI.h" +#include <immintrin.h> + +extern void foo(double x); + +__m256d add10(__m256d a0, __m256d a1, __m256d a2, __m256d a3, __m256d a4, __m256d a5, __m256d a6, __m256d a7, __m256d a8, __m256d a9) +{ + // Test whether the stack is 32-byte aligned. + // Under the System V ABI, the first eight parameters are passed in registers. + // Here, `a8` and `a9` are passed on the stack. + // We want the compiler to emit `vmovapd` so we can verify 32-byte stack alignment. + // We cannot use `_mm256_add_pd(a0, a9)` because the compiler may fuse the load and + // the addition and emit a `vaddpd m256` form, which permits an unaligned load of `a9`. + // Therefore, we ensure that both operands are read from the stack, + // so that at least one operand is loaded using `vmovapd`. + return _mm256_add_pd(a8, a9); +} + +__attribute__((noinline)) +void baz(void) +{ + // Make the stack pointer shift by 16 bytes on x86-64 + foo(2.2); +} + +__attribute__((noinline)) +void bar(void) +{ + // Make the stack pointer shift by 16 bytes on x86-64 + foo(1.1); + baz(); +} + +int main(int argc, char *argv[]) +{ + hs_init(&argc, &argv); + + foo(0.1); + bar(); + + hs_exit(); +} ===================================== testsuite/tests/simd/should_run/StackAlignment64.hs ===================================== @@ -0,0 +1,15 @@ +{-# LANGUAGE MagicHash, UnboxedTuples, UnliftedFFITypes #-} +module StackAlignment64 where +import GHC.Exts + +foreign import ccall unsafe add10 :: DoubleX8# -> DoubleX8# -> DoubleX8# -> DoubleX8# -> DoubleX8# -> DoubleX8# -> DoubleX8# -> DoubleX8# -> DoubleX8# -> DoubleX8# -> DoubleX8# + +foo :: Double -> IO () +foo (D# x) = do + let a = broadcastDoubleX8# x + b = packDoubleX8# (# 1.0##, 2.0##, 3.0##, 4.0##, 5.0##, 6.0##, 7.0##, 8.0## #) + c = add10 a a a a a a a a a b + (# c0, c1, c2, c3, c4, c5, c6, c7 #) = unpackDoubleX8# c + print (D# c0, D# c1, D# c2, D# c3, D# c4, D# c5, D# c6, D# c7) + +foreign export ccall foo :: Double -> IO () ===================================== testsuite/tests/simd/should_run/StackAlignment64.stdout ===================================== @@ -0,0 +1,3 @@ +(1.1,2.1,3.1,4.1,5.1,6.1,7.1,8.1) +(2.1,3.1,4.1,5.1,6.1,7.1,8.1,9.1) +(3.2,4.2,5.2,6.2,7.2,8.2,9.2,10.2) ===================================== testsuite/tests/simd/should_run/StackAlignment64_main.c ===================================== @@ -0,0 +1,42 @@ +#include "HsFFI.h" +#include <immintrin.h> + +extern void foo(double x); + +__m512d add10(__m512d a0, __m512d a1, __m512d a2, __m512d a3, __m512d a4, __m512d a5, __m512d a6, __m512d a7, __m512d a8, __m512d a9) +{ + // Test whether the stack is 64-byte aligned. + // Under the System V ABI, the first eight parameters are passed in registers. + // Here, `a8` and `a9` are passed on the stack. + // We want the compiler to emit `vmovapd` so we can verify 64-byte stack alignment. + // We cannot use `_mm512_add_pd(a0, a9)` because the compiler may fuse the load and + // the addition and emit a `vaddpd m512` form, which permits an unaligned load of `a9`. + // Therefore, we ensure that both operands are read from the stack, + // so that at least one operand is loaded using `vmovapd`. + return _mm512_add_pd(a8, a9); +} + +__attribute__((noinline)) +void baz(void) +{ + // Make the stack pointer shift by 16 bytes on x86-64 + foo(2.2); +} + +__attribute__((noinline)) +void bar(void) +{ + // Make the stack pointer shift by 16 bytes on x86-64 + foo(1.1); + baz(); +} + +int main(int argc, char *argv[]) +{ + hs_init(&argc, &argv); + + foo(0.1); + bar(); + + hs_exit(); +} ===================================== testsuite/tests/simd/should_run/all.T ===================================== @@ -92,7 +92,7 @@ test('T26411b', [], compile_and_run, ['-O']) # Even if the CPU we run on doesn't support *executing* those tests we should try to # compile them. # Currently even for compilation we only support 256+ bit on x86 -only_V32_plus_compilation_support = unless(arch('x86_64'), skip) +only_V32_plus_compilation_support = unless(arch('x86_64') or arch('i386'), skip) test('T25062_V32' , [ extra_hc_opts('-mavx2') @@ -116,17 +116,37 @@ test('T25486', [], compile_and_run, ['']) test('T26410_ffi' , [ only_ways(llvm_ways) # SIMD NCG TODO: support 512-bit wide vectors - , unless(arch('x86_64') and have_cpu_feature('avx512f'), skip) + , unless(arch('x86_64') or arch('i386'), skip) , extra_hc_opts('-mavx512f -optc -mavx512f -optlc -mcpu=penryn') - , when(opsys('mingw32'), fragile(26595)) ] , compile_and_run if have_cpu_feature('avx512f') else compile , ['T26410_ffi_c.c']) test('T26410_prim' , [ only_ways(llvm_ways) # SIMD NCG TODO: support 512-bit wide vectors - , unless(arch('x86_64') and have_cpu_feature('avx512f'), skip) + , unless(arch('x86_64') or arch('i386'), skip) , extra_hc_opts('-mavx512f -optlc -mcpu=penryn') ] , compile_and_run if have_cpu_feature('avx512f') else compile , ['']) + +test('FloatConstant', [], compile_and_run, ['']) +test('IntConstant', [], compile_and_run, ['']) + +test('StackAlignment32' + , [ only_ways(llvm_ways) # SIMD NCG TODO: support 256-bit wide vectors + , unless(arch('x86_64') or arch('i386'), skip) + , extra_hc_opts("-mavx -optc -mavx -no-hs-main") + ] + , compile_and_run if have_cpu_feature('avx') else compile + , ['StackAlignment32_main.c'] + ) + +test('StackAlignment64' + , [ only_ways(llvm_ways) # SIMD NCG TODO: support 512-bit wide vectors + , unless(arch('x86_64') or arch('i386'), skip) + , extra_hc_opts("-mavx512f -optc -mavx512f -no-hs-main") + ] + , compile_and_run if have_cpu_feature('avx512f') else compile + , ['StackAlignment64_main.c'] + ) View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/9f20f0e9f5009557f9ef0c3268ef4271... -- View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/9f20f0e9f5009557f9ef0c3268ef4271... You're receiving this email because of your account on gitlab.haskell.org.
participants (1)
-
Magnus (＠MangoIV)