Skip to content

Commit 0dba538

Browse files
phoebewangphoebewang
authored
[X86][AVX10.2] Support YMM rounding new instructions (#101825)
Ref.: https://cdrdv2.intel.com/v1/dl/getContent/828965
1 parent 3c483b8 commit 0dba538

File tree

17 files changed

+16458
-13
lines changed

17 files changed

+16458
-13
lines changed

‎clang/include/clang/Basic/BuiltinsX86.def‎

Lines changed: 120 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1966,6 +1966,126 @@ TARGET_BUILTIN(__builtin_ia32_mpsadbw512, "V32sV64cV64cIc", "ncV:512:", "avx10.2
19661966
TARGET_BUILTIN(__builtin_ia32_vaddpd256_round, "V4dV4dV4dIi", "nV:256:", "avx10.2-256")
19671967
TARGET_BUILTIN(__builtin_ia32_vaddph256_round, "V16xV16xV16xIi", "nV:256:", "avx10.2-256")
19681968
TARGET_BUILTIN(__builtin_ia32_vaddps256_round, "V8fV8fV8fIi", "nV:256:", "avx10.2-256")
1969+
TARGET_BUILTIN(__builtin_ia32_vcmppd256_round_mask, "UcV4dV4dIiUcIi", "nV:256:", "avx10.2-256")
1970+
TARGET_BUILTIN(__builtin_ia32_vcmpph256_round_mask, "UsV16xV16xIiUsIi", "nV:256:", "avx10.2-256")
1971+
TARGET_BUILTIN(__builtin_ia32_vcmpps256_round_mask, "UcV8fV8fIiUcIi", "nV:256:", "avx10.2-256")
1972+
TARGET_BUILTIN(__builtin_ia32_vcvtdq2ph256_round_mask, "V8xV8iV8xUcIi", "nV:256:", "avx10.2-256")
1973+
TARGET_BUILTIN(__builtin_ia32_vcvtdq2ps256_round_mask, "V8fV8iV8fUcIi", "nV:256:", "avx10.2-256")
1974+
TARGET_BUILTIN(__builtin_ia32_vcvtpd2dq256_round_mask, "V4iV4dV4iUcIi", "nV:256:", "avx10.2-256")
1975+
TARGET_BUILTIN(__builtin_ia32_vcvtpd2ph256_round_mask, "V8xV4dV8xUcIi", "nV:256:", "avx10.2-256")
1976+
TARGET_BUILTIN(__builtin_ia32_vcvtpd2ps256_round_mask, "V4fV4dV4fUcIi", "nV:256:", "avx10.2-256")
1977+
TARGET_BUILTIN(__builtin_ia32_vcvtpd2qq256_round_mask, "V4LLiV4dV4LLiUcIi", "nV:256:", "avx10.2-256")
1978+
TARGET_BUILTIN(__builtin_ia32_vcvtpd2udq256_round_mask, "V4UiV4dV4UiUcIi", "nV:256:", "avx10.2-256")
1979+
TARGET_BUILTIN(__builtin_ia32_vcvtpd2uqq256_round_mask, "V4ULLiV4dV4ULLiUcIi", "nV:256:", "avx10.2-256")
1980+
TARGET_BUILTIN(__builtin_ia32_vcvtph2dq256_round_mask, "V8iV8xV8iUcIi", "nV:256:", "avx10.2-256")
1981+
TARGET_BUILTIN(__builtin_ia32_vcvtph2pd256_round_mask, "V4dV8xV4dUcIi", "nV:256:", "avx10.2-256")
1982+
TARGET_BUILTIN(__builtin_ia32_vcvtph2psx256_round_mask, "V8fV8xV8fUcIi", "nV:256:", "avx10.2-256")
1983+
TARGET_BUILTIN(__builtin_ia32_vcvtph2qq256_round_mask, "V4LLiV8xV4LLiUcIi", "nV:256:", "avx10.2-256")
1984+
TARGET_BUILTIN(__builtin_ia32_vcvtph2udq256_round_mask, "V8UiV8xV8UiUcIi", "nV:256:", "avx10.2-256")
1985+
TARGET_BUILTIN(__builtin_ia32_vcvtph2uqq256_round_mask, "V4ULLiV8xV4ULLiUcIi", "nV:256:", "avx10.2-256")
1986+
TARGET_BUILTIN(__builtin_ia32_vcvtph2uw256_round_mask, "V16UsV16xV16UsUsIi", "nV:256:", "avx10.2-256")
1987+
