From c062fed4275c07e5c7a1a18d0779144519bec799 Mon Sep 17 00:00:00 2001 From: Thiago Macieira Date: Mon, 20 Dec 2021 21:11:17 -0300 Subject: [PATCH] qsimd_p.h: add a hack to allow AVX to work with MinGW GCC is unable to emit the SEH metadata about the stack aligning that is required to execute AVX aligned instructions (VMOVDQA, VMOVAPS, etc.), so it just doesn't align the stack. That causes crashes on a 50/50 chance every time the compiler attempts to address a stack-aligned variable. In a debug-mode build, because it always loads & saves everything on the stack, the chance of a crash happening is a near certainty. So we hack around it by going behind the compiler's back and instructing the assembler to emit the unaligned counterparts of the instructions every time the compiler wished to emit the aligned one. There's no performance penalty: if the variable is actually aligned, the unaligned instruction executes in the exact same time. Change-Id: Ib42b3adc93bf4d43bd55fffd16c29cac0da18972 Reviewed-by: Lars Knoll --- config.tests/x86_simd/main.cpp | 3 --- src/corelib/global/qsimd_p.h | 23 +++++++++++++++++++++++ 2 files changed, 23 insertions(+), 3 deletions(-) diff --git a/config.tests/x86_simd/main.cpp b/config.tests/x86_simd/main.cpp index 0e7ebed8d9c..29c8b97c123 100644 --- a/config.tests/x86_simd/main.cpp +++ b/config.tests/x86_simd/main.cpp @@ -161,9 +161,6 @@ attribute_target("sha") void test_shani() #endif #if T(AVX) -# if defined(__WIN64__) && defined(__GNUC__) && !defined(__clang__) -# error "AVX support is broken in 64-bit MinGW - https://gcc.gnu.org/bugzilla/show_bug.cgi?id=49001" -# endif attribute_target("avx") void test_avx() { __m256d a = _mm256_setzero_pd(); diff --git a/src/corelib/global/qsimd_p.h b/src/corelib/global/qsimd_p.h index fdf55291351..57ef30d567e 100644 --- a/src/corelib/global/qsimd_p.h +++ b/src/corelib/global/qsimd_p.h @@ -190,6 +190,29 @@ # define __SSE__ 1 # endif +# if defined(Q_OS_WIN) && defined(Q_CC_GNU) && !defined(Q_CC_INTEL) && !defined(Q_CC_CLANG) +// 64-bit GCC on Windows does not support AVX, so we hack around it by forcing +// it to emit unaligned loads & stores +// See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=49001 +asm( + ".macro vmovapd args:vararg\n" + " vmovupd \\args\n" + ".endm\n" + ".macro vmovaps args:vararg\n" + " vmovups \\args\n" + ".endm\n" + ".macro vmovdqa args:vararg\n" + " vmovdqu \\args\n" + ".endm\n" + ".macro vmovdqa32 args:vararg\n" + " vmovdqu32 \\args\n" + ".endm\n" + ".macro vmovdqa64 args:vararg\n" + " vmovdqu64 \\args\n" + ".endm\n" +); +# endif + # if defined(Q_CC_GNU) && !defined(Q_CC_INTEL) && !defined(Q_OS_WASM) // GCC 4.4 and Clang 2.8 added a few more intrinsics there # include