[clang] [CIR][X86] Add support for cmpps/cmppd* builtins (PR #228491)
Mariya Podchishchaeva via cfe-commits
cfe-commits at lists.llvm.org
Mon Nov 9 06:02:41 PST 2026
================
@@ -2637,10 +2638,195 @@ CIRGenFunction::emitX86BuiltinExpr(unsigned builtinID, const CallExpr *expr) {
case X86::BI__builtin_ia32_vcmpbf16512_mask:
case X86::BI__builtin_ia32_vcmpbf16256_mask:
case X86::BI__builtin_ia32_vcmpbf16128_mask:
+ isMaskFCmp = true;
+ [[fallthrough]];
case X86::BI__builtin_ia32_cmpps:
case X86::BI__builtin_ia32_cmpps256:
case X86::BI__builtin_ia32_cmppd:
- case X86::BI__builtin_ia32_cmppd256:
+ case X86::BI__builtin_ia32_cmppd256: {
+ mlir::Location loc = getLoc(expr->getExprLoc());
+
+ // The third argument is the comparison condition, an integer in the range
+ // [0, 31]. Predicates for 16-31 repeat the 0-15 predicates and differ only
+ // in signalling behavior.
+ unsigned cc =
+ ops[2].getDefiningOp<cir::ConstantOp>().getIntValue().getZExtValue() &
+ 0x1f;
+
+ // CIR has no unordered predicates, so they are expressed as the inverse of
+ // the corresponding ordered predicate, e.g. ULT is !OGE.
+ cir::CmpOpKind pred = cir::CmpOpKind::eq;
+ bool shouldInvert = false;
+ bool isTrue = false;
+ bool isFalse = false;
+ bool isSignaling = false;
+ switch (cc & 0xf) {
+ case 0x00:
+ pred = cir::CmpOpKind::eq;
+ isSignaling = false;
+ break;
+ case 0x01:
+ pred = cir::CmpOpKind::lt;
+ isSignaling = true;
+ break;
+ case 0x02:
+ pred = cir::CmpOpKind::le;
+ isSignaling = true;
+ break;
+ case 0x03:
+ pred = cir::CmpOpKind::uno;
+ isSignaling = false;
+ break;
+ case 0x04:
+ pred = cir::CmpOpKind::ne;
+ isSignaling = false;
+ break;
+ case 0x05: // UGE = !LT
+ pred = cir::CmpOpKind::lt;
+ shouldInvert = true;
+ isSignaling = true;
+ break;
+ case 0x06: // UGT = !LE
+ pred = cir::CmpOpKind::le;
+ shouldInvert = true;
+ isSignaling = true;
+ break;
+ case 0x07: // ORD = !UNO
+ pred = cir::CmpOpKind::uno;
+ shouldInvert = true;
+ isSignaling = false;
+ break;
+ case 0x08: // UEQ != ONE
+ pred = cir::CmpOpKind::one;
+ shouldInvert = true;
+ isSignaling = false;
+ break;
+ case 0x09: // ULT != GE
+ pred = cir::CmpOpKind::ge;
+ shouldInvert = true;
+ isSignaling = true;
+ break;
+ case 0x0a: // ULE != GT
+ pred = cir::CmpOpKind::gt;
+ shouldInvert = true;
+ isSignaling = true;
+ break;
+ case 0x0b: // FALSE
+ isFalse = true;
+ isSignaling = false;
+ break;
+ case 0x0c:
+ pred = cir::CmpOpKind::one;
+ isSignaling = false;
+ break;
+ case 0x0d:
+ pred = cir::CmpOpKind::ge;
+ isSignaling = true;
+ break;
+ case 0x0e: // GT
+ pred = cir::CmpOpKind::gt;
+ isSignaling = true;
+ break;
+ case 0x0f: // TRUE
+ isTrue = true;
+ isSignaling = false;
+ break;
+ default:
+ llvm_unreachable("Unhandled cc");
+ }
+
+ // Invert the signalling behavior for 16-31.
+ if (cc & 0x10)
+ isSignaling = !isSignaling;
+
+ auto opsTy = cast<cir::VectorType>(ops[0].getType());
+ unsigned numElts = opsTy.getSize();
+
+ // If the predicate is true or false, or the builtin is mask enabled, and
+ // we're using constrained intrinsics, there is no compare operation we can
+ // use. Just use the legacy X86 specific intrinsic.
+ if (builder.getIsFPConstrained() && (isTrue || isFalse || isMaskFCmp)) {
+ llvm::StringRef intrinsicName;
+ switch (builtinID) {
+ default: llvm_unreachable("Unexpected builtin");
+ case X86::BI__builtin_ia32_cmpps:
+ intrinsicName = "x86.sse.cmp.ps";
+ break;
+ case X86::BI__builtin_ia32_cmpps256:
+ intrinsicName = "x86.avx.cmp.ps.256";
+ break;
+ case X86::BI__builtin_ia32_cmppd:
+ intrinsicName = "x86.sse2.cmp.pd";
+ break;
+ case X86::BI__builtin_ia32_cmppd256:
+ intrinsicName = "x86.avx.cmp.pd.256";
+ break;
+ case X86::BI__builtin_ia32_cmpph128_mask:
+ intrinsicName = "x86.avx512fp16.mask.cmp.ph.128";
+ break;
+ case X86::BI__builtin_ia32_cmpph256_mask:
+ intrinsicName = "x86.avx512fp16.mask.cmp.ph.256";
+ break;
+ case X86::BI__builtin_ia32_cmpph512_mask:
+ intrinsicName = "x86.avx512fp16.mask.cmp.ph.512";
+ break;
+ case X86::BI__builtin_ia32_cmpps512_mask:
+ intrinsicName = "x86.avx512.mask.cmp.ps.512";
+ break;
+ case X86::BI__builtin_ia32_cmppd512_mask:
+ intrinsicName = "x86.avx512.mask.cmp.pd.512";
+ break;
+ case X86::BI__builtin_ia32_cmpps128_mask:
+ intrinsicName = "x86.avx512.mask.cmp.ps.128";
+ break;
+ case X86::BI__builtin_ia32_cmpps256_mask:
+ intrinsicName = "x86.avx512.mask.cmp.ps.256";
+ break;
+ case X86::BI__builtin_ia32_cmppd128_mask:
+ intrinsicName = "x86.avx512.mask.cmp.pd.128";
+ break;
+ case X86::BI__builtin_ia32_cmppd256_mask:
+ intrinsicName = "x86.avx512.mask.cmp.pd.256";
+ break;
+ }
----------------
Fznamznon wrote:
Same as another comment
https://github.com/llvm/llvm-project/pull/228491
More information about the cfe-commits
mailing list