TARGET_BUILTIN(__builtin_ia32_vcvtph2w256_round_mask, "V16sV16xV16sUsIi", "nV:256:", "avx10.2-256")
1988+
TARGET_BUILTIN(__builtin_ia32_vcvtps2dq256_round_mask, "V8iV8fV8iUcIi", "nV:256:", "avx10.2-256")
1989+
TARGET_BUILTIN(__builtin_ia32_vcvtps2pd256_round_mask, "V4dV4fV4dUcIi", "nV:256:", "avx10.2-256")
1990+
TARGET_BUILTIN(__builtin_ia32_vcvtps2phx256_round_mask, "V8xV8fV8xUcIi", "nV:256:", "avx10.2-256")
1991+
TARGET_BUILTIN(__builtin_ia32_vcvtps2qq256_round_mask, "V4LLiV4fV4LLiUcIi", "nV:256:", "avx10.2-256")
1992+
TARGET_BUILTIN(__builtin_ia32_vcvtps2udq256_round_mask, "V8UiV8fV8UiUcIi", "nV:256:", "avx10.2-256")
1993+
TARGET_BUILTIN(__builtin_ia32_vcvtps2uqq256_round_mask, "V4ULLiV4fV4ULLiUcIi", "nV:256:", "avx10.2-256")
1994+
TARGET_BUILTIN(__builtin_ia32_vcvtqq2pd256_round_mask, "V4dV4LLiV4dUcIi", "nV:256:", "avx10.2-256")
1995+
TARGET_BUILTIN(__builtin_ia32_vcvtqq2ph256_round_mask, "V8xV4LLiV8xUcIi", "nV:256:", "avx10.2-256")
1996+
TARGET_BUILTIN(__builtin_ia32_vcvtqq2ps256_round_mask, "V4fV4LLiV4fUcIi", "nV:256:", "avx10.2-256")
1997+
TARGET_BUILTIN(__builtin_ia32_vcvttpd2dq256_round_mask, "V4iV4dV4iUcIi", "nV:256:", "avx10.2-256")
1998+
TARGET_BUILTIN(__builtin_ia32_vcvttpd2qq256_round_mask, "V4LLiV4dV4LLiUcIi", "nV:256:", "avx10.2-256")
1999+
TARGET_BUILTIN(__builtin_ia32_vcvttpd2udq256_round_mask, "V4UiV4dV4UiUcIi", "nV:256:", "avx10.2-256")
2000+
TARGET_BUILTIN(__builtin_ia32_vcvttpd2uqq256_round_mask, "V4ULLiV4dV4ULLiUcIi", "nV:256:", "avx10.2-256")
2001+
TARGET_BUILTIN(__builtin_ia32_vcvttph2dq256_round_mask, "V8iV8xV8iUcIi", "nV:256:", "avx10.2-256")
2002+
TARGET_BUILTIN(__builtin_ia32_vcvttph2qq256_round_mask, "V4LLiV8xV4LLiUcIi", "nV:256:", "avx10.2-256")
2003+
TARGET_BUILTIN(__builtin_ia32_vcvttph2udq256_round_mask, "V8UiV8xV8UiUcIi", "nV:256:", "avx10.2-256")
2004+
TARGET_BUILTIN(__builtin_ia32_vcvttph2uqq256_round_mask, "V4ULLiV8xV4ULLiUcIi", "nV:256:", "avx10.2-256")
2005+
TARGET_BUILTIN(__builtin_ia32_vcvttph2uw256_round_mask, "V16UsV16xV16UsUsIi", "nV:256:", "avx10.2-256")
2006+
TARGET_BUILTIN(__builtin_ia32_vcvttph2w256_round_mask, "V16sV16xV16sUsIi", "nV:256:", "avx10.2-256")
2007+
TARGET_BUILTIN(__builtin_ia32_vcvttps2dq256_round_mask, "V8iV8fV8iUcIi", "nV:256:", "avx10.2-256")
2008+
TARGET_BUILTIN(__builtin_ia32_vcvttps2qq256_round_mask, "V4LLiV4fV4LLiUcIi", "nV:256:", "avx10.2-256")
2009+
TARGET_BUILTIN(__builtin_ia32_vcvttps2udq256_round_mask, "V8UiV8fV8UiUcIi", "nV:256:", "avx10.2-256")
2010+
TARGET_BUILTIN(__builtin_ia32_vcvttps2uqq256_round_mask, "V4ULLiV4fV4ULLiUcIi", "nV:256:", "avx10.2-256")
2011+
TARGET_BUILTIN(__builtin_ia32_vcvtudq2ph256_round_mask, "V8xV8UiV8xUcIi", "nV:256:", "avx10.2-256")
2012+
TARGET_BUILTIN(__builtin_ia32_vcvtudq2ps256_round_mask, "V8fV8UiV8fUcIi", "nV:256:", "avx10.2-256")
2013+
TARGET_BUILTIN(__builtin_ia32_vcvtuqq2pd256_round_mask, "V4dV4ULLiV4dUcIi", "nV:256:", "avx10.2-256")
2014+
TARGET_BUILTIN(__builtin_ia32_vcvtuqq2ph256_round_mask, "V8xV4ULLiV8xUcIi", "nV:256:", "avx10.2-256")
2015+
TARGET_BUILTIN(__builtin_ia32_vcvtuqq2ps256_round_mask, "V4fV4ULLiV4fUcIi", "nV:256:", "avx10.2-256")
2016+
TARGET_BUILTIN(__builtin_ia32_vcvtuw2ph256_round_mask, "V16xV16UsV16xUsIi", "nV:256:", "avx10.2-256")
2017+
TARGET_BUILTIN(__builtin_ia32_vcvtw2ph256_round_mask, "V16xV16sV16xUsIi", "nV:256:", "avx10.2-256")
2018+
TARGET_BUILTIN(__builtin_ia32_vdivpd256_round, "V4dV4dV4dIi", "nV:256:", "avx10.2-256")
2019+
TARGET_BUILTIN(__builtin_ia32_vdivph256_round, "V16xV16xV16xIi", "nV:256:", "avx10.2-256")
2020+
TARGET_BUILTIN(__builtin_ia32_vdivps256_round, "V8fV8fV8fIi", "nV:256:", "avx10.2-256")
2021+
TARGET_BUILTIN(__builtin_ia32_vfcmaddcph256_round_mask, "V8fV8fV8fV8fUcIi", "nV:256:", "avx10.2-256")
2022+
TARGET_BUILTIN(__builtin_ia32_vfcmaddcph256_round_maskz, "V8fV8fV8fV8fUcIi", "nV:256:", "avx10.2-256")
2023+
TARGET_BUILTIN(__builtin_ia32_vfcmaddcph256_round_mask3, "V8fV8fV8fV8fUcIi", "nV:256:", "avx10.2-256")
2024+
TARGET_BUILTIN(__builtin_ia32_vfcmulcph256_round_mask, "V8fV8fV8fV8fUcIi", "nV:256:", "avx10.2-256")
2025+
TARGET_BUILTIN(__builtin_ia32_vfixupimmpd256_round_mask, "V4dV4dV4dV4LLiIiUcIi", "nV:256:", "avx10.2-256")
2026+
TARGET_BUILTIN(__builtin_ia32_vfixupimmpd256_round_maskz, "V4dV4dV4dV4LLiIiUcIi", "nV:256:", "avx10.2-256")
2027+
TARGET_BUILTIN(__builtin_ia32_vfixupimmps256_round_mask, "V8fV8fV8fV8iIiUcIi", "nV:256:", "avx10.2-256")
2028+
TARGET_BUILTIN(__builtin_ia32_vfixupimmps256_round_maskz, "V8fV8fV8fV8iIiUcIi", "nV:256:", "avx10.2-256")
2029+
TARGET_BUILTIN(__builtin_ia32_vfmaddpd256_round_mask, "V4dV4dV4dV4dUcIi", "nV:256:", "avx10.2-256")
2030+
TARGET_BUILTIN(__builtin_ia32_vfmaddpd256_round_maskz, "V4dV4dV4dV4dUcIi", "nV:256:", "avx10.2-256")
2031+
TARGET_BUILTIN(__builtin_ia32_vfmaddpd256_round_mask3, "V4dV4dV4dV4dUcIi", "nV:256:", "avx10.2-256")
2032+
TARGET_BUILTIN(__builtin_ia32_vfmaddph256_round_mask, "V16xV16xV16xV16xUsIi", "nV:256:", "avx10.2-256")
2033+
TARGET_BUILTIN(__builtin_ia32_vfmaddph256_round_maskz, "V16xV16xV16xV16xUsIi", "nV:256:", "avx10.2-256")
2034+
TARGET_BUILTIN(__builtin_ia32_vfmaddph256_round_mask3, "V16xV16xV16xV16xUsIi", "nV:256:", "avx10.2-256")
2035+
TARGET_BUILTIN(__builtin_ia32_vfmaddps256_round_mask, "V8fV8fV8fV8fUcIi", "nV:256:", "avx10.2-256")
2036+
TARGET_BUILTIN(__builtin_ia32_vfmaddps256_round_maskz, "V8fV8fV8fV8fUcIi", "nV:256:", "avx10.2-256")
2037+
TARGET_BUILTIN(__builtin_ia32_vfmaddps256_round_mask3, "V8fV8fV8fV8fUcIi", "nV:256:", "avx10.2-256")
2038+
TARGET_BUILTIN(__builtin_ia32_vfmaddcph256_round_mask, "V8fV8fV8fV8fUcIi", "nV:256:", "avx10.2-256")
2039+
TARGET_BUILTIN(__builtin_ia32_vfmaddcph256_round_maskz, "V8fV8fV8fV8fUcIi", "nV:256:", "avx10.2-256")
2040+
TARGET_BUILTIN(__builtin_ia32_vfmaddcph256_round_mask3, "V8fV8fV8fV8fUcIi", "nV:256:", "avx10.2-256")
2041+
TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd256_round_mask, "V4dV4dV4dV4dUcIi", "nV:256:", "avx10.2-256")
2042+
TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd256_round_maskz, "V4dV4dV4dV4dUcIi", "nV:256:", "avx10.2-256")
2043+
TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd256_round_mask3, "V4dV4dV4dV4dUcIi", "nV:256:", "avx10.2-256")
2044+
TARGET_BUILTIN(__builtin_ia32_vfmaddsubph256_round_mask, "V16xV16xV16xV16xUsIi", "nV:256:", "avx10.2-256")
2045+
TARGET_BUILTIN(__builtin_ia32_vfmaddsubph256_round_maskz, "V16xV16xV16xV16xUsIi", "nV:256:", "avx10.2-256")
2046+
TARGET_BUILTIN(__builtin_ia32_vfmaddsubph256_round_mask3, "V16xV16xV16xV16xUsIi", "nV:256:", "avx10.2-256")
2047+
TARGET_BUILTIN(__builtin_ia32_vfmaddsubps256_round_mask, "V8fV8fV8fV8fUcIi", "nV:256:", "avx10.2-256")
2048+
TARGET_BUILTIN(__builtin_ia32_vfmaddsubps256_round_maskz, "V8fV8fV8fV8fUcIi", "nV:256:", "avx10.2-256")
2049+
TARGET_BUILTIN(__builtin_ia32_vfmaddsubps256_round_mask3, "V8fV8fV8fV8fUcIi", "nV:256:", "avx10.2-256")
2050+
TARGET_BUILTIN(__builtin_ia32_vfmsubpd256_round_mask3, "V4dV4dV4dV4dUcIi", "nV:256:", "avx10.2-256")
2051+
TARGET_BUILTIN(__builtin_ia32_vfmsubph256_round_mask3, "V16xV16xV16xV16xUsIi", "nV:256:", "avx10.2-256")
2052+
TARGET_BUILTIN(__builtin_ia32_vfmsubps256_round_mask3, "V8fV8fV8fV8fUcIi", "nV:256:", "avx10.2-256")
2053+
TARGET_BUILTIN(__builtin_ia32_vfmsubaddpd256_round_mask3, "V4dV4dV4dV4dUcIi", "nV:256:", "avx10.2-256")
2054+
TARGET_BUILTIN(__builtin_ia32_vfmsubaddph256_round_mask3, "V16xV16xV16xV16xUsIi", "nV:256:", "avx10.2-256")
2055+
TARGET_BUILTIN(__builtin_ia32_vfmsubaddps256_round_mask3, "V8fV8fV8fV8fUcIi", "nV:256:", "avx10.2-256")
2056+
TARGET_BUILTIN(__builtin_ia32_vfmulcph256_round_mask, "V8fV8fV8fV8fUcIi", "nV:256:", "avx10.2-256")
2057+
TARGET_BUILTIN(__builtin_ia32_vgetexppd256_round_mask, "V4dV4dV4dUcIi", "nV:256:", "avx10.2-256")
2058+
TARGET_BUILTIN(__builtin_ia32_vgetexpph256_round_mask, "V16xV16xV16xUsIi", "nV:256:", "avx10.2-256")
2059+
TARGET_BUILTIN(__builtin_ia32_vgetexpps256_round_mask, "V8fV8fV8fUcIi", "nV:256:", "avx10.2-256")
2060+
TARGET_BUILTIN(__builtin_ia32_vgetmantpd256_round_mask, "V4dV4dIiV4dUcIi", "nV:256:", "avx10.2-256")
2061+
TARGET_BUILTIN(__builtin_ia32_vgetmantph256_round_mask, "V16xV16xIiV16xUsIi", "nV:256:", "avx10.2-256")
2062+
TARGET_BUILTIN(__builtin_ia32_vgetmantps256_round_mask, "V8fV8fIiV8fUcIi", "nV:256:", "avx10.2-256")
2063+
TARGET_BUILTIN(__builtin_ia32_vmaxpd256_round, "V4dV4dV4dIi", "nV:256:", "avx10.2-256")
2064+
TARGET_BUILTIN(__builtin_ia32_vmaxph256_round, "V16xV16xV16xIi", "nV:256:", "avx10.2-256")
2065+
TARGET_BUILTIN(__builtin_ia32_vmaxps256_round, "V8fV8fV8fIi", "nV:256:", "avx10.2-256")
2066+
TARGET_BUILTIN(__builtin_ia32_vminpd256_round, "V4dV4dV4dIi", "nV:256:", "avx10.2-256")
2067+
TARGET_BUILTIN(__builtin_ia32_vminph256_round, "V16xV16xV16xIi", "nV:256:", "avx10.2-256")
2068+
TARGET_BUILTIN(__builtin_ia32_vminps256_round, "V8fV8fV8fIi", "nV:256:", "avx10.2-256")
2069+
TARGET_BUILTIN(__builtin_ia32_vmulpd256_round, "V4dV4dV4dIi", "nV:256:", "avx10.2-256")
2070+
TARGET_BUILTIN(__builtin_ia32_vmulph256_round, "V16xV16xV16xIi", "nV:256:", "avx10.2-256")
2071+
TARGET_BUILTIN(__builtin_ia32_vmulps256_round, "V8fV8fV8fIi", "nV:256:", "avx10.2-256")
2072+
TARGET_BUILTIN(__builtin_ia32_vrangepd256_round_mask, "V4dV4dV4dIiV4dUcIi", "nV:256:", "avx10.2-256")
2073+
TARGET_BUILTIN(__builtin_ia32_vrangeps256_round_mask, "V8fV8fV8fIiV8fUcIi", "nV:256:", "avx10.2-256")
2074+
TARGET_BUILTIN(__builtin_ia32_vreducepd256_round_mask, "V4dV4dIiV4dUcIi", "nV:256:", "avx10.2-256")
2075+
TARGET_BUILTIN(__builtin_ia32_vreduceph256_round_mask, "V16xV16xIiV16xUsIi", "nV:256:", "avx10.2-256")
2076+
TARGET_BUILTIN(__builtin_ia32_vreduceps256_round_mask, "V8fV8fIiV8fUcIi", "nV:256:", "avx10.2-256")
2077+
TARGET_BUILTIN(__builtin_ia32_vrndscalepd256_round_mask, "V4dV4dIiV4dUcIi", "nV:256:", "avx10.2-256")
2078+
TARGET_BUILTIN(__builtin_ia32_vrndscaleph256_round_mask, "V16xV16xIiV16xUsIi", "nV:256:", "avx10.2-256")
2079+
TARGET_BUILTIN(__builtin_ia32_vrndscaleps256_round_mask, "V8fV8fIiV8fUcIi", "nV:256:", "avx10.2-256")
2080+
TARGET_BUILTIN(__builtin_ia32_vscalefpd256_round_mask, "V4dV4dV4dV4dUcIi", "nV:256:", "avx10.2-256")
2081+
TARGET_BUILTIN(__builtin_ia32_vscalefph256_round_mask, "V16xV16xV16xV16xUsIi", "nV:256:", "avx10.2-256")
2082+
TARGET_BUILTIN(__builtin_ia32_vscalefps256_round_mask, "V8fV8fV8fV8fUcIi", "nV:256:", "avx10.2-256")
2083+
TARGET_BUILTIN(__builtin_ia32_vsqrtpd256_round, "V4dV4dIi", "nV:256:", "avx10.2-256")
2084+
TARGET_BUILTIN(__builtin_ia32_vsqrtph256_round, "V16xV16xIi", "nV:256:", "avx10.2-256")
2085+
TARGET_BUILTIN(__builtin_ia32_vsqrtps256_round, "V8fV8fIi", "nV:256:", "avx10.2-256")
2086+
TARGET_BUILTIN(__builtin_ia32_vsubpd256_round, "V4dV4dV4dIi", "nV:256:", "avx10.2-256")
2087+
TARGET_BUILTIN(__builtin_ia32_vsubph256_round, "V16xV16xV16xIi", "nV:256:", "avx10.2-256")
2088+
TARGET_BUILTIN(__builtin_ia32_vsubps256_round, "V8fV8fV8fIi", "nV:256:", "avx10.2-256")
19692089

19702090
// AVX-VNNI-INT16
19712091
TARGET_BUILTIN(__builtin_ia32_vpdpwsud128, "V4iV4iV4iV4i", "nV:128:", "avxvnniint16")

0 commit comments

Comments
 (0)